A Method of User Travel Mode Recognition Based on Convolutional Neural Network and Cell Phone Signaling Data

Abstract

As urbanization accelerates, traffic congestion in cities has become a problem. Therefore, accurately identifying urban residents’ travel patterns is crucial for urban traffic planning and intelligent transportation systems. In this study, a convolutional neural network (CNN) approach based on multichannel feature extraction using mobile phone signaling data to identify user travel modes is proposed. Here, a trajectory generation method was designed for five types of travel modes. By designing a spatiotemporal threshold screening method, anomalies were identified and processed, combined with the feature analysis method, key points in the signaling extracted, the travel trajectory sliced, and travel sub-trajectory data generated. Next, in the travel mode identification stage, road network information was introduced to improve localization accuracy, and the method for calculating feature values improved. A user travel feature dataset was generated by calculating the feature values, and the travel modes represented by each class were classified and recognized based on the CNN method. Satisfactory results were achieved through experiments using mobile phone signaling and field research data in Kunming, China. The experimental results showed that analysis based on mobile phone signaling data could classify, identify, and obtain different travel category modes. This method’s accuracy was 84.7%. The method provided a feasible way of identifying travel patterns in the context of smart cities and big data, providing strong support for urban transport planning and management, and has the potential for wider application.

1. Introduction

Mode of travel is a fundamental attribute for understanding a user’s mobile behavior. A user’s mobile state always shows certain regularity and identifying the mode of transport can reflect this regularity and provide specific services to the user. For example, if the user often drives through a certain place, a corresponding business push can be provided in that place. In addition, identifying traffic travel patterns enables the analysis and prediction of traffic conditions [1], and the importance of predicting traffic patterns in traffic management systems has increased. If a user takes a bus that does not move for a long time over a given period, it is possible to determine whether there may be traffic congestion. Identifying traffic patterns based on multiple users enables the detection of hotspots and classical routes in cities. The rapidly developing wireless location technology enables easier access to movement data that can be used not only for location and tracking but also to represent information spatially [2]. For example, arranging individual data points in chronological order provides a user’s trajectory, which reflects the individual’s movement status when changing between different modes of transport. Many studies have shown that people often combine multiple modes of transportation in their lives and work based on their purpose of travel [3], e.g., when going to work, they first walk from home to the bus stop, take the bus, alight, and walk to the office. Identifying multiple transport modes taken during travel has attracted the attention of many researchers.

This is an important basis for understanding urban transport systems, forecasting, and managing traffic demands [4]. In addition, it is of great theoretical and practical significance in the study of travel choice behavior, forecasting changes in the supply and demand of transport infrastructure, optimizing the allocation of transport infrastructure resources, and formulating targeted traffic management rules. Compared with previous traffic research data sources, which relied on human-acquired questionnaire data of low quality [5], signaling data contain rich spatiotemporal characteristics of mobility. More mature data pre-processing and data verification techniques have been developed, and more consistent industry understanding and analysis methods have been obtained from basic research of commuting behavior, origin-to-destination (OD) extraction and analysis [6], and other aspects of travel. Current research on urban residents’ travel mode selection is focused on travel surveys that analyze the characteristics of residents’ travel behavior at the micro level to optimize their travel modes. However, this research method has significant limitations, mainly because (1) it cannot fully use the characteristics of signaling data to study residents’ travel behavior, and (2) it cannot fully use big data mining technology to improve algorithm efficiency and recognition accuracy.

As a rich and valuable source of information, mobile phone signaling data [7] combined with the powerful analysis capabilities of machine learning (ML) can be explored and applied to the vast potential of mobile phone signaling data in the traffic field [8], enabling analysis and application in several traffic areas. Based on the location information and movement patterns in mobile phone signaling data, we can use a convolutional neural network (CNN) model [9,10] in ML to predict the distribution and changes in traffic flow. By training the CNN model, it can learn to extract spatial features and predict future traffic flow based on historical data, helping achieve a better-adjusted traffic resource allocation and deal with traffic congestion and bottlenecks in advance. By training the model, it can learn normal traffic patterns and determine abnormal conditions, such as traffic accidents and congestion occurrences, based on the current input data. This will help with the detection and response to traffic events on time, thereby improving traffic safety and efficiency. In addition, the location information in mobile phone signaling data can be used to build a road network map and analyze the road network using a CNN model. Using the CNN model, we can predict road topology, intersection flow, and road travel time, providing an important reference for traffic and route planning [6]. This will provide travelers with more intelligent and efficient navigation and route-planning services, thereby reducing traffic congestion and journey times.

Other frequently employed machine learning models might be hampered by a lack of global information, poor adaptability, and capacities to overfit and generalize. Through procedures like convolution and pooling, CNN models are able to capture traffic and behavioral patterns at various times and spatial places, and they have a great expressive ability for determining traffic modes. The CNN model can also automatically pick up end-to-end feature representation, maximizing local correlation in the data and enhancing traffic mode recognition precision. Therefore, this study examined the use of signaling data and combined it with a CNN to realize travel mode recognition of mobile phone users. The current approach to the analysis of residents’ travel mode adopts the research framework of trajectory data pre-processing as well as the slicing and recognition model; however, there are still some problems with signal processing and feature calculation. The main outstanding problems include the following three aspects:

• Data pre-processing does not fully consider data anomalies, which affects the quality of the basic data.
• It does not consider the problem of travel trajectory slicing.
• The feature calculation method of the recognition algorithm and the accuracy rate of travel mode recognition need to be improved.

Therefore, this study improves the existing travel mode recognition framework by improving the three aspects of data pre-processing, trajectory segmentation, and travel mode recognition to enhance the effectiveness of recognition.

In this study, an improved CNN method is proposed to utilize mobile phone signaling data for user travel mode identification. Compared with the traditional CNN model, this study makes corresponding adjustments in the design of the structure of each layer of the model and takes user travel features as multi-channel values. It addresses the problems of low efficiency and accuracy of data mining algorithms, and experiments are designed to verify the effectiveness of the proposed method in the main city of Kunming. The proposed method can improve the accuracy and efficiency of modeling the travel mode choice of residents in large cities under traffic congestion conditions and the ability to construct the behavioral features of travelers, as well as provide a reference for mining the characteristics and patterns of travel behavior of residents in large cities.

2. Related Work

2.1. Traffic Study Based on Mobile Phone Signaling

With the continuous accumulation of traffic information extraction technology from mobile phone location data and in-depth research on the extraction and analysis of traffic status information from mobile phones, researchers have begun to explore the acquisition of individual travel behavior based on mobile phone location data, and research on travel behavior based on mobile phone signaling data has begun.

In terms of transportation, some researchers focus on location traffic and trajectory studies of travelers. Lai et al. [11] used mobile phone signaling location information combined with rail line information to identify metro passenger travel paths. Caceres et al. [12] presented a set of models for inferring movement from one cell to another using anonymous call data from mobile phones. Dong et al. [13] proposed the idea of traffic semantics to extract information on the starting points and destinations of traffic commuters from mobile phone CDR data. To express urban traffic, Larijani et al. [14] used mobile phone data to investigate the behavior of origin–destination flows within Paris and its suburbs, aiming to explore different modes of transportation. Puntumapon et al. [15] proposed a method for classifying two types of mobility based on mobile phone information.

Meanwhile, other researchers have focused on congestion in transportation and some transportation planning. Wu et al. [16] focused on two aspects: (1) using mobile phone data to obtain accurate characteristics of commuting behavior, and (2) using an agent-based model to simulate the commuting behavior of residents and reverse the causes of congestion. Zhou et al. [17] provided a method for identifying travelers’ modes of transport by tracking mobile phone data, aiming to obtain accurate mode split rates to provide decision support for urban transport planning. Bassolas et al. [18] demonstrated the ability of mobile phone records to provide activity-based transport models and asserted the advantages of using activity-based models to estimate the effects of transport demand management policies. Kim et al. [19] investigated the relationship between urban-built environment factors and travel behavior of visually impaired people. Luo et al. [20] used mobile phone signaling data to build a monitoring model to rapidly estimate area occupancy and duration and deployed it in practice. Rose [21] identified issues that might affect the use of mobile phones as probes for obtaining real-time traffic information. Gariazzo et al. [22] focused on two aspects: using mobile phone data to obtain accurate characteristics of commuting behavior and simulating residents’ commuting behavior using agent-based models in order to reverse the causes of congestion.

Transportation mode detection is also a hot research topic. Bloch et al. [23] used location and sensor data from signaling data to build an ML model for identifying residents’ travel patterns. Byon et al. [24] proposed a method for identifying a traveler’s mode of transport by tracking global positioning system (GPS)-equipped mobile devices in the traffic stream. The simulation model included an access link to the Internet from an Internet service provider as well as background Internet phones and data traffic [25]. Hao et al. [26] proposed a cellular signaling-based user travel feature identification and traffic monitoring system for road networks. Wang et al. [27] proposed a cellular phone information collection method based on simulating real base station operating patterns, in which appropriate instruments were placed at the roadside to detect passing cellular phones. Neural network-based artificial intelligence was used to identify the mode of transport by detecting different physical profile patterns for each mode, including speed, acceleration, number of satellites in view, and electromagnetic levels [24]. Therefore, it is important to obtain accurate and valuable data to improve traffic conditions.

2.2. Machine-Learning-Based Traffic Pattern Detection

Research on travel behavior based on signaling data has increased. Existing research has used mobile phone signaling data to calculate feature parameters and identify them using methods such as ML. These recognition methods use manual tagging to classify the data in advance; these are then placed in a recognition model for training and the trained model is used to classify unknown data. However, not all mobile devices have their location modules switched on in real time. Signaling data, which reflect the location information of residents over a wide range of time and space, are used in urban computing, particularly in identifying travel modes. Methods based on mobile phone signaling data and supervised ML algorithms for travel mode identification require upfront work to add labels to the travel mode but are weak at analyzing residential travel over large time spans and spatial scales. This test dataset lacks adaptability to signaling data from other regions due to differences in factors such as topography and base station distribution in each region.

Wang et al. [28] used a division-based k-means clustering algorithm to cluster the travel durations of people with the same travel start and stop points, obtaining travel clusters with different durations and using the duration distribution of different travel modes provided by Google Maps to divide the travel modes of each cluster sample. However, this method lacked correlation analysis for different travel trajectories of people with the same start and end points and only correlated the travel durations of people with the durations provided by Google Maps. Das et al. [29] automated the adjustment process to produce an intelligent hybrid model that could use GPS trajectories for near-real-time pattern detection. A large amount of Dutch travel diary data from 2010 to 2012 was used and enriched with variables on the built and natural environments and weather conditions to make recommendations for model selection [30,31]. Das et al. [32] developed a novel knowledge-based approach for interpreting smartphone GPS trajectories by detecting various modes of transportation used in travel. Zhou et al. [33] investigated the spatiotemporal patterns of BSS and taxi trips in Chicago between 2014 and 2016. However, few studies have used ML to predict multimodal travel time. Servos et al. [34] applied machine learning algorithms such as extreme random trees, adaptive boosting, and support vector regression to this problem because of their ability to handle low data volumes and short processing times. Random forests (RF) provide good training and test data results and are considered the best predictors of user pattern choice in Kuantan. Borodinov et al. [35] investigated some of the best-known methods, such as support vector machines (SVM), decision trees (DT), RF, k-nearest neighbor algorithms, multilayer perceptron classifiers, and a method based on the estimation algorithm proposed by Salami et al. [36], to train four ML classifier algorithms using six years of historical dengue import data from 21 European countries and the mediating connectivity indices of imports and air transport network centrality measures. In addition, for the study of travel patterns, Liang et al. [37] designed and implemented a lightweight and energy-efficient traffic pattern detection application using only accelerometer sensors on smartphones, and built CNNs in this application to determine transport patterns. Dabiri et al. [38] used a CNN architecture to predict travel patterns based solely on the original GPS trajectory, where the patterns were labeled walking, cycling, bus, driving, and train. Dabiri et al. [39] proposed a deep semi-supervised convolutional autoencoder architecture that identified a traveler’s mode of transport based on the traveler’s GPS trajectory. Qin et al. [40] proposed a deep learning (DL)-based traffic pattern recognition algorithm, CL-TRANSMODE, which detected multiple traffic patterns. Nawaz et al. [41] proposed a DL-based convolutional long- and short-term memory (LSTM) model for transport pattern learning, in which deep high-level features were extracted using a CNN, and then the LSTM was used to learn sequential pattern data using both GPS and weather features, thus making full use of spatiotemporal operations. To evaluate these techniques, Bejani et al. [42] examined the results of three popular large-scale datasets and demonstrated their efficiency using two transport datasets.

3. Data Processing and Modeling

Figure 1 shows the general framework of the modeling approach used, with the corresponding subsections describing the relevant specific steps.

3.1. Data Pre-Processing

Signaling is the control signal required to ensure proper communication in a wireless communication system (in addition to transmitting user information) for the network to operate in an orderly manner. However, signaling is a system that allows each link to be analyzed and processed and to interact to form a series of operations and controls that ensure the efficient and reliable transmission of user information. Mobile phone signaling data is one form of signaling data, which is the data information generated by the control messages passed over the interfaces of different communication devices during the communication process and is regarded as a kind of spatiotemporal trajectory big data [43]. Raw signaling data cannot be used directly, and there is a lot of invalid data. Therefore, the data need to be preprocessed before using the signaling data. This study proposes a signaling data pre-processing method. We propose a pre-processing method for the signaling data. First, we use the user’s number to divide all the signaling into user trajectory datasets, and then we use the time and space thresholds to filter the signaling trajectory data of each user in turn to remove the noise, duplicate, “ping-pong switching” and “data drift” points. The data is then filtered to remove noise, duplicate, “ping-pong switching” and “data drift” points.

The signaling data [44,45] generated when a mobile phone user generates a communication signal with the base station when using their phone generates a record containing the user ID, cell location identifier, base station latitude and longitude, and time information defined as the j-th record of the i-th user and denoted as  ${\mathrm{s}}_{\mathrm{i},\mathrm{j}}=\{{\mathrm{uid}}_{\mathrm{j}},{\mathrm{lat}}_{\mathrm{j}},{\log }_{\mathrm{j}},{\mathrm{time}}_{\mathrm{j}}\}$ (the parameters represent the user’s unique identifier, longitude, latitude, and arrival time, respectively), the signaling dataset for user I is denoted as  ${\mathrm{S}}_{\mathrm{i}}=\{{\mathrm{s}}_{\mathrm{i},1},{\mathrm{s}}_{\mathrm{i},2},{\mathrm{s}}_{\mathrm{i},3},\dots , {\mathrm{s}}_{\mathrm{i},\mathrm{n}}\}$ and the set of signaling data for all users is denoted as  $\mathrm{S}=\left\{{\mathrm{S}}_1,{\mathrm{S}}_2,\dots ,{\mathrm{S}}_{\mathrm{N}}\right\}$. Here, the road network is referred to as the road network, which is defined by an undirected graph  $\mathrm{G}(\mathrm{V},\mathrm{E})$, where the set of road endpoints is denoted as  $\mathrm{V}$= (${\mathrm{v}}_{\mathrm{N}}$|$\mathrm{N}=1,2,\dots ,\mathrm{N})$, and the set of road line segments is represented as  $\mathrm{E}$= (${\mathrm{e}}_{\mathrm{M}}$|$\mathrm{M}=1,2,\dots ,\mathrm{M})$. The length of the road segment was defined without considering one-way or the weight of a directed edge, and the nodes of the road network contained information on the geographical coordinates.

Based on the characteristics of mobile phone signaling data, it was found that the original signaling data obtained from the base station contained data duplication, redundancy, drift, ping pong, and noise. Therefore, to obtain valid and usable data, traditional data cleaning and certain processing based on its unique problems are required. There are many invalid data points with missing and duplicate fields in the raw data. For the data with missing fields, conditional judgement is performed in the form of traversal to check the integrity of the data fields when reading the dataset. If the field is empty, it is deleted. For invalid duplicate data, the data need to be filtered initially according to the following formula:

$$\mathrm{n}=\left|{\mathrm{S}}_{\mathrm{i}}\right|$$

(1)

$${\mathrm{d}}_{\mathrm{s}\mathrm{e}}=\mathrm{D}\mathrm{i}\mathrm{s}\left({\mathrm{S}}_{\mathrm{i},0},{\mathrm{S}}_{\mathrm{i},\mathrm{n}}\right)$$

(2)

$$\overline{\mathrm{d}}=\frac{1}{\mathrm{n}-1}\sum _{\mathrm{j}=0}^{\mathrm{n}-1}\mathrm{D}\mathrm{i}\mathrm{s}\left({\mathrm{S}}_{\mathrm{i},\mathrm{j}},{\mathrm{S}}_{\mathrm{i},\mathrm{j}+1}\right)$$

(3)

where  $\left|{\mathrm{S}}_{\mathrm{i}}\right|$ was the number of signaling entries for user  $\mathrm{i}$ and  $\mathrm{Dis}\left(\mathrm{a},\mathrm{b}\right)$ is a function that calculates the straight-line distance from point a to point b. Thus,  ${\mathrm{d}}_{\mathrm{se}}$ is the Euclidean distance from the start point to the end point of user  $\mathrm{i}$, and  $\overline{\mathrm{d}}$ is the average distance between two consecutive trajectory points for user  $\mathrm{i}$. To reasonably filter the data, a threshold space of ($\mathrm{n}>10, {\mathrm{d}}_{\mathrm{se}}>100 \mathrm{m},\overline{\mathrm{d}}<500 \mathrm{m}$) is set.

When a mobile phone user was at the junction of two or more base station locations with comparable signal strengths, the mobile phone switched frequently between the two or more base stations over a short period, in turn generating multiple signaling records. This phenomenon is known as the ping-pong effect [46]. Ping-pong data are characterized by the fact that the location information (referring to  ${\mathrm{cid}}_{\mathrm{j}}, {\mathrm{lat}}_{\mathrm{j}}, {\log }_{\mathrm{j}}$) in the  $\mathrm{j}$-th data ${\mathrm{S}}_{\mathrm{i},\mathrm{j}}$ of user  $\mathrm{i}$ is not the same as that in the ($\mathrm{j}+1)$-th data  ${\mathrm{S}}_{\mathrm{i},\mathrm{j}+1}$ within a short interval. Therefore, a time threshold ${\mathrm{t}}_{\mathrm{limit}}$ was set, and when $\left|{\mathrm{time}}_{\mathrm{j}}-{\mathrm{time}}_{\mathrm{j}-1}\right|<{\mathrm{t}}_{\mathrm{limit}}$,  ${\mathrm{S}}_{\mathrm{i},\mathrm{j}+1}={\mathrm{S}}_{\mathrm{i},\mathrm{j}}$ and made continuous judgments on the ping-pong switching data. Within a short period, the mobile phone user’s base station data appeared to switch from an adjacent base station to a distant base station location, and data drift occurred. However, due to overlapping base station coverage, location switching, and base station load, multiple location data might be generated for one activity. The generated drifting data was treated as invalid. Therefore, the data processing process calculated the time interval and straight-line distances, azimuth angle, and average speed between adjacent points based on the user’s time, latitude, and longitude information (the attributes were denoted as difftime, nodedist, degree, and avgspd, and the formulae are presented below), such that the drift data were processed by simple modeling of the additional attributes calculated and the error data eliminated from the database.

$${\mathrm{difftime}}_{\mathrm{j}}={\mathrm{time}}_{\mathrm{j}-1}-{\mathrm{time}}_{\mathrm{j}}$$

(4)

$${\mathrm{nodedist}}_{\mathrm{j}} =\mathrm{calEuclidean}({\mathrm{lon}}_{\mathrm{j}-1},{\mathrm{lat}}_{\mathrm{j}-1},{\mathrm{lon}}_{\mathrm{j}},{\mathrm{lat}}_{\mathrm{j}})$$

(5)

$${\mathrm{degree}}_{\mathrm{j}} =\mathrm{Getdegree}({\mathrm{lon}}_{\mathrm{j}-1},{\mathrm{lat}}_{\mathrm{j}-1},{\mathrm{lon}}_{\mathrm{j}},{\mathrm{lat}}_{\mathrm{j}})$$

(6)

$${\overline{\mathrm{v}}}_{\mathrm{ij}}={\mathrm{avgspd}}_{\mathrm{j}}=\frac{{\mathrm{nodedist}}_{\mathrm{j}}\times 1.2}{{\mathrm{difftime}}_{\mathrm{j}}}$$

(7)

$$\mathrm{isDrift}=\mathrm{isDrift}(\mathrm{difftime},\mathrm{nodedist},\mathrm{degree},\mathrm{avgspd})$$

(8)

3.2. Track Point Analysis

Trajectory point analysis refers to the identification of the stopping and moving points of a user during the travel process. A stopping point is the user’s start and end during the trip, i.e., the OD point of the user’s trip. The movement point represents the position that the user passes through during the travel process, and its movement speed characterizes the user’s travel speed between two stopping points. Recognizing dwell points in spatial trajectories is a key step in transforming spatial trajectories into semantic traffic trajectories. A dwell is the period during which the user stays in a certain area, and a travel segment is the user’s movement between two different dwell points. Definitions related to trajectory point analysis include:

• Track points: The processed mobile phone signaling data are time-stamped location point records, and the travel track is a collection of multiple time-stamped location points.
• Stopping point: The stopping point is the user’s origin or destination during the trip, i.e., the trip’s start and end points. Transportation refers to the spatial movement of people and objects. Generally, people travel to reach a destination and conduct corresponding activities. Therefore, each journey comprises two or more stopping points.
• Minimum dwell time (MDT): The minimum dwell time is the shortest time over which a user stays at the origin or destination. People travel to destinations and perform their corresponding activities. Therefore, except for special professionals such as drivers and couriers, the vast majority of users will stay at their destination for a certain period after reaching it to conduct the corresponding activities. In mobile phone signaling data, the time spent at each location is an important feature for determining whether it is a stopping point.
• Maximum activity distance (MAD): The maximum activity distance is the maximum distance that a user can travel around the origin or destination. After arriving at a destination, users generally move around it. If the user has a small range of activity at the destination or if the number of base stations around the destination is small, the user may only communicate with one base station during the activity, i.e., only one signaling record is generated at the stopover point.
• Mobile points: Mobile points are the points where the user is positioned between the stopover points. The spatiotemporal characteristics of the mobile point represent the spatiotemporal characteristics of the user during a period of travel. For example, the location of the mobile point represents the location point passing through the user’s travel path, and the speed of the mobile point represents the travel speed of the user between two stopover points.

In mobile phone signaling data, because of the large coverage area of the base station, fewer location points are generated by the user when moving around the destination. However, their time at the dwell point does not decrease. Identifying user dwell points should be considered from the trajectory point time characteristics combined with the spatial characteristics of the trajectory points as constraints to analyze the purpose of analyzing user trajectory points. Therefore, we proposed a trajectory point analysis method based on a user’s shortest dwell time and maximum activity distance for mobile phone signaling data.

3.3. Track Slitting Method

A trajectory-slicing method was proposed for preprocessed signaling data. Furthermore, to refine the signaling data so that a single trip could correspond to multiple travel modes, this research identified key points in a trip trajectory using temporal and spatial thresholds combined with trajectory feature analysis and used these key points to slice a trip trajectory into several travel sub-trajectories to correspond to different travel modes. A single trip corresponded to multiple modes of travel, e.g., “walking”—“public transport”—“rail”—“rail”—“rail”—“walking”. The different modes of travel were connected by “stops” or “transfers.” Therefore, it was necessary to identify these “key points” (stop and moving points) in advance and then slice and dice the residents’ travel trajectories into sub-trajectories through these points. The key points were identified using commuter profiling, as the different modes of travel varied based on speed, time consumption, etc.

The user travel trajectory after waiting for exceptions was obtained through an exception handling operation track to remove noise, duplicate points, “ping-pong switching” and “data drift”. As a user might use multiple travel modes in a single trip, the user’s travel track trajectory track was transformed into a travel segment travel (sub-trajectory) in traffic travel mode identification. This process involved discovering key points (stop and transfer points) in the travel trajectory and then using them to slice a series of consecutive data trajectory points, splitting the user trajectory into several travel segments representing different modes of travel. Extracting travel sub-trajectories is an important part of identifying and analyzing traffic trips, providing the basis for later trajectory identification. We defined s to denote the instantaneous velocity of each point and d(a, b) to denote the distance function between adjacent points a and b. The velocity of the (i + 1)-th point was defined as:

$${\mathrm{s}}_{\mathrm{i}+1}=\frac{\mathrm{d}\left({\mathrm{point}}_{\mathrm{i}},{\mathrm{point}}_{\mathrm{i}+1}\right)}{{\mathrm{point}}_{\mathrm{i}+1}·\mathrm{t}-{\mathrm{point}}_{\mathrm{i}}·\mathrm{t}}$$

(9)

A key point contained a stop and move point, not an exact track point, but a set of points in a continuous time range. The velocity of a stopping point was zero, and the travel velocities before the first point and after the last point in the stop range were not zero. Therefore, the stop point was defined as any  $\mathrm{k}$ adjacent time-adjacent trajectory points in a track ${\mathrm{point}}_{\mathrm{i}+1}{,\mathrm{point}}_{\mathrm{i}+2},\dots ,{ \mathrm{point}}_{\mathrm{i}+\mathrm{k}}$, if  ${\mathrm{point}}_{\mathrm{i}+1}·\mathrm{s}={\mathrm{point}}_{\mathrm{i}+2}·\mathrm{s}=\dots ={ \mathrm{point}}_{\mathrm{i}+\mathrm{n}}·\mathrm{s}=0$, and if the time from  $\mathrm{i}+1$ to  $\mathrm{i}+\mathrm{n}$ exceeded the time threshold, then all points from  $\mathrm{i}+1$ to  $\mathrm{i}+\mathrm{n}$ were dwell points.

Once a certain set of key points was identified, the trajectory track before the first and after the last points were divided into two different travel. Thus, the user’s travel trajectory track was sliced into several sub-trajectories travel using key points l. The set of the user’s travel segments was then generated based on the user T, which contained the definition of the user’s travel sub-trajectories in chronological order and the uniquely numbered sub-trajectory travel tID = id_i, i.e., as a combination of user and sequential numbers. The set T contained all the travel sub-tracks of all users and was distinguished by tID. The five features of each track were calculated and included in the set.

Because the location accuracy of signaling data was not high, this study proposed a method of using road network data as a constraint to improve the calculation of trajectory features and reduce the impact of calculation errors caused by the low positioning accuracy of signaling data. To identify the travel mode, we performed a cluster analysis of the sub-trajectories, divided the number of clusters using the variability of the data itself so that each cluster represented a class of travel mode, and combined the real data situation of the sub-trajectories to identify the travel mode for each class of clusters. A travel mode was defined as the mode of movement used by a person or vehicle in a valid travel segment from origin to destination. Therefore, it was necessary to identify the stationary area and extract valid travel segments from the signaling data. A valid travel segment contained information such as the start location and time, end location and time, and intermediate trajectory points of a trip. The dwelling area was the area where people’s travel trajectory occurred for a long time, i.e., the start or end areas of the valid travel segment. Identifying dwell areas was critical for extracting valid trips from users. Because the strength of the signals received from the surrounding base stations might change over time when the user is stationed, they switched between multiple base stations and their locations constituted the station area. The identification steps were:

• Set the dwell range distance threshold, MAD, and the dwell range minimum time threshold, MDT.
• The first signaling sequence was placed in the comparison sequence. If the position of the second signaling was within the set distance threshold, MAD, the second signaling was placed in the comparison sequence seq, and the remaining signaling was judged against each of the signaling in seq in order of signaling time; if the signaling position was more than MAD from any of the signaling in seq, the judgement was interrupted.
• Calculate the time difference between the last and first signaling in seq. If the time difference was greater than the set minimum time threshold, MDT, the signaling location area in seq was considered the resident area, and the coordinates of the resident core points of these locations were calculated based on the time ratio, and the core point coordinates were used to represent the resident area. If the time difference was less than the set minimum time threshold, MAD, then these signaling location areas were not resident areas, and the signaling data in seq was released and the determination of the remaining signaling continued. The remaining signaling pathways were determined. After identifying the presence region, the signaling data between the two presence regions were used as the signaling data generated by one valid trip, based on the results, and the travel mode of one trip was identified using a CNN combined with navigation data.

3.4. How to Identify the Mode of Travel

3.4.1. Eigenvalues for Travel Mode Recognition

The raw signal data for each user consisted of time points collected over a period. First, if the time interval between two consecutive signal data points exceeded a predetermined threshold, the user trajectory was divided into trips. Each trip was divided into different segments based on the mode of transport. In a CNN architecture, all the samples have to have the same size [47]. Given the constant length of all instances, each line segment was either subdivided into a predetermined number of trajectory points or filled with zero values to obtain the same size as other line segments. After generating line segments of the same size, we used two consecutive points of signaling to calculate the motion characteristics of each trajectory point. Here, we calculated the distance, average velocity, acceleration, directional angle, and jitter between two adjacent valid signaling trajectory points based on the formulae listed (see

Table 1. Eigenvalues to be calculated.

User_id	Come_Time	lon	lat	dis	avg_spd	accel	dire
Dacp::++20⋯LK4Hjz+	20210517111249	102.723999	25.117001	⋯	⋯	⋯	⋯
Dacp::++20⋯LK4Hjz+	20210517111249	102.737129	25.117001		⋯	⋯	⋯
Dacp::++20⋯LK4Hjz+	20210517111249	102.740977	25.117001		⋯	⋯	⋯
Dacp::++20⋯LK4Hjz+	20210517111249	102.733002	25.117001		⋯	⋯	⋯
Dacp::++20⋯LK4Hjz+	20210517111249	102.723999	25.117001		⋯	⋯	⋯
⋯	⋯	⋯	⋯	⋯	⋯	⋯	⋯
Dacp::++20⋯LK4Hjz+	20210517111249	102.657997	25.117001		⋯	⋯	⋯
Dacp::++20⋯LK4Hjz+	20210517111249	102.650002	25.117001		⋯	⋯	⋯

).

Here, four feature variables, namely average speed, acceleration, directional angle, and jitter value, obtained after processing and analyzing user signaling data, were selected as input variables for the recognition model to identify the mode of travel (see Figure 2).

In the travel mode identification stage, the design introduces a road network to constrain the improved eigenvalue calculation method to solve the problem of poor positioning accuracy and then generates a user travel feature dataset through the eigenvalue calculation. The design uses unsupervised learning combined with empirical analysis to classify categories and identify the travel mode represented by each category.

3.4.2. Multi-Channel Feature Convolutional Neural Networks

In view of the excellent feature extraction ability of convolutional neural networks [48], this study uses convolutional neural networks for travel mode recognition. The structure of the convolutional neural network is shown in Figure 3. In the input part of the convolutional neural network, all the travel segments are uniformly shaped as 1 × M × 4, where M is the number of signaling points in the travel segments. There are two convolutional layers, each followed by a ReLU layer and a maximum pooling layer. The convolutional kernels in the convolutional layers are designed specifically for the travel segment variables, where the first convolutional layer has an output channel of 3, a convolutional kernel size of 1 × 3, a step size of 2, and a padding size of 1. The second convolutional layer has an output channel of 6, a convolutional kernel size of 3 × 5, a step size of 2, and a padding size of 1. After convolution, a flattened layer is added to the model to flatten the convolution results and feed them into the fully connected layer. The first layer has 36 input variables and 10 output variables, and the second layer has 10 input variables and 5 output variables. After the fully connected layer, the model uses the softmax layer to adjust the input results to between −1 and 1 to obtain the final travel mode recognition results.

3.4.3. Training Parameters

The purpose of the training process is to modify each layer’s settings in order to reduce the loss function. To determine the error in the output layer, we use categorical cross-entropy as the loss function. In most cases, an optimization technique is employed to update the weights and reduce the loss function’s error. For big datasets with many hyperparameters, the Adam optimization algorithm is a quick optimization technique that determines the learning rate for each individual parameter. This algorithm has recently seen broader adoption in deep learning applications [49]. We used a batch size equal to 64 and Adam’s default settings as provided in their paper: learning rate = 0.001, β₁ = 0.9, β₂ = 0.999, and ε = 10⁻⁸. The parameters of the convolutional and FC layers were set according to the parameter settings used in [50].

4. Experimentation and Discussion

4.1. Related Data

Figure 4 shows the study area—the main urban area of Kunming, Yunnan Province, China. The main urban area of Kunming City is suitable for the development of diversified transport and mobility modes because of its excellent natural geography, well-developed public transport system, well-constructed bicycle paths, beautiful walking environment, and policies that encourage green travel. The road network data were obtained from OpenStreetMap (www.openstreetmap.org, accessed on 10 May 2023). After pre-processing, 226,439 central line segments were obtained, generating 204,414 road network nodes. The original network data obtained from OSM had some problems: road segments were borderline, there were broken or hanging points in the segments, and intersecting roads were not connected. Therefore, to address these problems, the road edge lines were drawn and processed as road center lines before the experiment, and the intersection lines of the road line segments were interrupted after connecting all the road links. Thus, the road network was interrupted only at the intersection points, and the average vehicle speed on the road was estimated for attribute marking based on the road classification of the corresponding road class. Simultaneously, based on the original base station data provided by the operator on the road network to drop points, for the coverage of the base station signal to the base station surrounding road sections in the candidate set, the corresponding base station-road network node correspondence table was generated for the subsequent spatial signaling data to the matching road network.

Mobile phone signaling data were obtained from an operator in Kunming in May 2021. The dataset used contained hourly track point latitude and longitude data for 5,776,605 people, involving 9,983,083 users, with an average of 372 track points per person per day, reflecting the daily travel situation of residents.

Table 2. Raw cellular signaling data.

Code	Description
Timestamp	Signaling generation time, time stamped by the vendor on the capture card for successful signaling processes, accurate to the second.
imsi	User identification code, also known as mobile phone identification code, IMSI or the result of a single encryption by IMSI, uniquely identifies the phone.
mcc	Mobile Country Code
mnc	Mobile Network Code
lac	Base station location area code
cid	Cell identification code
lng	Longitude
lat	Latitude

lists the fields related to the signaling data. The raw data for the cell phone signaling portion of the user is shown in

Table 3. Raw data for a subscriber’s mobile phone signaling section.

User_id	Come_Time	lon	lat	County_Name	⋯
Dacp::++20⋯LK4Hjz+	20210517111249	102.723999	25.117001	Wuhua	⋯
Dacp::++20⋯LK4Hjz+	20210517111249	102.737129	25.117001	Wuhua	⋯
Dacp::++20⋯LK4Hjz+	20210517111249	102.740977	25.117001	Panlong	⋯
Dacp::++20⋯LK4Hjz+	20210517111249	102.733002	25.117001	Wuhua	⋯
Dacp::++20⋯LK4Hjz+	20210517111249	102.723999	25.117001	Wuhua	⋯
⋯	⋯	⋯	⋯	⋯	⋯
Dacp::++20⋯LK4Hjz+	20210517111249	102.657997	25.117001	Xishan	⋯
Dacp::++20⋯LK4Hjz+	20210517111249	102.650002	25.117001	Xishan	⋯

.

4.2. Data Cleaning

Analysis of the signaling data revealed that there were multiple duplicate locations of signaling data from the same base station over a continuous period. In addition, due to the influence of the mobile communication system’s mechanism, there were ping-pong effect data and misallocation data in the signaling data. Therefore, we conducted data pre-processing work, including compression and de-noising of the signaling data, which reduced the pressure of storage and analysis identification. The data cleaning process removed missing data from the original dataset. Here, the completed data with unique user identification, start time, latitude, and longitude coordinate fields were kept; the redundant or duplicate data of the same user were removed based on the kept complete data and the dataset obtained after calculating additional attributes based on the existing user attributes was the dataset for the experimental use phase. A trajectory generation method was designed. The spatiotemporal threshold screening method was designed to identify and deal with anomalies, then key points in the signaling were identified and combined with the feature analysis method, and these key points were used to slice and dice the travel trajectory and generate travel sub-trajectory data.

4.3. Testing

Here, data pre-processing was used to obtain the effective travel trajectories of several users on a certain day. The users’ effective trips were matched by trajectory matching and time association; the routes obtained were walking, bicycling, private car driving, and public metro routes (e.g.,

Table 4. Valid occurrence track segments.

Mode of Transport	Paragraphs	Average Speed (m/s)	Maximum Acceleration (m/s2)
Walking	10,233	4	3
Riding	5568	12	4
Driving	3362	46	10
Bus	3486	35	8
Underground	8977	60	6

).

4.4. Results and Discussion

The area of the road network data collected in the experiment was representative and covered the main urban areas in the geographic space, especially the main urban part of the city, with a high density of base stations. The selected mobile phone signaling data came from the main urban area of Kunming, which had a high density of mobile phone user data, allowing us to estimate the travel situation of the entire area more accurately through the road-matching algorithm and to conduct population travel analysis. In addition, it allowed us to understand trends in the overall travel mode choices of residents in the main urban area of Kunming, observe changes in the transport mode trends of the population living in the area, and understand the impact on the development of urban transport. Temporally, the time span of the mobile phone user data was appropriate, with multiple users traveling throughout the month of the experiment. Note that extracting base station location data had an impact on determining the true location of pedestrians, and due to the rigorous nature of the algorithm, there might be small deviations in the data processing. During the experimental process, and considering the model’s feasibility, we made situational assumptions and processed the base station location point data directly as user location data for the experimental calculations, thus avoiding inconsistent base stations and user data. To judge the effectiveness and reasonableness of the algorithm, 80% of the samples in the original data were used as the research data, and the rest were kept as the test data. The recognition method was applied to judge the travel mode, and the model recognition accuracy was evaluated using the check-all rate, check-accuracy rate, F1-score, and correct rate (see

Table 5. Accuracy evaluation indicators.

	Positive Prediction	Negative Prediction
Positive Class	TP	FN
Negative Class	FP	TN
Precision	P = TP/(TP + FP)
Recall	R = TP/(TP + FN)
F1-score	F1 = 2PR/(P + R)

).

• Recall rate: The ratio of the number of correct results to the actual data for a certain travel mode in this model.
• F1-measure: The F1-score is the weighted average of the sum of the completeness and accuracy rates. The F1-score was the weighted average of the full and accuracy rates and was used here to evaluate the two indicators together.
• Correctness rate: The ratio of the number of samples judged correctly to the total number of samples.

The above travel mode identification algorithm was used to make travel mode judgments and the algorithm’s accuracy was evaluated using the check-all rate, check-accuracy rate, F1 value, and correct rate.

Table 6. Sample identification results.

Mode	Class					Recall (%)	F-Score (%)
	Walk	Bicycle	Drive	Bus	Underground
walk	2075	53	7	46	2	92.7	89.6
bicycle	172	935	12	34	3	84.8	75.2
drive	147	43	1230	78	24	81.2	79.8
bus	71	18	153	607	32	67.5	79.8
underground	75	21	22	43	879	85.5	74.8
Accuracy (%)	81.6	90.3	80.7	86.7	86.3	-	-

shows the recall and accuracy associated with each mode of transport. The recall for each mode indicated the accuracy of the predictor variables for that mode. Accuracy indicated the proportion of true instances out of all predicted instances for each mode. Overall, the results showed that the CNN architecture was effective at inferring transport modes, with accuracy and recall rates exceeding 67%. Table 6 shows that there was a strong correlation between the model’s performance in predicting a pattern and the number of instances available for that pattern. Several samples and the walking mode achieved a perfect recall of 92.7%, whereas the driving mode with the fewest available road segments gave the lowest accuracy of 67.5%. Note that traveler behavior was more unpredictable in the driving mode, as taxis and drops were more flexible and freer in their travel choices compared with other vehicles. Buses were obliged to adhere to predefined routes and schedules; therefore, they had more predictable movement behavior. Table 6 shows that among the five modes of travel (walking, cycling, driving, public transport, and metro-walking trips), metro-walking trips had the highest check-all rates and F1 values, with both reaching over 89.6%, followed by metro trips, which had check-all rates of over 85%. Driving and cycling achieved recognition rates of over 80% with an overall accuracy rate of 80%. Walking and cycling had lower accuracy rates because they are slow-moving modes of transport that cover shorter distances. Therefore, they generated fewer signaling records during the journey, resulting in lower accuracy rates. The overall accuracy rate for the five transport modes was 84.6%. In the same context, this study identified the mode of travel for mobile phone signaling data without an interest-point analysis.

For example, in the literature, the author used single signaling data for travel identification; however, the literature used different travel modes to construct affiliation functions such as average travel speed, travel duration, and distance based on travel a priori knowledge and fuzzy identification algorithms to conduct travel mode identification but did not provide corresponding identification results. Using the same dataset, experiments were conducted based on the recognition scheme and relevant parameters provided in the above literature. The results showed that the overall accuracy of the a priori knowledge-based travel mode fuzzy recognition algorithm was 68%, while the accuracy of the proposed travel mode recognition method based on mobile phone signaling and navigation data improved by more than 15%. The proposed method was more accurate than the a priori knowledge-based travel mode fuzzy recognition algorithm in both accuracy and completeness. CNN used two convolutional layers, each followed by ReLU and maximum pooling layers. The convolutional kernel in the convolutional layer was designed for the trip segment variable.

Table 7. Sample identification results.

Model	Test Accuracy (%)	Average Precision (%)	Recall (%)	F-Score (%)
SVM	65.3	65.3	65.3	65.3
DT	75.2	75.2	75.2	75.2
RT	79.8	79.8	79.8	79.8
MLP	79.8	79.8	79.8	79.8
CNN	75.7	75.2	75.6	74.8
Best CNN	84.7	86.3	82.4	83.9

presents the results of the validation of the other algorithms. Here, we chose widely-used methods to learn traffic patterns based on handcrafted features, including radial basis function-based SVM and DT. We compared CNNs with RF, which represented integrated algorithms, and multilayer perceptrons, which are regular and fully-connected neural networks, where all models were implemented and evaluated using scikit-learn. The performance quality of the boosted CNN models compared with the supervised learning algorithms listed above on similar test data included four classification metrics: test accuracy, average precision, average recall, and average F1-score. This comparison demonstrated the superiority of the proposed CNN model, with a test accuracy nearly 16% higher than the average test accuracy of the other methods. The test accuracy of the RF model, which was the best classical model, was only 2.5% higher than that of the original CNN model. Some researchers [51] have compared the performance of CNN models with other traditional machine learning methods, including Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) under different window sizes. The results of this research prove the superiority of the CNN model.

This study classified residents’ travel modes into walking, bicycle, private car, bus, and metro, and analyzed the travel characteristics of different categories based on mobile phone signaling data. First, the effective number of trips was obtained by analyzing residents’ travel trajectory data. Existing studies identified residents’ travel features using a CNN-based method, which identified the user’s place of residence by dividing the area range and extracting feature points. Here, we extracted and analyzed the mobile trajectory data of residents based on mobile phone signaling data provided by the base station and obtained rich and reliable user trajectory information. Existing research has shown that the multichannel CNN approach used could identify residents’ travel patterns. It improved the accuracy and efficiency of modeling the travel mode choice of residents in large cities under traffic congestion, as well as the ability to construct the behavioral features of travelers and provided a reference for mining the characteristics and patterns of travel behavior of residents in large cities. Simultaneously, this study designed a trajectory generation method. First, by designing a spatiotemporal threshold screening method to identify and deal with anomalies. Second, by combining the method of feature analysis to identify key points in signaling and using these key points to slice and dice the travel trajectory and generate travel sub-trajectory data. In the travel mode identification stage, the design introduced a road network to constrain the method of improving feature value calculation to solve the problem of poor positioning accuracy and then generated user travel features. In the travel mode identification stage, we designed road networks to improve the eigenvalue calculation method to solve the problem of low positioning accuracy and then generated a user travel feature dataset through eigenvalue calculation and designed unsupervised learning combined with empirical analysis to classify categories and identify the travel mode represented by each category. It was feasible to classify travel modes based on mobile phone signaling data, and different categories of travel modes were obtained by this model; the accuracy rate of this experiment exceeded 84.7%.

5. Conclusions

This study was based on a multichannel CNN approach to traffic travel patterns, and the results of pattern clustering under different traffic behaviors were analyzed. The method was analyzed based on actual travel data and showed that the method had a better recognition effect on road sections with higher traffic demand. The experimental results showed that the algorithm had high classification accuracy for datasets with large differences in feature distribution and was more sensitive to abnormal data, which better solved the problem of human markers being prone to errors. This study improved the basic CNN algorithm, which was simple in principle, easy to implement, and fast in convergence, and the experimental results showed that the algorithm improved clustering accuracy. Simultaneously, the effectiveness of the algorithm in improving the prediction accuracy was demonstrated by comparing the experimental results with other classical classification algorithms. The accuracy of the proposed algorithm still needs to be improved. In the future, we will extend new classification methods using various integration techniques and classification algorithms to improve the accuracy of classification. According to our analysis, the training data used for the experiment is not large enough, which is the main reason why we were not able to improve the accuracy of the test, hence this cell phone signaling dataset is yet to be augmented. Simultaneously, the method used and the mathematical model identified provide reference values for the quantitative prediction and classification prediction of similar problems. In this era of traffic informatization, the application of new technologies and means of managing traffic infrastructure will play an important role in improving the level of informatization and intelligence in traffic infrastructure management. Accelerating the construction of urban traffic informatization will improve the level of comprehensive urban traffic management and service capacity and allow the general public to enjoy the benefits and convenience brought about by constructing such traffic information. Mobile phone signaling data can provide real-time information on traffic travel patterns, helping urban planning departments better understand people’s travel patterns and habits. These data can be used to assess the load on the transport network, identify traffic bottlenecks and hotspots, and conduct transport planning and optimization. With a better understanding of travel mode distribution and demand, planners can plan roads, public transport routes, and facilities to improve the efficiency and availability of transportation systems. Simultaneously, the method can provide objective data support for government departments and decision-makers. Using data analysis and model predictions, they can formulate more accurate and smarter transport policies and measures, providing a scientific basis for a city’s transport management and future development.