Automatic Identification of the Working State of High-Rise Building Machine Based on Machine Learning

Abstract

High-rise building machines (HBMs) play a crucial role in the construction of super-tall buildings, with their working states directly impacting safety, quality, and progress. Given their extensive floor coverage and complex internal structures, monitoring priorities should shift according to specific workflows. However, existing research has primarily focused on monitoring key HBM components during specific stages, neglecting the automated recognition of HBM workflows, which hinders adaptive monitoring strategies. This study investigates the critical states of HBM construction across various structural layers and proposes a method rooted in vibration signal analysis to determine the HBM’s working state. The method involves collecting vibration signals with a triaxial accelerometer, extracting five distinct vibration signal features, classifying these signals using a k-Nearest Neighbors (kNN) classifier, and finally, outputting the results through a classification rule that aligns with the actual workflow of the HBM. The method was implemented in super-high-rise buildings exceeding 350 m, achieving a measured accuracy of 97.4% in HBM working state recognition. This demonstrates its proficiency in accurately determining the construction state and facilitating timely feedback. Utilizing vibration signal analysis can enhance the efficiency and safety, with potential applications in monitoring large-scale formwork equipment construction processes. This approach provides a versatile solution for a wide range of climbing equipment used in the construction of super-tall buildings and towering structures.

1. Introduction

High-rise building machines (HBMs) serve as essential high-altitude integrated operation platforms in the construction of tall buildings [1,2,3]. These machines are ideal for the efficient construction of concrete core structures exceeding 200 m, such as the Shanghai International Aviation Service Center (230 m), the Shanghai Sinar Mas Plaza (320 m) [4], the Shanghai Tower (632 m), and the Canton Tower (610 m) [5]. Equipped with various construction devices, such as tower cranes, concrete placing booms, and construction elevators, HBMs are designed to climb in sync with the vertical construction flow of the building structure, significantly enhancing the overall construction efficiency. However, similar to the large tower cranes under high-risk working conditions, there are also considerable safety hazards resulting from the complex composition of the equipment and the sophisticated processes inherent to the HBMs [6,7]. Consequently, HBMs are consistently a primary focus for monitoring at construction sites.

During the construction of a structurally standard vertical section of a concrete core, the HBMs typically experience four typical working states: climbing, operating, pumping, and idling. In each state, the functional state of the HBM components requires tailored monitoring and onsite management, each with distinct areas of emphasis. For instance, during the climbing state, it is essential to monitor unauthorized personnel on the HBM, potential interactions or collisions between the HBM and the core wall, and the posture of the HBM’s primary load-bearing platform. In the operating state, the focus shifts to the HBM’s load-bearing capacity and structural integrity. Meanwhile, the pumping state emphasizes monitoring potential collisions between the tower crane and the placing boom. Moreover, by analyzing the duration of the idling periods, the HBM’s idle rate can be assessed, enabling the more effective management of the construction progress.

Different safety monitoring measures must be adopted for the HBM components depending on their construction state. Automated recognition of the HBM’s working state is a prerequisite for implementing the corresponding management and safety warning measures. Therefore, to enhance the intelligence and construction efficiency of super-high rise concrete structures and to ensure safety during the construction process, it is crucial to differentiate and accurately identify the working states of the HBM.

Currently, a considerable amount of research has been carried out regarding the recognition and prediction of the status of workers, equipment, and materials at construction sites [8,9], employing various methods and sensors to facilitate the process [10]. In addition, different machine learning (ML) methods such as Artificial Neural Network (ANN) [11], Support Vector Machine (SVM) [12], Decision Tree (DT) [13], k-Nearest Neighbors (kNN) [14], and Random Forest (RF) [15] have been widely utilized in construction site management. These methods have had a significant impact on improving the efficiency [16], safety [17,18], and predictive capabilities [19]. For example, Hosoda et al. [20] enhanced an ANN model for predicting the maximum thermal crack width in reinforced concrete abutment walls, demonstrating the potential of ML in improving the reliability and safety of concrete construction systems. Awe et al. [21] and Gondia et al. [22,23] used ML to assess the factors affecting building construction collapses and to predict the severity of building construction injuries. Bugalia et al. [24], Cavalcanti et al. [25], and Hayat and Morgado-Dias [26] utilized ML for the automated classification of safety reports, accident prevention, and safety helmet detection, respectively. Heide and Petereit [27] illustrated the role of ML in advancing autonomous operations in the construction industry, especially in hazardous environments. Mammadov et al. [28] constructed an artificial intelligence-based model using various machine learning algorithms to predict occupational accidents in cross-border pipeline construction projects. Shuang and Zhang [29] utilized ML to predict the different types of accidents on construction sites. Toğan et al. [30] developed an automated ML system that automatically trains and evaluates different MLs to select the most accurate predictor of building accident severity.

However, concerning HBM research, most existing studies have primarily focused on formwork system selection [31,32], performance analysis [33,34,35], the monitoring of key points or special working conditions [36,37], and operational assessment [38,39]. The absence of the automated recognition of HBM workflows hinders the ability to adjust monitoring strategies in response to dynamic changes in work conditions. Although the HBM is equipped with a control system that can identify the pressure and travel of the cylinder—thereby determining its climbing state—it is not capable of identifying the operating, pumping, or idling states. In addition, the manual comparison of HBM’s working states is often limited by human factors and errors, and cannot guarantee a consistent quality output. Therefore, technical solutions are needed to meet the challenge of automatically and accurately detecting the working state of the HBM.

As a real-time and noninvasive condition detection method, vibration signal analysis is widely used in engineering fields, mainly for the recognition and diagnosis of the conditions of mechanical equipment [40,41,42,43]. In the field of vibration analysis and monitoring in machining and construction, Fu et al. [44] presented an automatic feature construction method for machining state monitoring based on vibration signals, using a deep belief network to reveal the relationship between input signals and output states. Lu et al. [45] conducted onsite monitoring and analysis of vibration influence laws during double-shield tunnel boring machine construction, providing a reference for the vibration effects on surface structures. Manikumar et al. [46] utilized an empirical mode decomposition technique to study the vibration severity of a spur gearbox, decomposing the raw signals into intrinsic mode functions and extracting statistical parameters for fault diagnosis. Meng and Zhu [47] developed an Internet of Things sensing system for monitoring construction-induced vibrations and assessing their impact using a microcomputer, a microelectromechanical systems accelerometer, and 4G communication with cloud storage. In the realm of wearable technology for construction safety and equipment monitoring, Kim et al. [48] created a safety-helmet-wearing management system for construction workers using a triaxial accelerometer, achieving high accuracy in differentiating wearing statuses. Signal processing and recognition with neural networks is another significant area of focus. Li [49] proposed a method for the accurate recognition of nonstationary mechanical vibration signals based on convolutional neural networks. Sherafat et al. [50] developed a multilabel, multilevel sound classification method using a short-time Fourier transform and convolutional neural network to recognize the activities of multiple types of heavy construction equipment simultaneously. Liao et al. [51] proposed an inner pipeline method for detecting the support conditions of buried gas subsea pipelines using forced vibration signal analysis, with tests demonstrating the method’s applicability.

In practical applications of vibration signal analysis, real-time signal recognition may produce erroneous results due to factors such as signal noise, feature selection, and classifier choice. Therefore, noise reduction processes [52,53], time-domain and frequency-domain feature selection [54,55], and the choice of appropriate classifiers are crucial steps influencing the final recognition accuracy.

For large structures like the HBM, minimizing sensor deployment for state recognition is economically advantageous. In such cases, vibration signal monitoring presents a more favorable solution.

Most of the vibrations in the HBM exist in the forms of structural response, self-excited vibration, or the vibration response caused by environmental factors. These include vibrations from construction personnel operations, asynchronous vibrations of hydraulic cylinder telescopic cylinders, vibrations in response to encountered climbing obstacles, vibrations due to mechanical equipment operation, local structural vibrations, and vibrations in response to overhead wind loads. The monitoring and diagnosis of vibration signals plays a crucial role in enhancing the stabilization of an HBM. Developing a monitoring and diagnosis system for the HBM working state provides a scientific basis for accident early warning and precise control, thereby improving the safety of the HBM and its associated equipment.

As a result, we propose a real-time recognition method for the working state of the HBM. Triaxial accelerometers collect vibration signals from the HBM’s main force components in real time. The time-domain and frequency-domain characteristics of these signals are analyzed, and ML techniques are employed to monitor and identify the HBM’s critical operating conditions during the construction of the concrete core tube. Drawing on the working characteristics of the HBM, the recognition results of real-time vibration signals are further optimized and classified, enhancing the accuracy of the final output state. This will enable management personnel to accurately comprehend the current state of construction equipment, achieve a comprehensive evaluation of high-rise building construction progress, and ensure the safety and efficiency of the construction process.

2. Methodology

2.1. Main Working States of HBM

During the construction of a standard vertical section of a concrete core, the four typical working states of the HBM are illustrated in Figure 1.

Table 1. Descriptions of the four states of a high-rise building machine (HBM).

State		Description
A	Climbing	The HBM utilizes its mechanical power system to move itself upwards or downwards.
B	Operating	The HBM actively performs routine tasks such as tying reinforcement bars, opening and closing formwork, and other related construction activities.
C	Pumping	The concrete placing boom, which is deployed on the HBM, is actively pumping concrete to different floors of the building during the construction process.
D	Idling	The HBM is stationary and not in use, with no personnel or equipment activities.

provides a detailed explanation of the work content corresponding to each state. These four working states of the HBM—climbing, operating, pumping, and idling—are respectively represented by A, B, C, and D.

The average construction time for a 4.5 m-height standard layer of a concrete core ranges from 5 to 7 days. Within this period, state A (climbing) occurs only once and lasts for a brief duration. The HBM primarily uses an automated power system, such as hydraulic cylinders, to facilitate this climb. Although the HBM typically takes about 2 to 3 h to climb a standard structural layer, the climbing state is the critical state for HBM monitoring due to its active nature during this phase.

State B (operating) is lengthy, sometimes extending into the night in addition to the usual daytime hours. During this phase, the HBM supports workers from various trades. These workers primarily perform detailed operations such as stacking and transferring materials, tying rebar, opening and closing the formwork, and other related construction activities.

State C (pumping) also occurs only once, generally lasting more than 6 h. During state C, the pipe of the transfer pump is directly connected to the concrete placing boom located on top of the HBM, causing it to vibrate with a large amplitude and a vibration frequency that is close to the pumping frequency of the concrete transfer pump. During this process, concrete compaction is critical. Workers usually vibrate the concrete with a vibrator to ensure that it fills the corners and spaces in the formwork, as well as around the rebar.

State D (idling) frequently occurs in the construction process. This state is mainly due to temporary stoppages for various reasons such as shift changes, breaks, maintenance, or waiting for the next phase of construction to begin. These pauses typically happen during lunch, dinner, and at night. In this idling state, both personnel and machinery activity on the HBM are minimal, and the vibration amplitude is at its lowest, with only background noise present. Idling is a necessary stage in the HBM construction process and is often considered a nonproductive or downtime period, and reducing the length of the idling stage can help to improve the construction efficiency and productivity.

2.2. Workflow of the State Recognition Method

The basic flow of the state recognition method is illustrated in Figure 2. This process involves several steps, including signal acquisition, feature extraction and training, real-time recognition, and result output.

The first step of this methodology involves the installation of triaxial accelerometers on the key loaded structures of the HBM to monitor its critical working states. Subsequently, the signals from the three directions are merged into a single group of signals, and any anomalous signals are filtered. The obtained sensor signals are then partitioned into multiple data segments. During the signal preprocessing stage, various features are extracted from each data segment either in the time-domain, frequency-domain, or time–frequency-domain.

Following this, a kNN model is constructed, and the feature information is used for training and testing the state recognition classifier. Finally, during the actual operation of HBM, the vibration signal features are collected in real time. Based on the kNN classifier results, the decision tree is utilized to classify and obtain the HBM working state.

2.3. Vibration Signal Data Collection

In this study, vibration signals are collected using a triaxial accelerometer. To capture these signals at each stage of construction, accelerometers are installed on the web of the H-beam that supports the HBM. The triaxial accelerometer is designed to record acceleration data as analog signals, represented by voltage values from the sensor. The collected raw signals are subsequently amplified and filtered, converting the analog quantity into a stable output signal. This output signal starts as a voltage value and is ultimately converted into a digital signal using an analog-to-digital converter.

The vibration signals are collected using a data acquisition system with a sampling frequency of 125 Hz. This system is connected to an accelerometer for vibration signal acquisition, storage, and preprocessing. For each reading, the accelerometer generates three data points, each corresponding to the sensor’s measured values in the X, Y, and Z directions. This data acquisition system ensures the accuracy and reliability of the collected vibration signals, thus providing a solid foundation for subsequent signal analysis and processing. Figure 3 depicts the installation position of the accelerometer and the measured values in four states.

The 125 Hz sampling frequency was chosen based on two main considerations. First, given that the premium frequency band of the HBM structure and its ancillary equipment is about 0.5–40 Hz, 125 Hz is an effective sampling rate according to the Shannon sampling theorem. This ensures accurate measurements by adequately sampling signals at at least twice their highest frequency. Second, the system’s reliance on wireless data transmission and measurement required a frequency that would minimize packet loss. Empirical evidence from the pre-tests shows that a sampling rate of 125 Hz kept the wireless packet loss rate below 0.5%. In contrast, higher rates exacerbated packet loss, threatening the integrity and reliability of our data.

Samples were collected at the four states of the HBM construction stages for classifier training, including the climbing, operating, pumping, and idling states. During the data collection stage, construction personnel followed their standard operating procedures to avoid any abnormal or irregular construction activities. Vibration signals were continuously collected throughout all the construction stages, with any abnormal vibration signals or events being recorded and marked for further analysis.

Reliable construction site datasets are important for utilizing machine learning methods [56]. To ensure high data quality, a preliminary analysis of the vibration signals was conducted to remove any outliers or noises. The segmented data were then used to extract the time-domain, frequency-domain, and time–frequency-domain features.

2.4. State Recognition Method

During the model training stage, the measured acceleration values in each state were divided into sections with a duration of $∆t$. To avoid interference from nearby factors such as pedestrian traffic near the sensor, an interference threshold $\mathsf{\epsilon }$ was set to remove readings with values greater than $\mathsf{\epsilon }$. The acceleration time history signals for the three directions were integrated into a unified time history signal, $R$, using Equation (1). $R$ is denoted by $\left\{R_i\right\}$, with $i$ = 1, 2, 3, …, $m$. Here, $m$ represents the number of acceleration data points obtained in each interval in a single direction of the triaxial accelerometer, and $∆t$ represents the sampling interval. $a_{xi}$, $a_{yi}$, and $a_{zi}$ represent the real-time readings in the three directions of the accelerometer, respectively.

$$R_i=\sqrt{a_{xi}^2+a_{yi}^2+a_{zi}^2}$$

(1)

Three time-domain features of $R$ were extracted, namely, the mean ($\overline{R}$), the root mean square ($R_{rms}$), and the peak-to-peak ($R_{pp}$). $\overline{R}$ was calculated using Equation (2), $R_{rms}$ was calculated using Equation (3), and $R_{pp}$ was calculated using Equation (4).

$$\overline{R}=\frac{1}{m}{\sum }_{i=1}^m R_i$$

(2)

$$R_{rms}=\sqrt{\frac{1}{m}{\sum }_{i=1}^m R_i^2}$$

(3)

$$R_{pp}=max\left(R_i\right)-min\left(R_i\right)$$

(4)

R was first mean-centered and then transformed into the frequency domain using the Fast Fourier Transform (FFT). The relative power spectral density $P_i$ was obtained according to Equation (5); then, the relative power spectral entropy $R_H$ was determined by Equation (6). This process used Shannon entropy [57] from information theory to measure the complexity or randomness of the frequency domain power distribution.

$$P_i=\frac{{\left|\mathrm{f}\mathrm{f}\mathrm{t}\left(R_i\right)\right|}^2}{\sum _{j=1}^{m/2}{\left|\mathrm{f}\mathrm{f}\mathrm{t}\left(R_j\right)\right|}^2},$$

(5)

$$R_H=\frac{-\sum _{i=1}^{m/2}{P}_i\mathrm{l}\mathrm{o}{\mathrm{g}}_2{P}_i}{\mathrm{l}\mathrm{o}{\mathrm{g}}_2(m/2)},$$

(6)

where $\mathrm{f}\mathrm{f}\mathrm{t}\left(R_i\right)$ represents the performance of the Fast Fourier Transform on $R_i$.

To improve the recognition accuracy of the climbing state and minimize the environmental noise, the offset axis crossing count, denoted as $R_{cc}$, was introduced. As represented in Figure 4, $R_{cc}$ was calculated as:

$$R_{cc}={\sum }_i I[(R_i-\delta )\times (R_{i-1}-\delta )<0]$$

(7)

where $I[.]$ is an indicator function that equals 1 if the condition inside the brackets is true and 0 otherwise.

The offset magnitude, $\delta$, was determined by:

$$\delta =l_c/L_{max},\delta \in \left(\mathrm{0,1}\right),$$

(8)

where $l_c$ represents the horizontal distance between adjacent hydraulic cylinders at the location of the triaxial accelerometer, and $L_{max}$ signifies the horizontal distance between the two furthest hydraulic cylinders of the HBM.

The combined feature vector $F=\left\{F_j\right\}$ of the $n$ time-domain signals $R$ can be expressed as:

$$F_j=\left[R_{rms},\overline{R},R_{pp},R_H,R_{cc}\right],$$

(9)

where $j$ = 1, 2, 3, …, $n$. The feature vector F is composed of five distinct features, which provide a comprehensive representation of the signal characteristics and enable effective signal classification and recognition of the working state of the HBM.

Utilizing the combined feature vector F for the four distinct working states, a kNN classifier was trained and subsequently deployed within the monitoring system of the HBM. The classifier was utilized to categorize the real-time collected data into distinct working state categories, thereby enabling the recognition of the construction state of the equipment. The kNN classifier works by finding the k-nearest neighbors of a test sample in the feature space and assigning the test sample to the class that is most common among its k-nearest neighbors. The classification formula for kNN is:

$$\begin{array}{c}y=\mathrm{a}\mathrm{r}\mathrm{g}\underset{c_j}{\mathrm{m}\mathrm{a}\mathrm{x}} \sum _{i=1}^k w_{i,j}[y_i=c_j]\end{array},$$

(10)

where y is the class label assigned to the test sample, $c_j$ is the j-th class label, $w_{i,j}$ is the weight assigned to the i-th nearest neighbor of the test sample that belongs to class $c_j$, and $[y_i=c_j]$ is the indicator function that is equal to 1 if $y_i=c_j$, and 0 otherwise. The optimal value of k, a hyperparameter, can be determined through cross-validation.

In the real-time recognition step, the HBM monitoring system collects a segment of the acceleration data in the time-domain of duration $∆t$. Based on the above steps, the current acceleration signal $R_c$ and the current feature vector $F_c$ were obtained. The kNN classifier was employed for real-time comparison to output the results of the A, B, C, and D states for $R_c$.

To enhance the reliability of state recognition, a conditional output step, grounded in the actual working condition of the HBM, was incorporated into the process (see Figure 5). A time length, $∆T$, was established such that $∆T$ > $∆t$. If the counts of state A, B, C, and D identified within the time period $∆T$ are N_SA, N_SB, N_SC, and N_SD, respectively, the working state of the HBM during this $∆T$ period can be determined. This determination is based on the actual working process of the HBM and the conditional relationships among N_SA, N_SB, N_SC, and N_SD. In Figure 5, N represents the total number of recognized results during the $∆T$ period, which is the sum of N_SA, N_SB, N_SC, and N_SD. The scaling factor of the number of recognized results is represented by b, with b $\in$ (0,1]. The value of b can be adjusted according to the actual conditions at the project site. In the final state output, an additional state, state E (the error state), was introduced to account for conditions that are not possible during the actual operation process of the HBM.

For instance, the climbing and pumping states of HBM are mutually exclusive and cannot occur simultaneously in a short time span. Therefore, if the climbing and pumping states are identified within the $∆T$ period (i.e., N_SA > 0 and N_SC > 0), the output result should be state E. Furthermore, considering the continuity of the HBM states, transitions between various states are relatively smooth. As a result, if the recognition outcome during the current $∆T$ period does not correspond to any of the five states A, B, C, D, or E, as delineated in Figure 5, the system outputs the recognition result from the preceding period. When the output results in state E, it is critical to provide clear and understandable error messages in order to increase the availability of the HBM and reduce downtime. In this case, a simple text description can be used to describe the nature of the error, e.g., “Error: Climbing and pumping statuses at the same time.” This will enable managers to quickly understand what the problem is and take appropriate action to resolve it, thus improving the user experience and the overall efficiency of the HBM’s work.

3. Case Study

3.1. Project Overview

The aforementioned method was implemented in the Shanghai Xujiahui Center (SXC) super-high-rise construction project. The SXC, comprising 70 floors, features a concrete core tube structure that reaches a height of 350.10 m. The concrete core structure was built using an HBM, equipped with two tower cranes, two placing booms, and a construction elevator. The HBM had 46 hydraulic cylinders, $l_c$ = 3 m, and $L_{max}$ = 29.35 m. A triaxial accelerometer was installed on the belly plate of the steel beam. This location was recognized as the main force plane of the construction equipment. Real-time vibration data were collected, as depicted in Figure 6. In the feature extraction and training step, the sensor operated at a sampling frequency of f = 125 Hz, and $∆t$ was set to 30 s. Consequently, a total of m = 3750 sampling points were generated in a single direction over each $∆t$ period.

3.2. Vibration Signal Analysis

Figure 7, Figure 8 and Figure 9, respectively, present the time-domain, power spectral density, and time–frequency spectrograms of a segment of $R_i$, measured over $∆t$, in states A, B, C, and D.

In state A, the vibration signal deviated noticeably from the vibration neutral axis, despite the relatively small amplitude and overall smooth changes. This deviation can be attributed to the asynchronous drive cylinder expansion and contraction, which makes the main stress platform of the HBM tilt at a certain angle. Furthermore, during the climb, the HBM frame may encounter localized wall protrusions, generating relatively large vibrations or shocks. These shocks occasionally peak at more than 2 gal and occur at relatively large intervals. The power spectrum value of the vibration signal was relatively small, and the time–frequency spectrogram showed a certain frequency distribution.

In state B, the amplitude of the vibration signal was relatively large, with a power spectrum that was more evenly distributed and with larger power spectrum values compared to the climbing state. This can be attributed to the impact, friction, and vibration generated during the workers’ knocking operation, the movement of the tower crane and construction hoist, and other ancillary equipment during the normal working conditions of the HBM, resulting in the production of the vibration signal. During the operating state, the activities of people and mechanical equipment on the HBM are more frequent, so the peak-to-peak intervals of such signals are smaller.

In state C, the amplitude of the vibration signal exceeded 20 gals, and there was a clear pulse waveform. The power spectrum value of the vibration signal was the highest among the four states, and a clear pulse signal was evident in the time spectrum diagram. This can be attributed to the rigid or semi-rigid connection between the HBM and the placing boom. As a result, the vibration signals from the concrete transfer pump during concrete pumping are transmitted to the main load-bearing platform of the HBM. This transmission occurs via the base of the placing boom or the pump pipe. The accelerometer is mounted on the steel beams of the platform and is sensitive to the vibration of the placing boom.

In state D, the amplitude of the vibration signal typically did not exceed 0.5 gals. Additionally, the power spectrum of the signal was evenly distributed across all frequency bands, with relatively low power spectrum values. The time–frequency spectrogram exhibited relatively smooth temporal variation without any significant frequency changes. This is because, in a no-load state, the various components of the HBM remain unchanged, resulting in minimal vibration or shock. The vibration signal is mainly affected by the wind load, environmental noise, and accelerometer noise. Thus, the frequency distribution of the vibration signal remained relatively uniform and changed smoothly.

In summary, the main characteristics of the HBM vibration signals include spectrum, noise, and multimodal mixing. The signal manifests differently in different situations. These differences may be attributed to variations in the motion state of the HBM, pressure changes in the hydraulic system, and other influencing factors. Among the four states, the peak and frequency characteristics of the vibration signals in states C and D were more apparent, while the differences in the vibration signals in states A and B were relatively small. Therefore, the difficulty in real-time recognition of the HBM’s working state through the triaxial accelerometer sensor lies in distinguishing between states A and B.

3.3. Feature Extraction and Selection

To compare the relevant data features, n = 53 segments of acceleration time series data were extracted for each state. Using Equations (1)–(9), the combination feature vectors of the 53 segments were obtained, where $\mathsf{\epsilon }$ = 200 and $\delta$ = 0.102.

Table 2. Feature values in four states.

State	$\overline{\mathit{R}}$	${\mathit{R}}_{\mathit{r}\mathit{m}\mathit{s}}$	${\mathit{R}}_{\mathit{p}\mathit{p}}$	${\mathit{R}}_{\mathit{H}}$	${\mathit{R}}_{\mathit{c}\mathit{c}}$
A	1.10	1.11	1.84	0.82	0
B	0.44	0.52	3.66	0.81	142
C	3.24	4.85	29.3	0.85	2
D	0.14	0.16	0.40	0.93	1255

displays the typical results of the feature value calculations corresponding to states A, B, C, and D in the time series data.

State C exhibited the highest $\overline{R}$ = 3.07 and $R_{rms}$ = 4.51 values, signifying a greater average acceleration and variability, respectively. In contrast, state D displayed the lowest $\overline{R}$ = 0.15 and $R_{rms}$ = 0.16 values, indicating lower average acceleration and lesser variability. This corresponds with the forced excitation in the pumping state and the small ambient vibration in the idling state.

The $R_{pp}$ values portray considerable variation across states, with state C demonstrating the highest $R_{pp}$ = 29.3, and state D was the lowest, $R_{pp}$ = 0.40. State D also showed the highest $R_H$ = 0.93, suggesting a more homogeneously distributed power spectrum.

The initial comparison between states A and B, considering the first four features, revealed only minor differences. However, the feature $R_{cc}$ introduced a significant divergence. In state B, the value of $R_{cc}$ was 142, while state A recorded an $R_{cc}$ value of zero. This suggests that state B had a smaller offset magnitude, as evidenced by the higher frequency of intersections between the vibration curve and the offset axis. On the other hand, state A demonstrated a larger offset magnitude, with no intersections occurring between the vibration curve and the offset axis over 30 s. Consequently, $R_{cc}$ becomes a key factor in differentiating between states A and B. Among all states, state D showed the highest $R_{cc}$ value of 1255, indicating that its vibration curve was very close to the offset axis. Therefore, $R_{cc}$ is also a significant feature for distinguishing state D.

Figure 10 shows the scatter plots of feature values for 53 segments in each state. In state A, the segments were distributed as $\overline{R}$ = 1.0–3.5, $R_{rms}$ = 1.0–5.0, $R_{pp}$ = 0.5–12.0, $R_H$ = 0.5–0.9, and $R_{cc}$ = 0. This wide range indicates the climbing state has a scattered distribution, and will likely be confused with other states.

State B segments were distributed as $\overline{R}$ = 0.2–1.0, $R_{rms}$ = 0.2–2.1, $R_{pp}$ = 0.7–38, $R_H$ = 0.7–1.0, and $R_{cc}$ = 36–408. Except for an outlier of $R_{pp}$ = 37.4 in Figure 10a, $R_{pp}$ was 0.7 to 14. This consistent distribution signifies the HBM operating state, distinguishable from others.

In state C, the segments were distributed as $\overline{R}$ = 1.4–4.0, $R_{rms}$ = 1.9–6.0, $R_{pp}$ = 11–40, $R_H$ = 0.5–0.8, and $R_{cc}$ = 0–16. This indicates that under the pumping state, the segments showed relatively high root mean square values, mean values, and peak-to-peak values, exhibiting a large variation range. In Figure 10b, the overlapping ranges of $\overline{R}$, $R_{cc}$, and $R_H$ make it difficult to distinguish between states A and C. However, Figure 10a shows that due to the differing $R_{pp}$ ranges, states A and C can be relatively well differentiated.

The segments of state D were distributed within the following ranges: $\overline{R}$ = 0.14–0.15, $R_{rms}$ = 0.15–0.17, $R_{pp}$ = 0.3–0.6, $R_H$ = 0.92–0.94, and $R_{cc}$ = 1230–1416. In state D, the small changes indicate low amplitudes and minimal fluctuation. While in Figure 10a, state D may be easily confused with state B, Figure 10b provides a clear distinction between state D and the other three states, due to the large range of $R_{cc}$ values.

3.4. State Recognition Performance

In this study, the combination feature vectors were trained by the kNN classifier (k = 7), as detailed in Equation (10). During the state recognition step of the HBM, the triaxial accelerometer signal, $R_c$, is acquired in real-time, capturing one signal segment per second. Similar to the training step, each segment had a duration of $∆t$ = 30 s and comprised 3750 points. However, these segments overlapped in time with their adjacent counterparts. Each of the four states yielded 16,128 signals, collectively amassing nearly 242 million data points. Figure 11 presents the scatter plots of the $\overline{R}$, $R_{rms}$, $R_{pp}$, $R_H$, and $R_{cc}$ for the 538 representative signal segments in each state. It can be seen that the four states had rather obvious clustering characteristics. However, compared with Figure 10, the distributions of the vibration signal feature values measured under the actual working conditions were wider, and the values were more widely cross-distributed, which poses challenges for accurate classification.

Table 3. Classification prediction results of the measured signal $R_c$.

State		Predicted State				True Positive Rate (TPR) (%)	False Negative Rate (FNR) (%)
		A	B	C	D
True state	A	14,883	233	1012	0	92.3	7.7
	B	5	16,063	60	0	99.6	0.4
	C	21	1364	14,743	0	91.4	8.6
	D	5	474	0	15,649	97.0	3.0

presents the results of the classification predictions for the four states (A, B, C, D) using a confusion matrix, along with their corresponding True Positive Rate (TPR) and False Negative Rate (FNR), following the application of the kNN classifier. The predictions for state A yielded a TPR of 92.3% and an FNR of 7.7%, with 14,883 instances correctly classified. State B displayed exceptional precision, achieving an almost perfect TPR of 99.6%. The model produced a TPR of 91.4% for state C, with 14,743 correctly predicted instances. Finally, state D had a TPR of 97.0%, with 15,649 true positives.

These results suggest that the selection of $\overline{R}$, $R_{rms}$, $R_{pp}$, $R_H$, and $R_{cc}$ as the feature vectors of vibration signals can lead to a more accurate classification of the four states of HBM. This classification method improves the overall recognition rate, indicating a more effective approach toward identifying the working state of HBM under general conditions. However, there are still areas for improvement. For example, in the pumping state of the HBM, 8.6% of the signals were misclassified as originating from the climbing or operating states.

To further enhance the overall recognition rate, adjustments were made in the result output step, taking into account the actual working conditions of the HBM ($∆T$ = 90 s, N = 90, and $b$ = 0.7, as shown in Figure 5). Figure 12 provides a before-and-after comparison of the output state results for A, B, C, and D. With these adjustments, the number of reported errors, as shown in Figure 12a, stood at 3174 out of all signal segments. However, in Figure 12b, the number of predicted error points decreased to 1662, a reduction of 52.4%. Consequently, the total recognition accuracy increased from 95.1% to 97.4%, representing an increase of 2.3%. Clearly, incorporating the actual working conditions into the output smoothed the results for each state and significantly curtailed the false positives. It is important to note that a time delay, $T_d$, arises when comparing the actual working conditions during different state transitions, as all recognition results within the time length of $∆T$ were assembled for output. The duration of this delay depends on the settings of $∆T$ and $∆t$, and in this case, $T_d$ is calculated as $∆T-1$, resulting in a time delay of 89 s.

In actual construction, the transitions between the four states of the HBM are natural, continuous, and not particularly distinct. Once a transition from one state to another occurs, it remains stable for a period without sudden changes. Therefore, the smooth state prediction results displayed at the bottom of Figure 12, devoid of sudden changes, align more closely with the actual application of the HBM. It is worth noting that when applying this method in real-time monitoring, more timely state recognition results can be achieved by judiciously selecting the values of $∆T$ to minimize $T_d$ as much as possible.

4. Discussion

4.1. Impact of Vibration Signals

In the practical application of the HBM, the construction process is brimming with noise, which can affect the acceleration measurements in each direction. This study demonstrates that by combining the three measurements, this noise can be suppressed to some extent, thereby improving the signal-to-noise ratio of the data. During the four key states of the HBM construction process, the focus was more on the magnitude of acceleration rather than its specific direction. In this context, using the modulus $R_i$ can directly provide the magnitude of acceleration, eliminating the need to consider the vibration in each direction separately. This approach only requires handling a comprehensive vibration value, which simplifies the data processing and accelerates the processing speed. Importantly, these methods are insensitive to noise generated by surrounding human activities, underscoring the practical potential of this approach.

Wind loads have a significant impact on the construction of tall buildings. The HBM is a massive unified steel frame structure, typically weighing over 500 tons. The HBM connects to the core concrete structure through multiple support points, ensuring a high degree of integrity so that its local vibration is less affected by wind loads. In the practical construction management of an HBM, a safe operational wind speed, such as 18 m/s, is typically established. The HBM can function normally within this wind speed, with measurements indicating that it exhibits a cooperative pulsation with the attached building structure. The vibration’s fundamental frequency triggered at this moment is significantly different from the fundamental frequency triggered when the HBM is operational. No frequency-domain coupling interval exists between the two, meaning that the wind-induced vibrations minimally impact the identification of the HBM’s working state. When the wind speed exceeds 18 m/s, the HBM cannot be climbed, and its ancillary equipment must cease operations. In this state, due to the lack of measured data, the effect of wind load on the vibration of the HBM requires further investigation.

4.2. Implications

The findings of this study present substantial practical implications for the construction industry. The proposed method not only addresses the unquantifiable and non-real-time monitoring issues inherent in human judgment, but also facilitates the quantitative evaluation and automatic recognition of an HBM’s working conditions. By identifying the various working states of the HBM, the study enables the implementation of timely corrective actions, reducing the potential negative impacts on construction safety and the quality of the final structure. In practical engineering projects, this study uses machine learning to track and classify HBM workflows in real-time, improving the safety and efficiency of working at heights and ensuring that projects are completed on schedule. Notably, the method of this study is not only applicable to HBMs but also extends to other climbing equipment, such as auto-climbing formwork systems [58] and slip formwork systems [59]. It can be utilized in a wide range of construction projects, including high-rise buildings and towering structures.

Moreover, this study provides valuable insights for stakeholders and policymakers. By using accelerometers to accurately identify the critical working condition of the HBM, specialized subcontractors can grasp the performance status of the HBM and undertake timely maintenance work. This proactive approach helps reduce both human and financial costs associated with workplace accidents, and ensures the smooth operation of the HBM. The general contractors can use the information to adjust the construction progress plan promptly, coordinate the work of labor teams with different construction equipment, minimize the construction process conflicts, reduce the project downtime, and enhance the overall efficiency of the project. The project owners can leverage the HBM state data to monitor the construction progress of the building’s main structure. With this knowledge, they can make more informed decisions about construction operations and allocate resources effectively.

4.3. Limitations

There were some limitations to this study. Firstly, the employment of only a single acceleration sensor posed a challenge in adequately capturing the localized construction signals from a large-scale HBM. Furthermore, the sensor could not detect the states of other construction equipment such as tower cranes, placing booms, and elevators. These factors potentially limit the scope and accuracy of the data collected.

Secondly, the study was limited to the identification of four types of work states using vibration signal data. This approach limits the granularity of state identification. For instance, the construction state could be further subdivided into processes such as tying rebar, and closing and opening formwork. However, these intricate processes cannot be identified solely through vibration signals. Employing different types of sensors to capture varied signal characteristics may be necessary for further refinement.

Thirdly, the study used the kNN classifier, which is only a practical application case based on HBM working characteristics. Although satisfactory accuracy was achieved, alternative classifiers or deep learning algorithms may provide even higher accuracy rates.

Lastly, the study only explored engineering applications for one type of HBM equipped with a hydraulic climbing system. It did not include HBMs with different power systems. For HBMs with varying power systems, the vibration signals during climbing might exhibit unique characteristics. Thus, the applicability of the proposed method requires further exploration in relation to these different systems.

4.4. Future Works

Future research could increase the numbers and diversify the types of vibration sensors [60] in different parts of the HBM at the construction site, as well as on auxiliary equipment. This could provide more comprehensive data and potentially improve the accuracy of state recognition.

In addition to acceleration sensors, other sensor types such as cameras [61] and jacking cylinder stroke and pressure sensors could be integrated to complement the vibration signals. These would provide a more comprehensive dataset for the ML model, thereby improving recognition accuracy. This could also help to identify the more nuanced operating states of the HBM, such as formwork opening/closing and rebar tying states.

Improved kNN classifiers [62,63] or other types of classifiers such as ANN and SVM could be considered. The integration of deep learning algorithms [64], which can provide higher classification accuracy, is also a promising direction.

The applicability and improvements of the method can be extended to HBMs with different dynamical systems and structural types, and the analysis of the effect of larger wind loads on HBM vibration can also be carried out to increase the versatility of the method. These initiatives could lead to significant advances in the monitoring of HBM working states.

5. Conclusions

Accurately distinguishing and recognizing the working states of HBM is fundamental to ensure the safe and efficient construction of high-rise concrete structures. However, the current research on HBM monitoring mainly focuses on special components or working conditions, and lacks the automatic identification of HBM workflow. In this study, we developed an ML-based method for detecting the four working states of an HBM—operating, climbing, pumping, and idling—using triaxial accelerometers and verified it in a super-high-rise building project. The main conclusions are summarized as follows:

(1)

The comparison of the vibration signal characteristics for the four states of the HBM was completed and showed that the amplitude and frequency differences between the vibration signals in the idling and pumping states were more pronounced and therefore easier to distinguish. Conversely, the amplitude and frequency of the vibration signals in the operating and climbing states were more similar to each other. As a result, we developed an offset axis crossing count feature tailored to the working characteristics of the HBM. This feature improves our ability to distinguish between the vibration signals occurring during the operating and climbing states;

(2)

Five typical features such as the mean, root mean square, peak-to-peak, relative power spectral entropy, and the offset axis crossing count values of a single triaxial accelerometer signal were used, which can better recognize the four key states of the HBM. A classification output rule based on the actual workflow of the HBM was established, which can further improve the accuracy of recognizing the working states of the HBM. The number of reporting errors under different states was significantly reduced, and the output results were smoother;

(3)

The method was practically applied in the SXC super-high-rise construction project, and the overall accuracy rate of the HBM working state recognition reached 97.4%, which shows the strong potential of the method in engineering applications. The method can effectively solve the problem of the lack of real-time monitoring of traditional formwork equipment in the construction process by monitoring and recognizing the four working states of an HBM in real time. This provides a basis for the accurate control and action execution of each HBM component, and for reducing the equipment vacancy rate, ensuring the safe operation of equipment and the quality of concrete structure construction.