1. Introduction
The horizontal curve belt conveyor has become one of the main pieces of continuous conveying equipment in the field of coal mine and other industrial production because of its characteristics of long distance, large capacity, and high-efficiency conveying. Different from the straight section belt conveyor, the conveyor belt in the curved section turns naturally under the combined force of tension, support force, centripetal force, and other forces. Research shows that the key to improving the material conveying efficiency is to reduce the belt deviation in the curve section [
1,
2,
3,
4,
5]. However, the deviation phenomenon typically occurs in conveyor belts of some practical projects due to the installation error of the conveyor, manufacturing error, material distribution characteristics, and idler failure [
6,
7]. The conveyor belt will cause the material to pour after it deviates. Friction between the conveyor belt and roller idler intensifies, and the service life of the conveyor belt roller and idler significantly reduces. Belt deviation may even trigger an emergency stop switch to reduce the delivery efficiency of the system [
8,
9]. The research shows that most of the faults of the horizontal curve belt conveyor are caused by the belt deviation first. Therefore, it is necessary to evaluate the deviation state of the conveyor belt at the curved section.
At present, deviation detection of the conveyor belt can be divided into two types: contact and noncontact. Contact deviation detection involves the installation of deviation sensors on racks at both sides of the conveyor belt [
10,
11]. The sensor will trigger the first-stage knob of the deviation stroke switch and send out an alarm signal when the conveyor belt touches it. Meanwhile, the sensor will trigger the second-stage knob and emergency stop switch button when the conveyor belt continues to deviate. The structure and principle of the deviation device are simple. Sensors are often affected by environmental factors and fail to work due to rust. The machine vision is typically utilized in investigating the noncontact deviation of the conveyor belt. Yang et al. [
12] used linear CCD to collect conveyor belt images and detected the edge of the conveyor belt using an image segmentation algorithm to obtain the deviation of the conveyor belt. Yang et al. [
13] proposed the gray average method to realize the fast segmentation of the conveyor belt and background. The change rule of the binary image was explored to realize the deviation detection of the conveyor belt. Wang et al. [
14] put forward an improved Canny operator detection and morphological processing to detect the edge of the conveyor belt and applied Hough linear detection to locate the conveyor belt. However, the image processing method is complex and impractical. Zhang et al. [
7] established a laser-aided method to judge the deviation of the conveyor belt. Although the method of conveyor belt deviation detection based on deep learning introduced by Liu et al. [
15] shows superior performance in reliability and anti-interference ability, its calculated deviation degree (DD) error is large. Zhang et al. [
16] solved the problem of fast extraction of the edge features of conveyor belt and analysis of the deviation of the conveyor belt under complex backgrounds by improving the target detection network YOLOv5. Based on the detection method, it can be further expanded.
Many scholars have introduced machine learning algorithms into the fault prediction of belt conveyor on the basis of fault detection technology to obtain relevant fault information from the running state of conveyor and forecast the future fault trend. Agata et al. [
17] carried out nondestructive testing of conveyor belts using the DiagBelt system and predicted the remaining life of the conveyor belt on the basis of the linear regression model. Hu et al. [
18] established the prediction model of belt conveyor faults on the basis of gray least squares support vector machines and applied it to the prediction and early warning of belt conveyor fault characteristic parameters. Liu et al. [
19] fused the information of conveyor running state, reliability estimation of idlers, and state detection data to realize the predictive maintenance of the belt conveyor system. Mei et al. [
20] put forward an image enhancement algorithm for detecting the edge of the conveyor belt, calculating the DD of the conveyor belt, and estimating and predicting the deviation of the conveyor belt.
To sum up, existing machine vision methods have been successfully used to detect the deviation of conveyor belts in terms of detection speed and performance. However, studies on the prediction and correction of conveyor belt deviation are limited, and ensuring the correction effect of conveyor belt deviation in the curve section is difficult. To ensure the efficient conveying of the conveyor belt at the curve section, this study establishes an evaluation system for the deflection state of the conveyor belt at the curve section based on the ARIMA–LSTM combined prediction model, evaluates the state of the conveyor belt and puts forward predictive suggestions. This study provides an efficient and intelligent solution for deviation detection, prediction, early warning, and deviation correction of the curve belt conveyor.
4. Establishment of the Prediction Model of Conveyor Belt Deviation
Prediction methods commonly used in machine learning include Prophet, RNN, ARIMA, and LSTM [
27,
28,
29,
30]. The Prophet model is commonly used for general trend prediction despite its low accuracy. The susceptibility of the RNN model to the phenomenon of gradient explosion in the process of calculating the gradient of training set leads to a significant difference between the predicted value of the conveyor belt deviation and the actual value. Thus, the RNN model fails to predict the conveyor belt deviation for a long time. However, ARIMA and LSTM present satisfactory accuracy and prediction performance in time series. Therefore, data on conveyor belt deviation in
Section 1 are used as training samples of the prediction model in this work, and a prediction model suitable for conveyor belt deviation based on the ARIMA and LSTM prediction models is established.
4.1. ARIMA Prediction Model
Deviation data present a nonlinear trend in the case of conveyor failure because the deviation of the conveyor belt is subject to manual intervention, component life, material characteristics, and other factors during the operation of the conveyor. The auto regressive integrated moving average (ARIMA) model presents excellent advantages in processing nonstationary time series data [
29]. The parameter p is the autoregressive (AR) term that represents the lag of time series in conveyor belt training concentration, because ARIMA (p, q, d) improves the ARMA (p, q) model. The parameter d is the difference term that represents the order of the time series that differed. The parameter q is the moving average (MA) term that represents the lag of the prediction error. The ARIMA prediction model forecasts the deviation value of the conveyor belt using the sum of the weighted sum of the deviation value indicated at the latest time and the prediction error at the latest time. ARIMA is expressed as follows.
where t is the time value,
is the AR coefficient,
is the MA coefficient, and
is the prediction error.
Deviation data (70%) of the conveyor belt at the tail of the curved
Section 1 are taken as the training sample set of the ARIMA prediction model, and the remaining 30% is used as the test sample set. Detecting the stability of the actual deviation training set of the conveyor belt is necessary before using the ARIMA prediction model. The mean standard deviation of the actual deviation training set and the unite root method determines whether the data are stationary time series. The results are shown in
Figure 7a. The rolling mean and standard deviation of the length fluctuates randomly up and down within a range. Therefore, the time series data set is initially considered stable. The unit root method is used to determine stable data, and the results are listed in
Table 4. The
p-value is significantly less than the critical values at 1%, 5%, and 10%. Therefore, processing the data smoothly is unnecessary because conveyor belt data in this experiment are stable at the 99% confidence interval.
Orders
p and q of the ARIMA prediction model are determined using tailing characteristics of the autocorrelation (ACF) and partial correlation (PACF) coefficient images of stationary time series, as shown in
Figure 7b, with
p = 1 and q = 1. Finally, the predicted value of the ARIMA prediction model is subjected to differential reverse reduction to obtain the actual idler length value, as shown in
Figure 7c. The predicted results of the ARIMA model for the idler length are nearly consistent with the measured data, but the underfitting state also ignores the abnormal deviation. Hence, the ARIMA prediction model is only suitable for forecasting linear data sets because data loss may occur when used in nonlinear data sets. Note that predicting the deviation of the conveyor belt with only a single ARIMA model is impossible.
4.2. LSTM Prediction Model
Long short-term memory (LSTM) is an improvement of RNN that solves the problem of gradient disappearance and explosion in long time series training [
25]. LSTM is calculated as follows.
where
is input gate,
is forgotten gate,
is cell status,
is output gate,
W is weight coefficient matrix,
b is offset,
is the sigmoid function, tanh is the tangent function, and
is the final prediction results.
LSTM determines the output value of each forward unit through Equations (47)–(51) and then reversely calculates the error of each unit. The gradient of the weight matrix is calculated according to the error calculated at the upper level. The weight matrix is optimized through the Adam optimization algorithm, and the iterative method is used for point-by-point prediction.
Deviation data of the conveyor belt (80%) at the tail of the curved
Section 1 are taken as the training sample set of the LSTM deviation prediction model, and the remaining 20% is used as the test sample set.
Figure 8a–c present the prediction results of conveyor belt deviation via LSTM, three-step LSTM, and multi-LSTM models, respectively.
Figure 8c illustrates that the multi-LSTM prediction model is unsuitable for determining the fitting degree and prediction effect and exhibits an over- or underfitting state. The fitting results of LSTM and three-step LSTM are consistent with the measured values. However, the weak local fitting effect of the short prediction sequences requires further improvement.
4.3. ARIMA–LSTM Combined Prediction Model Based on Series–Parallel Weighting
According to the prediction results in
Section 4.1 and
Section 4.2, the ARIMA and LSTM models can complete basic prediction with some defects. On this basis, the ARIMA and LSTM prediction model is recombined in this work using series-parallel weighting. The error term of the ARIMA model is revised and predicted through the LSTM model. The revised predicted value of the ARIMA model and the predicted value of the LSTM model are then weighted and summed via the parallel connection. Finally, the predicted value of conveyor belt deviation is obtained after modified combined weighting. The calculation method of the ARIMA–LSTM combined forecasting model is presented as follows.
where
is the predicted value of the combined model,
are the weighting coefficients of LSTM and ARIMA prediction models, and
.
and
are the predicted values of the LSTM and ARIMA prediction models at time t.
The prediction error of the combined prediction model is expressed as follows.
where
is actual value of belt deviation.
,
are the prediction errors of the LSTM prediction model and ARIMA prediction model at time t, respectively.
The squared value of the prediction error is calculated as follows.
The sum of squares of prediction errors of the combined model is assumed as follows.
The weighted coefficient vector of the combined prediction model is expressed as follows:
. The error matrix is presented as follows.
where
,
.
The sum of squares of prediction errors of the combined model is calculated as follows.
Unit column vector is as follows:
. The constrains of weighting coefficients are determined as follows.
If the error variance of the combined forecasting model reaches the minimum under the constraint condition, then its minimum value is expressed as follows.
Both sides of Equation (59) are multiplied by to solve under minimum variance, where .
Equations for solving the weighting coefficient and the sum of squares of prediction errors are expressed as follows.
The prediction deviation value of the conveyor belt at time t can be obtained by substituting into Equation (52). The ARIMA–LSTM combined prediction model based on series-parallel combined weighting is shown in
Figure 9, where X represents the input data of conveyor belt deviation, C represents the state of LSTM cells at the current time, H represents the hidden output of LSTM cells at the current time, and P represents the data output.
The composite model is completely established in a Python environment. Deviation data of the conveyor belt (80%) at the tail of the curved
Section 1 are taken as the training sample set of the combined prediction model, and the remaining 20% is used as the test sample set. The prediction results of the combined model are shown in
Figure 10. The fitting curve of the combined model is closer to the actual value, and the local prediction curve has not been fitted or under fitted. The combination prediction model retains the advantages of the ARIMA and LSTM models to predict the deviation trend of the conveyor belt in the curve section accurately.
4.4. Comparison of Fitting Degree of Different Prediction Models
Figure 11a shows the loss function of the ARIMA–LSTM combined model, as well as RNN and LSTM models. The error degree between the RNN model and the measured value of conveyor belt deviation is larger than that of LSTM and the ARIMA–LSTM combined model.
Figure 11b presents the fitting curve of the LSTM, ARIMA, and ARIMA–LSTM combined models with actual values.
Figure 11c illustrates the error between predicted and actual values of the ARIMA, LSTM, and ARIMA–LSTM combined models. The error of the ARIMA–LSTM combined prediction model is smaller than that of the ARIMA and LSTM models. The results showed that the ARIMA–LSTM combined prediction model achieves higher prediction accuracy and stronger generalization ability than the single prediction model.
4.5. Performance Comparison of Different Prediction Models
Willmott et al. [
31] demonstrated that mean absolute error (MAE) is a natural average error index that can describe the average error of model performance accurately. Chai et al. [
32] indicated that root mean squared error (RMSE) can highlight the advantages and disadvantages of the prediction model. RMSE is very sensitive to sudden changes in the deviation between predicted and actual values of conveyor belt deviation and can reflect the prediction accuracy of conveyor belt deviation. RMSE is the key index of the conveyor belt deviation prediction model. MAE represents the average absolute error between predicted and actual deviation values of the conveyor belt and the basic index for investigating the deviation of the conveyor belt. Compared with a single performance index, scholars typically choose the combined performance index to reflect the performance of the prediction model comprehensively. RMSE and MAE are selected as evaluation indexes in this work. RMSE and MAE of the prediction model are calculated using Equations (62) and (63), respectively.
where
is the actual belt deviation value,
is the deviation prediction value of different prediction models, and n is the number of sample time points.
Figure 12 shows the MAE and RMSE values for different prediction models. The combination models of LSTM, three-LSTM, and ARIMA–LSTM demonstrate high precision in predicting the conveyor belt deviation. However, the prediction trend of the ARIMA–LSTM combined model shows a high fitting degree with the actual deviation trend. Meanwhile, RMSE is larger than MAE because RMSE amplifies the error between the predicted value of conveyor belt deviation and the actual value, while MAE reflects the actual error value. Therefore, small RMSE and MAE values of the prediction model indicate an enhanced prediction performance.
Compared with the three-LSTM and LSTM models, the ARIMA–LSTM combined prediction model demonstrated a decrease in RMSE by 14.48% and 21.68% and MAE by 20.47% and 13.96%, respectively.
Table 5 shows the comparison of the average error rate and the total time of the ARIMA, LSTM, and ARIMA–LSTM prediction models. The average error rate of the ARIMA–LSTM combined prediction model is significantly lower than that of other single prediction models. The average error rate of the ARIMA–LSTM model is 24.34% higher than that of the LSTM model. The ARIMA–LSTM combined prediction model is suitable for predicting the deviation degree of the conveyor belt given its accuracy, fitting degree, time, and performance.
6. Conclusions
A state evaluation system of conveyor belt deviation based on the ARIMA–LSTM combined prediction model is established in this study to realize the state evaluation of the conveyor belt deviation in the curve section. The proposed system can predict and warn operators according to deviation detection data of the conveyor belt and provide intelligent and efficient suggestions for correcting the deviation angle adjustment. This work can help promote the efficient and intelligent development of the curved conveyor belt, reduce the production costs, and realize green and sustainable development. Compared with a previous study, the main contributions of this study can be summarized as follows.
(1) The correctable deviation range was established. The mechanical model of conveyor belt deviation in the curve section was established and verified in this work. The priority of angle adjustment of the idler frame was determined, and the range of deviation correction of the idler frame was established.
(2) The ARIMA–LSTM combined prediction model was established. The ARIMA and LSTM models were combined through the series-parallel weighted method and their prediction performance was compared. The results showed that RMSE and MAE values of the combined prediction model reduced by 14.48% and 13.96%, respectively. Compared with LSTM, the average error rate increased by 24.34%. The ARIMA–LSTM combined prediction model is suitable for the prediction of conveyor belt deviation in the curved section in terms of accuracy, fitting degree, time, and performance.
(3) The evaluation system of the conveyor belt deviation state in the curved section was established. The correct deviation range of the idler angle was combined with the ARIMA–LSTM combined prediction model in this system. The OCSVM algorithm was used to detect the abnormal deviation of the conveyor belt accurately and efficiently. A visual interactive interface was then established to realize the application and visual fusion of the combined prediction model in the correctable deviation range of the idler frame. The system can predict the belt deviation according to deviation detection data, send out a warning signal according to the prediction results, and provide the corresponding angle adjustment suggestion.
(4) The method in this study was verified by an experimental platform. The experimental results show that the method based on data prediction is feasible for the curved conveyor belt condition evaluation. This study provides an efficient and intelligent solution for the state evaluation of the curve belt deviation.