Truck dynamic weighing method and system based on convolutional neural network

Abstract

The invention provides a truck dynamic weighing method and system based on a convolutional neural network, and the method comprises the steps: screening features highly related to weight through employing a Pearson's correlation coefficient method, and introducing two new features, namely, six-stage temperature and impact times, to compensate a model; therefore, the resistance of the model to the complex environment interference of the actual site is enhanced. Secondly, in the convolutional neural network model, by adding a noise reduction auto-encoder DAE, deep feature representation in the vehicle passing signals can be extracted layer by layer, and high-quality input features are provided for subsequent convolutional neural network prediction; meanwhile, a label distribution smoothing strategy is introduced, and a kernel density estimation method is used for weighting a loss function of the model, so that the influence of a low-density sample in model training can be improved; the problem that when a prediction model aims at minimizing the sum of prediction errors of all training samples, the proportion of samples in a low-density area is small, and consequently the model tends to learn samples in a high-density interval is solved.

Description

Technical Field

[0001] The present invention relates to a method for dynamic weighing of trucks. Specifically, it relates to a method and system for dynamic weighing of trucks based on a convolutional neural network. Background Art

[0002] Highway transportation plays a dominant role in the transportation system, and the overloading problem has always been a major challenge in the management of highway truck transportation. At present, truck weight detection mainly adopts two methods: static weighing and dynamic weighing. Although static weighing has high accuracy, its equipment scale is large and it is easy to cause traffic congestion. In contrast, dynamic weighing equipment is usually installed under the road surface and does not affect the normal passage of trucks. Therefore, it has become a key technology for solving truck road management problems in recent years. The dynamic weighing technology can automatically and real-time collect sensor waveform data during the driving process of trucks, process these data, obtain parameters such as vehicle speed and axle weight, and then use algorithms to predict the actual weight of trucks. Although the dynamic weighing system has the advantage of better management, it still has the problem that its accuracy is lower than that of the static weighing system.

[0003] With the booming development of artificial intelligence technology, machine learning algorithms have become increasingly mature and have shown remarkable results in the field of dynamic weighing. Currently, most of the research on truck dynamic weighing model algorithms focuses on the research of machine learning models adapted to the same site. Since it is difficult to collect label data of trucks at actual sites, whenever a new site is added, it is impossible to spend a lot of time obtaining training samples. Therefore, the research on algorithm models adapted to different sites has become particularly important.

[0004] In addition, in actual application scenarios, the measurement accuracy of the truck dynamic weighing system is easily affected by the coupling of multi-source interference: due to factors such as differences in sensor installation technology, road surface slope fluctuations, environmental noise, and road surface flatness at different sites, data distribution offsets occur when the same truck passes through different sites. Especially in the case of long-term operation, problems such as sensor fatigue and temperature changes will be further aggravated, thus affecting the prediction accuracy. These problems have brought challenges to the accuracy and stability of dynamic weighing.

[0005] To address the above problems, it is necessary to propose a general model with high robustness that can achieve real-time prediction of the weights of trucks at different sites. Summary of the Invention

[0006] The purpose of the present invention is to address the deficiencies of the prior art, and thus provide a method and system for dynamic weighing of trucks based on a convolutional neural network, which can be used for dynamic automatic weighing of trucks at different dynamic weighing sites.

[0007] To achieve the above purpose, the technical solution adopted by the present invention is:

[0008] The first aspect of the present invention provides a dynamic weighing method for trucks based on a convolutional neural network, comprising the following steps:

[0009] Collect the passing vehicle data of the truck, perform feature processing on the passing vehicle data to obtain pressure data features;

[0010] Send the pressure data features, six-stage temperature features, and impact frequency features into a convolutional neural network model for prediction to obtain the predicted weight of the truck;

[0011] Among them, the six-stage temperature feature is to first arrange the collected passing vehicle data in chronological order, divide 24 hours of a day into six time periods, namely 0:00 - 4:00, 4:00 - 8:00, 8:00 - 12:00, 12:00 - 16:00, 16:00 - 20:00, 20:00 - 24:00, and then perform feature marking on the passing vehicle data falling into the corresponding time periods; among them, 0:00 - 4:00 is marked as 1, 4:00 - 8:00 is marked as 2, 8:00 - 12:00 is marked as 3, 12:00 - 16:00 is marked as 4, 16:00 - 20:00 is marked as 5, and 20:00 - 24:00 is marked as 6;

[0012] The impact frequency feature is the total cumulative number of passing vehicles on the weighing platform lane within the corresponding time period.

[0013] Based on the above, the method for obtaining pressure data features is as follows:

[0014] The weighing platform includes 2 weighing platforms corresponding to the left and right wheels of the truck during driving, and 2 weighing sensors are arranged along the advancing direction of the truck for each weighing platform; the output data of the weighing sensors is used as the passing vehicle data;

[0015] For the collected passing vehicle data, select the average peak value, peak value, compensation value, time difference, and passing vehicle speed after each tire passes through each weighing sensor as feature data; among them, the average peak value is the arithmetic mean of 100 pressure values before and after the current sensor peak value, the compensation value is the pressure value of the sensor in front of the current sensor; the time difference is the time difference between the current sensor peak value and the first sensor peak value of this vehicle;

[0016] Summarize the feature data of the passing vehicle data of each vehicle and merge them in chronological order;

[0017] Use the Pearson correlation coefficient method to screen out the features that are most critical to the prediction result. Among them, the features with a Pearson correlation coefficient > 0.4 are selected, including the front average peak value, rear average peak value, front peak value, and rear peak value of weighing platforms one and two, as the final input features of the model.

[0018] Based on the above, the convolutional neural network model includes:

[0019] The DAE loading module includes an input data processing module, a model forward propagation module, a loss function calculation module, and a backpropagation and parameter update module;

[0020] The input data processing module is configured to: use the eigenvalue screened by the Pearson correlation coefficient and normalized as the input, and add Gaussian noise to the input data;

[0021] The model forward propagation module is configured to: map the data with Gaussian noise to reconstructed data through an encoder-decoder and add a BN layer;

[0022] The loss function calculation module is configured to: use the mean square error as the loss function to measure the gap between the reconstructed data and the original data;

[0023] The backpropagation and parameter update module is configured to: update the parameters to minimize the mean square error and iteratively obtain the optimal parameters of the DAE module;

[0024] The CNN module includes an input and output data module, a stacked multi-layer convolution module, a fully connected layer module, a weighted loss function calculation module, and a backpropagation and parameter update module;

[0025] The input and output data module is configured to: add the features obtained from the DAE loading module, the normalized impact number feature, and the six-stage temperature feature, concatenate them into an array by column, and use it as the input of the CNN module, and use the normalized actual weight as the label input;

[0026] The stacked multi-layer convolution module is configured to: use the value of each layer of convolution operation as the input of the next layer of convolution; it includes a stacked multi-layer convolution and a fully connected layer module; each layer of convolution includes a one-dimensional convolution module, an activation module, and a pooling module;

[0027] The one-dimensional convolution module is configured to: perform a convolution operation on the input data and the convolution kernel to generate a feature map;

[0028] The activation module is configured to: perform a non-linear transformation activation on the feature map generated by the convolution;

[0029] The pooling module is configured to: use max pooling on the feature map after non-linear transformation activation to capture the most valuable signal features;

[0030] The fully connected layer module is configured to: flatten the feature map generated by the stacked multi-layer convolution, and perform a fully connected output after flattening;

[0031] The weighted loss function calculation module is configured to: use the mean square error as the loss function to measure the gap between the reconstructed data and the original data; adopt a label distribution smoothing strategy to assign different weights to each sample in the training set during the loss function calculation;

[0032] The backpropagation and parameter update module is configured to update the parameters to minimize the mean square error and iteratively obtain the optimal parameters of the CNN module.

[0033] The second aspect of the present invention provides a truck dynamic weighing system based on a convolutional neural network, including

[0034] The pressure data feature acquisition module is configured to collect the passing vehicle data of the truck, perform feature processing on the passing vehicle data, and obtain pressure data features;

[0035] The weighing prediction module is configured to send the pressure data features, six-stage temperature features, and impact frequency features into a convolutional neural network model for prediction to obtain the predicted weight of the truck;

[0036] Among them, the six-stage temperature feature is to first arrange the collected passing vehicle data in chronological order, divide 24 hours of a day into six time periods, namely 0:00 - 4:00, 4:00 - 8:00, 8:00 - 12:00, 12:00 - 16:00, 16:00 - 20:00, 20:00 - 24:00, and then perform feature marking on the passing vehicle data falling into the corresponding time period; among them, 0:00 - 4:00 is marked as 1, 4:00 - 8:00 is marked as 2, 8:00 - 12:00 is marked as 3, 12:00 - 16:00 is marked as 4, 16:00 - 20:00 is marked as 5, and 20:00 - 24:00 is marked as 6;

[0037] The impact frequency feature is the total cumulative number of passing vehicles on the weighing platform lane within the corresponding time period.

[0038] The third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the aforementioned truck dynamic weighing method based on a convolutional neural network.

[0039] The fourth aspect of the present invention provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the aforementioned truck dynamic weighing method based on a convolutional neural network.

[0040] The present invention has prominent substantial features and significant progress compared with the prior art. Specifically:

[0041] By using the Pearson correlation coefficient method to screen the features highly correlated with weight from the collected multi-dimensional passing vehicle weight-related feature data and introducing two new features, six-stage temperature and impact frequency, to compensate the model, the present invention can enhance the resistance of the model to the interference of the complex environment of the actual site and ensure the accuracy of vehicle weight measurement.

[0042] In the convolutional neural network model of the method of the present invention, by adding a denoising autoencoder DAE, deep feature representations in the passing vehicle signals can be extracted layer by layer, providing high-quality input features for subsequent convolutional neural network prediction; at the same time, by introducing a label distribution smoothing strategy and using the method of kernel density estimation to weight the loss function of the model, the influence of low-density samples in model training can be increased to balance the processing of feature sample data with extremely unbalanced vehicle weight differences; when the prediction model aims to minimize the sum of prediction errors of all training samples, the problem that the proportion of samples in the low-density region is small, resulting in the model being more inclined to learn samples in the high-density interval, can be solved. Brief Description of the Drawings

[0043] Figure 1 is a composition diagram of the truck dynamic weighing system.

[0044] Figure 2 is the main body of the weighing plate and the installation diagram

[0045] Figure 3 is the H10B resistance strain gauge sensor.

[0046] Figure 4 is the flowchart of the truck dynamic weighing.

[0047] Figure 5 is the flowchart block diagram of the truck dynamic weighing method based on the convolutional neural network in Embodiment 1.

[0048] Figure 6 is the waveform diagram of the passing vehicle data of a 6-axle truck in the present invention.

[0049] Figure 7 is the peak value and compensation value obtained in the present invention.

[0050] Figure 8 is the time difference obtained in the present invention.

[0051] Figure 9 is the top view of the truck passing through the main weighing plate (abbreviated as weighing platform).

[0052] Figure 10 is the data analysis diagram in the present invention. Detailed Embodiments

[0053] The technical solutions of the present invention will be further described in detail below through specific embodiments.

[0054] Dynamic weighing system

[0055] The truck dynamic weighing system used in the present invention mainly includes: the main body of the weighing mainboard, weighing sensors, an information processing system, and a host computer. The composition of the truck dynamic weighing system is asFigure 1 as shown

[0056] (1) Weighing plate body

[0057] The weighing plate body adopts an integral structure design, and the load-bearing panel and the side force structure part are completely solidified. The weighing plate body covers the entire cross-section of the detection section, enabling blind-free detection of the entire section. The weighing plate body is parallel to the highway pavement and maintains the same horizontal plane, without affecting the normal passage of trucks. The weighing plate body is as Figure 2 shown, and the specific parameters are shown in Table 1

[0058] Table 1 Weighing plate body and specific parameters

[0059]

[0060] (2) Weighing sensor

[0061] The weighing sensor uses an H10B-D0.2-20t-25T-016 resistance strain type sensor. By measuring the change in resistance, the strain of an object can be calculated, and then the force or pressure exerted on the object can be deduced. The resistance strain type sensor used is as Figure 3 shown, and the specific parameters are shown in Table 2

[0062] Table 2 Resistance strain type sensor parameter table

[0063]

[0064] (3) Information processing system

[0065] The information processing system includes a measurement circuit and information processing. The main modules it includes are: signal collector, signal amplifier, filter, and A / D conversion module. Since the deformation of the weighing platform is small during dynamic weighing, the collected voltage signal is usually at the millivolt level. Therefore, an amplifier is needed to amplify the signal. The filter is used to remove high-frequency noise generated by trucks and sensors, and the A / D conversion module converts the filtered analog signal into a digital signal. Finally, the signal is transmitted to the upper computer through a data acquisition card for further processing and algorithm operations

[0066] (4) Upper computer

[0067] The upper computer is responsible for receiving signals from the data acquisition card, storing, processing, and analyzing them, and calculating information such as the weight, speed, and number of axles of the truck. It displays the processing results to the operator in real-time and provides feedback

[0068] The working principle of the weighing system is that sensors installed on the lane sense the driving state of the truck and convert the sensed pressure changes into electrical signals. These signals are processed and then transmitted to the host computer through a data acquisition card. Finally, the results are presented to the operator through calculation and display devices, thus completing the dynamic weighing process. In this process, the camera, as an auxiliary device, is responsible for recording and identifying truck information. Figure 4 It is the flowchart of truck dynamic weighing, and the truck dynamic weighing process is as follows:

[0069] a. The truck enters the weighing area: The truck drives into the lane equipped with the weighing system at a normal speed and enters the area where the sensors are laid;

[0070] b. The sensor senses the pressure change: The resistance strain type sensor laid in the weighing main board senses that the truck enters the weighing area. The pressure received by the sensor changes, and these pressure changes are converted into voltage signals;

[0071] c. Signal processing: The voltage signal output by the sensor undergoes amplification, filtering, and A / D conversion by the signal processing circuit. The purpose is to convert the analog signal into a digital signal for further processing;

[0072] d. Data acquisition and transmission: The processed digital signal is transmitted to the host computer through a data acquisition card;

[0073] e. Data analysis and processing: The host computer receives and processes the data from the sensor, and calculates the weight of the truck through a preset algorithm;

[0074] f. Output of weighing result: After calculation by the algorithm, the host computer obtains the real-time weight information of the truck and outputs the result through a display or other devices. The operator can make judgments or further processing based on the displayed information;

[0075] g. The truck leaves the weighing area: After the truck completes weighing, it continues to drive and leaves the weighing area;

[0076] The whole process is completed.

[0077] Embodiment 1

[0078] This embodiment provides a truck dynamic weighing method based on a convolutional neural network, as Figure 5 shown, including the following steps:

[0079] Step 1, collect the passing vehicle data of the truck, perform feature processing on the passing vehicle data to obtain pressure data features; at the same time, obtain six-stage temperature features and impact times features.

[0080] Specifically, the methods for pressure data features, six-stage temperature features, and impact times features are:

[0081] 1. When the truck passes through the weighing platform, the sensor generates a voltage signal. After signal amplification, the ADC of the signal collector samples to obtain the pressure waveform data generated when the truck passes (as Figure 6 shown). The waveform data is represented in hexadecimal form and consists of 816 bytes in total. Table 3 shows the waveform data format collected by the weighing platform. Among them, the first 16 bytes are the data header of the passing vehicle data, and the last 800 bytes are the data field of the passing vehicle data. The data field consists of 8 channels, and each channel is further divided into a high byte and a low byte. The combined data of the high byte and the low byte is the pressure value.

[0082] Table 3 Waveform data format collected by the weighing platform

[0083] Offset Length Field Name Field Description 0 4 Frame Header {57415645}, 'WAVE' indicates waveform data 4 1 Frame Sequence Number 0 to 255, incrementing, looping 5 1 Collector ID {1,8} 6 1 Number of Channels {1～8} 7 1 Data Format 2 - represents 2 - byte signed integer data 8 8 UNIX Timestamp Milliseconds since January 1, 1970 (little - endian) 16 N Data Field

[0084] 2. Export the collected valid passing vehicle data. According to the above data format, combine the high and low bytes and convert them into decimal pressure data.

[0085] 3. According to the pressure data, select the characteristic values of the sequence data that can reflect the weight of the truck. The characteristic values selected by this method are: the average peak value, peak value, compensation value, time difference, and passing vehicle speed after each tire passes through each weighing sensor. Among them, the average peak value is the arithmetic mean of 100 pressure values before and after the peak value of the current weighing sensor; the compensation value is the pressure value of the weighing sensor in front of the current weighing sensor; the time difference is the time difference between the peak value of the current weighing sensor and the peak value of the first weighing sensor of this vehicle, as Figure 7 and Figure 8 shown.

[0086] 4. As Figure 9 shown, the 2 weighing platforms respectively correspond to the left and right wheels of the truck. Each weighing platform has 2 weighing sensors. The weighing sensor that first contacts the wheel is the front weighing sensor, and the weighing sensor that finally contacts the wheel is the rear weighing sensor. Summarize the characteristic data of the passing vehicle data of each vehicle and combine them in chronological order. For example, the arrangement order of the passing vehicle characteristic data of a six-axle vehicle is:

[0087] The front average peak value 1, front peak value 1, front compensation value 1, front time difference 1... front average peak value 6, front peak value 6, front compensation value 6, front time difference 6 of weighing platform 1, and the rear average peak value 1, rear peak value 1, rear compensation value 1, rear time difference 1... rear average peak value 6, rear peak value 6, rear compensation value 6, rear time difference 6;

[0088] The front average peak 1, front peak 1, front compensation value 1, front time difference 1... front average peak 6, front peak 6, front compensation value 6, front time difference 6, rear average peak 1, rear peak 1, rear compensation value 1, rear time difference 1... rear average peak 6, rear peak 6, rear compensation value 6, rear time difference 6, and passing vehicle speed;

[0089] Among them, the front peak 1 is the peak value output by the front weighing sensor of the weighing platform for the first wheel of the truck; the rear peak 6 is the peak value output by the rear weighing sensor of the weighing platform for the sixth wheel of the truck. Finally, a total of 97 features are selected.

[0090] 5. For the selected 97 features, the Pearson correlation coefficient method is used to screen out the features that are most critical to the prediction result and eliminate redundant variables. Among them, the calculation formula of the Pearson correlation coefficient is as follows:

[0091]

[0092] Among them, is the average value of a certain feature, is the average value of the sample label.

[0093] Finally, the features with a Pearson correlation coefficient > 0.4 include the front average peak, rear average peak, front peak, and rear peak of weighing platforms one and two, a total of 48 features, which are used as the final input features of the model. Specifically, these 48 features are:

[0094] The front average peak 1, front average peak 2... front average peak 6 of weighing platform one, rear average peak 1, rear average peak 2... rear average peak 6;

[0095] The front peak 1, front peak 2... front peak 6 of weighing platform one, rear peak 1, rear peak 2... rear peak 6;

[0096] The front average peak 1, front average peak 2... front average peak 6 of weighing platform two, rear average peak 1, rear average peak 2... rear average peak 6;

[0097] The front peak 1, front peak 2... front peak 6 of weighing platform two, rear peak 1, rear peak 2... rear peak 6.

[0098] 6. Due to factors such as sensor fatigue and temperature changes, within a long time interval, even if the same vehicle passes through the sensor at the same passing vehicle speed, the collected passing vehicle data will be different. Therefore, this will lead to a problem that the prediction accuracy of the algorithm decreases after a period of time, thus having a greater impact on the accurate prediction of the passing vehicle weight. Based on this, the collected data is arranged in chronological order. Since each average peak actually represents the axle weight of the truck, the ratio obtained by dividing the sum of all average peaks and cumulative values by the true weight of the truck is used as the analysis data, and its result is visualized, such asFigure 10 As shown. It can be found from the figure that the data shows a periodic change pattern, and as time goes by, the periodic pattern of the error becomes more obvious and shows a gradually increasing trend. To more intuitively display this pattern, (a) is enlarged to (d), and data from two randomly selected days is visualized and shown in (b) and (c) respectively. Through analysis, it can be seen that the periodic decrease and increase mainly occur between 10:00 and 18:00. Specifically, the error starts to decrease around 10:00, reaches the trough around 14:00, and then the value increases until 18:00. This phenomenon may be related to the high temperature at noon, indicating that temperature has an impact on the measurement error. In addition, as time goes by, the overall value shows an increasing trend and the distribution becomes more scattered. Further analysis reveals that this may be related to the deformation of the weighing platform mechanical structure caused by the increase in the number of passing vehicles.

[0099] The performance of the model is enhanced by introducing two new features, namely the six-stage temperature feature and the impact number feature. Specifically, the six-stage temperature feature is used to replace the actual temperature feature, and the impact number feature is used to simulate the deformation of the weighing platform mechanical structure. By adding these two features to the convolutional neural network model for training, it is possible to better capture the influence of temperature changes and mechanical deformations on the model, thereby improving the prediction accuracy of the model and effectively alleviating the problem of model accuracy decline.

[0100] Among them, for the six-stage temperature feature, the collected passing vehicle data is first arranged in chronological order, and the 24 hours of a day are divided into six time periods, namely 0:00 - 4:00, 4:00 - 8:00, 8:00 - 12:00, 12:00 - 16:00, 16:00 - 20:00, 20:00 - 24:00. Then, the passing vehicle data falling into the corresponding time period is marked with features; among them, 0:00 - 4:00 is marked as 1, 4:00 - 8:00 is marked as 2, 8:00 - 12:00 is marked as 3, 12:00 - 16:00 is marked as 4, 16:00 - 20:00 is marked as 5, and 20:00 - 24:00 is marked as 6;

[0101] The impact number feature is the total cumulative number of passing vehicles on the weighing platform lane within the corresponding time period.

[0102] 7. Normalize the processed passing vehicle data, where x s is the normalized data, x is the original data, x min is the minimum value of the column where the original data is located, and x max is the maximum value of the column where the original data is located. Among them, the processed passing vehicle data here refers to the features selected by the Pearson correlation coefficient method, the six-stage temperature feature, and the impact number feature.

[0103] Step 2, model training;

[0104] The pressure data features, six-stage temperature features, and impact frequency features are fed into a convolutional neural network model for prediction to obtain the predicted weight of the truck.

[0105] The specific convolutional neural network model is as follows:

[0106] (1) Training of the DAE (denoising autoencoder) loading module

[0107] 1. Input data processing module: The feature values after being screened by the Pearson correlation coefficient and normalized are used as the input, and Gaussian noise is added to the input data to train the DAE loading module to recover clean original data from the noisy data, and it is restricted that after adding Gaussian noise, the data is still between 0 and 1. obeys the standard normal distribution, α is the noise factor for adjusting the noise intensity, clip [0,1] (1) Limit the result within the interval [0, 1].

[0108] X’ = clip [0,1] (X + α·∈),

[0109] 2. Model forward propagation module: The data with Gaussian noise is mapped to reconstructed data through the encoder-decoder, and a BN layer is added to accelerate the training of the DAE loading module and improve stability;

[0110]

[0111] where, μ B is the mean of the input data, σ B 2 is the variance of the input data, ε B is a very small constant value.

[0112] 3. Loss function calculation module: The mean square error (MSE) is used as the loss function to measure the gap between the reconstructed data and the original data;

[0113]

[0114] where, n is the batch size.

[0115] 4. Backpropagation and parameter update module: Continuously update the parameters to minimize the mean square error, and continuously iterate to obtain the optimal parameters, and the DAE training ends;

[0116]

[0117] where, γ is the learning rate, is the partial derivative of L DAE with respect to.

[0118] (2) Training of the CNN (Convolutional Neural Network) Module

[0119] 1. Input and Output Data Module: The features obtained from the DAE loading module are concatenated with the normalized number of impacts and the six-stage temperature column by column to form an array, which is used as the input of the CNN module. The normalized actual weight is used as the label to train the CNN model.

[0120] 2. Stack multiple convolutional modules, configured as: using the convolutional operation value of each layer as the input of the next layer; including stacking multiple convolutional and fully connected layer modules; each convolutional layer includes a one-dimensional convolutional module, an activation module, and a pooling module;

[0121] 2.1 One-dimensional Convolutional Module: Convolve the passing vehicle input data with the convolutional kernel to generate a feature map;

[0122]

[0123] Among them, is the output obtained by the input passing through the i-th convolutional kernel, * is the convolution operation without edge padding, is the j-th convolved region in the l-th layer, W k is the width of the convolutional kernel, is the m-th weight value in the i-th convolutional kernel, is the bias vector of the j-th feature map in the l-th convolutional layer.

[0124] 2.2 Activation Module: After the convolution operation, the activation function performs a non-linear transformation on the output of the convolutional layer. ReLU is used as the activation function of the convolutional neural network.

[0125]

[0126] 2.3 Pooling Module: The pooling operation follows the convolutional layer immediately, and max pooling is used to capture the most valuable signal features;

[0127] (j - 1)W p +1 ≤ n ≤ jW p

[0128] Among them, is the output value obtained through average pooling, W p is the width of the pooling operation region, is the feature matrix within the operation region.

[0129] 2.4 Fully Connected Layer Module: After multiple convolutional layers and pooling layers, first flatten the generated feature map, and then perform a fully connected output;

[0130]

[0131] Among them, is the output value of the j-th neuron in the (l + 1)-th layer.

[0132] 3. Weighted loss function calculation module: Using the mean square error as the loss function to measure the gap between the reconstructed data and the original data; adopting a label distribution smoothing strategy in the loss function calculation to assign different weights to each sample in the training set;

[0133] Specifically, first perform kernel density estimation on the distribution of labels y in the training set to generate a probability density function, and then based on this probability density function, assign different weights to each sample in the training set;

[0134]

[0135] Among them, is the kernel density function, and L is the weighted loss function.

[0136] 4. Backpropagation and parameter update module: Continuously update the parameters to minimize the mean square error, and continuously iterate to obtain the optimal parameters of the CNN module, and the CNN training ends;

[0137] δ * = argmin δ L(f(x k ; δ))

[0138] Among them, δ * is the optimal parameter of the CNN.

[0139] Regarding the truck dynamic weighing method based on the convolutional neural network in this embodiment, it should be noted that:

[0140] 1. Effects of adding six-stage temperature features and impact count features:

[0141] The processing method for the data input into the convolutional neural network model includes using the Pearson coefficient method for feature dimensionality reduction and adding two new feature variables, namely six-stage temperature features and impact count features. To verify the effectiveness of this data processing method in improving the prediction accuracy of the passing vehicle weight, a set of comparative experiments was set up. Among them, the input features of the DAE-CNN-KDE model were not processed at all; the input features of the DAE-CNN-KDE1 model were screened by the Pearson coefficient; the input of the DAE-CNN-KDE2 model was the combination of two new features, namely six-stage temperature features and impact count features, and the extracted features of the DAE as the input of the CNN on the basis of the features being screened by the Pearson coefficient. Table 4 shows the quantity and proportion of each error range in the prediction results.

[0142] Table 4 Evaluation results of data processing effectiveness on the test set

[0143]

[0144]

[0145] 2. Effectiveness of adding label distribution smoothing distribution:

[0146] To verify the effectiveness of the label distribution smoothing strategy, a set of comparative experiments were designed to compare the model prediction results with and without data imbalance processing respectively. Among them, EMAE and EMAPE represent the average of the mean absolute error (MAE) and the mean absolute percentage error (MAPE) of samples in a certain area. The evaluation index results are shown in Table 5.

[0147] Table 5 Evaluation results of the effectiveness of the label distribution smoothing strategy

[0148]

[0149] According to the data in Table 5, it can be seen that the model processed by label distribution smoothing has a significant improvement in the accuracy of samples in the low-density area.

[0150] Specifically, in terms of the EMAE index:

[0151] In May, for the model processed by label distribution smoothing (LDS), the EMAE of all samples decreased by 85.6 Kg compared with the model without LDS processing. The EMAE of samples in the high-density area increased by 5.91 Kg, the EMAE of samples in the medium-density area decreased by 78.51 Kg, and the EMAE of samples in the low-density area decreased by 111.42 Kg;

[0152] In June, the EMAE of all samples of the model processed by LDS decreased by 131.28 Kg compared with the model without LDS processing. The EMAE of samples in the high-density area decreased by 12.35 Kg, the EMAE of samples in the medium-density area decreased by 122.53 Kg, and the EMAE of samples in the low-density area decreased by 494.91 Kg;

[0153] In July, the EMAE of all samples of the model processed by LDS decreased by 113.14 Kg compared with the model without LDS processing. The EMAE of samples in the high-density area decreased by 24.73 Kg, the EMAE of samples in the medium-density area decreased by 101.63 Kg, and the EMAE of samples in the low-density area decreased by 1353.41 Kg;

[0154] In August, the EMAE of all samples of the model processed by LDS decreased by 138.08 Kg compared with the model not processed by LDS. The EMAE of samples in the high-density area decreased by 98.94 Kg, the EMAE of samples in the medium-density area decreased by 317.18 Kg, and the EMAE of samples in the low-density area decreased by 40.4 Kg.

[0155] Regarding the EMAPE index:

[0156] In May, the EMAPE of all samples of the model processed by LDS decreased by 0.24% compared with the model not processed by LDS. The EMAPE of samples in the high-density area increased by 0.02%, the EMAPE of samples in the medium-density area decreased by 0.17%, and the EMAPE of samples in the low-density area decreased by 1.86%;

[0157] In June, the EMAPE of all samples of the model processed by LDS decreased by 0.38% compared with the model not processed by LDS. The EMAPE of samples in the high-density area decreased by 0.04%, the EMAPE of samples in the medium-density area decreased by 0.38%, and the EMAPE of samples in the low-density area decreased by 1.52%;

[0158] In July, the EMAPE of all samples of the model processed by LDS decreased by 0.47% compared with the model not processed by LDS. The EMAPE of samples in the high-density area decreased by 0.05%, the EMAPE of samples in the medium-density area decreased by 0.3%, and the EMAPE of samples in the low-density area decreased by 4.52%;

[0159] In August, the EMAPE of all samples of the model processed by LDS decreased by 0.47% compared with the model not processed by LDS. The EMAPE of samples in the high-density area decreased by 0.2%, the EMAPE of samples in the medium-density area decreased by 1.08%, and the EMAPE of samples in the low-density area decreased by 0.57%.

[0160] It can be seen from these data that whether it is EMAE or EMAPE, after being processed by LDS, the indicators of samples in the medium-density area and the low-density area have decreased significantly, showing higher precision. While the indicator changes of samples in the high-density area are small or slightly increased. Therefore, the model processed by LDS can effectively improve the precision, especially in the low-density and medium-density areas.

[0161] Note: In this comparative experiment, the range of passing vehicle weights is divided at equal intervals of 1 ton. The ratio of the total number of training set samples N to the number of intervals M is calculated as the reference density C. Subsequently, according to the ratio relationship between the actual number of samples in each interval and the reference density, the sample distribution is divided into three density levels: when the number of samples exceeds 2C, it is defined as the high-density area; when it is between 0.2C and 2C, it is the medium-density area; and when it is less than 0.2C, it is the low-density area. When the prediction model aims to minimize the sum of prediction errors of all training samples, the proportion of samples in the low-density area is relatively small, resulting in the model being more inclined to learn the samples in the high-density interval. Therefore, a label distribution smoothing strategy is introduced, and the kernel density estimation method is used to weight the loss function of the model to enhance the influence of low-density samples in model training.

[0162] 3. Effectiveness of adding DAE:

[0163] As an important feature extraction and noise reduction method, the denoising autoencoder has unique advantages in the task of truck dynamic weighing: it introduces a noise reduction mechanism to reconstruct the original signal features by adding noise, suppressing environmental noise interference to a certain extent. Through the multi-layer denoising autoencoder structure, the DAE can extract deep feature representations in the passing vehicle signal layer by layer, providing high-quality input features for subsequent convolutional neural network prediction.

[0164] Table 6 Experimental results

[0165]

[0166] As can be seen from Table 6, adding the DAE module on the basis of the convolutional neural network has a good effect. Specifically, compared with CNN-KDE without using the DAE module, DAE-CNN-KDE has obvious improvements in all indicators:

[0167] In May, the RMSE, MAE, MAPE, and R of DAE-CNN-KDE 2 decreased by 330.73 Kg, 144.95 Kg, 0.185%, and increased by 0.0241 compared with CNN-KDE respectively;

[0168] In June, the RMSE, MAE, MAPE, and R of DAE-CNN-KDE 2 decreased by 282.2 Kg, 165.82 Kg, 0.327%, and increased by 0.0234 compared with CNN-KDE respectively;

[0169] The RMSE, MAE, MAPE, and R in July 2 decreased by 488.72 Kg, 200.36 Kg, 0.5975%, and increased by 0.0446 compared with CNN-KDE respectively;

[0170] In August, the RMSE, MAE, MAPE, and R of DAE-CNN-KDE 2 decreased by 306.85 Kg, 252.32 Kg, 0.3374%, respectively, compared with CNN-KDE, and R increased by 0.0444;

[0171] The DAE-CNN-KDE model performs excellently in the feature extraction part, and there are obvious improvements in various performance indicators compared with CNN-KDE without using the DAE module.

[0172] Example 2

[0173] This example provides an implementation case of a truck dynamic weighing method based on a convolutional neural network.

[0174] (1) Construction of the neural network model for predicting the passing vehicle weight at dynamic weighing site A

[0175] Collect the passing vehicle data of 6-axle trucks at a certain dynamic weighing site A in March and April 2023, and construct a vehicle weight prediction model based on a convolutional neural network according to the method steps in the foregoing Section 2 and Section 3. The specific parameters are shown in Table 4.

[0176] Table 4: Neural network model parameters of dynamic weighing site A

[0177]

[0178] (2) Prediction of the passing vehicle weight at dynamic weighing site A

[0179] Collect the passing vehicle data of 6-axle trucks at dynamic weighing site A in May, June, July, and August 2023; sort out the passing vehicle feature data according to the method in Section 2; use the already constructed neural network model of dynamic weighing site A to predict the truck weight. The predicted results of the passing vehicle weights at site A in May, June, July, and August are shown in Table 5.

[0180] Table 5: Predicted results of the passing vehicle weights at site A in May, June, July, and August

[0181]

[0182] Example 3

[0183] Based on the same inventive concept, the embodiment of the present application also provides a truck dynamic weighing system based on a convolutional neural network. The implementation solution provided by the truck dynamic weighing system based on a convolutional neural network is similar to the implementation solution described in the method of Example 1. Therefore, the specific limitations in one or more embodiments of the truck dynamic weighing system based on a convolutional neural network provided below can refer to the limitations on the method in Example 1, and will not be repeated here.

[0184] This embodiment provides a truck dynamic weighing system based on a convolutional neural network, including:

[0185] A pressure data feature acquisition module configured to: collect passing vehicle data of a truck, perform feature processing on the passing vehicle data, and obtain pressure data features;

[0186] A weighing prediction module configured to: send the pressure data features, six-stage temperature features, and impact count features into a convolutional neural network model for prediction to obtain the predicted weight of the truck;

[0187] Among them, the six-stage temperature feature is obtained by first arranging the collected passing vehicle data in chronological order, dividing 24 hours of a day into six time periods, namely 0:00 - 4:00, 4:00 - 8:00, 8:00 - 12:00, 12:00 - 16:00, 16:00 - 20:00, 20:00 - 24:00, and then performing feature marking on the passing vehicle data falling into the corresponding time periods; where 0:00 - 4:00 is marked as 1, 4:00 - 8:00 is marked as 2, 8:00 - 12:00 is marked as 3, 12:00 - 16:00 is marked as 4, 16:00 - 20:00 is marked as 5, and 20:00 - 24:00 is marked as 6;

[0188] The impact count feature is the total cumulative number of passing vehicles on the weighing platform lane within the corresponding time period.

[0189] Embodiment 4

[0190] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the truck dynamic weighing method based on a convolutional neural network described in Embodiment 1.

[0191] Embodiment 5

[0192] This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the program is executed by a processor, it implements the truck dynamic weighing method based on a convolutional neural network described in Embodiment 1.

[0193] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that: it is still possible to modify the specific implementation manners of the present invention or perform equivalent replacements on some technical features; without departing from the spirit of the technical solutions of the present invention, they should all be covered within the scope of the technical solutions claimed by the present invention.

Claims

1. A truck dynamic weighing method based on a convolutional neural network, characterized in that It includes the following steps: Collect the passing vehicle data of the freight car, perform feature processing on the passing vehicle data, and obtain the pressure data features; Send the pressure data features, six-stage temperature features, and impact frequency features into a convolutional neural network model for prediction to obtain the predicted weight of the freight car; Among them, for the six-stage temperature features, first arrange the collected passing vehicle data in chronological order, divide 24 hours of a day into six time periods, namely 0:00 - 4:00, 4:00 - 8:00, 8:00 - 12:00, 12:00 - 16:00, 16:00 - 20:00, 20:00 - 24:00, and then perform feature marking on the passing vehicle data falling into the corresponding time periods; among them, 0:00 - 4:00 is marked as 1, 4:00 - 8:00 is marked as 2, 8:00 - 12:00 is marked as 3, 12:00 - 16:00 is marked as 4, 16:00 - 20:00 is marked as 5, and 20:00 - 24:00 is marked as 6; The impact frequency feature is the total cumulative number of passing vehicles on the weighing platform lane within the corresponding time period.

2. The truck dynamic weighing method based on a convolutional neural network according to claim 1, wherein The method for obtaining the pressure data features is as follows: The weighing platform includes 2 weighing platforms corresponding to the left and right wheels of the freight car during driving respectively, and 2 weighing sensors are arranged along the advancing direction of the freight car on each weighing platform; the output data of the weighing sensors is used as the passing vehicle data; For the collected passing vehicle data, select the average peak value, peak value, compensation value, time difference, and passing vehicle speed after each tire passes each weighing sensor as feature data; among them, the average peak value is the arithmetic mean of 100 pressure values before and after the current sensor peak value, and the compensation value is the pressure value of the sensor in front of the current sensor; the time difference is the time difference between the current sensor peak value and the first sensor peak value of this vehicle; Summarize the feature data of the passing vehicle data of each vehicle and merge them in chronological order; Use the Pearson correlation coefficient method to screen out the features that are most critical to the prediction result. Among them, the features with a Pearson correlation coefficient > 0.4 are selected, including the front average peak value, rear average peak value, front peak value, and rear peak value of weighing platforms one and two, as the final input features of the model.

3. The truck dynamic weighing method based on a convolutional neural network according to claim 1 or 2, characterized in that, The convolutional neural network model includes: The DAE loading module, including an input data processing module, a model forward propagation module, a loss function calculation module, and a backpropagation and parameter update module; The input data processing module is configured to: use the feature values screened by the Pearson correlation coefficient and normalized as the input, and add Gaussian noise to the input data; The model forward propagation module is configured to: map the data with Gaussian noise to reconstructed data through an encoder-decoder and add a BN layer; The loss function calculation module is configured to: use the mean square error as the loss function to measure the gap between the reconstructed data and the original data; The backpropagation and parameter update module is configured to: update the parameters to minimize the mean square error and iteratively obtain the optimal parameters of the DAE module; The CNN module, including an input and output data module, a stacked multi-layer convolution module, a fully connected layer module, a weighted loss function calculation module, and a backpropagation and parameter update module; Input and output data module, configured to: concatenate the features obtained by the DAE loading module with the normalized impact number feature and the six-stage temperature feature column-wise into an array, and use it as the input of the CNN module, and use the normalized actual weight as the label input; Stacked multi-layer convolution module, configured to: use the convolution operation value of each layer as the input of the next layer of convolution; includes stacked multi-layer convolution and fully connected layer modules; each layer of convolution includes a one-dimensional convolution module, an activation module, and a pooling module; One-dimensional convolution module, configured to: perform a convolution operation on the input data and the convolution kernel to generate a feature map; Activation module, configured to: perform a non-linear transformation activation on the feature map generated by the convolution; Pooling module, configured to: use max pooling on the feature map after non-linear transformation activation to capture the most valuable signal features; Fully connected layer module, configured to: flatten the feature map generated by the stacked multi-layer convolution, and perform a fully connected output after flattening; Weighted loss function calculation module, configured to: use the mean squared error as the loss function to measure the difference between the reconstructed data and the original data; adopt a label distribution smoothing strategy in the loss function calculation to assign different weights to each sample in the training set; Backpropagation and parameter update module, configured to: update the parameters to minimize the mean squared error, and iteratively obtain the optimal parameters of the CNN module.

4. A truck dynamic weighing system based on a convolutional neural network, characterized in that, including Pressure data feature acquisition module, configured to: collect the passing vehicle data of the truck, perform feature processing on the passing vehicle data, and obtain the pressure data features; Weighing prediction module, configured to: send the pressure data features, the six-stage temperature features, and the impact number features into the convolutional neural network model for prediction to obtain the predicted weight of the truck; Among them, the six-stage temperature feature is obtained by first arranging the collected passing vehicle data in chronological order, dividing 24 hours of a day into six time periods, namely 0:00~4:00, 4:00~8:00, 8:00~12:00, 12:00~16:00, 16:00~20:00, 20:00~24:00, and then performing feature marking on the passing vehicle data falling into the corresponding time period; among them, 0:00~4:00 is marked as 1, 4:00~8:00 is marked as 2, 8:00~12:00 is marked as 3, 12:00~16:00 is marked as 4, 16:00~20:00 is marked as 5, and 20:00~24:00 is marked as 6; The impact number feature is the total cumulative number of passing vehicles on the weighing platform lane within the corresponding time period.

5. The truck dynamic weighing system based on a convolutional neural network according to claim 4, characterized in that, The method for obtaining the pressure data features is: The weighing platform includes 2 weighing platforms corresponding to the left and right wheels of the truck respectively, and 2 weighing sensors are arranged along the advancing direction of the truck; the output data of the weighing sensors is used as the passing vehicle data; For the collected vehicle passing data, the average peak value, peak value, compensation value, time difference, and vehicle passing speed after each tire passes through each weighing sensor are selected as feature data; among them, the average peak value is the arithmetic mean of 100 pressure values before and after the current sensor peak, the compensation value is the pressure value of the sensor in front of the current sensor; the time difference is the time difference between the current sensor peak and the first sensor peak of this vehicle; Summarize the feature data of the vehicle passing data of each vehicle and merge them in chronological order; Use the Pearson correlation coefficient method to screen out the features that are most critical to the prediction result. Among them, the features with a Pearson correlation coefficient > 0.4 are selected, including the front average peak value, rear average peak value, front peak value, and rear peak value of weighing platforms 1 and 2, as the final input features of the model.

6. The truck dynamic weighing system based on a convolutional neural network according to claim 4 or 5, characterized in that, The convolutional neural network model includes: DAE loading module, including input data processing module, model forward propagation module, loss function calculation module, backpropagation and parameter update module; Input data processing module, configured to: take the feature values screened by the Pearson correlation coefficient and normalized as input, and add Gaussian noise to the input data; Model forward propagation module, configured to: map the data with Gaussian noise to reconstructed data through an encoder-decoder and add a BN layer; Loss function calculation module, configured to: use the mean square error as the loss function to measure the gap between the reconstructed data and the original data; Backpropagation and parameter update module, configured to: update the parameters to minimize the mean square error and iteratively obtain the optimal parameters of the DAE module; CNN module, including input and output data module, stacked multi-layer convolution module, fully connected layer module, weighted loss function calculation module, backpropagation and parameter update module; Input and output data module, configured to: add the impact number feature and six-stage temperature feature after normalization to the features obtained by the DAE loading module, concatenate them into an array by column, and use it as the input of the CNN module, and use the normalized actual weight as the label input; Stacked multi-layer convolution module, configured to: use the convolution operation value of each layer as the input of the next layer convolution; including stacked multi-layer convolution and fully connected layer module; each layer of convolution includes a one-dimensional convolution module, an activation module, and a pooling module; One-dimensional convolution module, configured to: perform convolution operation on the input data and the convolution kernel to generate a feature map; Activation module, configured to: perform a non-linear transformation activation on the feature map generated by the convolution; Pooling module, configured to: use max pooling on the feature map after non-linear transformation activation to capture the most valuable signal features; Fully connected layer module, configured to: flatten the feature map generated by the stacked multi-layer convolution, and perform a fully connected output after flattening; Weighted loss function calculation module, configured to: use the mean square error as the loss function to measure the gap between the reconstructed data and the original data; adopt a label distribution smoothing strategy to assign different weights to each sample in the training set during the loss function calculation; Backpropagation and parameter update module, configured to: update the parameters to minimize the mean square error and iteratively obtain the optimal parameters of the CNN module.

7. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the truck dynamic weighing method based on convolutional neural network according to any one of claims 1 to 3.

8. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, and when the program is executed by a processor, it implements the truck dynamic weighing method based on convolutional neural network according to any one of claims 1 to 3.

Images