Vehicle weighing compensation method based on ECDOA-RGBPNN

Abstract

The invention relates to a vehicle weighing compensation method based on ECDOA-RGBPNN. The method comprises the steps of obtaining weighing data of a target vehicle; the weighing data are input into a preset ECDOA-RGBPNN weighing compensation model, the compensated weighing data are output, the ECDOA-RGBPNN weighing compensation model is obtained based on training of a training set, the training set comprises vehicle weighing data collected by a plurality of non-stop detection stations and overload control stations, and the vehicle weighing data are acquired by the non-stop detection stations and the overload control stations. According to the ECDOA-RGBPNN weighing compensation model, residual connection and gating are introduced into a BP neural network, and an improved group decision optimization algorithm is adopted to initialize weight and bias construction. The ECDOA-RGBPNN weighing compensation model constructed by the invention can effectively correct the weighing data and compensate the measurement error of the weighing sensor, and has potential application value in the fields of logistics transportation, road management and the like.

Description

Technical Field

[0001] The present invention relates to the technical fields of logistics, transportation and highway management, and in particular to a vehicle weighing compensation method based on ECDOA-RGBPNN. Background Art

[0002] Traditional methods of overweight vehicle enforcement rely primarily on on-site enforcement, requiring significant manpower and resources, and often resulting in suboptimal results. With technological advancements and the increasing level of informatization, non-site enforcement has become a new option. Traditional overweight vehicle enforcement often relies on static weighing. While static weighing technology is relatively mature and offers high accuracy, it is slow, complex to install, and can easily cause traffic congestion. Vehicle weigh-in-motion (WIM) is a key tool for non-site enforcement of overweight vehicles, effectively improving enforcement accuracy and efficiency. However, the accuracy of WIM is generally lower than that of static weighing, limiting its practical application. WIM accuracy is primarily affected by both vehicular and non-vehicular factors. WIM is primarily affected by interference from multiple sources, such as speed fluctuations, body vibration, and tire pressure, while WIM stems primarily from sensor accuracy, sensor installation, temperature, and road surface roughness. Because WIM systems are influenced by numerous factors, and their relationships lack linear characteristics, it is difficult to determine the specific relationship between the WIM system and these factors. To reduce weighing errors, software filtering is often employed to filter the weighing signal. Traditional methods for weighing data compensation mainly revolve around data mapping and function approximation. The table lookup method achieves compensation by establishing a discrete mapping table between the original data and the standard output. It is suitable for scenarios with small data volumes and low precision requirements, but storage requirements increase significantly with the increase in data volume. The interpolation rule divides the data into several intervals and dynamically corrects the data within the interval through methods such as linear interpolation and spline interpolation, which can ensure high accuracy when hardware resources are limited. The curve fitting rule uses basis functions (such as polynomials) to perform global or piecewise fitting on the original data. The least squares method is a typical representative, but in highly nonlinear scenarios, local optimality or endpoint discontinuity problems are prone to occur, and it is often necessary to combine piecewise fitting optimization.

[0003] Intelligent methods are now commonly used to compensate weighing data. For example, Shi et al. developed a temperature compensation model based on a BP neural network for pressure sensors, improving compensation accuracy by optimizing the network structure and training algorithm. While these models can effectively address the error issues of piezoresistive pressure sensors under temperature fluctuations, they are only applicable to specific scenarios and lack universal applicability.

[0004] At present, most models are designed specifically for a certain type of pressure sensor, and the error factors considered are limited to a few factors such as temperature and humidity. This weighing compensation model developed for a specific sensor and a few influencing factors may have a better compensation effect in a specific application scenario, but the environment in actual application is often not so ideal. In addition, in existing research, it is usually just a simple implementation of the swarm intelligence optimization algorithm to initialize the BPNN weights, without making some adaptive adjustments to this task, and the performance of the currently used swarm intelligence optimization algorithm in the task of vehicle weighing compensation still has room for improvement. Therefore, the present invention proposes a vehicle weighing compensation method based on ECDOA-RGBPNN. Summary of the Invention

[0005] The purpose of the present invention is to provide a vehicle weighing compensation method based on ECDOA-RGBPNN to address the measurement error problem caused by various complex factors in the vehicle weighing measurement process. By introducing residual connection and gating into the BP neural network, and using an improved group decision optimization algorithm to initialize weights and biases to construct an ECDOA-RGBPNN weighing compensation model, the relationship between vehicle weighing and other factors is learned, and the accuracy of weighing measurement is improved.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] The vehicle weighing compensation method based on ECDOA-RGBPNN includes:

[0008] Obtain the weighing data of the target vehicle;

[0009] The weighing data is input into a preset ECDOA-RGBPNN weighing compensation model, and the compensated weighing data is output, wherein the ECDOA-RGBPNN weighing compensation model is obtained based on training of a training set, and the training set includes vehicle weighing data collected by several non-stop inspection stations and overweight control stations. The ECDOA-RGBPNN weighing compensation model is constructed by introducing residual connections and gating into the BP neural network, and using an improved group decision optimization algorithm to initialize weights and biases.

[0010] Optionally, before training the ECDOA-RGBPNN weighing compensation model based on the training set, the method further includes preprocessing the training set, where the preprocessing includes:

[0011] Obtain weighing data collected by several non-stop inspection stations and overweight control stations, and match them according to license plate numbers;

[0012] Calculate the weighing error of the matching weighing data by using the preset weighing compensation target value, and remove the matching weighing data when the weighing error is greater than the preset error value, or when the matching weighing data exceeds the preset time range;

[0013] The matching weighing data after elimination are processed for missing values ​​and outliers, and are one-hot encoded to construct the training set.

[0014] Optionally, the weighing compensation target value is obtained based on weighing data of the same vehicle at a traffic control station within a preset time range and a weighing deviation range.

[0015] Optionally, the ECDOA-RGBPNN weighing compensation model is constructed by introducing residual connections and gating into the BP neural network and using an improved group decision optimization algorithm to initialize weights and biases, including:

[0016] By introducing residual connections and gating into the BP neural network, the RGBPNN model is obtained;

[0017] The improved group decision optimization algorithm is used to initialize the weights and biases of the RGBPNN model, and the model is trained through forward propagation and back propagation to construct the ECDOA-RGBPNN weighing compensation model.

[0018] Optionally, introducing residual connection and gating into the BP neural network includes:

[0019] A BP neural network consisting of three fully connected layers connected in sequence is set up, a gating is added after the second fully connected layer, and the output of the first fully connected layer and the output of the second fully connected layer are spliced ​​and transmitted to the third fully connected layer to construct the RGBPNN model.

[0020] Optionally, the gating is used to perform sigmoid calculation on the output of the second fully connected layer, obtain a 0-1 gated output, and then multiply it by the output of the second fully connected layer to obtain the final output of the second fully connected layer.

[0021] Optionally, the improved group decision optimization algorithm is obtained by adopting He initialization to replace the random population initialization of the extended group decision optimization algorithm.

[0022] Optionally, the process of the improved group decision optimization algorithm is:

[0023] S1. Set initial parameters and termination conditions;

[0024] S2. Generate an initial population using He initialization and the initial parameters and calculate the fitness value;

[0025] S3, update the step size according to the current number of iterations and generate an evolution sequence;

[0026] S4. Based on the current individual and population fitness values, determine whether it is the current optimal position. If not, update the position based on experience, other agents, group thinking, leaders, and innovation according to the evolutionary sequence order; if so, randomly generate candidate solutions near the current optimal position;

[0027] S5. Evaluate the new solution and update it to determine whether the termination condition is met. If so, output the best candidate solution; if not, return to S3 until the termination condition is met.

[0028] The beneficial effects of the present invention are:

[0029] Based on the original three-layer BPNN, this invention adds gating to the deep feature output for feature weighting. It then adds residual connections from the shallow feature output to the deep feature output to create an RGBPNN. The RGBPNN weights and biases are initialized using an improved group decision optimization algorithm. The vehicle type, number of axles, number of tires, vehicle speed, detection point code, and detection point sensor type are used as inputs. After RGBPNN training, an ECDOA-RGBPNN weighing compensation model is developed, which effectively compensates for weighing data. The ECDOA-RGBPNN model converges faster, achieves better data compensation, and effectively reduces the measurement error of weighing sensors. This provides a new approach and method for non-stop weighing measurement, with potential application value in logistics, transportation, highway management, and other fields. In logistics and transportation, vehicle weighing can be more accurately monitored, avoiding overloaded transportation and improving transportation safety and efficiency. In highway management, vehicle weighing can be more accurately monitored to protect highway infrastructure. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0031] Figure 1 A flow chart for constructing the ECDOA-RGBPNN weighing compensation model according to an embodiment of the present invention;

[0032] Figure 2 Flowchart of the ECDOA algorithm according to an embodiment of the present invention;

[0033] Figure 3 This is a schematic diagram of the RGBPNN structure of an embodiment of the present invention;

[0034] Figure 4This is a graph showing the results before and after weighing compensation of the ECDOA-RGBPNN test set according to an embodiment of the present invention;

[0035] Figure 5 It is an iterative comparison diagram of the fitness of each model in an embodiment of the present invention;

[0036] Figure 6 2 is a comparison chart of ablation experiments according to an embodiment of the present invention. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0038] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0039] BPNN (Backpropagation Neural Network) is a multi-layer feedforward neural network trained using the backpropagation algorithm. It consists of an input layer, a hidden layer, and an output layer, with each layer containing multiple neurons. The core idea of ​​BP neural networks is to learn the mapping relationship between input and output, automatically adjust the network weights and biases, and use activation functions to perform nonlinear transformations, thereby modeling the complex nonlinear relationships between multiple features and target values.

[0040] The training process of the BP neural network mainly includes two stages: a) Forward propagation: the input signal is passed through the input layer to the hidden layer, and after weighting and activation function processing, the prediction result is finally output. b) Backward propagation: the error of the output layer is calculated, and the error is passed back to each layer of the network layer by layer through back propagation. The weights and biases are adjusted according to the error to minimize the total error of the network. The BP neural network has powerful nonlinear fitting capabilities and can approximate arbitrarily complex nonlinear functions. It is suitable for dealing with complex system modeling problems. In addition, the BP neural network can self-learn and adapt, and automatically adjust the network parameters through training data without the need to define the function form in advance. The BP neural network also has a certain generalization ability. Based on the training data, it can predict new data that has not been seen.

[0041] However, BP neural networks also have some limitations. Because the initial weights and biases are randomly generated, they may fall into local optimal solutions during training, resulting in poor model performance. In addition, traditional gradient descent methods can lead to slow training and low convergence speed, especially in complex high-dimensional data. Therefore, this embodiment improves the BP neural network and proposes a vehicle weighing compensation method based on ECDOA-RGBPNN, including the following:

[0042] Obtain the weighing data of the target vehicle;

[0043] The weighing data is input into a preset ECDOA-RGBPNN weighing compensation model, and the compensated weighing data is output, wherein the ECDOA-RGBPNN weighing compensation model is obtained based on training of a training set, and the training set includes vehicle weighing data collected by several non-stop inspection stations and overweight control stations. The ECDOA-RGBPNN weighing compensation model is constructed by introducing residual connections and gating into the BP neural network, and using an improved group decision optimization algorithm to initialize weights and biases.

[0044] Specifically, this embodiment adds gating to the deep feature outputs of the original three-layer BPNN for feature weighting. A residual connection is then added between the shallow feature outputs and the deep feature outputs to create an RGBPNN. The RGBPNN weights and biases are initialized using an improved group decision optimization algorithm. The RGBPNN is trained using the vehicle type, number of axles, number of tires, speed, detection point codes, and detection point sensor types as inputs. The resulting ECDOA-RGBPNN weighing compensation model effectively compensates for weighing data. The ECDOA-RGBPNN model converges faster, achieves better data compensation, and effectively reduces weighing sensor measurement errors. This provides a new approach and method for non-stop weighing measurement, with potential applications in logistics, transportation, highway management, and other fields. In logistics, this allows for more accurate monitoring of vehicle weighing, preventing overloads and improving transportation safety and efficiency. In highway management, it enables more precise monitoring of vehicle weighing and protects highway infrastructure.

[0045] Furthermore, before training the ECDOA-RGBPNN weighing compensation model based on the training set, the training set is preprocessed, and the preprocessing includes:

[0046] Obtain weighing data collected by several non-stop inspection stations and overweight control stations, and match them according to license plate numbers;

[0047] Calculate the weighing error of the matching weighing data by using the preset weighing compensation target value, and remove the matching weighing data when the weighing error is greater than the preset error value, or when the matching weighing data exceeds the preset time range;

[0048] The matching weighing data after elimination are processed for missing values ​​and outliers, and are one-hot encoded to construct the training set.

[0049] The weighing compensation target value is obtained based on the weighing data of the same vehicle passing the overweight control station within a preset time range and a weighing deviation range.

[0050] Specifically, the data items collected from non-stop inspection points and overweight control stations are first matched and filtered, and then feature selection is performed. Compared with traditional static weighing methods, the existing non-stop weighing process often faces many complex situations. The type of vehicle, number of axles, number of tires, and speed will affect the vehicle weighing results and produce different errors. Different weighing equipment sensors and weighing equipment at different inspection sites will affect the error of vehicle weighing. Therefore, the vehicle type, number of axles, number of tires, speed, inspection point code and weighing sensor type are selected as input features, and the training set and test set are divided into 8:2 ratio.

[0051] This embodiment uses the vehicle weighing inspection data collected from multiple non-stop inspection stations in Jiangxi Province and the weighing data collected from the overweight control station provided by Jiangxi Transportation Investment and Maintenance Technology Group as the original data. Because the data collection and implementation of static weighing of vehicles is very time-consuming and costly, the weighing data of the same vehicle passing through the overweight control station within a reasonable time and weighing deviation range is used as the target value for weighing compensation. In the data table of the non-stop inspection station, the weighing data of the overweight control station is matched according to the license plate number. Taking into account the problem of loading goods between the two stations, the data with a deviation of more than 15% in the two weighing data are eliminated, among which the weighing error of the non-stop inspection station is between 5%-10%, and the weighing error of the overweight control station is within 5%. At the same time, considering the possibility of loading and unloading goods between the two measurements, the data entries whose two weighing data exceed a specific time range will also be eliminated.

[0052] Python code was written to handle missing values ​​and outliers in the matched data, and to perform one-hot encoding on the categorical data, resulting in 121 input values. The 9,425 sample dataset was then divided into a training set and a test set in an 8:2 ratio, and the numerical data in both the training and test sets were normalized.

[0053] Furthermore, by introducing residual connection and gating into the BP neural network and using the improved group decision optimization algorithm to initialize weights and biases, the ECDOA-RGBPNN weighing compensation model is constructed, including:

[0054] By introducing residual connections and gating into the BP neural network, the RGBPNN model is obtained;

[0055] The improved group decision optimization algorithm is used to initialize the weights and biases of the RGBPNN model, and the model is trained through forward propagation and back propagation to construct the ECDOA-RGBPNN weighing compensation model.

[0056] Specifically, such as Figure 1 As shown in the figure, the ECDOA-RGBPNN weighing compensation model consists of two parts: ECDOA weight initialization and RGBPNN error compensation. Specifically, ECDOA searches the search space for suitable initial weights and biases for the neural network. These initial weights and biases are used as the initial weights and biases for the RGBPNN, which are then adjusted through forward and backpropagation. ECDOA is used to initialize the neural network weight parameters, replacing the original random initialization. ECDOA initialization brings the weight parameters closer to the optimal weight parameters and reduces the randomness of the initial weight parameters. Switching to RGBPNN for model training improves the training performance and makes training more stable.

[0057] Furthermore, introducing residual connection and gating into the BP neural network includes:

[0058] A BP neural network consisting of three fully connected layers connected in sequence is set up, a gating is added after the second fully connected layer, and the output of the first fully connected layer and the output of the second fully connected layer are spliced ​​and transmitted to the third fully connected layer to construct the RGBPNN model.

[0059] The gating is used to perform sigmoid calculation on the output of the second fully connected layer, obtain a 0-1 gated output, and then multiply it by the output of the second fully connected layer to obtain the final output of the second fully connected layer.

[0060] Specifically, BPNN is a multi-layer feedforward neural network widely used in machine learning and artificial intelligence. It trains the network through the backpropagation algorithm, enabling it to learn the complex mapping relationship between input and output.

[0061] BPNN completes data compensation by fitting the functional relationship between input features and target values. In this embodiment, it fits the complex functional relationship between multiple input features and weighing compensation target values. The weighing data fitted by the model is used as the weighing result after model compensation. This embodiment improves the original BPNN network model to enhance the ability of BPNN to fit data. Specific improvements are as follows: Figure 3As shown in the figure, first, residual connections are introduced to the original three-layer BPNN. The feature outputs of the shallow layer are concatenated with the feature outputs of the deep layer and passed to the next layer. This allows the network to simultaneously learn and map information from both shallow and deep layers to the output layer, and can also reduce the gradient vanishing phenomenon to a certain extent. Then, a gate is added to the output of the deep feature extraction to obtain the RGBPNN. The deep features are sigmoid-ed to a 0-1 gated output, which is then multiplied with the deep features. This is a process of weighting the importance of the deep features, which can better filter out important features and weaken unimportant ones.

[0062] The forward propagation calculation formula using the RGBPNN model is as follows:

[0063]

[0064] Among them, a1 is the result output of the first layer, X represents the input feature vector (including vehicle type, number of axles, number of tires, speed, detection point code, sensor type), W i ,b i are the weight, bias and activation function of the i-th layer, g gate represents the gate output, σ represents the sigmoid function, W g ,b g are the weights and biases of the gated unit, Concat(·) represents the data concatenation function, and ⊙ represents element-by-element multiplication.

[0065] Furthermore, the improved group decision optimization algorithm is obtained by adopting He initialization to replace the random population initialization of the extended group decision optimization algorithm.

[0066] The process of the improved group decision-making optimization algorithm is as follows:

[0067] S1. Set initial parameters and termination conditions;

[0068] S2. Generate an initial population using He initialization and the initial parameters and calculate the fitness value;

[0069] S3, update the step size according to the current number of iterations and generate an evolution sequence;

[0070] S4. Based on the current individual and population fitness values, determine whether it is the current optimal position. If not, update the position based on experience, other agents, group thinking, leaders, and innovation according to the evolutionary sequence order; if so, randomly generate candidate solutions near the current optimal position;

[0071] S5. Evaluate the new solution and update it to determine whether the termination condition is met. If so, output the best candidate solution; if not, return to S3 until the termination condition is met.

[0072] Specifically, CDOA (Collective Decision Optimization Algorithm) optimizes problems by simulating personal experience, other people's opinions, group thinking, leadership opinions and innovation. However, the fixed evolutionary order of CDOA may cause the population to form fixed thinking patterns, leading to premature population growth. ECDOA has been improved on the basis of CDOA, mainly through the following improvements to optimize the algorithm performance: ① Multi-step position selection scheme: By randomly combining existing operators, a new evolutionary sequence is provided for each search individual, increasing the diversity of solutions and avoiding premature convergence. ② Improved operators: The operators of the experience-based stage, the other-based stage, the group thinking-based stage, the leadership-based stage and the innovation-based stage have been adjusted to better balance the exploration and development capabilities. The specific algorithm flow of ECDOA is as follows: Figure 2 shown.

[0073] In this embodiment, the ECDOA algorithm is used to initialize the weight parameters of the RGBPNN to reduce the impact of random parameter initialization on the instability of the RGBPNN training results and poor convergence effect. It should be noted that, unlike the original ECDOA implementation, the population initialization of the ECDOA in this embodiment uses He initialization instead of the original random population initialization. This can generate an initial population that is more consistent with this embodiment, improving the convergence speed and effect of ECDOA. The optimal population obtained by ECDOA, as the result of the RGBPNN initial parameters, is also more suitable for RGBPNN training.

[0074] The calculation formula for initializing the weight parameters of RGBPNN using the ECDOA algorithm is as follows:

[0075] (W * ,b * )=ECDOA(W,b;X train ,y train ,θ);

[0076] Among them, W * ,b * are the initial weights and bias parameters optimized by ECDOA, W,b are the weights and biases initialized by He, θ represents the hyperparameters of ECDOA and RGBPNN, X train ,y train Represents the training set features and weighing compensation target values.

[0077] The fitness value in ECDOA is calculated in the same way as the loss L in RGBPNN. It includes the mean square error and L2 regularization. The former measures the deviation between the predicted value and the true value, while the latter is often used to prevent overfitting. The calculation formula is as follows:

[0078]

[0079] Among them, f RGBPNN It represents the abstract representation of the forward propagation process of the RGBPNN model. When calculating the fitness, W,b is the weight and bias decomposed from the position of the current population. When calculating the loss, it is the weight and bias of the neural network determined by the current iteration. s is the sth learnable weight parameter in RGBPNN, λ is the regularization coefficient, N is the total number of data samples, and S is the total number of weight matrices.

[0080] Furthermore, model training and testing are performed, including the following:

[0081] (1) Model training:

[0082] The training set is input into the constructed ECDOA-RGBPNN weighing compensation model for model training.

[0083] The ECDOA fitness function was set to the sum of MSE and L2 regularization. The population size N was set to 50, the number of iterations T was set to 100, the innovation factor MF was set to 0.45, the mutation dimension MD was set to 100, and the problem dimension was equal to the learnable parameter size of the neurons, which was 24001. For the RGBPNN, the input dimension was 121, the number of neurons in the first layer was set to 128, the number of neurons in the second layer was set to 64, the number of inputs to the fully connected output layer was the sum of the neurons in the first two layers, the output layer output was the compensated weight, the number of neurons in the output layer was 1, and the total learnable parameter size was 24001. The ECDOA fitness function and the BPNN loss function were both set to MSE plus L2 regularization. The learning rate was set to 0.0001, and the regularization coefficient was 0.00001.

[0084] The training set data was input into the constructed ECDOA-RGBPNN model to train the model. The saved best model was tested using the test set. The weighing data after model compensation was recorded, and the RMSE, model accuracy MAPE, and model fit R were calculated. 2 There are three evaluation criteria: RMSE measures the standard deviation of the error between the predicted value and the true value, which can effectively reflect the dispersion of the predicted value. The smaller the value, the smaller the prediction deviation. MAPE represents the average value of the absolute percentage error between the predicted value and the true value, which intuitively reflects the average deviation of the prediction. The smaller the value, the more accurate the prediction. 2 Used to measure the goodness of fit of the model to the data. The closer the value is to 1, the better the model fits the data. The calculation formulas for these three evaluation indicators are as follows:

[0085]

[0086] in, To predict the weight of the i-th sample after compensation, y i is the weighing target value of the i-th sample, SSE is the total sum of squares of the residuals, and SST is the total sum of squares of the actual values.

[0087] (2) Model testing;

[0088] Make predictions on the weighing data in the test set to evaluate the performance of the model.

[0089] The test samples of the 1885 test set data were input into the trained ECDOA-RGBPNN model to obtain the model-compensated weighing data. The weighing data before and after compensation were plotted against the weighing data reference value data, as shown in the figure. Figure 4 As shown, the horizontal axis is the sample number, the vertical axis is the weight (unit, kg), and the reference weighing weight is the weighing compensation target value. It can be seen that the original weighing measurement data is generally lower than the reference weighing weight, especially for samples with a reference weighing weight greater than 3000kg, indicating that the weighing results of the weighing sensor for vehicles with heavier weights are usually lower than the weight of static weighing, and vehicles with heavier weights are often more concerned about the possibility of overloading. ECDOA-RGBPNN is used to compensate the original weighing data. The compensated results are closer to the reference weighing data, indicating that the model proposed in this embodiment can compensate for the measurement error of the weighing sensor to a certain extent. In particular, for some samples with a reference weighing weight greater than 3000kg, the ECDOA-RGBPNN model can provide greater compensation so that these data with larger measurement deviations are closer to the reference static weighing data after compensation, making the weighing data closer to the actual vehicle weighing data, and achieving more accurate vehicle overload detection. At the same time, we can see several samples with significant deviations from the reference weight due to excessive compensation. A closer look reveals that the original weight data for these samples differs significantly from the majority of samples with the same weight. These are most likely abnormal data, possibly due to changes in the cargo weight caused by loading and unloading between the vehicle's non-stop inspection station and the vehicle's overweight control station.

[0090] Comparative test and result analysis:

[0091] Comparative experiments were conducted with different weight initialization schemes, including GA-BPNN, MEA-BPNN, and traditional He-BPNN. To ensure fair comparison, the neuron parameters of GA-BPNN, MEA-BPNN, and He-BPNN were set to 128, 65, and 1, respectively, for a total of 24067 parameters, 65 more than the RGBPNN. In the GA, the mutation probability (MR) was set to 0.01, and the population size was 80; in the MEA, the population size was set to 160, with both the optimal subpopulation and the temporary subpopulation set to 10. The number of iterations for each population was set to 100, the learning rate to 0.0001, and the regularization coefficient to 0.00001. It should be noted that to ensure fair comparison, the population sizes of GA, MEA, and ECDOA were controlled to ensure that the execution times of the three algorithms were roughly the same.

[0092] The fitness values ​​of GA-BPNN, MEA-BPNN and ECDOA-BPNN are compared. The comparison results are as follows Figure 5 As shown in the figure, the horizontal axis represents the number of iterations, and the vertical axis represents the fitness value. Under the same number of iterations, the smaller the fitness value, the stronger the algorithm convergence ability. Under the same fitness, the fewer the number of iterations, the faster the algorithm converges. Figure 5 As can be seen from the graph, ECDOA-BPNN converges the fastest, having essentially converged by the 40th iteration, while the other two algorithms have yet to converge. Furthermore, ECDOA-BPNN achieves the best convergence, consistently outperforming the other two algorithms after the tenth iteration. This demonstrates that ECDOA finds appropriate weights and biases faster than GA and MEA, and that the weights and biases found by ECDOA are more optimal.

[0093] By changing the random seed, the model training and validation experiments were repeated 20 times, and the mean and standard deviation of each evaluation index were calculated to obtain the experimental results in Table 1.

[0094] Table 1

[0095]

[0096] It can be clearly seen from the data in Table 1 that the best fitness after ECDOA optimization is 0.006, which is much smaller than GA and MEA. This shows that ECDOA has advantages in optimizing neural networks, and the weight parameters and biases solved are better. At the same time, the initial weight values ​​solved by ECDOA are more stable than those of GA and MEA, and will not fluctuate too much, making subsequent neural network training more stable. The weighing error RMSE after the compensation algorithm of this method is optimized is 1287kg, which is reduced by 8.52%, 5.43% and 0.73% compared with He-BPNN, GA-BPNN and MEA-BPNN respectively, and is better in MAPE and R 2 The ECDOA-RGBPNN model also achieved superior performance in the weight compensation results. Both RMSE and MAPE were at the lowest levels, and its predicted values ​​were closer to the true values. The standard deviations of various indicators in this method were also low, indicating that the model trained with this method is more stable and less likely to fall into local optimal solutions.

[0097] Overall, among the compared methods, the ECDOA-RGBPNN model is able to find more optimal initial weight parameters through ECDOA, significantly reducing the risk of the BP neural network falling into a local optimum due to poor initial weights. It can also better capture the relationship between vehicle weighing data features, achieve more accurate predictions, and compensate for measurement errors to a greater extent.

[0098] Ablation test and result analysis:

[0099] The ablation experiment mainly verifies the effectiveness of the ECDOA weight initialization module, the gating mechanism in the BPNN, and the residual connection design, as well as their contribution to the overall performance of the model. The ablation experiment sets up the following three models for ablation comparison, with the specific settings as follows:

[0100] (1) He-BPNN: Using the He weight initialization method, BPNN adopts the original three-layer BP neural network;

[0101] (2) ECDOA-BPNN: This model uses ECDOA to optimize the original three-layer BP neural network;

[0102] (3) ECDOA-BPNN*: Introducing residual connections into the original three-layer BP neural network;

[0103] (4) ECDOA-RGBPNN: The solution proposed in this embodiment uses ECDOA to initialize the BPNN weights and biases, and adds residual connections and gating to the original BPNN.

[0104] In order to ensure the reliability and stability of the experimental results, the random seed was changed and the experiment was repeated 20 times. The average value and standard deviation of each evaluation index of the model test were calculated. The experimental results are as follows: Figure 6 As shown. It can be seen that in terms of RMSE and MAPE indicators, ECDOA-RGBPNN has the smallest values, the R2 value is closest to 1, and the standard deviation of each indicator is also the smallest. This shows that the ECDOA-RGBPNN model has the smallest prediction error, the highest accuracy, the best data fitting effect, and the most stable model training. Specific comparison of each experimental group: a) Compared with He-BPNN, ECDOA-BPNN introduces ECDOA for weight initialization on the basis of the original BPNN, which reduces RMSE by 7.55%, MAPE is also smaller, and R 2 It is closer to 1, indicating that the ECDOA weight initialization method can produce better initial weights than He initialization, effectively improving the effect of model training; b) Compared with ECDOA-BPNN*, although the numerical differences of various indicators are not large, the standard deviations of various indicators are reduced, which means that the introduction of residual connections can make model training more stable and less likely to fall into local optimality; c) ECDOA-RGBPNN adds a gating mechanism on the basis of ECDOA-BPNN*, and the model performance is further improved. The standard deviations of various indicators are further reduced, indicating that adding gating can optimize the training effect of the model to a certain extent.

[0105] In summary, ablation experiments demonstrate that ECDOA weight initialization can effectively mitigate the problem of BPNNs being easily trapped in local optima due to poor initial weight parameters. Residual connections and gating mechanisms can help BPNNs better learn relationships between features and better fit the data. Furthermore, residual connections and gating mechanisms can make model training more stable, reducing the risk of the model falling into local optima.

[0106] The embodiments described above are merely descriptions of preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by persons skilled in the art should fall within the scope of protection defined by the claims of the present invention.

Claims

1. The vehicle weighing compensation method based on ECDOA-RGBPNN is characterized by: include: Obtain the weighing data of the target vehicle; The weighing data is input into a preset ECDOA-RGBPNN weighing compensation model, and the compensated weighing data is output, wherein the ECDOA-RGBPNN weighing compensation model is obtained based on training of a training set, and the training set includes vehicle weighing data collected by several non-stop inspection stations and overweight control stations. The ECDOA-RGBPNN weighing compensation model is constructed by introducing residual connections and gating into the BP neural network, and using an improved group decision optimization algorithm to initialize weights and biases.

2. The vehicle weighing compensation method based on ECDOA-RGBPNN according to claim 1, characterized in that: Before training the ECDOA-RGBPNN weighing compensation model based on the training set, the training set is also preprocessed, and the preprocessing includes: Obtain weighing data collected by several non-stop inspection stations and overweight control stations, and match them according to license plate numbers; Calculate the weighing error of the matching weighing data by using the preset weighing compensation target value, and remove the matching weighing data when the weighing error is greater than the preset error value, or when the matching weighing data exceeds the preset time range; The matching weighing data after elimination are processed for missing values ​​and outliers, and are one-hot encoded to construct the training set.

3. The vehicle weighing compensation method based on ECDOA-RGBPNN according to claim 2, characterized in that: The weighing compensation target value is obtained based on the weighing data of the same vehicle passing the overweight control station within a preset time range and a weighing deviation range.

4. The vehicle weighing compensation method based on ECDOA-RGBPNN according to claim 1, characterized in that: The ECDOA-RGBPNN weighing compensation model is constructed by introducing residual connections and gating into the BP neural network and using an improved group decision optimization algorithm to initialize weights and biases. The model includes: By introducing residual connections and gating into the BP neural network, the RGBPNN model is obtained; The improved group decision optimization algorithm is used to initialize the weights and biases of the RGBPNN model, and the model is trained through forward propagation and back propagation to construct the ECDOA-RGBPNN weighing compensation model.

5. The vehicle weighing compensation method based on ECDOA-RGBPNN according to claim 4 is characterized in that: Introducing residual connection and gating into the BP neural network includes: A BP neural network consisting of three fully connected layers connected in sequence is set up, a gating is added after the second fully connected layer, and the output of the first fully connected layer and the output of the second fully connected layer are spliced ​​and transmitted to the third fully connected layer to construct the RGBPNN model.

6. The vehicle weighing compensation method based on ECDOA-RGBPNN according to claim 5, characterized in that: The gate is used to perform sigmoid calculation on the output of the second fully connected layer, obtain a 0-1 gated output, and then multiply it by the output of the second fully connected layer to obtain the final output of the second fully connected layer.

7. The vehicle weighing compensation method based on ECDOA-RGBPNN according to claim 4, characterized in that: The improved group decision optimization algorithm is obtained by adopting He initialization to replace the random population initialization of the extended group decision optimization algorithm.

8. The vehicle weighing compensation method based on ECDOA-RGBPNN according to claim 7, characterized in that: The process of the improved group decision optimization algorithm is as follows: S1. Set initial parameters and termination conditions; S2. Generate an initial population using He initialization and the initial parameters and calculate the fitness value; S3, update the step size according to the current number of iterations and generate an evolution sequence; S4. Based on the current individual and population fitness values, determine whether it is the current optimal position. If not, update the position based on experience, other agents, group thinking, leaders, and innovation according to the evolutionary sequence order; if so, randomly generate candidate solutions near the current optimal position; S5. Evaluate the new solution and update it to determine whether the termination condition is met. If so, output the best candidate solution; if not, return to S3 until the termination condition is met.

Images