Method for improving metering precision of total mass detection device of large transport vehicle

Abstract

The invention discloses a method for improving the metering precision of a total mass detection device of a large transport vehicle, and belongs to the technical field of calibration of detection devices. The method comprises the following steps: firstly, taking axle load data, obtained under different test conditions, of a to-be-tested large transport vehicle as a data set for training a deep fitting neural network model, and performing cleaning, data enhancement and data set division operation; then, constructing a deep fitting neural network model, fully learning sample data in the data set, and adjusting training parameters in real time according to training conditions; and finally, the total mass of the to-be-detected large transport vehicle can be predicted only by inputting the load data, collected in real time, of each axle of the to-be-detected large transport vehicle into the trained deep fitting neural network model. By adopting the method, the identification error can be effectively reduced, the dependence on the identification precision of the axle load detection device is reduced, and the identification accuracy under the conditions of different vehicle speeds and severe road surfaces is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of calibration of detection devices, and specifically relates to a method for improving the measurement accuracy of a gross mass detection device for large-scale transport vehicles based on a deep fitting neural network. Background Art

[0002] In recent years, driven by continued advancements in power grid upgrades, rail transit construction, deployment of new energy equipment, and chemical plant expansion, the scale of oversized cargo transportation has seen exponential growth. Statistics show that nationwide, oversized cargo transportation operations exceeded 1.5 million in 2024, creating an urgent need for technological upgrades and management optimization within the transportation regulatory system. Accurate measurement of gross vehicle and cargo mass parameters is a core element of oversized cargo transportation management. This parameter is not only the legal basis for transport permit approval but also a key technical indicator for determining overloading violations. In current transportation regulatory practices, traditional measurement methods suffer from significant measurement errors, leading to extended approval cycles and reduced reliability, which in turn impacts the progress of related projects. Therefore, there is an urgent need to develop an intelligent measurement system that can achieve the coordinated optimization of rapid safety assessments and efficient approvals. With breakthrough developments in intelligent detection technology, data-driven nonlinear regression models offer innovative solutions for the accurate prediction of mass parameters. Summary of the Invention

[0003] In order to solve the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide a method for improving the measurement accuracy of the total mass detection device of large-scale transport vehicles based on deep fitting neural network. The present invention focuses on the fusion innovation of intelligent deep fitting algorithm and large-scale axle weight detection device, and constructs a measurement accuracy improvement method with adaptive measurement error compensation function by establishing a multi-dimensional feature coupling data set. This method not only solves the problem of low accuracy of traditional measurement methods, but also provides reliable technical support for the construction of intelligent large-scale measurement equipment.

[0004] In order to achieve the above object, the present invention adopts the following technical solutions: A method for improving the measurement accuracy of a large-scale transport vehicle gross mass detection device comprises the following steps: Step 1: The axle load detection device measures the axle load data of all large-scale transport vehicles under different test conditions to form a source data set with the same axle number information; Step 2: Clean and enhance all sample data in the source dataset to form a qualified dataset; Step 3: Obtain the total number of load axes of the qualified data set and construct a deep fitting neural network model whose input dimension matches the total number of load axes; Step 4: Randomly extract some sample data from the qualified data set as the training set, and use it to train the weight parameters in the deep fitting neural network model; Step 5: Adjust the training parameters and activation function types in real time to avoid gradient vanishing, gradient exploding, falling into local optimal points, overfitting, and underfitting problems in the deep fitting neural network model during training; Step 6: Use some sample data from the qualified dataset as a test set, input the sample data in the test set into the trained deep fitting neural network model and evaluate the relative error between its output value and the label value. If the relative error is less than 1.5%, the deep fitting neural network model training is complete; otherwise, repeat steps 4 to 5. Step 7: deriving weight parameters from the deep fitting neural network model trained in step 6 and obtaining its dimensionality information; then using the derived weight parameters to update the weight parameters of the deep fitting neural network model with the same input dimensions in the data processing module of the axle load detection device, thereby obtaining a calibrated axle load detection device, wherein the calibrated axle load detection device is deployed with multiple deep fitting neural network models with different input dimensions; Step 8: Match the deep fitting neural network model of the corresponding dimension in the calibrated axle weight detection device according to the total number of loaded axles of the large-scale transport vehicle to be tested, and then input the axle weight data of the large-scale transport vehicle to be tested into the successfully matched deep fitting neural network model to obtain the final predicted value of the total mass of the large-scale transport vehicle.

[0005] The different test conditions include driving speed, ramp type, road surface flatness, and changes in cargo position.

[0006] The axle weight detection device includes a weighing platform, a ramp and a data processing module deployed with a deep fitting neural network model.

[0007] To match each sample data with the deep fitting neural network model, we first need to determine the total number of load axes in each qualified sample. The input dimension of the constructed deep fitting neural network model must be the same as this total number.

[0008] The deep fitting neural network model includes 1 input layer, 3 hidden layers and 1 output layer; wherein the input layer receives the axle weight data of each load axle of the large-scale transport vehicle, and the output layer directly outputs the total mass of the large-scale transport vehicle.

[0009] The method of randomly extracting part of the sample data from the qualified data set as the training set is as follows: using a simple random sampling algorithm to extract part of the sample data from the qualified data set as the training set, and the rest as the test set; wherein the proportion of the number of training set samples to the total number of samples is 80%-92%.

[0010] The real-time adjustment of training parameters and activation function types is specifically carried out as follows: first, replacing the activation function type and the gradient clipping strategy are adopted to avoid the gradient vanishing or gradient exploding problems; a dynamic learning rate parameter adjustment mechanism is adopted to avoid falling into the local optimal point problem; and the number of nodes in each hidden layer is adapted to avoid falling into the overfitting or underfitting problem.

[0011] The activation function is Sigmoid, Tanh or ReLU.

[0012] Compared with the prior art, the present invention has the following advantages: 1. This invention improves the estimation accuracy of the total mass of large transport vehicles by traditional axle load detection devices. At present, the estimation error of the total mass of large transport vehicles based on traditional axle load measurement technology is still at However, the present invention collects the axle weight data from actual tests and inputs it into a deep fitting neural network model for continuous iterative training, so that the total mass prediction error can be controlled within 2.5%.

[0013] 2. The present invention eliminates the important steps of calibration and error correction in traditional axle load measurement systems, which reduces production costs to a certain extent. This is because the embedded deep fitting neural network model can achieve error compensation and calibration by continuously modifying weight parameters during the training process. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 This is an overall flow chart of the method for improving the measurement accuracy of the gross mass detection device for large-scale transport vehicles of the present invention.

[0015] Figure 2 This is a physical picture of the axle weight detection device.

[0016] Figure 3 This is the flow chart of the sample data compliance verification algorithm.

[0017] Figure 4 This is the flow chart of the data enhancement algorithm. DETAILED DESCRIPTION

[0018] like Figure 1 As shown, a method for improving the measurement accuracy of a large-scale transport vehicle gross mass detection device includes the following steps: Step 1: Construct a dataset for training and testing, where each sample in the dataset is all the axle weight data of a large transport vehicle obtained from an axle weight detection device under different test conditions; Step 2: Perform preprocessing operations such as cleaning, data enhancement, and data set partitioning on all sample data in the dataset to form a qualified dataset; Step 3: Obtain the total number of load axes of the qualified data set and construct a deep fitting neural network model whose input dimension matches the total number of load axes; Step 4: Randomly extract some sample data from the qualified data set as the training set, and use it to train the weight parameters in the deep fitting neural network model; Step 5: Adjust the training parameters and activation function types in real time to avoid problems such as gradient vanishing, gradient exploding, falling into local optimal points, overfitting, and underfitting during the training of deep neural network models. Step 6: Use some sample data from the qualified dataset as a test set. Input the sample data from the test set into the trained deep fitting neural network model and evaluate the relative error between its output value and the label value. If the relative error is less than 1.5%, the deep fitting neural network model training is complete. Otherwise, repeat steps 4 to 5. Step 7: deriving relevant weight parameters from the deep fitting neural network model trained in step 6 and obtaining its dimensionality information; then applying the derived relevant weight parameters to the weight parameters of the deep fitting neural network model with the same input dimensionality information in the data processing module of the axle load detection device to obtain a calibrated axle load detection device, wherein the calibrated axle load detection device is deployed with multiple deep fitting neural network models with different input dimensions; Step 8: Match the deep fitting neural network model of the corresponding dimension in the calibrated axle weight detection device according to the total number of loaded axles of the large-scale transport vehicle to be tested, and then input the axle weight data of the large-scale transport vehicle to be tested into the successfully matched deep fitting neural network model to obtain the final predicted value of the total mass of the large-scale transport vehicle.

[0019] Example 1. Build datasets for training and testing To ensure sufficient generalization of the trained deep neural network model, diversity must be ensured when constructing the dataset. Therefore, this paper focuses on acquiring relevant sample data based on different test environments, such as driving speed, ramp type, road surface smoothness, and cargo position. Once the raw data is collected, it is necessary to clean the data to handle missing values, duplicate data, and outliers.

[0020] Figure 2The axle weight detection device used for data collection in the present invention is mainly composed of three parts: a ramp, a weighing platform and a data processing module. The function of the ramp is to reduce the drop between the weighing platform and the road surface, so that the load axle of the large-scale transport vehicle to be tested will not experience severe bumps or vibrations when passing through the weighing platform. The weighing sensor in the weighing platform uses the U10F dynamic force sensor manufactured by HBM of Germany, and its range can reach 15 tons. In order to meet the testing requirements of common large-scale transport vehicles, 10 U10F dynamic force sensors are installed in parallel inside the weighing platform, so that the total range reaches 150 tons. In the data processing module, a deep fitting neural network model with corresponding input dimensions is deployed for large-scale transport vehicles with different numbers of load axles.

[0021] The data information for a sample in the source dataset can be directly used to obtain the axle weight data for each loaded axle through the weighing platform data output interface provided by the axle load detection device. After all loaded axles of the large-scale transport vehicle to be tested have passed through the weighing platform, the data for these axles is packaged into a single sample data set. By varying the test conditions, multiple sample data sets can be obtained.

[0022] For example, keeping the cargo placement unchanged and at a fixed speed under the same road conditions, five samples in the source data set are obtained from five test environments: no incline, short incline, medium incline, long incline, and extra-long incline, as shown in Table 1: Table 1 (Unit: kg)

[0023] The i-th sample uses uppercase letters Then, the j-th axle weight data in the i-th sample is expressed as In order to obtain the label value of each group of samples, a large-scale weighing device with higher precision will be used to directly weigh the total mass of the vehicle to be tested. Since the mass of the loaded cargo has not changed, the label values of all samples in Table 1 are are the same, denoted as =45000kg.

[0024] While keeping the other three environmental conditions unchanged, the present invention allows the test vehicle to randomly pass through the weighing platform at a speed of 3-7 kilometers per hour, and collects 7 samples from the source data set, as shown in Table 2: Table 2 (Unit: kg)

[0025] In addition, different test conditions can be set through permutations and combinations to obtain a wide range of test sample data under different label data. For example, taking a large transport truck with a total mass label value of 78,400 kg as the test object, by changing two of the four variables: driving speed, approach slope type, road surface smoothness, and cargo position, the following data can be obtained in Table 3: Table 3 (Unit: kg)

[0026] Next, the present invention further uses a large transport truck with a total mass label value of 138,000 kg as the test object. After combining different parameters for ramp type, road surface flatness, and cargo position at a fixed speed of 5.1 km / h, and after the cleaning and data enhancement processing in step 2, the five sets of qualified training sample data shown in Table 4 were obtained: Table 4 (Unit: kg)

[0027] 2. Data cleaning and data enhancement preprocessing After obtaining the source data set, the authenticity of the individual axle weight data and the final accumulated total mass data is verified. The data verification algorithm process is as follows: Figure 3 As shown: 1. If the relative error between the sum of the axle weight data of each sample and the label value If the data rate is greater than 13%, it is considered abnormal data and its availability needs further verification; 2. Randomly select a group of data from the qualified sample data with the label value still being 45,000 kg and calculate the mean of their axle weight data. ; 3. The sample to be verified Each axle weight data in and mean Compare and calculate the relative error ; 4. Statistics of samples to be verified middle Greater than 35% of the total ; 5. If this total If it is less than or equal to 2, the adjacent mean value replacement algorithm is used to replace the outliers and the updated samples can be used as qualified samples; otherwise, the samples to be verified are directly discarded. .

[0028] It should be noted that certain cargo with uneven centers of mass (such as wind turbine blades) or high-density cargo often results in excessively concentrated loads on certain axles of large-scale transport vehicles. Therefore, large deviations in axle load data may be reasonable. Furthermore, it is recommended that the error limit for axle load data be set within 25% to 35%. Furthermore, the relative error of conventional axle load scales for measuring the gross weight of large-scale transport vehicles is currently controlled within 15%. Therefore, the error limit for gross mass can be set between 10% and 20%. After the deep fitting neural network model is trained on a qualified dataset, more anomalous data with an error limit of 15%-20% can be further used to train the deep fitting neural network model, further improving its generalization and ability to handle unknown disturbances.

[0029] Using the original sample data in Table 2, the calculation process of the abnormal data verification algorithm is introduced in detail. , the sum of its axle weights for (1) Further calculation of the sum of axle loads With label value The relative error R1 is: (2) because It is less than the specified maximum error parameter of 13%, so we continue to check whether each axle weight data in the sample is an abnormal value. It should be noted that in order to ensure that the original test data has reference, the present invention will manually check a group of qualified sample data for different labels. The sample data with the same label value collected later will be used as follows Figure 3 The verification algorithm shown automatically verifies the compliance of the samples, thereby improving the efficiency and accuracy of data collection. Here, it can be assumed that samples 2 to 4 in Table 1 are qualified data after manual or automatic verification. Then, if the normal distribution random sampling algorithm is used to extract sample 2 in Table 1 as the standard data, its axle weight average can be calculated. Then calculate the relative deviation between the first load shaft load data and the standard value in sample 1: (3) Similarly, the relative errors of other load axes can be calculated as: , , , , , , , , , It can be found that to The values of are all less than or equal to 35%, so the data of sample 1 is qualified.

[0030] Next, continue to check sample 5 and calculate With label value The relative error is: (4) The total mass error can be found It is qualified. Continuing to use sample 2 as the standard data, the axle load data is checked and the relative error data of the 2nd, 3rd, 7th and 8th axes are found to be , , , The data significantly exceeded the specified error limit of 35%. The abnormal data may be due to a sensor failure, so sample 5 needs to be discarded.

[0031] As shown in Table 2, due to sensor failure or unstable transmission line network, the load data corresponding to axle 5 in Sample 2 was lost. In this case, the data dimension of Sample 2 changed from 11 to 10. Data with such inconsistent dimensions cannot be directly used to train a neural network. Therefore, to address this situation, the present invention uses an arithmetic mean algorithm to fill in the gaps. That is, summing up the remaining axle weight data in the sample and dividing it by the total number of axles will yield the following result: (5) Therefore, the load data for axle 5 corresponding to sample 2 is padded to 3958 kg. Furthermore, digital verification reveals that the axle weight data for each axle in samples 6 and 7 are identical. Therefore, sample 7 is eliminated, resulting in six samples that can be used for training.

[0032] In order to further expand the number of relevant samples in the data set, the present invention adopts data enhancement technology. That is, the randn function in MATLAB software that can generate data that obeys normal distribution is used to randomly generate the load data of the relevant axis in each sample. The calculation process of the specific data enhancement algorithm is as follows: Figure 4 First, the j-th axle weight data in sample i is As the mean parameter of the normal distribution, then, a larger variance value can be set for sample i under the premise of ensuring that the generated data can pass the automatic verification algorithm. , which can further improve the diversity of training sample data. The variance value recommended by this invention is The allowed range is 5-150. Next, the j-th axle weight data in the new sample i It can be calculated using the following formula: (6) Here, randn will randomly generate a specific value between 0 and 1. Similarly, the random number randn is regenerated and the next axle weight data in the standard sample is selected as the mean, and then the specific data of the corresponding loaded axis is calculated using formula (6).

[0033] The present invention uses the data enhancement algorithm to generate 5 groups of qualified sample data as shown in Table 5. Next, the calculation process of the data enhancement algorithm is introduced in detail in combination with the enhanced sample data related to sample 1 in Table 5. First, the first axle weight data corresponding to sample 1 in Table 4 is As the mean parameter of the normal distribution; then use the randn function to generate a random number of 0.8; then the variance value Set to 100; finally, use formula (6) to calculate the relevant data of axis 1 corresponding to sample 1 in Table 5: (7) Table 5 (Unit: kg)

[0034] 3. Build a deep fitting neural network model The present invention designs a deep fitting neural network with three hidden layers. For the sample data provided in this embodiment, it can be found that each sample contains load data of 11 axles. Therefore, the input layer of the deep fitting neural network model is set with 11 nodes, which are respectively used to receive the data of each axle weight. In addition, it is necessary to reasonably determine the node parameters of each hidden layer according to the complexity of the fitting problem. Too many node parameters can easily cause overfitting problems. However, too few node parameters can also easily cause underfitting problems, so the node parameters of each hidden layer need to be continuously debugged during the training process. Finally, the number of nodes in the first hidden layer is set to 8, the number of nodes in the second hidden layer is set to 16, and the number of nodes in the third hidden layer is set to 32. It should be noted that the number of nodes in the hidden layer is recommended to be set to an integer multiple of 2, which is conducive to improving the computational efficiency of the neural network. Since the output layer directly outputs the predicted value of the total mass of the large-scale transport vehicle, it can be set to 1.

[0035] 4. Dataset Segmentation In order to make the extracted training set more representative and reduce sampling errors, the present invention adopts a stratified sampling method characterized by different total mass label values. This embodiment collects a total of 30 sets of qualified data set samples, including enhanced sample data. Among them, the sample data with a total mass label value of 45,000 kg totals 10, the sample data with a total mass label value of 78,400 kg totals 10, and the sample data with a total mass label value of 138,000 kg totals 10. First, the sampling ratio of the test set and the training set is set to 1:9; then, 1 data is extracted from the sample sets with label values of 45,000 kg, 78,400 kg and 138,000 kg respectively to form a test set as shown in Table 6. Of course, the remaining samples are used as training sets for training neural networks. If the prediction performance of the trained neural network does not meet the given performance indicators, the data set can be re-split and the training operation can be performed again.

[0036] Table 6 (Unit: kg)

[0037] 5. Adjust parameters based on training results First, the backpropagation process uses the Levenberg-Marquardt optimization algorithm to correct weight parameters. It has advantages such as fast convergence and strong real-time performance for handling nonlinear fitting problems. Next, a maximum number of training cycles (epochs) is set, which is the maximum number of iterations during network training. A larger epoch count increases training time. Therefore, a smaller value, such as 20, can be used in the initial stages of training. However, significant errors were still observed when testing the trained neural network using a test set, so epochs was set to 100. Next, the learning rate is set. Its primary function is to control the step size for adjusting network weights. A larger learning rate can accelerate convergence but may lead to oscillation or instability; a smaller learning rate results in more stable training but slower convergence. In this paper, the learning rate is set to 0.1. Finally, learning rate decay dynamically adjusts the learning rate to control the speed of parameter updates. Using a larger learning rate in the early stages of training allows for rapid convergence to the optimal solution. As training progresses, gradually reducing the learning rate allows the network to conduct a more refined search near the optimal solution, avoiding missing the optimal solution due to excessively large step sizes. Therefore, in this embodiment, the learning decay rate is set to 0.8. In the experiment, it was found that the deep fitting neural network model training can achieve better training results when the above training parameters are used for training.

[0038] 6. Prediction Performance Evaluation The prediction performance of the deep fitting neural network model is evaluated using the test set data shown in Table 6. Inputting the axle weight data of each sample in Table 6 into the deep fitting neural network model, the total mass can be predicted in turn as follows: , =78428.95708kg and =137370.2722kg, as shown in the third row of Table 7. The traditional method of calculating the total mass is the sum of all axle load data. The total mass can be calculated based on the axle load data of each sample in Table 6. , =72660kg and =136910kg, as shown in the second row of Table 7.

[0039] Table 7 (Unit: kg)

[0040] The relative error generated by the total mass prediction algorithm provided by the present invention is calculated using the following formula: (8) in Represents the label value corresponding to sample i. The calculation results are =0.16%, =0.04%, =0.46%, as shown in the third row of Table 8. It can be found that each relative error is less than 1.5%, so the approximate fitting capability of the deep fitting neural network model has met the specified performance requirements and can be deployed in the data processing module.

[0041] Next, the following formula can be used to calculate the relative error caused by the traditional algorithm in calculating the total mass: (9) The calculation results are =5.56%, =7.32%, =0.79%, as shown in the second row of Table 8.

[0042] Table 8

[0043] For a certain test set, an average cumulative error index can be defined to describe the estimation accuracy of the total quality estimation algorithm. Taking the test set provided in Table 6 as an example, the cumulative error of the traditional quality estimation algorithm can be calculated as (10) The cumulative error caused by the quality estimation algorithm provided by the present invention is (11) Therefore, the accuracy of estimating the total mass of large-scale transport vehicles by using deep fitting neural networks is improved compared with traditional algorithms. =95.18%.

Claims

1. A method for improving the measurement accuracy of a large-scale transport vehicle gross mass detection device, characterized by: The steps include: Step 1: The axle load detection device measures the axle load data of all large-scale transport vehicles under different test conditions to form a source data set with the same axle number information; Step 2: Clean and enhance all sample data in the source dataset to form a qualified dataset; Step 3: Obtain the total number of load axes of the qualified data set and construct a deep fitting neural network model whose input dimension matches the total number of load axes; Step 4: Randomly extract some sample data from the qualified data set as the training set, and use it to train the weight parameters in the deep fitting neural network model; Step 5: Adjust the training parameters and activation function types in real time to avoid gradient vanishing, gradient exploding, falling into local optimal points, overfitting, and underfitting problems in the deep fitting neural network model during training; Step 6: Use some sample data from the qualified dataset as a test set. Input the sample data in the test set into the trained deep fitting neural network model and evaluate the relative error between its output value and the label value. If the relative error is less than 1.5%, the deep fitting neural network model training is complete. Otherwise, repeat steps 4 to 5; Step 7: Export the weight parameters from the deep fitting neural network model trained in step 6 and obtain its dimension information; The derived weight parameters are then used to update the weight parameters of a deep fitting neural network model with the same input dimensions in a data processing module of the axle load detection device, thereby obtaining a calibrated axle load detection device, wherein the calibrated axle load detection device is deployed with multiple deep fitting neural network models with different input dimensions; Step 8: Match the deep fitting neural network model of the corresponding dimension in the calibrated axle weight detection device according to the total number of loaded axles of the large-scale transport vehicle to be tested, and then input the axle weight data of the large-scale transport vehicle to be tested into the successfully matched deep fitting neural network model to obtain the final predicted value of the total mass of the large-scale transport vehicle.

2. A method for improving the measurement accuracy of a large-cargo transport vehicle gross mass detection device according to claim 1, characterized in that: The different test conditions include driving speed, ramp type, road surface smoothness, and changes in cargo position.

3. The method for improving the measurement accuracy of a large-cargo transport vehicle gross mass detection device according to claim 1, characterized in that: The axle weight detection device includes a weighing platform, a ramp and a data processing module deployed with a deep fitting neural network model.

4. The method for improving the measurement accuracy of a gross mass detection device for a large transport vehicle according to claim 1, characterized in that: To match each sample data with the deep fitting neural network model, we first need to determine the total number of load axes in each qualified sample. The input dimension of the constructed deep fitting neural network model must be the same as this total number.

5. The method for improving the measurement accuracy of a large-cargo transport vehicle gross mass detection device according to claim 1, characterized in that: The deep fitting neural network model includes 1 input layer, 3 hidden layers and 1 output layer; wherein the input layer receives the axle weight data of each load axle of the large-scale transport vehicle, and the output layer directly outputs the total mass of the large-scale transport vehicle.

6. The method for improving the measurement accuracy of a gross mass detection device for a large transport vehicle according to claim 1, characterized in that: The method of randomly extracting part of the sample data from the qualified data set as the training set is as follows: using a simple random sampling algorithm to extract part of the sample data from the qualified data set as the training set, and the rest as the test set; wherein the proportion of the number of training set samples to the total number of samples is 80%-92%.

7. The method for improving the measurement accuracy of a large-cargo transport vehicle gross mass detection device according to claim 1, characterized in that: The real-time adjustment of training parameters and activation function types is specifically carried out as follows: first, replacing the activation function type and the gradient clipping strategy are adopted to avoid the gradient vanishing or gradient exploding problems; a dynamic learning rate parameter adjustment mechanism is adopted to avoid falling into the local optimal point problem; and the number of nodes in each hidden layer is adapted to avoid falling into the overfitting or underfitting problem.

8. The method for improving the measurement accuracy of a large-cargo transport vehicle gross mass detection device according to claim 7, characterized in that: The activation function is Sigmoid, Tanh or ReLU.

Images