Self-learning-based load identification and confidence evaluation method for commercial vehicle

Abstract

Disclosed in the present invention is a self-learning-based load identification and confidence evaluation method for a commercial vehicle. The method comprises: first, using big data extraction technology to extract, from a multi-source data set, vehicle operation data required by a model; then, identifying and preprocessing abnormal frames in the vehicle operation data according to a data cleaning strategy, and completing expansion operations of vehicle operation features and road structure features; and using a recursive least squares algorithm with a forgetting factor to construct a load reference generation model based on vehicle longitudinal dynamics, and using an optimized Informer neural network to construct a load identification output and confidence evaluation model, so as to complete a vehicle load identification and confidence evaluation scheme. By means of the apparatus in the present invention, Internet-of-Vehicles data is used to realize the real-time identification of a vehicle load, thereby facilitating the standardization of in-transit full-process supervision of heavy-duty freight vehicles, and providing a reliable basis for the optimization of the utilization rate of a vehicle travel.

Description

A self-learning method for identifying and assessing the load capacity of commercial vehicles Technical Field

[0001] This invention belongs to the field of vehicle parameter identification, specifically relating to a self-learning method for identifying the load capacity of commercial vehicles and assessing confidence levels. Background Technology

[0002] With the development of technologies such as big data science, 5G communication, and artificial intelligence, the electrification, connectivity, and intelligence of automobiles are gradually becoming the development direction of the automotive industry. Intelligent connected vehicles refer to vehicles equipped with advanced onboard sensors, controllers, actuators, and other devices, enabling them to perceive complex environments, make intelligent decisions, and coordinate control, greatly facilitating the acquisition of various parameters under vehicle operating conditions. Combined with modern communication and network technologies, the data storage capacity and accuracy of vehicle-to-everything (V2X) networks have significantly improved. Under this development trend, vehicle parameter identification and data mining based on massive amounts of V2X data have gradually become research hotspots.

[0003] In recent years, with the rapid development of the transportation industry, road freight transport has accounted for 73.8% of the total freight volume in my country. Currently, there are two main measures to achieve full-process on-the-road monitoring of freight vehicles: one is to install additional gravity sensors, and the other is to have drivers manually input and upload data. However, both methods have certain problems: installing gravity sensors will increase vehicle costs, making it difficult to widely implement for general vehicles and small fleets; while the authenticity and accuracy of data uploaded manually by drivers will be affected by factors such as individual driver factors.

[0004] With the increasing maturity of vehicle-to-everything (V2X) technology, identifying and estimating the vehicle weight based on vehicle operating condition measurement data from intelligent connected vehicles is one of the important application solutions for V2X data. Through the research and development of vehicle load estimation algorithms, the real-time cargo transportation status of vehicles can be obtained, enabling full-process monitoring of the vehicle's on-road status. This allows for effective supervision of vehicle overloading, thus preventing traffic accidents to a certain extent. Simultaneously, by obtaining real-time vehicle load information, abnormal loading and unloading behavior of trucks can be avoided, and the utilization rate of vehicles in the corresponding area can also be reflected, thereby helping freight companies to rationally allocate trucks, avoid wasting vehicle transport capacity resources, and improve transportation efficiency. Summary of the Invention

[0005] Purpose of the invention: This invention provides a self-learning method for identifying the load capacity of commercial vehicles and assessing their confidence level, which can realize full-process monitoring of freight vehicles en route.

[0006] Technical solution: The self-learning commercial vehicle load identification and confidence assessment method of the present invention includes the following steps:

[0007] (1) Vehicle operation initialization: includes GPS module, vehicle management system, vehicle communication module and vehicle static parameter initialization. The vehicle static parameters include the following fields: vehicle identification number, vehicle model, engine type, vehicle width, vehicle height, rear axle speed ratio, gear ratio of each gear, wheel rolling radius, and engine maximum torque.

[0008] (2) Multi-source data acquisition: Data acquired through the GPS module includes timestamps, elevation, and vehicle speed; data acquired through the CAN bus includes engine speed, engine torque, and gear information.

[0009] (3) Operation data detection and cleaning: Implement corresponding cleaning strategies for missing, noisy or other erroneous frames in the operation data, and then take preprocessing measures such as interpolation and filtering for operation data that do not meet the continuity conditions;

[0010] (4) Vehicle operation characteristics extension: Based on the vehicle system transmission characteristics and road structure characteristics, the preprocessed vehicle operation data is used to obtain the road slope characteristics, vehicle acceleration and jerk characteristics of the entire road segment at all times;

[0011] (5) Load joint identification system design: Establish a load benchmark generation model based on recursive least squares method, and obtain the distribution benchmark of estimated load through vehicle operation data iteration; establish a load regression identification model based on improved Informer encoder, and output load identification results through GPS source data and extended operation features;

[0012] (6) Confidence assessment and model fine-tuning: Based on the dynamic load distribution benchmark, the confidence assessment of the load identification result is completed by combining the output of the neural network load regression model; store the vehicle operation data and total weight benchmark data at each moment of the whole journey, and perform offline data training and fine-tuning of the load regression identification model after the current journey ends, and redeploy the model after training is completed;

[0013] Furthermore, the implementation process of step (3) is as follows:

[0014] Considering the differences in the operating characteristics of non-same-source data, erroneous frames in GPS source data and CAN bus source data are manually feature extracted and labeled. Based on the continuity criterion, erroneous frame data samples are separated, and the following multi-source data erroneous frame feature and cleaning strategy are formulated.

[0015] GPS source data: Error frames mainly originate from data interruptions. Based on the data feedback timestamps, they can be categorized into short-term data loss and long-term signal interruptions. A time threshold should be set accordingly. For short-term data loss less than this threshold, linear interpolation is performed using the nearest non-error data value before and after the error frame. For long-term signal interruptions exceeding this threshold, the data during that period is set to zero and discarded.

[0016] CAN bus source data: Error frame data is mainly characterized by short-term data loss, high-frequency measurement noise, and discontinuous abrupt changes.

[0017] Short-term data missing mainly targets vehicle gear data detection. Based on the principle of continuous change in vehicle gears, short-term single-point missing data is filled with non-abnormal data from the previous moment.

[0018] High-frequency measurement noise is mainly aimed at detecting engine speed and engine torque data. A Savitzky-Golay filter is designed to process the noise data.

[0019] Discontinuous abrupt changes are primarily targeted at engine torque data detection, and the characteristic detection standard is set as follows:

[0020]

[0021] Where R represents the engine torque data at a certain moment, and T represents the timestamp at the corresponding moment. To calculate the time step, z is the time step. The detection threshold is set; for the data points obtained according to the above detection standards, a linear interpolation method is used to cover the data values ​​at the corresponding time points.

[0022] Furthermore, the implementation process of step (4) is as follows:

[0023] After the vehicle operation data is cleaned, based on the vehicle parameter attributes and operating characteristics, the preprocessed vehicle operation data is used to obtain vehicle operating speed characteristics:

[0024]

[0025] This allows us to obtain the vehicle's acceleration and jerk characteristics:

[0026]

[0027]

[0028] Where n is the engine speed and r is the wheel rolling radius. This refers to the gear ratio of the transmission. Rear axle speed ratio For time step, This is the vehicle speed value. This is the acceleration value. The degree of urgency.

[0029] Based on the construction standards and changing characteristics of highways, the road slope is obtained by preprocessing vehicle speed and elevation data.

[0030]

[0031] In the formula, This is the time step for backward calculation; Let be the elevation value at time t. for Elevation value at time, in units of ; This is the vehicle speed value, in units of .

[0032] Slope calculations are performed at points with a distance span of 100m, ensuring that elevation data changes monotonically within a single calculation step, and the time step k satisfies:

[0033]

[0034]

[0035] Based on the above calculation strategy and formula, the data of all running segments are traversed to obtain the road gradient characteristics, vehicle acceleration and jerk characteristics of the entire journey.

[0036] Furthermore, the load-bearing joint identification system in step (5) is designed as follows:

[0037] Based on the vehicle dynamics model, a reasonable distribution range of the estimated load is obtained, and the longitudinal dynamics model of the vehicle under driving conditions is adopted:

[0038]

[0039] Where g is the acceleration due to gravity, M is the mass of the car, and f is the coefficient of friction. For road slope, C D The air drag coefficient, Let ρ be the vehicle's frontal area and ρ be the air density. This is the conversion factor for the vehicle's moment of inertia. For vehicle speed, To accelerate the vehicle, This refers to engine torque. This refers to the gear ratio of the transmission. Rear axle speed ratio The total efficiency of the transmission system is given by r, where r is the rolling radius of the wheel.

[0040] A recursive least squares algorithm with a forgetting factor is used to obtain the distribution benchmark of the estimated load based on the above vehicle dynamics model filtering, and the loss function of the filtering algorithm is constructed as follows:

[0041]

[0042] in, Indicates the forgetting factor, Indicates the actual output of the system. Indicates system input, express The vector of parameters to be estimated at time t.

[0043] The load reference generation model based on recursive least squares is designed as follows:

[0044]

[0045] A load regression identification model was built using an Informer-based neural network architecture. The model input parameters were selected as: vehicle speed, acceleration, jerk, road elevation, and road slope. The model output parameters were the load regression identification results.

[0046] Furthermore, the implementation process of step (6) is as follows:

[0047] Construct using the load distribution output by the recursive least squares algorithm The distribution interval is used, and the interval average is used as the confidence evaluation benchmark. The confidence of the load regression identification model output of the Informer encoder is evaluated, and MAE is used as the evaluation index.

[0048]

[0049] in, The output of the load regression identification model. This represents the average value within the load reference range;

[0050] The confidence level assessment criteria are as follows:

[0051]

[0052] The load regression identification results and confidence assessment results iterate over time and are continuously output as the vehicle runs until the end of the vehicle's working journey; the load regression identification results and confidence assessment results at each time dimension are recorded in the vehicle's storage unit;

[0053] After the vehicle's current working journey ends, the confidence level of the load identification results throughout the entire journey is statistically analyzed. Load identification data with a confidence level below 10% are marked. The average load identification benchmark at the corresponding time point is used as the fine-tuning label to generate the fine-tuning dataset of the Informer encoder for offline training. After the model is fine-tuned, it is redeployed.

[0054] Through the above model calculations and identification, the load identification results and confidence evaluation values ​​can be output, and the dataset generation, model self-learning and model redeployment can be completed after the vehicle trip ends.

[0055] The present invention also provides a self-learning commercial vehicle load identification and confidence assessment system, the system comprising:

[0056] Initialization module: Used for initializing the vehicle GPS module, vehicle management system, vehicle communication module, and vehicle static parameters;

[0057] Multi-source data acquisition module: used to acquire data including timestamp, elevation, and vehicle speed through GPS module, and data including engine speed, engine torque, and gear information through CAN bus;

[0058] Data cleaning and feature expansion module: It is used to implement corresponding cleaning strategies for missing, noisy and other erroneous frame data in the running data, and to take preprocessing measures such as interpolation and filtering for running data that does not meet the continuity condition. Based on the transmission characteristics of the vehicle system and the characteristics of the road structure, it uses the preprocessed vehicle running data to obtain the road slope characteristics, vehicle acceleration and jerk characteristics of the entire road segment at all times.

[0059] The load joint identification and model fine-tuning module is used to obtain the distribution benchmark of the estimated load based on the load benchmark generation model based on the recursive least squares method, output the load identification result based on the load regression identification model based on the improved Informer encoder, evaluate the confidence of the load identification result based on the dynamic load distribution benchmark and the output of the neural network load regression model, store the vehicle operation data and total weight benchmark data at all times of the whole journey, and perform offline data training, fine-tuning and redeployment of the load regression identification model.

[0060] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement any of the above-described self-learning commercial vehicle load identification and confidence assessment methods.

[0061] Beneficial Effects: Compared with existing technologies, the beneficial effects of this invention are as follows: 1. This invention considers the technological advantages and levels of existing intelligent connected freight vehicles, combines the content and characteristics of vehicle network data, and adopts mature data preprocessing methods for typical vehicle network data types. Based on the known vehicle and road status information, it accurately identifies the total vehicle weight; 2. This invention uses a recursive least squares algorithm with a forgetting factor and an optimized Informer neural network model, which can achieve real-time identification of the total vehicle load and high-precision output of the result confidence level; 3. This invention can achieve automatic learning and fine-tuning of the model. With the accumulation of application scope and application time, the accuracy and versatility of the identification regression model will continue to improve; 4. This invention can be applied to many fields related to commercial vehicles. In the vehicle design stage, understanding the vehicle operating load conditions in the segmented market makes the vehicle development goals more accurate. In the vehicle usage stage, it enables precise matching of cargo and transport capacity, optimization of cargo loading rate, improvement of vehicle energy consumption assessment, calculation of intelligent maintenance strategies, calculation of residual value of used vehicles, and even participation in the calculation of intelligent predictive driving strategies. Attached Figure Description

[0062] Figure 1 is a flowchart illustrating the self-learning commercial vehicle load identification and confidence assessment method provided by the present invention.

[0063] Figure 2 is a schematic diagram illustrating the effect of real-time data acquisition preprocessing provided by the present invention.

[0064] Figure 3 is a schematic diagram illustrating the effect of real-time acquisition of elevation data and preprocessing of slope data provided by the present invention.

[0065] Figure 4 shows the output results of the recursive least squares load estimation model provided by this invention.

[0066] Figure 5 is a training loss diagram of the load regression identification model provided by the present invention.

[0067] Figure 6 is an error diagram of the output results of the multi-stroke load regression identification model provided by the present invention.

[0068] Figure 7 is a schematic diagram of the structure of the electronic device provided by the present invention. Embodiments of the present invention

[0069] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0070] The present invention provides a self-learning method for identifying and assessing the load capacity of commercial vehicles, comprising the following steps:

[0071] (1) Vehicle operation initialization: includes GPS module, vehicle management system, vehicle communication module and vehicle static parameter initialization. The vehicle static parameters include the following fields: vehicle identification number, vehicle model, engine type, vehicle width, vehicle height, rear axle speed ratio, gear ratio of each gear, wheel rolling radius, and engine maximum torque.

[0072] (2) Multi-source data acquisition: Data acquired through the GPS module includes timestamps, elevation, and vehicle speed; data acquired through the CAN bus includes engine speed, engine torque, and gear information.

[0073] (3) Operation data detection and cleaning: Implement corresponding cleaning strategies for missing, noisy or other erroneous frames in the operation data, and then take preprocessing measures such as interpolation and filtering for operation data that do not meet the continuity conditions;

[0074] (4) Vehicle operation characteristics extension: Based on the vehicle system transmission characteristics and road structure characteristics, the preprocessed vehicle operation data is used to obtain the road slope characteristics, vehicle acceleration and jerk characteristics of the entire road segment at all times;

[0075] (5) Load joint identification system design: Establish a load benchmark generation model based on recursive least squares method, and obtain the distribution benchmark of estimated load through vehicle operation data iteration; establish a load regression identification model based on improved Informer encoder, and output load identification results through GPS source data and extended operation features;

[0076] (6) Confidence assessment and model fine-tuning: Based on the dynamic load distribution benchmark, the confidence assessment of the load identification result is completed by combining the output of the neural network load regression model; store the vehicle operation data and total weight benchmark data at each moment of the whole journey, and perform offline data training and fine-tuning of the load regression identification model after the current journey ends, and redeploy the model after training is completed;

[0077] Furthermore, the implementation process of step (1) is as follows:

[0078] A UBLOX F9P high-precision GNSS receiver is used to establish a spatiotemporal reference. A communication link is established with the vehicle-mounted industrial control computer through a CAN-FD interface compatible with the SAE J1939 protocol, and a power-on self-diagnosis process is executed. After the device self-diagnosis is successful, the vehicle's inherent parameter set is loaded from the encrypted storage medium. The parameters are stored in JSON-LD format, and the transmission security is ensured by the TLS 1.3 protocol, which meets the ISO 21434 network security standard.

[0079] Table 1 provides detailed information on the vehicle's static parameters.

[0080]

[0081] Furthermore, the implementation process of step (2) is as follows:

[0082] The vehicle is equipped with a self-developed Xbox device from Yukuai Innovation, which uses a combination of UBLOX F9P and TDK ICM-42688-P IMU to acquire the vehicle's 3D position, speed, and UTC timestamp at a sampling frequency of 1 Hz. Engine speed, torque, and gear information are analyzed via the CAN bus. The PTPv2 (IEEE 1588) protocol is used to synchronize the GPS and CAN bus clocks, with a time deviation ≤1ms; CRC32 checksum is applied to all CAN frames.

[0083] Table 2 shows the detailed information of real-time data collected by the XBox device.

[0084]

[0085] Table 3 shows the vehicle data collected for multiple routes.

[0086]

[0087] Furthermore, the implementation process of step (3) is as follows:

[0088] Considering the differences in the operating characteristics of non-same-source data, erroneous frames in GPS source data and CAN bus source data are manually feature extracted and labeled. Based on the continuity criterion, erroneous frame data samples are separated, and the following multi-source data erroneous frame feature and cleaning strategy are formulated.

[0089] GPS source data: Error frames mainly originate from data interruptions. Based on the data feedback timestamps, they can be categorized into short-term data loss and long-term signal interruptions. A time threshold should be set accordingly. For short-term data loss less than this threshold, linear interpolation is performed using the nearest non-error data value before and after the error frame. For long-term signal interruptions exceeding this threshold, the data during that period is set to zero and discarded.

[0090] CAN bus source data: Error frame data is mainly characterized by short-term data loss, high-frequency measurement noise, and discontinuous abrupt changes.

[0091] Short-term data missing mainly targets vehicle gear data detection. Based on the principle of continuous change in vehicle gears, short-term single-point missing data is filled with non-abnormal data from the previous moment.

[0092] High-frequency measurement noise is mainly aimed at detecting engine speed and engine torque data. A Savitzky-Golay filter is designed to process the noise data. The filter is constructed using a quadratic polynomial, and the filter window length is set to 19.

[0093] Discontinuous abrupt changes are primarily targeted at engine torque data detection, and the characteristic detection standard is set as follows:

[0094]

[0095] Where R represents the engine torque data at a certain moment, and T represents the timestamp at the corresponding moment. To calculate the time step, z is the time step. The detection threshold is set; for the data points obtained according to the above detection standards, a linear interpolation method is used to cover the data values ​​at the corresponding time points.

[0096] Taking the vehicle driving data with a length of 180s collected in step (2) above as an example, the time step is set. Engine speed detection threshold Engine torque detection threshold The data preprocessing measures applied in step (3) are visualized as shown in Figure 3. As shown in Figure 3, the blue curve is the visualization curve of the original data. During the change process, there is significant noise data and missing data that jumps to 0 at a single point. The data after preprocessing is shown as the red curve in the figure. It can be clearly seen that after applying the data preprocessing measures proposed in this invention, the noise of the original data can be reduced to the greatest extent while effectively ensuring the original trend of data change, and the missing data can be accurately located and filled.

[0097] Furthermore, the implementation process of step (4) is as follows:

[0098] After the vehicle operation data is cleaned, based on the vehicle parameter attributes and operating characteristics, the preprocessed vehicle operation data is used to obtain vehicle operating speed characteristics:

[0099]

[0100] This allows us to obtain the vehicle's acceleration and jerk characteristics:

[0101]

[0102]

[0103] Where n is the engine speed and r is the wheel rolling radius. This refers to the gear ratio of the transmission. Rear axle speed ratio For time step, This is the vehicle speed value. This is the acceleration value. The degree of urgency.

[0104] Based on the construction standards and changing characteristics of highways, the road slope is obtained by preprocessing vehicle speed and elevation data.

[0105]

[0106] In the formula, This is the time step for backward calculation; Let be the elevation value at time t. for Elevation value at time, in units of ; This is the vehicle speed value, in units of .

[0107] Slope calculations are performed at points with a distance span of 100m, ensuring that elevation data changes monotonically within a single calculation step, and the time step k satisfies:

[0108]

[0109]

[0110] Based on the above calculation strategy and formula, the data of all running segments are traversed to obtain the road gradient characteristics, vehicle acceleration and jerk characteristics of the entire journey.

[0111] Taking a 300s segment of vehicle driving data collected in step (2) as an example, the slope data within the vehicle driving section is obtained using the method described in step (4), as shown in Figure 3. As shown in Figure 3, the real-time elevation data changes continuously within a timing cycle. At each corresponding moment, the slope obtained by the preprocessing in step (4) changes continuously with the real-time elevation, and the slope and elevation are highly consistent in terms of change trend and numerical characteristics. It can be clearly seen that the slope preprocessing measures proposed in this invention can accurately and continuously obtain the slope information corresponding to the real-time elevation.

[0112] Furthermore, the load-bearing joint identification system in step (5) is designed as follows:

[0113] Based on the vehicle dynamics model, a reasonable distribution range of the estimated load is obtained, and the longitudinal dynamics model of the vehicle under driving conditions is adopted:

[0114]

[0115] Where g is the acceleration due to gravity, M is the mass of the car, and f is the coefficient of friction. For road slope, C D The air drag coefficient, Let ρ be the vehicle's frontal area and ρ be the air density. This is the conversion factor for the vehicle's moment of inertia. For vehicle speed, To accelerate the vehicle, This refers to engine torque. This refers to the gear ratio of the transmission. Rear axle speed ratio The total efficiency of the transmission system is given by r, where r is the rolling radius of the wheel.

[0116] A recursive least squares algorithm with a forgetting factor is used to obtain the distribution benchmark of the estimated load based on the above vehicle dynamics model filtering, and the loss function of the filtering algorithm is constructed as follows:

[0117]

[0118] in, Indicates the forgetting factor, Indicates the actual output of the system. Indicates system input, express The vector of parameters to be estimated at time t.

[0119] The load reference generation model based on recursive least squares is designed as follows:

[0120]

[0121] in:

[0122]

[0123]

[0124]

[0125] Taking the four segments of vehicle driving data collected in step (2) above as an example, the recursive least squares load estimation method described in step (5) is used, and the output results are visualized as shown in Figure 4. As shown in Figure 4, in the estimation process of the four segments with different real loads, there is a stage of change from zero to close to the real load. Subsequently, the estimated value oscillates slightly around the true value. The final estimated result is used as the output result, and the model estimation error is less than 5%. The recursive least squares estimation method proposed in this invention is accurate and effective for load estimation.

[0126] A load regression identification model was built using an Informer-based neural network architecture. The model input parameters were selected as: vehicle speed, acceleration, jerk, road elevation, and road slope. The model output parameters were the load regression identification results.

[0127] The Informer encoder structure is designed as follows: encoding feature dimension (120, 7), extended feature dimension (120, 128), 4 encoder layers, 256 fully connected layers, 10 sampling factors, 8 multi-head attention points, and output feature dimension (1, 1). The specific application steps of the model are as follows:

[0128] (1) Network initialization

[0129] The feature expansion layer weights are initialized using a Xavier normal distribution, mapping the 7-dimensional input features to 128 dimensions, resulting in a 7×128 weight matrix. In the 4-layer encoder, the multi-head attention projection matrix is ​​initialized using a Kaiming uniform distribution, with each head having a dimension of 16, and the projection matrix dimensions are all 128×128. The fully connected layer contains a 256-dimensional intermediate layer, with its weight matrix and bias terms initialized to random values ​​within the range of 0.01. The scaling factor of each LayerNorm is initialized to 1, and the offset is initialized to 0.

[0130] (2) Input data preprocessing

[0131] The input data is a time-series feature sequence with dimensions (120, 7), containing 7-dimensional features at 120 time steps. Preprocessing first uses Z-score standardization to eliminate the influence of dimensions, calculates the mean and standard deviation of each feature based on the training set, and converts each feature value into the form of (feature value - mean) / standard deviation.

[0132] (3) Forward propagation calculation

[0133] The input features are expanded in dimension through a linear layer, mapping from 7 dimensions to 128 dimensions. The expanded features are then fed into a 4-layer stacked encoder structure, with each layer containing a ProbSparse Attention mechanism and residual connections. The 4-layer stacked encoder outputs (120, 128) encoded features. The features at the last time step are fed into a fully connected layer, which, after passing through a 256-dimensional intermediate layer and ReLU activation, outputs the weight prediction value.

[0134] (4) Loss function calculation

[0135] Mean squared error is used as the loss function to measure the difference between the predicted value and the actual load label. It is calculated as the squared mean of the prediction errors of all samples within the batch. The actual load label unit is tons, and the loss value corresponds to the square of tons, thus quantifying the prediction bias.

[0136] (5) Parameter update

[0137] The gradient of the loss function with respect to all learnable parameters is calculated using an automatic differentiation tool, and the parameters are updated using the Adam optimizer. The initial learning rate of the optimizer is set to... Momentum parameters , The step size is dynamically adjusted and updated by estimating the first and second moments.

[0138] (6) Iteration and convergence judgment

[0139] Training is performed using batch random sampling. In each iteration, a batch of data is randomly selected for forward propagation, loss calculation, and parameter updates. The validation set loss is calculated every 100 iterations to monitor generalization ability. If the validation set loss does not decrease for five consecutive iterations, the learning rate is decayed to 0.5 times its current value. Convergence is achieved when the validation set loss is below 0.01 (tons squared) or the number of iterations reaches 500; training terminates when either condition is met.

[0140] (7) Model saving and deployment

[0141] After training, the network parameters and normalized statistics are saved. During the deployment phase, 120 real-time time-series features are input, and after standardization preprocessing, the load estimate is output through forward propagation.

[0142] Taking the six vehicle driving data segments collected in step (2) as an example, the model training process described in step (5) is used, and the model training loss is shown in Figure 5. The output result of applying the trained model to the six vehicle driving data segments is shown in Figure 6. As shown in the box plot in Figure 6, the final output error of the load regression identification model converges to Within the scope of this invention, the load regression identification model provides the required accuracy for estimating the load of vehicles in real-time operation.

[0143] Furthermore, the implementation process of step (6) is as follows:

[0144] Construct using the load distribution output by the recursive least squares algorithm The distribution interval is used, and the interval average is used as the confidence evaluation benchmark. The confidence of the load regression identification model output of the Informer encoder is evaluated, and MAE is used as the evaluation index.

[0145]

[0146] in, The output of the load regression identification model. This represents the average value within the load reference range;

[0147] The confidence level assessment criteria are as follows:

[0148]

[0149] The load regression identification results and confidence assessment results iterate over time and are continuously output as the vehicle runs until the end of the vehicle's working journey; the load regression identification results and confidence assessment results at each time dimension are recorded in the vehicle's storage unit;

[0150] After the current working journey of the vehicle is completed, the confidence level of the load identification results during the entire journey is statistically analyzed, and the load identification data with a confidence level of less than 10% are marked. The average load identification benchmark at the corresponding time is used as the fine-tuning label to generate the fine-tuning dataset of the Informer encoder for offline training. The above steps (4) and (5) are repeated to fine-tune the existing model. After the model fine-tuning is completed, it is redeployed.

[0151] Figure 7 illustrates a schematic diagram of the physical structure of an electronic device. As shown in Figure 7, the electronic device may include: a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus. The processor can call logical instructions in the memory to execute a self-learning commercial vehicle load identification and confidence assessment method. This method includes: an onboard GPS module, a vehicle management system, a vehicle communication module, and vehicle static parameter initialization; acquiring data including timestamps, elevation, and vehicle speed through the GPS module, and acquiring data including engine speed, engine torque, and gear information through the CAN bus; implementing corresponding cleaning strategies for missing, noisy, and other erroneous frames in the operational data; and taking preprocessing measures such as interpolation and filtering for operational data that does not meet the continuity condition; based on the vehicle system transmission characteristics... Based on road structure characteristics, preprocessed vehicle operation data is used to obtain road slope characteristics, vehicle acceleration, and jerk characteristics for the entire road segment at all times. A load baseline generation model based on recursive least squares method is used to obtain the distribution baseline of the estimated load. A load regression identification model based on an improved Informer encoder outputs the load identification results. Based on the dynamic load distribution baseline, the output of the neural network load regression model is combined to evaluate the confidence of the load identification results. Vehicle operation data and total weight baseline data at all times of the entire journey are stored, and the load regression identification model is trained, fine-tuned, and redeployed offline.

[0152] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0153] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0154] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0155] 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 foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A self-learning method for identifying and assessing the load capacity of commercial vehicles, characterized in that, Includes the following steps: (1) Vehicle operation initialization: including the initialization of the on-board GPS module, vehicle management system, vehicle communication module and vehicle static parameters; (2) Multi-source data acquisition: Data acquired through the GPS module includes timestamps, elevation, and vehicle speed; data acquired through the CAN bus includes engine speed, engine torque, and gear information. (3) Operation data detection and cleaning: Implement corresponding cleaning strategies for missing, noisy or other erroneous frames in the operation data, and then take preprocessing measures such as interpolation and filtering for operation data that do not meet the continuity conditions; (4) Vehicle operation characteristics extension: Based on the vehicle system transmission characteristics and road structure characteristics, the preprocessed vehicle operation data is used to obtain the road slope characteristics, vehicle acceleration and jerk characteristics of the entire road segment at all times; (5) Load joint identification system design: Establish a load benchmark generation model based on recursive least squares method, and obtain the distribution benchmark of estimated load through vehicle operation data iteration; establish a load regression identification model based on improved Informer encoder, and output load identification results through GPS source data and extended operation features; (6) Confidence assessment and model fine-tuning: Based on the dynamic load distribution benchmark, the confidence assessment of the load identification results is completed by combining the output of the neural network load regression model; Store vehicle operation data and total weight baseline data at each moment of the entire journey. After the current journey ends, perform offline data training and fine-tuning of the load regression identification model, and redeploy the model after training is completed.

2. The self-learning commercial vehicle load identification and confidence assessment method according to claim 1, characterized in that, The implementation process of step (3) is as follows: Considering the differences in the operating characteristics of non-same-source data, the error frames in GPS source data and CAN bus source data are manually extracted and labeled. Based on the continuity criterion, the error frame data samples are separated, and the error frame features and cleaning strategies for multi-source data are formulated as follows. GPS source data: Error frames mainly originate from data interruptions. Based on the data feedback timestamps, they can be categorized into short-term data loss and long-term signal interruptions. A time threshold should be set accordingly. ； For short-term data missing values ​​less than this threshold, linear interpolation is performed using the non-error data value that is closest to the frame before and after the error. For long-term signal interruptions exceeding this threshold, the data during that period is set to zero and discarded. CAN bus source data: Error frame data is mainly characterized by short-term data loss, high-frequency measurement noise, and discontinuous abrupt changes; Short-term data missing mainly targets vehicle gear data detection. Based on the principle of continuous change in vehicle gears, short-term single-point missing data is filled with non-abnormal data from the previous moment. High-frequency measurement noise is mainly aimed at engine speed and engine torque data detection. A Savitzky-Golay filter is designed to process the noise data. Discontinuous abrupt changes are primarily targeted at engine torque data detection, and the characteristic detection standard is set as follows: Where R represents the engine torque data at a certain moment, and T represents the timestamp at the corresponding moment. To calculate the time step, z is the time step. The detection threshold is set; for the data points obtained according to the above detection standards, a linear interpolation method is used to cover the data values ​​at the corresponding time.

3. The self-learning commercial vehicle load identification and confidence assessment method according to claim 1, characterized in that, The implementation process of step (4) is as follows: After the vehicle operation data is cleaned, based on the vehicle parameter attributes and operating characteristics, the preprocessed vehicle operation data is used to obtain vehicle operating speed characteristics: This allows us to obtain the vehicle's acceleration and jerk characteristics: Where n is the engine speed and r is the wheel rolling radius. This refers to the gear ratio of the transmission. Rear axle speed ratio For time step, This is the vehicle speed value. This is the acceleration value. The degree of urgency; Based on the construction standards and changing characteristics of highways, the road slope is obtained by preprocessing vehicle speed and elevation data. In the formula, This is the time step for backward calculation; Let be the elevation value at time t. for Elevation value at time, in units of ； This is the vehicle speed value, in units of ； Slope calculations are performed at points with a distance span of 100m, ensuring that elevation data changes monotonically within a single calculation step, and the time step k satisfies: Based on the above calculation strategy and formula, the data of all running segments are traversed to obtain the road gradient characteristics, vehicle acceleration and jerk characteristics of the entire journey.

4. The self-learning commercial vehicle load identification and confidence assessment method according to claim 1, characterized in that, The load-bearing joint identification system designed in step (5) is as follows: Based on the vehicle dynamics model, a reasonable distribution range of the estimated load is obtained, and the longitudinal dynamics model of the vehicle under driving conditions is adopted: Where g is the acceleration due to gravity, M is the mass of the car, and f is the coefficient of friction. For road slope, C D The air drag coefficient, Where ρ is the vehicle's frontal area, and ρ is the air density. This is the conversion factor for the vehicle's moment of inertia. For vehicle speed, To accelerate the vehicle, This refers to engine torque. This refers to the gear ratio of the transmission. Rear axle speed ratio The total efficiency of the transmission system is given by r, where r is the rolling radius of the wheel. A recursive least squares algorithm with a forgetting factor is used to obtain the distribution benchmark of the estimated load based on the above vehicle dynamics model filtering, and the loss function of the filtering algorithm is constructed as follows: in, Indicates the forgetting factor, Indicates the actual output of the system. Indicates system input, express The vector of parameters to be estimated at time t; The load reference generation model based on recursive least squares is designed as follows: A load regression identification model was built using an Informer-based neural network architecture. The model input parameters were selected as: vehicle speed, acceleration, jerk, road elevation, and road slope. The model output parameters were the load regression identification results.

5. The self-learning commercial vehicle load identification and confidence assessment method according to claim 1, characterized in that, The implementation process of step (6) is as follows: Construct using the load distribution output by the recursive least squares algorithm The distribution interval is used, and the interval average is used as the confidence evaluation benchmark. The confidence of the load regression identification model output of the Informer encoder is evaluated, and MAE is used as the evaluation index. in, The output of the load regression identification model. This represents the average value within the load reference range; The confidence level assessment criteria are as follows: The load regression identification results and confidence assessment results iterate over time and are continuously output as the vehicle runs until the end of the vehicle's working journey; the load regression identification results and confidence assessment results at each time dimension are recorded in the vehicle's storage unit; After the vehicle's current working journey ends, the confidence level of the load identification results throughout the entire journey is statistically analyzed. Load identification data with a confidence level below 10% are marked. The average load identification benchmark at the corresponding time point is used as the fine-tuning label to generate the fine-tuning dataset of the Informer encoder for offline training. After the model is fine-tuned, it is redeployed.

6. A self-learning commercial vehicle load identification and confidence assessment system based on the method of any one of claims 1-5, characterized in that, The system includes: Initialization module: Used for initializing the vehicle GPS module, vehicle management system, vehicle communication module, and vehicle static parameters; Multi-source data acquisition module: used to acquire data including timestamp, elevation, and vehicle speed through GPS module, and data including engine speed, engine torque, and gear information through CAN bus; Data cleaning and feature expansion module: It is used to implement corresponding cleaning strategies for missing, noisy and other erroneous frame data in the running data, and to take preprocessing measures such as interpolation and filtering for running data that does not meet the continuity condition. Based on the transmission characteristics of the vehicle system and the characteristics of the road structure, it uses the preprocessed vehicle running data to obtain the road slope characteristics, vehicle acceleration and jerk characteristics of the entire road segment at all times. The load joint identification and model fine-tuning module is used to obtain the distribution benchmark of the estimated load based on the load benchmark generation model based on the recursive least squares method, output the load identification result based on the load regression identification model based on the improved Informer encoder, evaluate the confidence of the load identification result based on the dynamic load distribution benchmark and the output of the neural network load regression model, store the vehicle operation data and total weight benchmark data at all times of the whole journey, and perform offline data training, fine-tuning and redeployment of the load regression identification model.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the self-learning commercial vehicle load identification and confidence assessment method as described in any one of claims 1 to 5.

Images