Two-stage vehicle load estimation method, system and equipment based on residual fusion

Abstract

The invention relates to the technical field of vehicle control, in particular to a two-stage vehicle load estimation method, system and equipment based on residual fusion, and the method comprises the steps: obtaining the inherent information of a vehicle and the real-time measurement data of a sensor in the vehicle operation process, forming an original data set, and carrying out the preprocessing of the original data set, and obtaining a preprocessed data set; performing long and short time domain feature extraction on the preprocessed data set to obtain a multi-time domain feature data set; according to the multi-time-domain feature data set and the vehicle kinetic parameters, performing first-stage estimation by using a multi-time-domain forgetting factor dynamic estimation model to obtain load residual error distribution; and based on the load residual distribution, performing second-stage optimization by using a multi-time-domain adaptive evolution estimation method to obtain an optimal estimated value of the vehicle load. According to the method, technologies such as the multi-time-domain feature data set and the vehicle dynamic parameters are combined, vehicle load optimal estimation based on load residual error distribution is achieved, and the accuracy and robustness of load estimation are improved.

Description

Technical Field

[0001] The present invention relates to the field of vehicle control technology, and in particular to a two-stage vehicle load estimation method, system and device based on residual fusion. Background Art

[0002] In the field of vehicle design and research and development, vehicle load is a core parameter for evaluating the durability of the vehicle structure. Its accurate acquisition plays a vital role in optimizing user-related design, clarifying load targets, setting design strength benchmarks, and formulating road and bench test specifications. However, vehicle load is essentially a static state attribute of the vehicle and does not directly reflect the dynamic attributes of the vehicle's operating conditions. Therefore, it is very challenging to capture its precise value in real time through conventional measurement methods.

[0003] At present, the industry mainly uses three methods to estimate vehicle load. The first method is to directly measure the load by adding special sensors to the vehicle. However, this method is not only complicated to operate and increases the installation and maintenance costs, but also has a certain impact on the overall performance and structure of the vehicle. The second method is to directly calculate the vehicle load based on the dynamic formula. This method has obvious limitations in practical applications. Due to the friction of transmission components, heat loss and other loss factors that are difficult to accurately quantify during vehicle operation, the calculation method based on the dynamic formula is often insufficient in accuracy in practical applications and difficult to meet engineering requirements. Although subsequent researchers have tried to optimize the dynamic formula through algorithms such as Kalman filtering, To improve the estimation accuracy, but this type of method still cannot completely overcome the problem that the loss is difficult to quantify, resulting in the performance in practical applications often being limited by the difference between the simulation environment and the real vehicle data, and the effect is not ideal; the third method is to use a learning method for load estimation, especially for the highly electronic and integrated characteristics of new energy vehicles, extracting rich operating information from the vehicle CAN bus as input, and building a neural network model to learn the inherent relationship between the data to achieve load estimation. However, most of the existing learning methods rely on deep neural networks, which have high requirements on the amount of training data, and have a long model training cycle and huge consumption of computing resources, which makes it difficult to meet the needs of real-time deployment and computing on the vehicle side.

[0004] In summary, the existing technology has many shortcomings in vehicle load estimation. Therefore, developing a method that can accurately estimate the vehicle's own load using the vehicle's existing sensor data has important research significance and broad application prospects. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a two-stage vehicle load estimation method, system and device based on residual fusion.

[0006] In a first aspect, the present invention provides a two-stage vehicle load estimation method based on residual fusion, the method comprising the following steps:

[0007] Acquire vehicle inherent information and real-time sensor measurement data during vehicle operation to form an original data set, and preprocess the original data set to obtain a preprocessed data set;

[0008] Extracting long and short time domain features from the preprocessed data set to obtain a multi-time domain feature data set;

[0009] According to the multi-time domain feature data set and vehicle dynamics parameters, a multi-time domain forgetting factor dynamic estimation model is used to perform a first-stage estimation to obtain a load residual distribution;

[0010] Based on the load residual distribution, a multi-time domain adaptive evolutionary estimation method is used to perform the second stage optimization to obtain the optimal estimated value of the vehicle load.

[0011] In a further embodiment, the vehicle inherent information includes the frontal area, air resistance coefficient and wheel radius of the vehicle; the real-time sensor measurement data includes wheel side torque, lateral and longitudinal speeds, and lateral and longitudinal accelerations; the step of acquiring the vehicle inherent information and the real-time sensor measurement data during vehicle operation to form the original data set includes:

[0012] The inherent information of the vehicle and the real-time measurement data of the sensors during the operation of the vehicle are obtained, and the existing load data is matched with the real-time measurement data of the sensors according to the timestamp, and a true value label is added to each real-time measurement data of the sensors to form an original data set with a true value label.

[0013] In a further embodiment, the step of extracting long and short time domain features from the preprocessed data set to obtain a multi-time domain feature data set comprises:

[0014] Based on the preprocessed data set, a dynamic analysis of the vehicle is performed using Newton's second law to calculate the power, rolling resistance, air resistance and transmission friction of the vehicle during operation;

[0015] The mass loss calculation value is calculated based on the power, rolling resistance, air resistance and transmission friction of the vehicle during operation;

[0016] adding the quality loss calculated value to the preprocessed data set to obtain an expanded data set;

[0017] Performing a short-term dynamic behavior analysis on the expanded data set at a first time scale to extract a short-time domain feature vector;

[0018] Performing a vehicle long-term behavior trend analysis on the expanded data set at a second time scale to extract a long-time domain feature vector; wherein the second time scale is greater than the first time scale;

[0019] The short-time domain feature vector and the long-time domain feature vector are integrated to form a multi-time domain feature data set.

[0020] In a further embodiment, the short-time domain feature vector includes mean feature data, variance feature data, maximum feature data, and minimum feature data;

[0021] The long-term feature vector includes moving average feature data, exponential smoothing mean feature data and historical operating condition statistical feature data.

[0022] In a further embodiment, the step of performing a first-stage estimation using a multi-time domain forgetting factor dynamic estimation model based on the multi-time domain feature data set and vehicle dynamics parameters to obtain a load residual distribution comprises:

[0023] The least squares estimated load of the vehicle is used as the parameter to be estimated, and a dynamic least squares linear estimation model with a forgetting factor is constructed according to the vehicle dynamics parameters and the real-time measurement data of the sensor, and a multi-time domain forgetting factor dynamic estimation model is obtained;

[0024] Recursively estimating the multi-time domain feature data set using the multi-time domain forgetting factor dynamic estimation model to obtain a vehicle load linear analysis estimation value;

[0025] The load estimation error is calculated based on the difference between the vehicle load linear analysis estimation value and the vehicle load actual value, and the load estimation error is statistically analyzed to obtain the load residual distribution.

[0026] In a further embodiment, the step of performing a first-stage estimation using a multi-time domain forgetting factor dynamic estimation model based on the multi-time domain feature data set and the vehicle dynamics parameters to obtain a load residual distribution further includes:

[0027] In the process of recursively estimating the multi-time domain feature data set using the multi-time domain forgetting factor dynamic estimation model, a residual mean square error is calculated based on the load estimation error;

[0028] According to the residual mean square error, a residual cost function with a forgetting factor is defined using a gradient descent method;

[0029] According to the forgetting factor gradient information of the residual cost function, an adaptive mechanism is introduced to dynamically update the forgetting factor.

[0030] In a further embodiment, the step of performing the second-stage optimization based on the load residual distribution using a multi-time domain adaptive evolutionary estimation method to obtain the optimal estimated value of the vehicle load includes:

[0031] Dividing the multi-time domain feature data set into a training set to obtain a multi-time domain feature sample set to be trained;

[0032] According to the multi-time domain feature sample set to be trained, the decision tree is encoded using a binary encoding method to form an initial encoding decision tree population;

[0033] Determining a residual fitness factor according to the load residual distribution, and constructing a fitness function using the residual fitness factor;

[0034] Based on the fitness function, a roulette wheel selection method is used to select individuals whose fitness is greater than a preset fitness threshold from the initial encoding decision tree population as parents;

[0035] Randomly select two parent trees from the parent generation to perform a subtree exchange operation to form a new tree structure, and randomly mutate the nodes in the new tree structure. After a preset number of iterations, output the coding decision tree population with the highest fitness as the optimal tree population;

[0036] The multi-time domain feature data set is input into the decision tree in the optimal tree population for prediction, and the optimal estimated value of the vehicle load is obtained by fitness weighted average calculation.

[0037] In a further embodiment, the residual fitness factor includes a residual factor, a multi-time domain prediction robustness factor, a time domain feature partitioning effectiveness factor and an uncertainty factor; the step of determining the residual fitness factor according to the load residual distribution and constructing a fitness function using the residual fitness factor includes:

[0038] Calculating the average prediction error of all decision trees in the coding decision tree population for the multi-time domain feature sample set to be trained to obtain a basic error metric, and weighting the basic error metric using the volatility of the load residual distribution to obtain a residual factor;

[0039] Measuring the volatility of the load prediction estimation value of the multi-time domain feature sample set to be trained for all decision trees in the coding decision tree population at different time periods to obtain a multi-time domain prediction robustness factor;

[0040] Performing a weighted summation on the average prediction errors of all decision trees in the coding decision tree population on the multi-time domain feature sample set to be trained at the first time scale and the second time scale, respectively, to calculate a time domain feature division effectiveness factor;

[0041] Calculating the high-order skewness and kurtosis of the load residual distribution, and performing weighted summation on the high-order skewness and kurtosis of the load residual distribution to obtain an uncertainty factor;

[0042] The residual factor, the multi-time domain prediction robustness factor, the time domain feature division effectiveness factor and the uncertainty factor are weighted and summed to construct a fitness function.

[0043] In a second aspect, the present invention provides a two-stage vehicle load estimation system based on residual fusion, the system comprising:

[0044] A data processing module is used to obtain vehicle inherent information and real-time sensor measurement data during vehicle operation to form an original data set, and preprocess the original data set to obtain a preprocessed data set;

[0045] A feature extraction module, used for extracting long and short time domain features from the preprocessed data set to obtain a multi-time domain feature data set;

[0046] A residual analysis module, used to perform a first-stage estimation using a multi-time domain forgetting factor dynamic estimation model according to the multi-time domain feature data set and vehicle dynamics parameters to obtain a load residual distribution;

[0047] The optimal estimation module is used to perform second-stage optimization based on the load residual distribution using a multi-time domain adaptive evolutionary estimation method to obtain an optimal estimate of the vehicle load.

[0048] In a third aspect, the present invention further provides a computer device, comprising a processor and a memory, wherein the processor is connected to the memory, the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory so that the computer device performs the steps of implementing the above method.

[0049] The present invention provides a two-stage vehicle load estimation method, system and device based on residual fusion. The method forms an original data set by acquiring vehicle inherent information and real-time sensor measurement data during vehicle operation, and preprocesses the original data set to obtain a preprocessed data set; extracts long and short time domain features from the preprocessed data set to obtain a multi-time domain feature data set; performs a first-stage estimation based on the multi-time domain feature data set and vehicle dynamics parameters using a multi-time domain forgetting factor dynamic estimation model to obtain a load residual distribution; and performs a second-stage optimization based on the load residual distribution using a multi-time domain adaptive evolutionary estimation method to obtain an optimal vehicle load estimation value. Compared with the prior art, the method achieves high-precision and high-robustness prediction of vehicle load through a multi-time domain forgetting factor dynamic estimation model and a multi-time domain adaptive evolutionary estimation method, combined with a multi-time domain feature data set and vehicle dynamics parameters, can effectively adapt to load changes under different working conditions, and provide strong support for safe driving and load management of vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic flow chart of a two-stage vehicle load estimation method based on residual fusion provided by an embodiment of the present invention;

[0051] Figure 2 is a block diagram of a two-stage vehicle load estimation process based on residual fusion provided by an embodiment of the present invention;

[0052] Figure 3 is a schematic diagram of vehicle force analysis provided by an embodiment of the present invention;

[0053] Figure 4 is a block diagram of a two-stage vehicle load estimation system based on residual fusion provided by an embodiment of the present invention;

[0054] Figure 5 It is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following specifically illustrates the implementation mode of the present invention in conjunction with the accompanying drawings. The embodiments are provided for illustrative purposes only and cannot be understood as limiting the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0056] refer to Figure 1 , the embodiment of the present invention provides a two-stage vehicle load estimation method based on residual fusion, such as Figure 1 As shown, the method comprises the following steps:

[0057] S1. Acquire vehicle inherent information and real-time sensor measurement data during vehicle operation to form an original data set, and preprocess the original data set to obtain a preprocessed data set.

[0058] In this embodiment, the vehicle inherent information includes the vehicle's frontal area, air resistance coefficient, and wheel radius; the sensor real-time measurement data includes wheel side torque, lateral and longitudinal speeds, and lateral and longitudinal accelerations; the step of acquiring the vehicle inherent information and the sensor real-time measurement data during vehicle operation to form an original data set includes:

[0059] The inherent information of the vehicle and the real-time measurement data of the sensors during the operation of the vehicle are obtained, and the existing load data is matched with the real-time measurement data of the sensors according to the timestamp, and a true value label is added to each real-time measurement data of the sensors to form an original data set with a true value label.

[0060] Specifically, Figure 2 As shown, this embodiment implements a two-stage method for vehicle load estimation based on the residual fusion principle. The method receives vehicle inherent information and real-time sensor measurement data during vehicle operation from a data source, and preprocesses the vehicle inherent information and real-time sensor measurement data during vehicle operation to remove unlabeled data points and outliers to ensure data quality. Then, a multi-time domain forgetting factor dynamic estimation model (Multi-Timescale Forgetting Factor Dynamic Estimation, M-TFFDE model) is constructed using the dynamics principle. The multi-time domain forgetting factor dynamic estimation model is used to perform the first stage of load estimation. The goal of this stage is to obtain the residual distribution of the vehicle under different operating conditions to provide a basis for subsequent analysis; then, in order to more accurately reflect the characteristics of the data and the uncertainty of the prediction, this embodiment sets a multi-dimensional weighted fitness function, which comprehensively considers multiple dimensions such as the residual distribution of the data, multi-time domain characteristics, robustness and uncertainty of the prediction. In the second stage, this embodiment sets a decision tree encoding and self-evolution mutation method based on the idea of ​​the tree model, and uses the multi-dimensional weighted fitness function constructed above as the convergence criterion to construct a multi-time domain adaptive evolution estimation model (Multi-Timescale Adaptive Residual Evolutionary Tree, M-TARET model), this multi-time domain adaptive evolutionary estimation model is used to achieve accurate load estimation in the second stage.

[0061] In the initial stage, the present embodiment extracts vehicle-specific information from the database. The vehicle-specific information may include frontal area, air resistance coefficient, wheel radius, etc. At the same time, real-time sensor measurement data with time information during vehicle operation is collected. The real-time sensor measurement data may include wheel-side torque, lateral and longitudinal speeds, and lateral and longitudinal accelerations. Then, the present embodiment matches the original data with corresponding real load labels according to the time correspondence. It should be noted that for different types of vehicles, if the wheel-side torque data is difficult to obtain, the present embodiment can use engine output power as a substitute.

[0062] S2. Extract long and short time domain features from the preprocessed data set to obtain a multi-time domain feature data set.

[0063] In this embodiment, the step of extracting long and short time domain features from the preprocessed data set to obtain a multi-time domain feature data set includes:

[0064] Based on the preprocessed data set, a dynamic analysis of the vehicle is performed using Newton's second law to calculate the power, rolling resistance, air resistance and transmission friction of the vehicle during operation;

[0065] The mass loss calculation value is calculated based on the power, rolling resistance, air resistance and transmission friction of the vehicle during operation;

[0066] adding the quality loss calculated value to the preprocessed data set to obtain an expanded data set;

[0067] Performing a short-term dynamic behavior analysis on the expanded data set at a first time scale to extract a short-time domain feature vector;

[0068] Performing a vehicle long-term behavior trend analysis on the expanded data set at a second time scale to extract a long-time domain feature vector; wherein the second time scale is greater than the first time scale;

[0069] The short-time domain feature vector and the long-time domain feature vector are integrated to form a multi-time domain feature data set.

[0070] Specifically, this embodiment uses Newton's second law F=ma (where F is force, m is mass, and a is acceleration) to perform dynamic analysis on the vehicle. When the vehicle is running, its propulsion force F is the power. Figure 3 The propulsion force of the vehicle during operation is the power F t , and the resistance includes rolling resistance F f , air resistance F w and transmission friction F p , so Newton's second law can be transformed into:

[0071]

[0072] At the same time, since the loss value (such as rolling resistance coefficient, air resistance coefficient, etc.) cannot be accurately given, the mass m calculated in this embodiment is a mass loss calculation value, but this loss calculation value can provide an indication of the relationship between each parameter and the mass. Therefore, in this embodiment, the mass loss calculation value is added to the preprocessing data set as a new feature to obtain an extended data set, and then a statistical analysis method is used to analyze the short-term dynamic behavior of the vehicle on the extended data set on a first time scale to extract a short-time domain feature vector. The first time scale is within a shorter time window (such as a few seconds or minutes). The short-time domain feature vector includes average feature data, variance feature data, maximum feature data and minimum feature data These features can reflect the dynamic behavior characteristics of the vehicle in a short period of time. At the same time, the long-term behavior trend of the vehicle is analyzed on the expanded data set at the second time scale to extract the long-term feature vector. The second time scale is within a longer time window (such as a few hours, a few days or longer). The long-term feature vector includes moving average feature data, exponential smoothing mean feature data and historical operating condition statistical feature data. These features can reflect the behavior trend and change law of the vehicle in the long term. Finally, this embodiment integrates the extracted short-term feature vectors and long-term feature vectors to form a multi-time domain feature data set containing multiple time scale features. This data set will be used for subsequent model prediction.

[0073] S3. Based on the multi-time domain feature data set and vehicle dynamics parameters, a multi-time domain forgetting factor dynamic estimation model is used to perform a first-stage estimation to obtain a load residual distribution.

[0074] In this embodiment, the step of performing a first-stage estimation using a multi-time domain forgetting factor dynamic estimation model based on the multi-time domain feature data set and the vehicle dynamics parameters to obtain a load residual distribution includes:

[0075] The least squares estimated load of the vehicle is used as the parameter to be estimated, and a dynamic least squares linear estimation model with a forgetting factor is constructed according to the vehicle dynamics parameters and the real-time measurement data of the sensor, and a multi-time domain forgetting factor dynamic estimation model is obtained;

[0076] Recursively estimating the multi-time domain feature data set using the multi-time domain forgetting factor dynamic estimation model to obtain a vehicle load linear analysis estimation value;

[0077] The load estimation error is calculated based on the difference between the vehicle load linear analysis estimation value and the vehicle load actual value, and the load estimation error is statistically analyzed to obtain the load residual distribution.

[0078] Specifically, this embodiment uses the least squares estimated load of the vehicle as the parameter to be estimated θ(k), which is the estimated value of the vehicle load and needs to be dynamically estimated and updated through the model. Then, this embodiment constructs a coefficient matrix based on the vehicle dynamics parameters and the real-time measurement data of the sensor. Coefficient Matrix It is mainly related to parameters such as gravitational acceleration g and rolling resistance coefficient f, reflecting the relationship between vehicle dynamics and load, and obtaining measurement information y(k) at time k. The measurement information y(k) is mainly related to the real-time measurement data of sensors such as the wheel torque and transmission loss torque and speed v of the vehicle, providing real-time observation data for the model. This embodiment starts from the basic definition and constructs a dynamic least squares linear estimation model with a forgetting factor λ, λ∈(0,1). In this embodiment, the least squares method is a convenient and reliable method for estimating unknown data in linear and nonlinear systems. Assuming that θ is the unknown parameter to be estimated and y is the real-time measurement information, the least squares method expression is as follows:

[0079]

[0080] Where y(k) is the measurement information at time k; is the coefficient matrix; θ(k) is the parameter to be estimated; v(k) is the measurement error vector.

[0081] This embodiment defines a residual cost function, and the value of θ that minimizes the residual cost function is the estimated value. The mathematical expression of the residual cost function is:

[0082]

[0083] make but:

[0084]

[0085] Considering that the influence of the residual information in the past on the current stage gradually decreases during the estimation process, the least squares expression after introducing the forgetting factor is:

[0086]

[0087] Where K(k) is the recursive gain matrix; P(k) is the covariance matrix.

[0088] This embodiment enables the model to dynamically estimate the load parameters by recursive updating. At the same time, this embodiment can set the initial value of the parameter to be estimated θ(k) to be a zero vector, that is, θ(0)=0. The initial value P0 of the covariance matrix is ​​a sufficiently large positive number δ multiplied by the unit matrix I. Therefore, the expression of the recursive gain matrix is:

[0089]

[0090] In this embodiment, the multi-time domain feature data set x t Input into the multi-time domain forgetting factor dynamic estimation model, and use the constructed multi-time domain forgetting factor dynamic estimation model for recursive estimation. In each step of recursion, the multi-time domain forgetting factor dynamic estimation model calculates the coefficient matrix of the current moment. The current measurement information y(k), the estimated parameters θ(k-1) at the previous moment, and the covariance matrix P at the previous moment k-1 , the linear analysis estimated value θ(k) of the vehicle load at the current moment is calculated by the recursive formula. At the same time, this embodiment introduces an adaptive mechanism to define the residual cost function R(λ) according to the residual mean square error and gradient descent method. k ), using the gradient descent method to dynamically update and adjust the forgetting factor λ t , in order to adapt to the time-varying characteristics of the system, so that the model can better capture the system characteristics under different working conditions and obtain a more stable residual distribution. Specifically, in this embodiment, in the process of recursively estimating the multi-time domain feature data set using the multi-time domain forgetting factor dynamic estimation model, the residual mean square error is calculated according to the load estimation error; based on the residual mean square error, a residual cost function with a forgetting factor is defined using a gradient descent method; based on the forgetting factor gradient information of the residual cost function, an adaptive mechanism is introduced to dynamically update the forgetting factor. The specific formula is as follows:

[0091]

[0092] Where α is the learning rate and R(λ) is the residual mean square error.

[0093] This embodiment calculates the load estimation error, i.e., the residual term, based on the difference between the linear analysis estimated value θ(k) of the vehicle load and the actual value of the vehicle load. The residual term can reflect the strength of the linear connection between the feature and the value to be solved. The load estimation error is statistically analyzed to obtain the distribution of the load residuals. For example, the distribution of the load residuals can be analyzed by calculating statistical quantities such as the mean, variance, maximum and minimum values ​​of the error. The distribution of the residuals will serve as the basis for constructing the fitness factor, which is used to evaluate the estimation accuracy and stability of the model, and provide an important basis for subsequent model optimization and adaptive adjustment.

[0094] S4. Based on the load residual distribution, a multi-time domain adaptive evolutionary estimation method is used to perform a second-stage optimization to obtain an optimal estimate of the vehicle load.

[0095] In this embodiment, the step of performing the second-stage optimization based on the load residual distribution by using the multi-time domain adaptive evolutionary estimation method to obtain the optimal estimated value of the vehicle load includes:

[0096] Dividing the multi-time domain feature data set into a training set to obtain a multi-time domain feature sample set to be trained;

[0097] According to the multi-time domain feature sample set to be trained, the decision tree is encoded using a binary encoding method to form an initial encoding decision tree population;

[0098] Determining a residual fitness factor according to the load residual distribution, and constructing a fitness function using the residual fitness factor;

[0099] Based on the fitness function, a roulette wheel selection method is used to select individuals whose fitness is greater than a preset fitness threshold from the initial encoding decision tree population as parents;

[0100] Randomly select two parent trees from the parent generation to perform a subtree exchange operation to form a new tree structure, and randomly mutate the nodes in the new tree structure. After a preset number of iterations, output the coding decision tree population with the highest fitness as the optimal tree population;

[0101] The multi-time domain feature data set is input into the decision tree in the optimal tree population for prediction, and the optimal estimated value of the vehicle load is obtained by fitness weighted average calculation.

[0102] The residual fitness factor includes a residual factor, a multi-time domain prediction robustness factor, a time domain feature division effectiveness factor and an uncertainty factor; the step of determining the residual fitness factor according to the load residual distribution and constructing a fitness function using the residual fitness factor includes:

[0103] Calculating the average prediction error of all decision trees in the coding decision tree population for the multi-time domain feature sample set to be trained to obtain a basic error metric, and weighting the basic error metric using the volatility of the load residual distribution to obtain a residual factor;

[0104] Measuring the volatility of the load prediction estimation value of the multi-time domain feature sample set to be trained for all decision trees in the coding decision tree population at different time periods to obtain a multi-time domain prediction robustness factor;

[0105] Performing a weighted summation on the average prediction errors of all decision trees in the coding decision tree population on the multi-time domain feature sample set to be trained at the first time scale and the second time scale, respectively, to calculate a time domain feature division effectiveness factor;

[0106] Calculating the high-order skewness and kurtosis of the load residual distribution, and performing weighted summation on the high-order skewness and kurtosis of the load residual distribution to obtain an uncertainty factor;

[0107] The residual factor, the multi-time domain prediction robustness factor, the time domain feature division effectiveness factor and the uncertainty factor are weighted and summed to construct a fitness function.

[0108] Specifically, this embodiment randomly divides the multi-time domain feature data set according to a certain ratio (such as 70% as a training set and 30% as a test set) to obtain a multi-time domain feature sample set to be trained. These feature data sets include acceleration, speed, timestamp, etc. According to the multi-time domain feature sample set to be trained, the internal nodes (division features and thresholds) and leaf nodes (prediction values) of each decision tree are represented by binary codes. For example, for an internal node, this embodiment can use a binary string to represent the selection of its division feature (such as a short-time domain feature or a long-time domain feature), and another binary string to represent the value range of its division threshold; for a leaf node, a binary string can be used to represent the range of its prediction value. In this way, Formula, converting the structure and parameter information of the decision tree into binary coding form to form an initial coding decision tree population, and then this embodiment determines the residual fitness factor according to the load residual distribution, and these residual fitness factors include residual correlation factors, which are used to consider the residual distribution characteristics, multi-time domain prediction robustness and uncertainty, so as to construct a fitness function using the residual fitness factor. The fitness function comprehensively considers the basic error metric, weighted error term, multi-time domain volatility penalty, short-long time domain division effectiveness and residual distribution uncertainty metric. By selecting the corresponding weight coefficient, the fitness function can be flexibly adjusted to meet actual needs. In this embodiment, assuming that a tree is T, the fitness function is defined as:

[0109] F(T)=w1f res (T)+w2f stab (T)+w3f time (T)+w4f uncert (T)

[0110] Among them, w1, w2, w3, w4 are all weight coefficients. In this embodiment, the decision tree T can be set to the sample set {(x i ,y i )} for prediction, where x i is the eigenvector, y i is the corresponding target value (load), and the mathematical expression of the residual factor is:

[0111]

[0112] In the formula, f res (T) is the residual factor; For decision tree T to sample x i The predicted value set of g(ε t ) is the weight function dynamically adjusted according to the residual distribution; σε is the standard deviation of the distribution of loading residuals.

[0113] At the same time, this embodiment penalizes the volatility of the model in multiple time domains, mainly by measuring the prediction variance of different time periods, reducing excessive fluctuations in predictions within a time period, and encouraging stable predictions. The mathematical expression of the multi-time domain prediction robustness factor is:

[0114]

[0115] In the formula, f stab (T) is the multi-time domain prediction robustness factor; is the set of predicted values ​​of tree T in time period τ; is the standard deviation of the set of prediction values ​​of tree T in time period τ.

[0116] This embodiment considers the effectiveness of the decision tree when using short-term or long-term features for partitioning. If the residual of the tree is significantly reduced after the short-term feature dimension partitioning or the trend error is smaller after the long-term feature partitioning, the corresponding time domain score is increased. The mathematical expression of the time domain feature partition effectiveness factor is:

[0117] f time (T) = -(α S Err S (T)+α L Err L (T))

[0118]

[0119] In the formula, f time (T) is the time domain feature partition effectiveness factor; α S is the short-time domain balance coefficient; α L is the long-term balance coefficient; Err S (T) is the error caused by prediction based on short-term features; Err L (T) is the error caused by prediction based on long-term features; N is the number of samples.

[0120] This embodiment measures uncertainty based on the two statistics of high-order skewness and kurtosis of the residual distribution, which are used to suppress the influence of long tail or abnormal working conditions on the model. The mathematical expression of the uncertainty factor is:

[0121] f uncert (T) = -(γ1|Skew(ε t )|+γ2|Kurt(ε t )|)

[0122] In the formula, f uncert(T) is the uncertainty factor; γ1 and γ2 are the weight factors of high-order skewness and kurtosis respectively; Skew(ε t ) is the higher-order skewness based on the residual distribution; Kurt(ε t ) is the kurtosis based on the residual distribution.

[0123] This embodiment combines the above-mentioned residual fitness factors to obtain a fitness function, and flexibly adjusts it by selecting appropriate weights according to actual needs. Based on the fitness function, this embodiment adopts a roulette wheel selection method to select individuals with fitness greater than a preset fitness threshold from the initial coding decision tree population as parents. Specifically, this embodiment calculates the probability of each individual being selected as a parent according to its fitness value. The higher the fitness, the greater the probability of the individual being selected, ensuring that excellent individuals have a higher chance of being selected. The selection probability method is:

[0124]

[0125] In the formula, P(T i ) is the individual T i The probability of being selected as a parent; F(T i ) is the individual T i The fitness value of ; I is the number of individuals.

[0126] Then, according to these probabilities, random selection is performed to obtain a certain number of parent individuals, and two parent trees are randomly selected from the parent generation to perform subtree exchange operations. Specifically, two nodes in the two parent trees are randomly selected, and then the two nodes and their subtrees are exchanged to form a new tree structure. This step is to introduce a new tree structure through a crossover operation, increase the diversity of the population, and help explore a better solution space. Subsequently, the nodes in the new tree structure are randomly mutated. The random mutation includes feature mutation (changing the division feature or threshold) and subtree mutation (randomly adding or deleting subtrees). These mutation operations help explore new solution spaces and avoid falling into local optimality. After a preset number of iterations, the evolution process is stopped according to the preset iteration termination condition. Finally, the encoded decision tree population with the highest fitness is output as the optimal tree population. These decision trees have higher fitness values ​​and can better adapt to multi-time domain information and residual distribution characteristics. The multi-time domain feature data set to be predicted is input into each decision tree in the optimal tree population for prediction. Each decision tree predicts the input feature vector to obtain the corresponding prediction result. Then, the prediction results are integrated and fused by weighted averaging based on the tree's fitness, that is, the weight of each decision tree in the final prediction result is calculated according to its fitness value, and then the prediction value of each decision tree is weighted averaged according to the weight to obtain the optimal estimate of the vehicle load. The weight allocation method is based on the fitness value of each tree, that is, the weight is proportional to the fitness. The decision tree with higher fitness has a greater weight in the final prediction result, ensuring that the prediction result of the excellent tree has a greater impact on the final estimate. Finally, the fused prediction value is the optimal estimate of the vehicle load output by the evolutionary tree method based on the residual fitness factor and multi-time domain information. The mathematical expression for integration and fusion using weighted averaging based on the tree's fitness is:

[0127]

[0128] in,

[0129]

[0130] In the formula, is the fused prediction value; w i is the prediction value weight of each decision tree; For each decision tree The predicted value obtained by predicting x*; P is the number of samples; F(T i G ) is the fitness value of the decision tree.

[0131] It should be noted that, when inputting data, this embodiment uses a sliding window method to input multiple continuous data as a sample, so as to enhance the model's ability to understand feature relationships in the time series dimension; when outputting data, the sliding window method is also used to average several continuous results in time as the load estimation result at that moment, so as to reduce the error of the estimated value. When estimating the actual load, the input parameters of the model mainly come from the data directly provided by the CAN bus, so that the system can achieve the task of real-time estimation of the vehicle load without adding new sensors, relying solely on its own sensor data.

[0132] The embodiment of the present invention provides a two-stage vehicle load estimation method based on residual fusion, wherein the method forms an original data set by acquiring vehicle inherent information and real-time sensor measurement data during vehicle operation, and preprocesses the original data set to obtain a preprocessed data set; extracts long and short time domain features from the preprocessed data set to obtain a multi-time domain feature data set; performs a first-stage estimation based on the multi-time domain feature data set and vehicle dynamics parameters using a multi-time domain forgetting factor dynamic estimation model to obtain a load residual distribution; and performs a second-stage optimization based on the load residual distribution using a multi-time domain adaptive evolutionary estimation method to obtain an optimal vehicle load estimation value. Compared with the prior art, the method achieves high-precision and high-robustness prediction of vehicle load through a multi-time domain forgetting factor dynamic estimation model and a multi-time domain adaptive evolutionary estimation model, combined with a multi-time domain feature data set and vehicle dynamics parameters, and can effectively adapt to load changes under different working conditions, providing strong support for safe driving and load management of vehicles.

[0133] It should be noted that the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0134] In one embodiment, Figure 4 As shown, an embodiment of the present invention provides a two-stage vehicle load estimation system based on residual fusion, the system comprising:

[0135] The data processing module 101 is used to obtain the inherent information of the vehicle and the real-time measurement data of the sensor during the operation of the vehicle to form an original data set, and pre-process the original data set to obtain a pre-processed data set;

[0136] A feature extraction module 102 is used to extract long- and short-time domain features from the preprocessed data set to obtain a multi-time domain feature data set;

[0137] A residual analysis module 103 is used to perform a first-stage estimation using a multi-time domain forgetting factor dynamic estimation model according to the multi-time domain feature data set and vehicle dynamics parameters to obtain a load residual distribution;

[0138] The optimal estimation module 104 is used to perform the second stage optimization based on the load residual distribution by using a multi-time domain adaptive evolutionary estimation method to obtain the optimal estimation value of the vehicle load.

[0139] For the specific limitations of a two-stage vehicle load estimation system based on residual fusion, please refer to the above-mentioned limitations on a two-stage vehicle load estimation method based on residual fusion, which will not be repeated here. A person of ordinary skill in the art will appreciate that the various modules and steps described in conjunction with the embodiments disclosed in this application can be implemented in hardware, software, or a combination of both. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0140] The embodiment of the present invention provides a two-stage vehicle load estimation system based on residual fusion. The system obtains vehicle inherent information and real-time sensor measurement data during vehicle operation through a data processing module to form an original data set, and preprocesses the original data set to obtain a preprocessed data set; the feature extraction module extracts long and short time domain features from the preprocessed data set to obtain a multi-time domain feature data set; the residual analysis module uses a multi-time domain forgetting factor dynamic estimation model to perform a first-stage estimation based on the multi-time domain feature data set and vehicle dynamics parameters to obtain a load residual distribution; the optimal estimation module uses a multi-time domain adaptive evolutionary estimation method to perform a second-stage optimization based on the load residual distribution to obtain the optimal vehicle load estimation value. Compared with the prior art, the system method realizes high-precision and high-robustness prediction of vehicle load through a multi-time domain forgetting factor dynamic estimation model and a multi-time domain adaptive evolutionary estimation model, combined with a multi-time domain feature data set and vehicle dynamics parameters, can effectively adapt to load changes under different working conditions, and provide strong support for safe driving and load management of vehicles.

[0141] Figure 5 A computer device provided by an embodiment of the present invention includes a memory, a processor and a transceiver, which are connected via a bus; the memory is used to store a set of computer program instructions and data, and can transmit the stored data to the processor, and the processor can execute the program instructions stored in the memory to perform the steps of the above method.

[0142] The memory may include a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories; the processor may be a central processing unit, a microprocessor, an application-specific integrated circuit, a programmable logic device, or a combination thereof. By way of example but not limitation, the programmable logic device may be a complex programmable logic device, a field programmable gate array, a general purpose array logic, or any combination thereof.

[0143] Additionally, the memory may be a physically separate unit or may be integrated with the processor.

[0144] It can be understood by those skilled in the art that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have the same component arrangement.

[0145] The above-mentioned embodiments only express several preferred implementation modes of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in the technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention, and these improvements and substitutions should also be regarded as the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be based on the protection scope of the claims.

Claims

1. A two-stage vehicle load estimation method based on residual fusion, characterized in that: The following steps are involved: Acquire vehicle inherent information and real-time sensor measurement data during vehicle operation to form an original data set, and preprocess the original data set to obtain a preprocessed data set; Extracting long and short time domain features from the preprocessed data set to obtain a multi-time domain feature data set; According to the multi-time domain feature data set and vehicle dynamics parameters, a multi-time domain forgetting factor dynamic estimation model is used to perform a first-stage estimation to obtain a load residual distribution; Based on the load residual distribution, a multi-time domain adaptive evolutionary estimation method is used to perform the second stage optimization to obtain the optimal estimated value of the vehicle load.

2. A two-stage vehicle load estimation method based on residual fusion as claimed in claim 1, characterized in that: The inherent information of the vehicle includes the frontal area, air resistance coefficient and wheel radius of the vehicle; the real-time measurement data of the sensor includes wheel torque, lateral and longitudinal speed, and lateral and longitudinal acceleration; The step of acquiring vehicle inherent information and real-time sensor measurement data during vehicle operation to form an original data set includes: The inherent information of the vehicle and the real-time measurement data of the sensors during the operation of the vehicle are obtained, and the existing load data is matched with the real-time measurement data of the sensors according to the timestamp, and a true value label is added to each real-time measurement data of the sensors to form an original data set with a true value label.

3. A two-stage vehicle load estimation method based on residual fusion as claimed in claim 1, characterized in that: The step of extracting long and short time domain features from the preprocessed data set to obtain a multi-time domain feature data set comprises: Based on the preprocessed data set, a dynamic analysis of the vehicle is performed using Newton's second law to calculate the power, rolling resistance, air resistance and transmission friction of the vehicle during operation; The mass loss calculation value is calculated based on the power, rolling resistance, air resistance and transmission friction of the vehicle during operation; Adding the quality loss calculated value to the preprocessed data set to obtain an expanded data set; Performing a short-term dynamic behavior analysis on the expanded data set at a first time scale to extract a short-time domain feature vector; Performing a vehicle long-term behavior trend analysis on the expanded data set at a second time scale to extract a long-time domain feature vector; wherein the second time scale is greater than the first time scale; The short-time domain feature vector and the long-time domain feature vector are integrated to form a multi-time domain feature data set.

4. A two-stage vehicle load estimation method based on residual fusion as claimed in claim 3, characterized in that: The short-time domain feature vector includes average feature data, variance feature data, maximum feature data and minimum feature data; The long-term feature vector includes moving average feature data, exponential smoothing mean feature data and historical operating condition statistical feature data.

5. The two-stage vehicle load estimation method based on residual fusion according to claim 1, characterized in that: The step of performing a first-stage estimation using a multi-time domain forgetting factor dynamic estimation model according to the multi-time domain feature data set and the vehicle dynamics parameters to obtain a load residual distribution comprises: The least squares estimated load of the vehicle is used as the parameter to be estimated, and a dynamic least squares linear estimation model with a forgetting factor is constructed according to the vehicle dynamics parameters and the real-time measurement data of the sensor, and a multi-time domain forgetting factor dynamic estimation model is obtained; Recursively estimating the multi-time domain feature data set using the multi-time domain forgetting factor dynamic estimation model to obtain a vehicle load linear analysis estimation value; The load estimation error is calculated based on the difference between the vehicle load linear analysis estimation value and the vehicle load actual value, and the load estimation error is statistically analyzed to obtain the load residual distribution.

6. A two-stage vehicle load estimation method based on residual fusion as claimed in claim 5, characterized in that: The step of performing a first-stage estimation using a multi-time domain forgetting factor dynamic estimation model according to the multi-time domain feature data set and the vehicle dynamics parameters to obtain a load residual distribution also includes: In the process of recursively estimating the multi-time domain feature data set using the multi-time domain forgetting factor dynamic estimation model, a residual mean square error is calculated based on the load estimation error; According to the residual mean square error, a residual cost function with a forgetting factor is defined using a gradient descent method; According to the forgetting factor gradient information of the residual cost function, an adaptive mechanism is introduced to dynamically update the forgetting factor.

7. A two-stage vehicle load estimation method based on residual fusion as claimed in claim 3, characterized in that: The step of performing the second-stage optimization based on the load residual distribution using the multi-time domain adaptive evolutionary estimation method to obtain the optimal estimated value of the vehicle load includes: Dividing the multi-time domain feature data set into a training set to obtain a multi-time domain feature sample set to be trained; According to the multi-time domain feature sample set to be trained, the decision tree is encoded using a binary encoding method to form an initial encoding decision tree population; Determining a residual fitness factor according to the load residual distribution, and constructing a fitness function using the residual fitness factor; Based on the fitness function, a roulette wheel selection method is used to select individuals whose fitness is greater than a preset fitness threshold from the initial encoding decision tree population as parents; Randomly select two parent trees from the parent generation to perform a subtree exchange operation to form a new tree structure, and randomly mutate the nodes in the new tree structure. After a preset number of iterations, output the coding decision tree population with the highest fitness as the optimal tree population; The multi-time domain feature data set is input into the decision tree in the optimal tree population for prediction, and the optimal estimated value of the vehicle load is obtained by fitness weighted average calculation.

8. A two-stage vehicle load estimation method based on residual fusion as claimed in claim 7, characterized in that: The residual fitness factor includes a residual factor, a multi-time domain prediction robustness factor, a time domain feature division effectiveness factor and an uncertainty factor; the step of determining the residual fitness factor according to the load residual distribution and constructing a fitness function using the residual fitness factor includes: Calculating the average prediction error of all decision trees in the coding decision tree population for the multi-time domain feature sample set to be trained to obtain a basic error metric, and weighting the basic error metric using the volatility of the load residual distribution to obtain a residual factor; Measuring the volatility of the load prediction estimation value of the multi-time domain feature sample set to be trained for all decision trees in the coding decision tree population at different time periods to obtain a multi-time domain prediction robustness factor; Performing a weighted summation on the average prediction errors of all decision trees in the coding decision tree population on the multi-time domain feature sample set to be trained at the first time scale and the second time scale, respectively, to calculate a time domain feature division effectiveness factor; Calculating the high-order skewness and kurtosis of the load residual distribution, and performing weighted summation on the high-order skewness and kurtosis of the load residual distribution to obtain an uncertainty factor; The residual factor, the multi-time domain prediction robustness factor, the time domain feature division effectiveness factor and the uncertainty factor are weighted and summed to construct a fitness function.

9. A two-stage vehicle load estimation system based on residual fusion, characterized in that: The system comprises: A data processing module is used to obtain vehicle inherent information and real-time sensor measurement data during vehicle operation to form an original data set, and preprocess the original data set to obtain a preprocessed data set; A feature extraction module, used for extracting long and short time domain features from the preprocessed data set to obtain a multi-time domain feature data set; A residual analysis module, used to perform a first-stage estimation using a multi-time domain forgetting factor dynamic estimation model according to the multi-time domain feature data set and vehicle dynamics parameters to obtain a load residual distribution; The optimal estimation module is used to perform second-stage optimization based on the load residual distribution using a multi-time domain adaptive evolutionary estimation method to obtain an optimal estimate of the vehicle load.

10. A computer device, characterized in that: The computer device comprises a processor and a memory, wherein the processor is connected to the memory, the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory, so that the computer device executes the method according to any one of claims 1 to 8.

Images