New energy automobile thermal management system based on AI technology and intelligent control method thereof

Abstract

The invention relates to the field of new energy automobile control, in particular to a new energy automobile thermal management system based on the AI technology and an intelligent control method thereof, and the method comprises the steps: building a multi-field coupling feature set based on a multi-source heterogeneous original signal in the operation process of a new energy automobile; a kernel partial least squares algorithm is used as a prediction framework, a local linear mechanism is introduced to carry out dimension reduction processing, and a final thermal management key parameter prediction value is obtained through an embedded fuzzy reasoning strategy; and in combination with one-by-one simulation of the thermal management system simulation model, an optimal control strategy set is obtained. According to the method, the optimal neighbor sample is screened through the Euclidean distance measurement feature sample similarity, the local embedded weight is locked, the low-dimensional space dimension is dynamically adapted in combination with kernel density analysis, and iterative optimization is performed through the gradient descent method, so that invalid redundant information in the multi-field coupling features is eliminated, the calculation complexity of a subsequent prediction model is reduced, and the prediction efficiency is improved. And a local nonlinear correlation structure which plays a key role in thermal management state change is reserved to the greatest extent.

Description

Technical Field

[0001] This invention relates to the field of new energy vehicle control, specifically to a new energy vehicle thermal management system based on AI technology and its intelligent control method. Background Technology

[0002] The thermal management system of new energy vehicles is crucial for ensuring the stable operation of core components such as power batteries and drive motors. Its operating status directly affects the vehicle's range, lifespan, and safety performance. Currently, new energy vehicles generate heterogeneous signals from multiple sources during operation, including power battery voltage / current, motor temperature, environmental parameters, and vehicle operating status. These signals are characterized by inconsistent timing, complex correlations, and strong coupling. Existing technologies struggle to accurately uncover the deep coupling relationships between these signals, leading to inaccurate selection of key influencing signals and susceptibility to interference from redundant information.

[0003] In the feature processing and parameter prediction stages, traditional methods face multiple bottlenecks: on the one hand, the high dimensionality of multi-field coupled features makes it easy for conventional dimensionality reduction techniques to lose the local nonlinear correlation structure that plays a key role in the changes of thermal management status, increasing the computational burden of subsequent prediction models; on the other hand, single prediction algorithms are not well adapted to the nonlinear changes of thermal management parameters, making it difficult to dynamically track parameter fluctuations, and lacking an effective error correction mechanism, resulting in limited prediction accuracy, weak anti-interference ability, and inability to achieve efficient and optimized operation of the thermal management system, and may even affect the stability of the entire system due to overload of local components. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a new energy vehicle thermal management system and its intelligent control method based on AI technology.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: An intelligent control method for a thermal management system of a new energy vehicle based on AI technology, the method comprising: S101. Based on the multi-source heterogeneous raw signals during the operation of new energy vehicles, including but not limited to the timing signals of voltage / current / internal resistance of power battery cells, timing signals of motor stator / rotor temperature, evaporation / condensation pressure signals of air conditioning system, coolant flow / temperature signals, external ambient temperature / wind speed / solar intensity signals, vehicle operating status parameters such as vehicle speed / acceleration / braking frequency signals, and battery SOC / SOF / SOH status signals, construct a coupling correlation matrix between multi-source signals to obtain the key signal subset that plays a dominant role in the stability of the thermal management system. Based on the time domain, frequency domain, and time-frequency domain, extract multi-dimensional features of the key signal subset to obtain a multi-field coupling feature set. S102. Based on the kernel partial least squares algorithm as the prediction framework, a local linear mechanism is introduced to reduce the dimensionality of the obtained multi-field coupling feature set as input, resulting in a low-dimensional feature set that retains the local nonlinear correlation structure. This low-dimensional feature set is then input into the prediction framework to obtain preliminary predicted values ​​of key thermal management parameters. The preliminary predicted values ​​of key thermal management parameters are then dynamically corrected using an embedded fuzzy inference strategy to obtain the error correction amount. Finally, the predicted values ​​of key thermal management parameters are obtained based on the preliminary predicted values ​​of key thermal management parameters and the error correction amount. S103. Based on the final predicted values ​​of key thermal management parameters, and combined with the simulation of each thermal management system model, the final predicted values ​​of key thermal management parameters are used as the rise values ​​of each actuator of the new energy vehicle. Based on the constraints of the actuators, the optimal control strategy set is obtained. S104. The obtained optimal control strategy set is parsed into specific control instructions for each execution component by the central controller, and the specific control instructions are communicated in real time with the local execution controller through the CAN bus to complete the thermal management control.

[0006] In a preferred embodiment, after obtaining the multi-source heterogeneous original signal, S101 uses the pulse signal of the new energy vehicle's driving mileage as the time reference to extract the timestamp information of all the multi-source heterogeneous original signals, fills in the missing sampling points of the signal by linear interpolation, and then unifies and normalizes all the multi-source heterogeneous original signals to the same sampling frequency, thereby completing the time axis alignment of the input multi-source heterogeneous original signals and ensuring that different signals can be compared and analyzed in the same time dimension. Any two sets of aligned multi-source heterogeneous original signals are divided into data segments of equal length. The concept density distribution of each data segment is calculated using the kernel density estimation method. Then, the mutual information entropy value is solved by integration. The mutual information entropy value of any two sets of aligned multi-source heterogeneous original signals is calculated. Based on the mutual information entropy values ​​of all two sets of aligned multi-source heterogeneous original signals, a correlation strength matrix is ​​constructed. The element values ​​of the correlation strength matrix correspond to the mutual information entropy values ​​of the two sets of aligned multi-source heterogeneous original signals. The larger the mutual information entropy value, the higher the nonlinear correlation strength between the two sets of signals. Granger causality test is performed on any two sets of aligned multi-source heterogeneous original signals whose mutual information entropy values ​​are higher than the mean mutual information entropy in the correlation strength matrix. That is, it is to determine which signal change will dominate the change of the other signal and generate a causal relationship matrix. Element-level product operations are used to fuse the correlation strength matrix and the causal relationship matrix to form a coupling correlation matrix among multi-source signals. The matrix element value of the coupling correlation matrix is ​​the product of the corresponding element value of the correlation strength matrix and the direction coefficient of the corresponding element of the causal relationship matrix. The direction coefficient of the corresponding element of the causal relationship matrix includes 1 and -1, representing two diametrically opposed causal relationships. Based on the coupling correlation matrix, signals that affect the coupling state of the thermal field, electric field and flow field are screened to obtain a subset of key signals.

[0007] In a preferred embodiment, after obtaining the key signal subset, S101 extracts statistical features from the key signal subset in the time domain, including peak value, valley value, mean value, variance, kurtosis, waveform factor, etc., extracts frequency domain features including characteristic frequency, frequency amplitude, and frequency band energy through fast Fourier transform, and extracts time-frequency domain features including wavelet coefficient energy and entropy value through wavelet packet transform, and performs feature dimension correction and normalization processing to obtain a multi-field coupling feature set; Considering that the difference in the dimensions of different physical quantity signals can lead to an imbalance in feature weights, an adaptive dimensional analysis method is used to unify the dimensions of the extracted multi-dimensional features. At the same time, to address the non-stationarity of feature distribution, kernel density estimation is used to adaptively determine the effective value range of the features. Finally, by detecting local outliers and removing outliers from the normalized feature set, a high-quality multi-field coupling feature set is obtained.

[0008] In a preferred embodiment, step S102 measures the similarity between each feature sample in the multi-field coupling feature set and all other feature samples based on Euclidean distance, and selects at least ten candidate samples that are most similar to the feature sample based on the similarity ranking result, i.e., the top ten in the similarity ranking result. Then, the optimal number of nearest neighbor samples, i.e., the optimal number of candidate samples, is determined through cross-validation. The cross-validation includes dividing the multi-field coupling feature set into a training subset and a validation subset in a 7:3 ratio, and performing subsequent dimensionality reduction processes for different numbers of candidate nearest neighbor samples. The goal is to minimize the prediction error of the feature set after dimensionality reduction in the kernel partial least squares model, select the optimal number of nearest neighbor samples, and determine the optimal number of nearest neighbor samples and the corresponding candidate samples for the multi-field coupling feature set. A local linear reconstruction matrix is ​​constructed based on the number of optimal nearest neighbor samples and the total number of feature samples. The dimension of the local linear reconstruction matrix is ​​set to be the product of the number of optimal nearest neighbor samples and the total number of feature samples in the multi-field coupled feature set. Each matrix element in the local linear reconstruction matrix corresponds to the linear reconstruction coefficient of the candidate sample to the feature sample. The specific process for generating linear reconstruction coefficients is as follows: When constructing the local linear reconstruction matrix, the initial values ​​of all elements in the local linear reconstruction matrix are uniformly set to zero to simplify the calculation and focus on local correlations. Only the matrix elements corresponding to the feature samples and candidate samples are set to non-zero values ​​to be solved, while the remaining elements are fixed to zero values ​​to eliminate interference from non-nearest neighbor samples. With the goal of minimizing the reconstruction error of the linear combination of feature samples and candidate samples, an error loss function is constructed. The reconstruction error is the L2 norm of the difference between the feature vector of the feature sample and the weighted sum of the feature vectors of the candidate samples. At the same time, a constraint condition is added that the sum of the weights of the reconstruction coefficients corresponding to all candidate samples is 1. The linear reconstruction coefficients of each candidate sample to the feature sample are finally obtained by solving the least squares method and used as local embedding weights.

[0009] In a preferred embodiment, after obtaining the local embedding weights, the weights of each feature sample and the corresponding candidate sample are matched to form a feature sample-candidate sample weight set, and the weight set is stored and locked so that the weight parameters are not adjusted in the subsequent low-dimensional embedding process. By analyzing the original dimensionality distribution of all feature samples in the multi-field coupling feature set through kernel density analysis, the probability density integral value of the overlapping area of ​​all feature sample distributions is calculated to obtain the degree of redundancy and overlap between feature samples. If the degree of redundancy and overlap is greater than or equal to 60%, one-third of the dimension of the feature sample is taken; if the degree of redundancy and overlap is less than 40%, half of the dimension of the feature sample is taken, thus completing the setting of the low-dimensional space. The algorithm iteratively searches for corresponding low-dimensional feature samples for each feature sample based on gradient descent. Specifically, the initial low-dimensional coordinates are obtained by using the principal component analysis (PCA) results of the feature samples as initial low-dimensional coordinates. PCA is a dimensionality reduction method used to extract core features of data. Here, it is used to quickly determine the starting point of the low-dimensional search. The goal is to minimize the deviation between the low-dimensional reconstruction weights of the feature samples in the low-dimensional space and the fixed local embedding weights. The initial low-dimensional coordinates are gradually adjusted until the deviation is minimized. Finally, all feature samples in the multi-field coupled feature set are matched one by one with the corresponding low-dimensional feature samples. This results in a low-dimensional feature set that preserves the local nonlinear correlation structure, thus completing the introduction of the local linear mechanism and the dimensionality reduction processing of the multi-field coupled feature set.

[0010] Furthermore, traditional kernel partial least squares algorithms or other single prediction algorithms suffer from insufficient nonlinear adaptation, lagging dynamic fluctuation tracking, and weak anti-interference ability in predicting thermal management parameters. Therefore, this application introduces a local linear mechanism to solve the high-dimensional redundancy problem of single prediction algorithms, eliminate redundant information in the multi-field coupling feature set, reduce the computational complexity of the subsequent kernel partial least squares model, and at the same time retain the local feature correlation information that plays a key role in the change of thermal management state to the maximum extent.

[0011] In a preferred embodiment, step S102 inputs the low-dimensional feature set into the kernel partial least squares algorithm, selects the radial basis function (RBF) kernel function as the mapping carrier from low-dimensional features to high-dimensional feature space, aims to minimize the prediction error of the kernel partial least squares algorithm, sets the search range of the RBF kernel function, adjusts the weights and step size through an adaptive weighted particle swarm optimization algorithm, iteratively updates the RBF kernel function value until the prediction error of the kernel partial least squares algorithm is minimized and / or the maximum number of iterations is reached, and finally obtains the kernel parameters of the optimal RBF kernel function. The low-dimensional feature vectors of the low-dimensional feature set are then processed with the optimal RBF kernel function to complete the nonlinear mapping from low-dimensional features to high-dimensional feature space. The core purpose of this mapping is to transform the thermal management feature correlation relationship that is difficult to linearly fit in the low-dimensional space into a linear correlation relationship in the high-dimensional space, laying the foundation for the subsequent linear model construction. All high-dimensional feature vectors obtained after mapping low-dimensional features to high-dimensional features are arranged in rows to construct a high-dimensional feature matrix. Each element in the high-dimensional feature matrix corresponds to a single high-dimensional feature value of a single sample. The true values ​​of the thermal management key parameters corresponding to each high-dimensional feature vector are extracted, which are the parameter values ​​of the multi-source heterogeneous original signals, forming a thermal management key parameter vector. The loss function is to minimize the fitting error between the high-dimensional feature vector and the true values ​​of the thermal management key parameters. The fitting error is quantified as the L2 norm of the difference between the product of the high-dimensional feature matrix and the linear reconstruction matrix and the thermal management key parameter vector. The loss function is solved by the least squares method, and the linear coefficient matrix with the minimum fitting error is obtained as the linear prediction model. The high-dimensional feature matrix is ​​multiplied with the constructed linear coefficient matrix to obtain the preliminary predicted value of the key thermal management parameters corresponding to each feature sample.

[0012] In a preferred embodiment, step S102 aligns the preliminary predicted values ​​of key thermal management parameters, the low-dimensional feature set, and the actual values ​​of key thermal management parameters on the time axis based on a fuzzy inference strategy, and completes the normalization process. By detecting and removing outliers through local anomaly factors, the input dataset is obtained. The fuzzy mean clustering algorithm is used to determine the optimal number of clusters through cross-validation. The preprocessed input dataset is input into the fuzzy mean clustering algorithm to calculate the membership degree of each sample to each cluster center and generate fuzzy rules. Each rule corresponds to a cluster category. The antecedent of the rule is the fuzzy partition of the input variable, and the consequent of the rule is a linear expression of the error correction amount. Finally, a fuzzy rule base covering the entire input scenario is formed. Among them, if the antecedent of the rule has a large error and significant low-dimensional features, the rule will be selected. The preprocessed input dataset is divided into a training subset and a validation subset. The gradient descent method is used to iteratively optimize the membership function parameters used to calculate the membership degree from each sample to each cluster center in the training subset. The optimization stops when the training subset reaches the maximum number of iterations, and the optimal parameter combination is obtained and solidified into the fuzzy rule base. The preliminary predicted values ​​of key thermal management parameters and the low-dimensional features of the low-dimensional feature set are matched with corresponding rules according to the fuzzy rule base. The fuzzy inference results are transformed into membership degrees. The error correction amount is obtained according to the rule corresponding to the membership degree. The preliminary predicted values ​​of key thermal management parameters are adjusted based on the error correction amount. Specifically, the sum of the preliminary predicted values ​​of key thermal management parameters and the error correction amount is used as the corrected predicted values. Finally, the predicted values ​​of key thermal management parameters and the predicted value set of key thermal management parameters are obtained.

[0013] In a preferred embodiment, step S103 assigns decision priorities to the execution components of the thermal management system based on the analytic hierarchy process (AHP). For example, in a fast charging scenario, the priority of the electric heater is greater than that of the air conditioning compressor, and in a high-speed scenario, the priority of the radiator is greater than that of the electric heater. Based on the hardware performance constraints of the execution components and the safety constraints of the key thermal management parameters, and based on the predicted values ​​of the key thermal management parameters, the predicted values ​​of the key thermal management parameters that exceed the hardware performance constraints and safety constraints are eliminated. The simulation model of the thermal management system is used to simulate each execution component one by one. Execution components whose simulation results exceed the predicted values ​​of the key thermal management parameters, hardware performance constraints, and safety constraints are marked for control. According to the assigned decision priorities, the execution components marked for control are shut down and / or reduced in sequence.

[0014] The present invention also provides a new energy vehicle thermal management system based on AI technology, the system comprising: Data acquisition module and processing module: The user acquires multi-source heterogeneous raw signals during the operation of new energy vehicles, constructs a coupling correlation matrix between multi-source signals based on the acquired multi-source heterogeneous raw signals, and extracts features to obtain a multi-field coupling feature set; Feature Dimensionality Reduction and Parameter Prediction Module: This module introduces a local linear mechanism, using Euclidean distance to select the optimal nearest neighbor samples of feature samples, constructs a local linear reconstruction matrix, and solves for the linear reconstruction coefficients as local embedding weights. Iterative optimization based on gradient descent yields a low-dimensional feature set that preserves the local nonlinear correlation structure. This low-dimensional feature set is then input into a kernel partial least squares prediction framework. A radial basis function kernel is selected as the mapping carrier, and an adaptive weighted particle swarm optimization algorithm is used to determine the optimal kernel parameters, completing the nonlinear mapping from low-dimensional features to a high-dimensional feature space. A linear prediction model is constructed, and preliminary predicted values ​​of key thermal management parameters are output. Finally, based on a fuzzy inference strategy, the final predicted values ​​of key thermal management parameters are output.

[0015] The control strategy optimization module assigns scenario-based decision priorities to each actuator in the thermal management system based on the analytic hierarchy process (AHP). It uses the predicted values ​​of the final key thermal management parameters as target reference values ​​for each actuator and performs individual simulations of each actuator using the thermal management system simulation model. Actuators whose simulation results exceed predicted values, hardware performance constraints, and safety constraints are flagged for control purposes. Based on decision priorities, shutdown or reduction operations are executed sequentially, ultimately forming the optimal control strategy set.

[0016] The beneficial effects of this invention are as follows: By using Euclidean distance to measure the similarity of feature samples, the optimal nearest neighbor samples are selected and the local embedding weights are locked. Combined with kernel density analysis, the low-dimensional space dimension is dynamically adapted. Then, through gradient descent iterative optimization, invalid and redundant information in multi-field coupling features is eliminated, reducing the computational complexity of subsequent prediction models. At the same time, the local nonlinear correlation structure that plays a key role in the change of thermal management status is preserved to the maximum extent. Using the radial basis kernel function as the mapping carrier, the feature correlation that is difficult to linearly fit in the low-dimensional space is transformed into a linear relationship in the high-dimensional space. Combined with the adaptive weighted particle swarm optimization algorithm to dynamically adjust the kernel parameters, the defects of single algorithms in adapting to nonlinear changes in thermal management parameters and lagging in dynamic fluctuation tracking are effectively solved, and the prediction accuracy and anti-interference ability of the model are significantly improved. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0020] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0021] As attached Figure 1 As shown, this embodiment provides: an intelligent control method for a new energy vehicle thermal management system based on AI technology, including the following steps: The methods include: S101. Based on the multi-source heterogeneous raw signals during the operation of new energy vehicles, including but not limited to the timing signals of voltage / current / internal resistance of power battery cells, timing signals of motor stator / rotor temperature, evaporation / condensation pressure signals of air conditioning system, coolant flow / temperature signals, external ambient temperature / wind speed / solar intensity signals, vehicle operating status parameters such as vehicle speed / acceleration / braking frequency signals, and battery SOC / SOF / SOH status signals, construct a coupling correlation matrix between multi-source signals to obtain the key signal subset that plays a dominant role in the stability of the thermal management system. Based on the time domain, frequency domain, and time-frequency domain, extract multi-dimensional features of the key signal subset to obtain a multi-field coupling feature set. After obtaining the multi-source heterogeneous raw signals, S101 uses the pulse signal of the driving range of new energy vehicles as the time reference to extract the timestamp information of all multi-source heterogeneous raw signals, fills in the missing sampling points of the signals through linear interpolation, and then unifies and normalizes all multi-source heterogeneous raw signals to the same sampling frequency, thereby completing the time axis alignment of the input multi-source heterogeneous raw signals and ensuring that different signals can be compared and analyzed in the same time dimension. Any two sets of aligned multi-source heterogeneous original signals are divided into data segments of equal length. The concept density distribution of each data segment is calculated using the kernel density estimation method. Then, the mutual information entropy value is solved by integration. The mutual information entropy value of any two sets of aligned multi-source heterogeneous original signals is calculated. Based on the mutual information entropy values ​​of all two sets of aligned multi-source heterogeneous original signals, a correlation strength matrix is ​​constructed. The element values ​​of the correlation strength matrix correspond to the mutual information entropy values ​​of the two sets of aligned multi-source heterogeneous original signals. The larger the mutual information entropy value, the higher the nonlinear correlation strength between the two sets of signals. Granger causality test is performed on any two sets of aligned multi-source heterogeneous original signals whose mutual information entropy values ​​are higher than the mean mutual information entropy in the correlation strength matrix. That is, it is to determine which signal change will dominate the change of the other signal and generate a causal relationship matrix. Element-level product operations are used to fuse the correlation strength matrix and the causal relationship matrix to form a coupling correlation matrix among multi-source signals. The matrix element value of the coupling correlation matrix is ​​the product of the corresponding element value of the correlation strength matrix and the direction coefficient of the corresponding element of the causal relationship matrix. The direction coefficient of the corresponding element of the causal relationship matrix includes 1 and -1, representing two diametrically opposed causal relationships. Based on the coupling correlation matrix, signals that affect the coupling state of the thermal field, electric field and flow field are screened to obtain a subset of key signals.

[0022] After acquiring the key signal subset, S101 extracts the statistical features of the key signal subset in the time domain, including peak value, valley value, mean, variance, kurtosis, waveform factor, etc. It extracts the frequency domain features including characteristic frequency, frequency amplitude, and frequency band energy through fast Fourier transform, and extracts the time-frequency domain features including wavelet coefficient energy and entropy value through wavelet packet transform. After feature dimension correction and normalization, a multi-field coupling feature set is obtained. Considering that the difference in the dimensions of different physical quantity signals can lead to an imbalance in feature weights, an adaptive dimensional analysis method is used to unify the dimensions of the extracted multi-dimensional features. At the same time, to address the non-stationarity of feature distribution, kernel density estimation is used to adaptively determine the effective value range of the features. Finally, by detecting local outliers and removing outliers from the normalized feature set, a high-quality multi-field coupling feature set is obtained.

[0023] S102. Based on the kernel partial least squares algorithm as the prediction framework, a local linear mechanism is introduced to reduce the dimensionality of the obtained multi-field coupling feature set as input, resulting in a low-dimensional feature set that retains the local nonlinear correlation structure. This low-dimensional feature set is then input into the prediction framework to obtain preliminary predicted values ​​of key thermal management parameters. The preliminary predicted values ​​of key thermal management parameters are then dynamically corrected using an embedded fuzzy inference strategy to obtain the error correction amount. Finally, the predicted values ​​of key thermal management parameters are obtained based on the preliminary predicted values ​​of key thermal management parameters and the error correction amount. S102 uses Euclidean distance to measure the similarity between each feature sample in the multi-field coupling feature set and all other feature samples. Based on the similarity ranking results, at least ten candidate samples that are most similar to the feature sample are selected, i.e., the top ten in the similarity ranking results. The optimal number of nearest neighbor samples, i.e., the optimal number of candidate samples, is determined through cross-validation. The cross-validation includes dividing the multi-field coupling feature set into a training subset and a validation subset in a 7:3 ratio. Subsequent dimensionality reduction processes are performed for different numbers of candidate nearest neighbor samples. The goal is to minimize the prediction error of the feature set after dimensionality reduction in the kernel partial least squares model. The optimal number of nearest neighbor samples is selected, and the optimal number of nearest neighbor samples and the corresponding candidate samples for the multi-field coupling feature set are determined. A local linear reconstruction matrix is ​​constructed based on the number of optimal nearest neighbor samples and the total number of feature samples. The dimension of the local linear reconstruction matrix is ​​set to be the product of the number of optimal nearest neighbor samples and the total number of feature samples in the multi-field coupled feature set. Each matrix element in the local linear reconstruction matrix corresponds to the linear reconstruction coefficient of the candidate sample to the feature sample. The specific process for generating linear reconstruction coefficients is as follows: When constructing the local linear reconstruction matrix, the initial values ​​of all elements in the local linear reconstruction matrix are uniformly set to zero to simplify the calculation and focus on local correlations. Only the matrix elements corresponding to the feature samples and candidate samples are set to non-zero values ​​to be solved, while the remaining elements are fixed to zero values ​​to eliminate interference from non-nearest neighbor samples. With the goal of minimizing the reconstruction error of the linear combination of feature samples and candidate samples, an error loss function is constructed. The reconstruction error is the L2 norm of the difference between the feature vector of the feature sample and the weighted sum of the feature vectors of the candidate samples. At the same time, a constraint condition is added that the sum of the weights of the reconstruction coefficients corresponding to all candidate samples is 1. The linear reconstruction coefficients of each candidate sample to the feature sample are finally obtained by solving the least squares method and used as local embedding weights.

[0024] After obtaining the local embedding weights, the weights of each feature sample and the corresponding candidate sample are matched to form a feature sample-candidate sample weight set. The weight set is stored and locked so that the weight parameters are not adjusted in the subsequent low-dimensional embedding process. By analyzing the original dimensionality distribution of all feature samples in the multi-field coupling feature set through kernel density analysis, the probability density integral value of the overlapping area of ​​all feature sample distributions is calculated to obtain the degree of redundancy and overlap between feature samples. If the degree of redundancy and overlap is greater than or equal to 60%, one-third of the dimension of the feature sample is taken; if the degree of redundancy and overlap is less than 40%, half of the dimension of the feature sample is taken, thus completing the setting of the low-dimensional space. The algorithm iteratively searches for corresponding low-dimensional feature samples for each feature sample based on gradient descent. Specifically, the initial low-dimensional coordinates are obtained by using the principal component analysis (PCA) results of the feature samples as initial low-dimensional coordinates. PCA is a dimensionality reduction method used to extract core features of data. Here, it is used to quickly determine the starting point of the low-dimensional search. The goal is to minimize the deviation between the low-dimensional reconstruction weights of the feature samples in the low-dimensional space and the fixed local embedding weights. The initial low-dimensional coordinates are gradually adjusted until the deviation is minimized. Finally, all feature samples in the multi-field coupled feature set are matched one by one with the corresponding low-dimensional feature samples. This results in a low-dimensional feature set that preserves the local nonlinear correlation structure, thus completing the introduction of the local linear mechanism and the dimensionality reduction processing of the multi-field coupled feature set.

[0025] Furthermore, traditional kernel partial least squares algorithms or other single prediction algorithms suffer from insufficient nonlinear adaptation, lagging dynamic fluctuation tracking, and weak anti-interference ability in predicting thermal management parameters. Therefore, this application introduces a local linear mechanism to solve the high-dimensional redundancy problem of single prediction algorithms, eliminate redundant information in the multi-field coupling feature set, reduce the computational complexity of the subsequent kernel partial least squares model, and at the same time retain the local feature correlation information that plays a key role in the change of thermal management state to the maximum extent.

[0026] S102 inputs the low-dimensional feature set into the kernel partial least squares algorithm, selects the radial basis function (RBF) kernel function as the mapping carrier from low-dimensional features to high-dimensional feature space, aims to minimize the prediction error of the kernel partial least squares algorithm, sets the search range of the RBF kernel function, adjusts the weights and step size through an adaptive weighted particle swarm optimization algorithm, iteratively updates the RBF kernel function value until the prediction error of the kernel partial least squares algorithm is minimized and / or the maximum number of iterations is reached, and finally obtains the kernel parameters of the optimal RBF kernel function. The low-dimensional feature vectors of the low-dimensional feature set are operated with the optimal RBF kernel function to complete the nonlinear mapping from low-dimensional features to high-dimensional feature space. The core purpose of this mapping is to transform the thermal management feature correlation relationship that is difficult to linearly fit in low-dimensional space into a linear correlation relationship in high-dimensional space, laying the foundation for the subsequent linear model construction. All high-dimensional feature vectors obtained after mapping low-dimensional features to high-dimensional features are arranged in rows to construct a high-dimensional feature matrix. Each element in the high-dimensional feature matrix corresponds to a single high-dimensional feature value of a single sample. The true values ​​of the thermal management key parameters corresponding to each high-dimensional feature vector are extracted, which are the parameter values ​​of the multi-source heterogeneous original signals, forming a thermal management key parameter vector. The loss function is to minimize the fitting error between the high-dimensional feature vector and the true values ​​of the thermal management key parameters. The fitting error is quantified as the L2 norm of the difference between the product of the high-dimensional feature matrix and the linear reconstruction matrix and the thermal management key parameter vector. The loss function is solved by the least squares method, and the linear coefficient matrix with the minimum fitting error is obtained as the linear prediction model. The high-dimensional feature matrix is ​​multiplied with the constructed linear coefficient matrix to obtain the preliminary predicted value of the key thermal management parameters corresponding to each feature sample.

[0027] S102 uses a fuzzy inference strategy to align the preliminary predicted values ​​of key thermal management parameters, low-dimensional feature sets, and actual values ​​of key thermal management parameters along the time axis and performs normalization processing. It then detects and removes outliers through local anomaly detection to obtain the input dataset. The fuzzy mean clustering algorithm is used to determine the optimal number of clusters through cross-validation. The preprocessed input dataset is input into the fuzzy mean clustering algorithm to calculate the membership degree of each sample to each cluster center and generate fuzzy rules. Each rule corresponds to a cluster category. The antecedent of the rule is the fuzzy partition of the input variable, and the consequent of the rule is a linear expression of the error correction amount. Finally, a fuzzy rule base covering the entire input scenario is formed. Among them, if the antecedent of the rule has a large error and significant low-dimensional features, the rule will be selected. The preprocessed input dataset is divided into a training subset and a validation subset. The gradient descent method is used to iteratively optimize the membership function parameters used to calculate the membership degree from each sample to each cluster center in the training subset. The optimization stops when the training subset reaches the maximum number of iterations, and the optimal parameter combination is obtained and solidified into the fuzzy rule base. The preliminary predicted values ​​of key thermal management parameters and the low-dimensional features of the low-dimensional feature set are matched with corresponding rules according to the fuzzy rule base. The fuzzy inference results are transformed into membership degrees. The error correction amount is obtained according to the rule corresponding to the membership degree. The preliminary predicted values ​​of key thermal management parameters are adjusted based on the error correction amount. Specifically, the sum of the preliminary predicted values ​​of key thermal management parameters and the error correction amount is used as the corrected predicted values. Finally, the predicted values ​​of key thermal management parameters and the predicted value set of key thermal management parameters are obtained.

[0028] S103. Based on the final predicted values ​​of key thermal management parameters, and combined with the simulation of each thermal management system model, the final predicted values ​​of key thermal management parameters are used as the rise values ​​of each actuator of the new energy vehicle. Based on the constraints of the actuators, the optimal control strategy set is obtained. S103 assigns decision priorities to the execution components of the thermal management system based on the analytic hierarchy process (AHP). For example, in fast charging scenarios, the priority of the electric heater is greater than that of the air conditioning compressor, and in high-speed scenarios, the priority of the radiator is greater than that of the electric heater. Based on the hardware performance constraints of the execution components and the safety constraints of the key thermal management parameters, and based on the predicted values ​​of the key thermal management parameters, the predicted values ​​of the key thermal management parameters that exceed the hardware performance constraints and safety constraints are eliminated. Combined with the thermal management system simulation model, each execution component is simulated one by one. Execution components whose simulation results exceed the predicted values ​​of the key thermal management parameters, hardware performance constraints, and safety constraints are marked for control. According to the assigned decision priorities, the execution components marked for control are shut down and / or reduced in sequence.

[0029] S104. The obtained optimal control strategy set is parsed into specific control instructions for each execution component by the central controller, and the specific control instructions are communicated in real time with the local execution controller through the CAN bus to complete the thermal management control.

[0030] This invention also provides a new energy vehicle thermal management system based on AI technology, the system comprising: Data acquisition module and processing module: The user acquires multi-source heterogeneous raw signals during the operation of new energy vehicles, constructs a coupling correlation matrix between multi-source signals based on the acquired multi-source heterogeneous raw signals, and extracts features to obtain a multi-field coupling feature set; Feature Dimensionality Reduction and Parameter Prediction Module: This module introduces a local linear mechanism, using Euclidean distance to select the optimal nearest neighbor samples of feature samples, constructs a local linear reconstruction matrix, and solves for the linear reconstruction coefficients as local embedding weights. Iterative optimization based on gradient descent yields a low-dimensional feature set that preserves the local nonlinear correlation structure. This low-dimensional feature set is then input into a kernel partial least squares prediction framework. A radial basis function kernel is selected as the mapping carrier, and an adaptive weighted particle swarm optimization algorithm is used to determine the optimal kernel parameters, completing the nonlinear mapping from low-dimensional features to a high-dimensional feature space. A linear prediction model is constructed, and preliminary predicted values ​​for key thermal management parameters are output. Finally, based on a fuzzy inference strategy, the final predicted values ​​for key thermal management parameters are output.

[0031] The control strategy optimization module assigns scenario-based decision priorities to each actuator in the thermal management system based on the analytic hierarchy process (AHP). It uses the predicted values ​​of the final key thermal management parameters as target reference values ​​for each actuator and performs individual simulations of each actuator using the thermal management system simulation model. Actuators whose simulation results exceed predicted values, hardware performance constraints, and safety constraints are flagged for control purposes. Based on decision priorities, shutdown or reduction operations are executed sequentially, ultimately forming the optimal control strategy set.

[0032] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0033] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0034] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0035] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0036] These computer program instructions can also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0037] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0038] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. An intelligent control method for a thermal management system of a new energy vehicle based on AI technology, characterized in that, The method includes: S101. Based on the multi-source heterogeneous raw signals during the operation of new energy vehicles, construct the coupling correlation matrix between the multi-source signals to obtain key signal subsets. Extract multi-dimensional features of the key signal subsets based on the time domain, frequency domain, and time-frequency domain to obtain a multi-field coupling feature set. S102. Based on the kernel partial least squares algorithm as the prediction framework, a local linear mechanism is introduced to reduce the dimensionality of the obtained multi-field coupling feature set as input, resulting in a low-dimensional feature set that retains the local nonlinear correlation structure. This low-dimensional feature set is then input into the prediction framework to obtain preliminary predicted values ​​of key thermal management parameters. The preliminary predicted values ​​of key thermal management parameters are then dynamically corrected using an embedded fuzzy inference strategy to obtain the error correction amount. Finally, the predicted values ​​of key thermal management parameters are obtained based on the preliminary predicted values ​​of key thermal management parameters and the error correction amount. S103. Based on the final predicted values ​​of key thermal management parameters, and combined with the simulation of each thermal management system model, the final predicted values ​​of key thermal management parameters are used as the rise values ​​of each actuator of the new energy vehicle. Based on the constraints of the actuators, the optimal control strategy set is obtained. S104. The obtained optimal control strategy set is parsed into specific control instructions for each execution component by the central controller, and the specific control instructions are communicated in real time with the local execution controller through the CAN bus to complete the thermal management control.

2. The intelligent control method for a new energy vehicle thermal management system based on AI technology according to claim 1, characterized in that, After obtaining the multi-source heterogeneous original signal, S101 uses the pulse signal of the new energy vehicle driving mileage as the time reference to extract the timestamp information of all multi-source heterogeneous original signals, fills in the missing sampling points of the signal by linear interpolation, and then unifies and regularizes all multi-source heterogeneous original signals to the same sampling frequency to complete the time axis alignment of the input multi-source heterogeneous original signal. Any two sets of aligned multi-source heterogeneous original signals are divided into data segments of equal length. The concept density distribution of each data segment is calculated using the kernel density estimation method. Then, the mutual information entropy value is solved by integration. The mutual information entropy value of any two sets of aligned multi-source heterogeneous original signals is calculated. Based on the mutual information entropy values ​​of all two sets of aligned multi-source heterogeneous original signals, a correlation strength matrix is ​​constructed. The element values ​​of the correlation strength matrix correspond to the mutual information entropy values ​​of the two sets of aligned multi-source heterogeneous original signals. Granger causality test is performed on any two pairs of aligned multi-source heterogeneous original signals whose mutual information entropy values ​​are higher than the mean mutual information entropy in the correlation strength matrix to generate a causality matrix. Element-level product operations are used to fuse the correlation strength matrix and the causal relationship matrix to form a coupling correlation matrix among multi-source signals. The matrix element value of the coupling correlation matrix is ​​the product of the direction coefficient of the corresponding element value of the correlation strength matrix and the corresponding element of the causal relationship matrix. Based on the coupling correlation matrix, signals that affect the coupling state of the thermal field, electric field and flow field are screened to obtain a subset of key signals.

3. The intelligent control method for a new energy vehicle thermal management system based on AI technology according to claim 1, characterized in that, After obtaining the key signal subset, S101 extracts the statistical features of the key signal subset in the time domain, extracts the frequency domain features including feature frequency, frequency amplitude, and frequency band energy through fast Fourier transform, and extracts the time-frequency domain features including wavelet coefficient energy and entropy value through wavelet packet transform. Then, it performs feature dimension correction and normalization processing to obtain a multi-field coupling feature set.

4. The intelligent control method for a new energy vehicle thermal management system based on AI technology according to claim 1, characterized in that, S102 measures the similarity between each feature sample in the multi-field coupling feature set and all other feature samples based on Euclidean distance, and selects at least ten candidate samples that are most similar to the feature sample based on the similarity ranking results. The optimal number of nearest neighbor samples, i.e. the optimal number of candidate samples, is determined through cross-validation, thus determining the optimal number of nearest neighbor samples and the corresponding candidate samples for the multi-field coupling feature set. A local linear reconstruction matrix is ​​constructed based on the number of optimal nearest neighbor samples and the total number of feature samples. The dimension of the local linear reconstruction matrix is ​​set to be the product of the number of optimal nearest neighbor samples and the total number of feature samples in the multi-field coupled feature set. Each matrix element in the local linear reconstruction matrix corresponds to the linear reconstruction coefficient of the candidate sample to the feature sample. The specific process for generating linear reconstruction coefficients is as follows: When constructing the local linear reconstruction matrix, the initial values ​​of all elements in the local linear reconstruction matrix are uniformly set to zero. The matrix elements corresponding to the feature samples and candidate samples are set to non-zero values ​​to be solved, while the remaining elements are fixed to zero. With the goal of minimizing the reconstruction error of the linear combination of feature samples and candidate samples, an error loss function is constructed. The reconstruction error is the L2 norm of the difference between the feature vector of the feature sample and the weighted sum of the feature vectors of the candidate samples. At the same time, a constraint condition is added that the sum of the reconstruction coefficient weights corresponding to all candidate samples is 1. The linear reconstruction coefficients of each candidate sample to the feature sample are finally obtained by solving the least squares method and used as local embedding weights.

5. The intelligent control method for a new energy vehicle thermal management system based on AI technology according to claim 4, characterized in that, After obtaining the local embedding weights, the weights of each feature sample and the corresponding candidate sample are matched to form a feature sample-candidate sample weight set, and the weight set is stored and locked. By analyzing the original dimensionality distribution of all feature samples in the multi-field coupling feature set through kernel density analysis, the probability density integral value of the overlapping area of ​​all feature sample distributions is calculated to obtain the degree of redundancy and overlap between feature samples. If the degree of redundancy and overlap is greater than or equal to 60%, one-third of the dimension of the feature sample is taken; if the degree of redundancy and overlap is less than 40%, half of the dimension of the feature sample is taken, thus completing the setting of the low-dimensional space. Based on the gradient descent method, the corresponding low-dimensional feature samples are iteratively found for each feature sample. Specifically, the principal component analysis results of the feature samples are used as the initial low-dimensional coordinates. The goal is to minimize the deviation between the low-dimensional reconstruction weights of the feature samples in the low-dimensional space and the fixed local embedding weights. The initial low-dimensional coordinates are gradually adjusted until the deviation is minimized. Finally, all feature samples in the multi-field coupling feature set are matched one by one with the corresponding low-dimensional feature samples, and a low-dimensional feature set that preserves the local nonlinear correlation structure is obtained. This completes the introduction of the local linear mechanism and the dimensionality reduction of the multi-field coupling feature set.

6. The intelligent control method for a new energy vehicle thermal management system based on AI technology according to claim 4, characterized in that, In step S102, the low-dimensional feature set is input into the kernel partial least squares algorithm. The radial basis function (RBF) is selected as the mapping carrier from low-dimensional features to high-dimensional feature space. With the goal of minimizing the prediction error of the kernel partial least squares algorithm, the search range of the RBF is set. The weights and step size are adjusted by the adaptive weighted particle swarm optimization algorithm. The RBF values ​​are iteratively updated until the prediction error of the kernel partial least squares algorithm is minimized and / or the maximum number of iterations is reached. Finally, the kernel parameters of the optimal RBF are obtained. The low-dimensional feature vector of the low-dimensional feature set is operated on with the optimal RBF to complete the nonlinear mapping from low-dimensional features to high-dimensional feature space. All high-dimensional feature vectors obtained after mapping low-dimensional features to high-dimensional features are arranged in rows to construct a high-dimensional feature matrix. Each element in the high-dimensional feature matrix corresponds to a single high-dimensional feature value of a single sample. The true values ​​of the thermal management key parameters corresponding to each high-dimensional feature vector are extracted to form a thermal management key parameter vector. The loss function is to minimize the fitting error between the high-dimensional feature vector and the true values ​​of the thermal management key parameters. The fitting error is quantified as the L2 norm of the difference between the product of the high-dimensional feature matrix and the linear reconstruction matrix and the thermal management key parameter vector. The loss function is solved by the least squares method, and the linear coefficient matrix with the minimum fitting error is obtained as the linear prediction model. The high-dimensional feature matrix is ​​multiplied with the constructed linear coefficient matrix to obtain the preliminary predicted value of the key thermal management parameters corresponding to each feature sample.

7. The intelligent control method for a new energy vehicle thermal management system based on AI technology according to claim 1, characterized in that, S102 uses a fuzzy inference strategy to align the preliminary predicted values ​​of key thermal management parameters, low-dimensional feature sets, and actual values ​​of key thermal management parameters along the time axis, and performs normalization processing. It then detects and removes outliers to obtain the input dataset. The fuzzy mean clustering algorithm is used to determine the optimal number of clusters through cross-validation. The preprocessed input dataset is input into the fuzzy mean clustering algorithm to calculate the membership degree of each sample to each cluster center and generate fuzzy rules. Each rule corresponds to a cluster category, and finally a fuzzy rule library covering the entire input scenario is formed. The preprocessed input dataset is divided into a training subset and a validation subset. The gradient descent method is used to iteratively optimize the membership function parameters used to calculate the membership degree from each sample to each cluster center in the training subset. The optimization stops when the training subset reaches the maximum number of iterations, and the optimal parameter combination is obtained and solidified into the fuzzy rule base. The preliminary predicted values ​​of key thermal management parameters and the low-dimensional features of the low-dimensional feature set are matched with corresponding rules according to the fuzzy rule base. The fuzzy inference results are transformed into membership degrees. The error correction amount is obtained according to the rule corresponding to the membership degree. The preliminary predicted values ​​of key thermal management parameters are adjusted based on the error correction amount. Specifically, the sum of the preliminary predicted values ​​of key thermal management parameters and the error correction amount is used as the corrected predicted values. Finally, the predicted values ​​of key thermal management parameters and the predicted value set of key thermal management parameters are obtained.

8. The intelligent control method for a new energy vehicle thermal management system based on AI technology according to claim 1, characterized in that, S103 assigns decision priorities to the execution components of the thermal management system based on the analytic hierarchy process (AHP). Based on the hardware performance constraints and safety constraints of the key thermal management parameters of the execution components, and based on the predicted values ​​of the key thermal management parameters, it eliminates the predicted values ​​of the key thermal management parameters that exceed the hardware performance constraints and safety constraints. Combined with the thermal management system simulation model, it simulates each execution component one by one. Execution components whose simulation results exceed the predicted values ​​of the key thermal management parameters, hardware performance constraints, and safety constraints are marked for control. According to the assigned decision priorities, the execution components marked for control are shut down and / or reduced in sequence.

9. A new energy vehicle thermal management system based on AI technology, applied to the intelligent control method of the new energy vehicle thermal management system based on AI technology as described in any one of claims 1-8, characterized in that, The system includes: Data acquisition module and processing module: The user acquires multi-source heterogeneous raw signals during the operation of new energy vehicles, constructs a coupling correlation matrix between multi-source signals based on the acquired multi-source heterogeneous raw signals, and extracts features to obtain a multi-field coupling feature set; Feature Dimensionality Reduction and Parameter Prediction Module: This module introduces a local linear mechanism, using Euclidean distance to select the optimal nearest neighbor samples of feature samples, constructs a local linear reconstruction matrix, and solves for the linear reconstruction coefficients as local embedding weights. Iterative optimization based on gradient descent yields a low-dimensional feature set that preserves the local nonlinear correlation structure. This low-dimensional feature set is then input into a kernel partial least squares prediction framework. A radial basis function kernel is selected as the mapping carrier, and an adaptive weighted particle swarm optimization algorithm is used to determine the optimal kernel parameters, completing the nonlinear mapping from low-dimensional features to a high-dimensional feature space. A linear prediction model is constructed, and preliminary predicted values ​​for key thermal management parameters are output. Finally, based on a fuzzy inference strategy, the final predicted values ​​for key thermal management parameters are output. Control strategy optimization module: It is used to assign scenario-based decision priorities to each execution component of the thermal management system based on the analytic hierarchy process. The final predicted value of the key thermal management parameters is used as the target reference value of each execution component. Combined with the thermal management system simulation model, each execution component is simulated one by one. Execution components whose simulation results exceed the predicted value, hardware performance constraints and safety constraints are marked for control. According to the decision priority, shutdown or reduction operations are executed in sequence to finally form the optimal control strategy set.

Images