A power battery safety risk assessment method based on real vehicle data

Abstract

The present invention relates to a power battery safety risk assessment method based on real vehicle data, including the data preparation stage: obtaining operation data; data sorting and feature extraction stage: preprocessing the operation data and extracting relevant features; algorithm model construction stage: coupling the extracted relevant feature data with an LSTM model and an equivalent circuit model through an Adaboost algorithm to construct an algorithm model; algorithm model prediction stage: training risk-free pure electric vehicle data with the constructed algorithm model, and predicting the normal voltage of risky pure electric vehicles through the trained model; safety risk assessment stage. The online parameter identification method of the equivalent circuit model is combined with big data technology and machine learning and statistical learning algorithms to fully mine the hidden battery normal voltage information and its change law from vehicle data and battery data, so as to perform safety risk assessment according to the deviation between the normal voltage and the actual voltage.

Description

Technical Field

[0001] The present invention relates to the field of safety risk assessment of power batteries for pure electric vehicles, and in particular to a power battery safety risk assessment method based on real vehicle data. Background Art

[0002] The automotive industry is developing in the direction of the new "four trends" represented by electrification, networking, intelligence, and sharing. However, there are still considerable challenges in conducting safety risk assessments on batteries for pure electric vehicles.

[0003] The safety risk assessment of power batteries is to conduct a reasonable assessment of the risks during the operation of the vehicle to provide guidance to the driver and determine whether the vehicle can continue to drive safely. The safety risk of power batteries is not only related to the electrochemical system inside the battery and the manufacturing process of the battery, but also to the driving status of the vehicle, the use environment and the environment inside the battery. Summary of the invention

[0004] The purpose of the present invention is to provide a power battery safety risk assessment method based on real vehicle data to overcome the deficiencies in the prior art.

[0005] The technical solution of the present invention to solve the above technical problem is as follows: a power battery safety risk assessment method based on real vehicle data, the specific steps are as follows:

[0006] Data preparation stage: Obtaining the operating data of risk-free pure electric vehicles and risky pure electric vehicles;

[0007] Data collation and feature extraction stage: pre-process the operation data of pure electric vehicles and extract relevant features;

[0008] Algorithm model construction stage: Based on the extracted relevant feature data, the LSTM model and the equivalent circuit model are coupled through the Adaboost algorithm to construct the algorithm model;

[0009] Algorithm model prediction stage: train the risk-free pure electric vehicle data according to the constructed algorithm model, and predict the normal voltage of risky pure electric vehicles through the trained model;

[0010] Safety risk assessment stage: Safety risk assessment is performed based on the deviation between the predicted normal voltage and the actual voltage of the risky pure electric vehicle. The maximum error of the risk-free pure electric vehicle model training is used as the deviation threshold. The power battery safety risk is divided into multiple risk levels based on whether the deviation between the normal voltage and the actual voltage of the risky pure electric vehicle when the power battery is working exceeds the threshold, and then the safety risk assessment is performed based on this.

[0011] Furthermore, the pure electric vehicle operation data mainly includes vehicle status data, vehicle operation data, power battery data itself and vehicle abnormal alarm data;

[0012] The vehicle status data includes vehicle status data and charging status data;

[0013] Vehicle operation data includes vehicle operation mode, speed, gear position and accumulated mileage;

[0014] The data of the power battery itself includes total current, total voltage, battery SOC, maximum value of battery cell voltage, maximum voltage battery, maximum voltage battery subsystem number, minimum value of battery cell voltage, minimum voltage battery subsystem number, minimum voltage battery, maximum temperature subsystem number, maximum temperature value, maximum temperature probe, minimum temperature value, minimum temperature probe, total number of single cell temperature probes, insulation resistance and DC-DC status;

[0015] Vehicle abnormality alarm data includes common alarm signs.

[0016] Furthermore, data preprocessing includes data exploratory analysis, segmentation, missing data processing, and abnormal data processing.

[0017] Furthermore, data exploratory analysis is used to discover whether there are a large number of data anomalies, including but not limited to: the vehicle driving status is displayed as a parking charging status, and the vehicle parking charging status is displayed as a driving status;

[0018] Also, whether there are a large number of missing values ​​in the mileage, speed, vehicle status, temperature, voltage, DC-DC status and other data, as well as general alarm information that appears in a certain discharge segment and abnormal data with delayed collection;

[0019] The segment division is used to divide the operation data of pure electric vehicles into three vehicle state segments: driving segment, parking charging segment, and stationary segment;

[0020] The missing data are in speed, cumulative mileage, charging status, maximum cell voltage, minimum cell voltage, maximum temperature, minimum temperature, DC-DC status, vehicle status, and total number of cell temperature probes. The missing values ​​of mileage and speed are in the vehicle charging status, and the speed is filled with a value of 0;

[0021] The accumulated mileage is filled with the mileage value in the segment. If the mileage value in the segment is completely missing, the mileage value at the end of the previous segment is used to fill it.

[0022] The highest cell voltage, the lowest cell voltage, the highest temperature, and the lowest temperature are filled in using the random forest algorithm;

[0023] For missing DC-DC state values, if the DC-DC state values ​​before and after the missing value are the same, the DC-DC state value at the previous moment is used to fill it, otherwise it is directly deleted;

[0024] For missing values ​​of vehicle status, speed, current and voltage are used to fill them;

[0025] For the total number of single cell temperature probes, if all data are missing, they will be deleted, otherwise they will be filled with the majority;

[0026] All abnormal data with a value of 0 are processed using the missing value method;

[0027] The abnormal data with delayed acquisition is corrected by shifting the speed data;

[0028] If the vehicle status mark is abnormal, the existing vehicle status is corrected according to the characteristics such as vehicle speed and current;

[0029] For general alarm abnormal data, data containing a large amount of coupled battery alarm information will be deleted.

[0030] Furthermore, the feature extraction stage is specifically as follows: extracting features from the data obtained in the data sorting stage, analyzing the correlation between the extracted related features using the Pearson coefficient (PCC), and extracting parameters with higher correlation.

[0031] Furthermore, the algorithm model construction stage is specifically as follows: the relevant feature data obtained in the feature extraction stage is recorded as x, the normal battery voltage value at time t is recorded as Y, and a model of Y=f(x) is established, wherein f() is the conversion relationship between x and Y obtained based on model learning; the input of the model is the feature extracted data at time t, and the output of the model is the normal battery voltage value of the battery at time t; f() is constructed by coupling the LSTM model and the equivalent circuit model with Adaboost.

[0032] Furthermore, the LSTM model construction process is:

[0033] Loading dataset: importing the input amount of data;

[0034] Define the model: In Keras, a neural network is defined through a series of layers. First, the input and output quantities are determined. Multiple inputs and one output are set. The input is the strongly correlated feature extracted by PCC, and the output is the normal voltage value.

[0035] Compile the model:

[0036] Compile the defined model. During the compilation process, select the parameters required for training the network for prediction. The loss function is MSE.

[0037] Train the model:

[0038] The model is trained on the imported data through the fit() function. During the training process, the number of iterations on the data set is called epochs, and the number of samples contained in each batch during gradient descent is batch.

[0039] Furthermore, the equivalent circuit model construction process is:

[0040] The Theven in model is established, and the normal voltage is predicted by online parameter identification. The specific calculation process is as follows:

[0041]

[0042] Among them, μ is the forgetting factor; K Ls,i is the gain of the algorithm; P Ls,i is the error covariance matrix of the state estimate, Φ i Is the data variable of the system; is the parameter variable of the system; the subscript i indicates the data value at the i-th moment;

[0043] The predicted value of normal voltage can be obtained as follows:

[0044]

[0045] Furthermore, the construction process of the Adaboost algorithm coupling construction algorithm model is as follows:

[0046] The voltage prediction matrix obtained by LSTM and equivalent circuit model is as follows:

[0047]

[0048] The error rates of the LSTM model and the equivalent circuit model calculated based on the Adaboost algorithm are as follows:

[0049]

[0050] Among them, U i,c represents the real voltage value at time i, represents the maximum error of training; n represents the number of samples

[0051] The weight α of the LSTM model and the equivalent circuit model is calculated as follows:

[0052]

[0053] Where η represents the learning rate hyperparameter;

[0054] According to the above formula, the weight α of the LSTM model is obtained respectively LSTMand the weight α of the equivalent circuit model E , so as to calculate the final weights of LSTM and equivalent circuit model and ensure that the sum of the two weights is 1. The specific calculation method is as follows:

[0055]

[0056] The final voltage prediction value is as follows:

[0057]

[0058] in is the normal voltage prediction value of the equivalent circuit, is the normal voltage prediction value of the LSTM model.

[0059] Furthermore, MSE is used to evaluate the accuracy of model prediction in the algorithm model prediction stage.

[0060] Furthermore, in the security risk assessment stage, the risk levels are divided into three levels, namely 1, 2, and 3.

[0061] The present invention has the following beneficial effects: according to the operation data of pure electric vehicles, the normal voltage information of the battery and its variation law are fully mined, and strong correlation features are extracted through PCC. The online parameter identification method of the equivalent circuit model is combined with big data technology and machine learning and statistical learning algorithms to fully mine the hidden normal voltage information of the battery and its variation law from the vehicle data and battery data, so as to perform safety risk assessment according to the deviation between the normal voltage and the actual voltage;

[0062] In addition, the data type is clarified and the data is cleaned (preprocessed). At the same time, the data features are extracted on this basis. Finally, the characterized data is used for model training and algorithm evaluation, and the prediction accuracy of the model is improved through model coupling. Moreover, with the accumulation of time and the increase of data volume, the prediction effect and accuracy of the model will continue to improve. On this basis, the maximum error of the risk-free pure electric vehicle model training is used as the deviation threshold, and the power battery safety risk is divided into three risk levels of 1, 2, and 3 according to whether the deviation between the normal voltage and the actual voltage of the risky pure electric vehicle power battery when working exceeds the threshold, so as to conduct safety risk assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Implement methods for power battery safety risk assessment;

[0064] Figure 2 is a system block diagram of the present invention;

[0065] Figure 3 It is the framework diagram of the LSTM model of the present invention;

[0066] Figure 4 Identify the structural diagram for equivalent circuit parameters. DETAILED DESCRIPTION

[0067] The principles and features of the present invention are described below. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0068] Example 1

[0069] like Figure 1 to Figure 4 As shown in the figure, a power battery safety risk assessment method based on real vehicle data is shown in the figure. The specific steps are as follows:

[0070] In the data preparation stage, the operating data of risk-free pure electric vehicles and risky pure electric vehicles are obtained;

[0071] Data collation and feature extraction stage: pre-process the operation data of pure electric vehicles and extract relevant features;

[0072] Algorithm model construction stage: Based on the extracted relevant feature data, the LSTM model and the equivalent circuit model are coupled through the Adaboost algorithm to construct the algorithm model;

[0073] Algorithm model prediction stage: train the risk-free pure electric vehicle data according to the constructed algorithm model, and predict the normal voltage of risky pure electric vehicles through the trained model;

[0074] Safety risk assessment stage: Safety risk assessment is performed based on the deviation between the predicted normal voltage and the actual voltage of the risky pure electric vehicle. The maximum error of the risk-free pure electric vehicle model training is used as the deviation threshold. The power battery safety risk is divided into multiple risk levels based on whether the deviation between the normal voltage and the actual voltage of the risky pure electric vehicle when the power battery is working exceeds the threshold, and then the safety risk assessment is performed based on this.

[0075] Example 2

[0076] This embodiment is a further improvement and optimization based on the embodiment 1, and specifically includes:

[0077] The pure electric vehicle operation data is the operation history data of risk-free pure electric vehicles and risky pure electric vehicles;

[0078] The operation data of pure electric vehicles mainly includes vehicle status data, vehicle operation data, power battery data and vehicle abnormal alarm data;

[0079] The vehicle status data includes vehicle status data and charging status data;

[0080] Vehicle operation data includes vehicle operation mode, speed, gear position and accumulated mileage;

[0081] The data of the power battery itself includes total current, total voltage, battery SOC, maximum value of battery cell voltage, maximum voltage battery, maximum voltage battery subsystem number, minimum value of battery cell voltage, minimum voltage battery subsystem number, minimum voltage battery, maximum temperature subsystem number, maximum temperature value, maximum temperature probe, minimum temperature value, minimum temperature probe, total number of single cell temperature probes, insulation resistance and DC-DC status;

[0082] Vehicle abnormality alarm data includes common alarm signs.

[0083] Example 3

[0084] This embodiment is a further improvement and optimization based on Embodiment 1 or 2, and specifically includes:

[0085] Data preprocessing includes data exploratory analysis, segmentation, missing data processing, and general alarm abnormal data processing.

[0086] Among them, data exploratory analysis is used to find out whether there are a large number of data anomalies;

[0087] Specific examples include but are not limited to:

[0088] The vehicle driving status is displayed as parking and charging, but the vehicle is displayed as driving when it is parking and charging. In addition, there are a large number of missing values ​​in the mileage, speed, vehicle status, temperature, voltage, DC-DC status and other data. There are general alarm messages in a certain discharge segment and abnormal data with delayed collection.

[0089] In order to facilitate the processing of outliers, the segments need to be divided, so that the operation data of pure electric vehicles can be divided into three vehicle state segments: driving segment, parking charging segment, and static segment;

[0090] The missing data are in speed, cumulative mileage, charging status, maximum cell voltage, minimum cell voltage, maximum temperature, minimum temperature, DC-DC status, vehicle status, and total number of cell temperature probes. The missing values ​​of mileage and speed are in the vehicle charging status, so the speed is filled with 0 value;

[0091] The accumulated mileage is filled with the mileage value in the segment. If the mileage value in the segment is completely missing, the mileage value at the end of the previous segment is used to fill it.

[0092] The highest cell voltage, the lowest cell voltage, the highest temperature, and the lowest temperature are filled in using the random forest algorithm;

[0093] For missing DC-DC state values, if the DC-DC state values ​​before and after the missing value are the same, the DC-DC state value at the previous moment is used to fill it, otherwise it is directly deleted;

[0094] For missing values ​​of vehicle status, speed, current and voltage are used to fill them;

[0095] For the total number of single cell temperature probes, if all data are missing, they will be deleted, otherwise they will be filled with the majority;

[0096] All abnormal data with a value of 0 are processed using the missing value method;

[0097] The abnormal data with delayed acquisition is corrected by shifting the speed data;

[0098] If the vehicle status mark is abnormal, the existing vehicle status is corrected according to the characteristics such as vehicle speed and current;

[0099] For general alarm abnormal data, data containing a large amount of coupled battery alarm information will be deleted.

[0100] Example 4

[0101] This embodiment is a further improvement and optimization based on any one of Embodiments 1 to 3, and specifically includes:

[0102] The feature extraction stage is as follows:

[0103] Feature extraction is performed on the data obtained in the data sorting stage, and the correlation between the extracted relevant features is analyzed using the Pearson coefficient (PCC) to extract parameters with higher correlation.

[0104] Example 5

[0105] This embodiment is a further improvement and optimization based on any one of Embodiments 1 to 4, and specifically, it can be:

[0106] Specifically, the algorithm model construction stage records the relevant feature data obtained in the feature extraction stage as x, the normal battery voltage value at time t as Y, and establishes a model of Y=f(x), where f() is the conversion relationship between x and Y obtained based on model learning; the input of the model is the feature extracted data at time t, and the output of the model is the normal battery voltage value of the battery at time t; f() is constructed by coupling the LSTM model and the equivalent circuit model with Adaboost.

[0107] Among them, the LSTM model construction process is:

[0108] (1) Load the dataset:

[0109] The input volume of the imported data;

[0110] (2) Define the model:

[0111] In Keras, a neural network is defined through a series of layers. First, the input and output quantities are determined. This time, there are multiple inputs and one output. The input is the strongly correlated feature extracted by PCC, and the output is the normal voltage value.

[0112] (3) Compile the model:

[0113] Compile the defined model, select the parameters required for training the network to predict, and select the loss function as MSE;

[0114] (4) Training model:

[0115] The model is trained on the imported data through the fit() function. During the training process, the number of iterations on the data set is called epochs, and the number of samples contained in each batch during gradient descent is called batch. In this embodiment, the number of iterations is set to 50 and the batch is set to 32;

[0116] Example 6

[0117] This embodiment is a further improvement and optimization based on any one of Embodiments 1 to 5, and specifically includes:

[0118] The equivalent circuit model construction process is:

[0119] Common power battery equivalent circuit models include the Ri nt model, Theven in model, and dual polarization model. The Ri nt model and Theven in model are widely used in SOC estimation and voltage prediction and identification. Since the Theven in model has high modeling accuracy and low computational complexity, this equivalent circuit model uses this model to predict the normal battery voltage.

[0120] The equivalent circuit model can use online or offline data to identify parameters. Online parameter identification is mainly carried out through the real-time current, temperature and voltage data of the power battery to update the parameters, while offline parameter identification is to update and calibrate the parameters through the data stored in the BMS system. This time, the online parameter identification method is used to predict the normal voltage. The specific calculation process is as follows:

[0121]

[0122] Among them, μ is the forgetting factor; K Ls,i is the gain of the algorithm; P Ls,i is the error covariance matrix of the state estimate, Φi Is the data variable of the system; is the parameter variable of the system; the subscript i indicates the data value at the i-th moment.

[0123] The predicted value of normal voltage can be obtained as follows:

[0124]

[0125] Among them, the construction process of the Adaboost algorithm coupling construction algorithm model is:

[0126] The voltage prediction matrix obtained by LSTM and equivalent circuit model is as follows:

[0127]

[0128] The error rates of the LSTM model and the equivalent circuit model calculated based on the Adaboost algorithm are as follows:

[0129]

[0130] Among them, U i,c represents the real voltage value at time i, represents the maximum error of training; n represents the number of samples

[0131] The weight α of the LSTM model and the equivalent circuit model is calculated as follows:

[0132]

[0133] Among them, η represents the learning rate hyperparameter, which defaults to 1;

[0134] According to the above formula, the weight α of the LSTM model is obtained respectively LSTM and the weight α of the equivalent circuit model E , thereby calculating the final weights of the LSTM model and the equivalent circuit model and ensuring that the sum of the two weights is 1. The specific calculation method is as follows:

[0135]

[0136] The final voltage prediction value is as follows:

[0137]

[0138] in is the normal voltage prediction value of the equivalent circuit, is the normal voltage prediction value of the LSTM model.

[0139] Example 7

[0140] This embodiment is a further improvement and optimization based on any one of Embodiments 1 to 6, and specifically includes:

[0141] MSE is used to evaluate the accuracy of model prediction in the algorithm model prediction stage.

[0142] Example 8

[0143] This embodiment is a further improvement and optimization based on any one of Embodiments 1 to 7. Specifically, the risk level in the security risk assessment stage is divided into three levels, namely, risk levels 1, 2, and 3. Of course, it does not exclude the possibility of being divided into other levels.

[0144] A step of power battery safety risk assessment based on real vehicle data in a specific implementation of the present invention, wherein:

[0145] Obtain the relevant data of pure electric vehicle operation. The pure electric vehicle operation data mainly includes vehicle status data, vehicle operation data, power battery data and general alarm data; the relevant data is shown in the following table:

[0146]

[0147]

[0148] Clean the data related to the operation of pure electric vehicles and extract features from the cleaned data to prepare for subsequent model learning;

[0149] In this embodiment, it is mainly achieved through data processing, and ensuring data quality is the premise of improving the accuracy of the results. Therefore, it is necessary to clean the relevant data. The present invention converts low-quality data into data that meets the quality requirements, including:

[0150] Segment division: Data loss is very likely to occur during data transmission. In order to better handle abnormal and missing data, the vehicle operation data needs to be segmented. The vehicle operation data is divided into three vehicle state segments: driving segment, parking charging segment, and stationary segment.

[0151] Missing value processing: Missing data are mainly in speed, cumulative mileage, charging status, highest cell voltage, lowest cell voltage, highest temperature, lowest temperature, DC-DC status, vehicle status, and total number of cell temperature probes. The missing values ​​of mileage and speed are in the vehicle charging status, so the speed is filled with 0 value, and the cumulative mileage is filled with the mileage value in the segment. If the mileage value in the segment is all missing, the mileage value at the end of the previous segment is used to fill it; the highest cell voltage, the lowest cell voltage, the highest temperature and the lowest temperature are filled by the random forest algorithm; for the missing values ​​of DC-DC status, if the DC-DC status values ​​before and after the missing value are the same, the DC-DC status value at the previous moment is used to fill it, otherwise it is directly deleted; for the missing values ​​of vehicle status, speed, current and voltage are used to fill it; for the total number of cell temperature probes, if all the data are missing, it is deleted, otherwise it is filled with the majority;

[0152] Abnormal value processing: Abnormal data with a value of 0 are all processed using the missing value method; abnormal data with delayed collection are corrected by shifting the speed data; abnormal vehicle status marks are corrected based on characteristics such as speed and current; and data with a large number of coupling battery alarm messages are deleted for general alarm abnormal data.

[0153] After cleaning the data, summarize and extract features to obtain characterized data. Since the subsequent steps require data processing and calculation, in order to facilitate the model to better identify and calculate the features of the data, the data needs to be characterized so as to display the features of the data and facilitate the recognition and learning of the model.

[0154] The relevant feature data obtained in the feature construction stage is recorded as x, the normal battery voltage value at time t is recorded as Y, and a model of Y=f(x) is established, where f() is the conversion relationship between x and Y obtained based on model learning; the input of the model is the feature extracted data at time t, and the output of the model is the normal battery voltage value of the battery at time t; f() is constructed by coupling the LSTM model and the equivalent circuit model with Adaboost.

[0155] Next, the risk-free pure electric vehicle data is trained according to the constructed algorithm model. The trained model is used to predict the normal voltage of risky pure electric vehicles and the MSE is used to evaluate the accuracy of the model prediction.

[0156] Finally, a safety risk assessment is performed based on the deviation between the predicted normal voltage and the actual voltage of the risky pure electric vehicle. The maximum error of the risk-free pure electric vehicle model training is used as the deviation threshold. The power battery safety risk is divided into three risk levels of 1, 2, and 3 according to whether the deviation between the normal voltage and the actual voltage of the risky pure electric vehicle when the power battery is working exceeds the threshold to perform safety risk assessment.

[0157] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A power battery safety risk assessment method based on real vehicle data, characterized in that: The specific steps are as follows: Data preparation stage: Obtaining the operating data of risk-free pure electric vehicles and risky pure electric vehicles; Data collation and feature extraction stage: pre-process the operation data of pure electric vehicles and extract relevant features; Algorithm model construction stage: Based on the extracted relevant feature data, the LSTM model and the equivalent circuit model are coupled through the Adaboost algorithm to construct the algorithm model; Algorithm model prediction stage: train the risk-free pure electric vehicle data according to the constructed algorithm model, and predict the normal voltage of risky pure electric vehicles through the trained model; The algorithm model construction stage is specifically as follows: record the relevant feature data obtained in the feature extraction stage as x, the normal battery voltage value at time t as Y, and establish a model of Y = f(x), where f() is the conversion relationship between x and Y obtained based on model learning; the input of the model is the data at time t, and the output of the model is the normal battery voltage value of the battery at time t; f() is constructed by coupling the LSTM model and the equivalent circuit model with Adaboost; The LSTM model construction process is: Loading dataset: importing the input amount of data; Define the model: In Keras, a neural network is defined through a series of layers. First, the input and output quantities are determined. Multiple inputs and one output are set. The input is the strongly correlated feature extracted by PCC, and the output is the normal voltage value. Compile the model: Compile the defined model. During the compilation process, select the parameters required for training the network for prediction. The loss function is MSE. Train the model: The model is trained on the imported data through the fit() function. During the training process, the number of iterations on the data set is called epochs, and the number of samples contained in each batch during gradient descent is batch; The equivalent circuit model construction process is: The Thevenin model is established, and the normal voltage is predicted by online parameter identification. The specific calculation process is as follows: Among them, μ is the forgetting factor; K Ls,i is the gain of the algorithm; P Ls,i is the error covariance matrix of the state estimate, Φ i Is the data variable of the system; is the parameter variable of the system; the subscript i indicates the data value at the i-th moment; The predicted value of normal voltage can be obtained as follows: Algorithm model prediction stage: train the risk-free pure electric vehicle data according to the constructed algorithm model, and predict the normal voltage of risky pure electric vehicles through the trained model; Safety risk assessment stage: Safety risk assessment is performed based on the deviation between the predicted normal voltage and the actual voltage of the risky pure electric vehicle. The maximum error of the risk-free pure electric vehicle model training is used as the deviation threshold. The power battery safety risk is divided into multiple risk levels based on whether the deviation between the normal voltage and the actual voltage of the risky pure electric vehicle when the power battery is working exceeds the threshold, and then the safety risk assessment is performed based on this.

2. A power battery safety risk assessment method based on real vehicle data according to claim 1, characterized in that: Pure electric vehicle operation data includes vehicle status data, vehicle operation data, power battery data and vehicle abnormal alarm data; The vehicle status data includes vehicle status data and charging status data; The vehicle operation data includes vehicle operation mode, vehicle speed, gear position and accumulated mileage; The power battery data includes total current, total voltage, battery SOC, maximum value of battery cell voltage, maximum voltage battery, maximum voltage battery subsystem number, minimum value of battery cell voltage, minimum voltage battery subsystem number, minimum voltage battery, maximum temperature subsystem number, maximum temperature value, maximum temperature probe, minimum temperature value, minimum temperature probe, total number of single cell temperature probes, insulation resistance and DC-DC status; The vehicle abnormality alarm data includes a general alarm mark.

3. The power battery safety risk assessment method based on real vehicle data according to claim 1 is characterized in that: Data preprocessing includes data exploratory analysis, segmentation, missing data processing, and general alarm abnormal data processing.

4. The power battery safety risk assessment method based on real vehicle data according to claim 3 is characterized in that: Data exploratory analysis is used to discover whether there are a large number of data anomalies, including but not limited to: the vehicle driving status is displayed as parking and charging status, and the vehicle parking and charging status is displayed as driving status; Also, whether there are a large number of missing values ​​in the mileage, speed, vehicle status, temperature, voltage, DC-DC status data, as well as general alarm information that appears in a certain discharge segment and abnormal data with delayed collection; The segment division is used to divide the operation data of pure electric vehicles into three vehicle state segments: driving segment, parking charging segment, and stationary segment; The missing data are in speed, cumulative mileage, charging status, maximum cell voltage, minimum cell voltage, maximum temperature, minimum temperature, DC-DC status, vehicle status, and total number of cell temperature probes. The missing values ​​of mileage and speed are in the vehicle charging status, and the speed is filled with a value of 0; The accumulated mileage is filled with the mileage value in the segment. If the mileage value in the segment is completely missing, the mileage value at the end of the previous segment is used to fill it. The highest cell voltage, the lowest cell voltage, the highest temperature, and the lowest temperature are filled in using the random forest algorithm; For missing DC-DC state values, if the DC-DC state values ​​before and after the missing value are the same, the DC-DC state value at the previous moment is used to fill it, otherwise it is directly deleted; For missing values ​​of vehicle status, speed, current and voltage are used to fill them; For the total number of single cell temperature probes, if all data are missing, they will be deleted, otherwise they will be filled with the majority; All abnormal data with a value of 0 are processed using the missing value method; The abnormal data with delayed acquisition is corrected by shifting the speed data; If the vehicle status mark is abnormal, the existing vehicle status is corrected according to the vehicle speed and current characteristics; For general alarm abnormal data, data containing a large amount of coupled battery alarm information will be deleted.

5. The power battery safety risk assessment method based on real vehicle data according to claim 1 is characterized in that: The feature extraction stage is specifically as follows: extracting features from the data obtained in the data sorting stage, using the Pearson coefficient to analyze the correlation between the extracted related features, and extracting parameters with higher correlation.

6. The power battery safety risk assessment method based on real vehicle data according to claim 1 is characterized in that: The construction process of the Adaboost algorithm coupling construction algorithm model is as follows: The voltage prediction matrix obtained by LSTM and equivalent circuit model is as follows: The error rates of the LSTM model and the equivalent circuit model calculated based on the Adaboost algorithm are as follows: Among them, U i,c represents the real voltage value at time i, represents the maximum error of training; n represents the number of samples; The weight α of the LSTM model and the equivalent circuit model is calculated as follows: Where η represents the learning rate hyperparameter; According to the above formula, the weight α of the LSTM model is obtained respectively LSTM and the weight α of the equivalent circuit model E , thereby calculating the final weights of the LSTM model and the equivalent circuit model and ensuring that the sum of the two weights is 1. The specific calculation method is as follows: The final voltage prediction value is as follows: in is the normal voltage prediction value of the equivalent circuit, is the normal voltage prediction value of the LSTM model.

7. A power battery safety risk assessment method based on real vehicle data according to claim 6, characterized in that: MSE is used to evaluate the accuracy of model prediction in the algorithm model prediction stage.

Images