Power battery abnormity identification method based on greedy-relaxation hyper-parameter optimization

Abstract

The invention discloses a battery abnormity identification method based on greedy-relaxation hyper-parameter optimization. The method comprises the following steps: step 1, preprocessing vehicle operation data; step 2, selecting data with strong correlation as an input item of the model; step 3, determining a hyper-parameter category needing to be optimized, and performing hyper-parameter optimization by adopting a greedy-relaxation hyper-parameter optimization method to obtain a comprehensive optimal hyper-parameter; 4, applying the comprehensive optimal hyper-parameter to the neural network prediction model to obtain an optimized neural network prediction model; inputting the data with strong correlation into the optimized neural network prediction model to obtain an output value, namely a predicted value; and step 5, calculating a residual error between a predicted value and a true value of the neural network prediction model, and when a difference value between the true value and the predicted value is greater than an abnormal threshold, determining that the model is abnormal, otherwise, determining that the model is normal. According to the method, the balance between the model performance and the calculation efficiency is sought, the battery anomaly recognition neural network model is quickly and accurately constructed, and the efficiency of the model is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power battery safety, and relates to a method for identifying battery anomalies, and specifically to a method for identifying battery anomalies based on greedy-relaxed hyperparameter optimization. Background Art

[0002] In recent years, power battery technology has achieved rapid development with the rapid growth of electric vehicles. Power batteries not only provide energy for vehicles, but also affect the safety of vehicles. Before serious failure, battery parameters such as voltage and temperature will change abnormally, so accurate identification of battery abnormalities is of great significance to ensure the safe operation of vehicles.

[0003] Neural network models have been widely used in power battery anomaly detection. Hyperparameters have a very important impact on neural networks, affecting the robustness, stability and generalization ability of the network to a certain extent. Therefore, choosing appropriate hyperparameters is of great significance to improving the performance of the network.

[0004] The commonly used hyperparameter optimization algorithms cannot achieve a balance between optimization time and optimization performance, which leads to certain instability of neural network models in large-scale applications. With the continuous and rapid growth of the number of electric vehicles, how to obtain a more accurate hyperparameter combination in a short time is of great significance for the use of neural network models in the big data framework. Summary of the invention

[0005] The purpose of the present invention is to provide a battery anomaly identification method to solve the problem of hyperparameter optimization in battery anomaly identification of neural network models under the big data framework, and to achieve rapid optimization of hyperparameters of battery anomaly identification models. The method seeks a balance between model performance and computational efficiency based on vehicle cloud platform monitoring data, and achieves rapid and accurate construction of a battery anomaly identification neural network model, thereby improving the efficiency of the model.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions:

[0007] A battery anomaly identification method based on greedy-relaxed hyperparameter optimization specifically includes the following steps:

[0008] Step 1, preprocessing the vehicle operation data to obtain preprocessed vehicle operation data;

[0009] Step 2: Perform correlation analysis on different feature items and output features, and select data with strong correlation as input items of the model;

[0010] Step 3: Determine the hyperparameter categories that need to be optimized, and use the greedy-relaxed hyperparameter optimization method to optimize the hyperparameters to obtain the comprehensive optimal hyperparameters. The hyperparameters that need to be optimized include four categories: learning rate lr, batch size batch_size, number of neural network layers, and number of neurons. The loss function needs to be selected as the evaluation indicator. The specific steps are as follows:

[0011] Step 31, hyperparameter initialization: determine multiple hyperparameter categories that need to be optimized according to task requirements, list possible hyperparameter values ​​for each category of hyperparameters, and randomly select a group as the initial hyperparameters;

[0012] Step 32, hyperparameter pre-tuning: for each hyperparameter category in the initial hyperparameters, iterate all possible values ​​thereof while keeping the hyperparameter values ​​of other categories unchanged, train the model on such hyperparameter combination and calculate the loss obtained in each iteration, and obtain the loss of the model under all possible values ​​of each hyperparameter category;

[0013] Step 33, hyperparameter category importance evaluation: calculate the range of the model loss obtained under all possible values ​​of the same hyperparameter category to characterize the importance of the hyperparameter category;

[0014] Step 34, search for local optimal hyperparameters: based on the initial hyperparameters obtained in step 31, first optimize the hyperparameter category with the highest importance obtained in step 33;

[0015] Step 35, searching for the overall optimal hyperparameter: according to the local optimal candidate hyperparameter of the current hyperparameter category, the parameter value under the optimized parameter category is fixed unchanged, and the local optimal hyperparameter of the most important category among the remaining hyperparameter categories is searched; when the hyperparameter category with the lowest importance is optimized, the hyperparameter with the smallest loss is selected as the overall optimal hyperparameter;

[0016] Step 4, applying the comprehensive optimal hyperparameters to the neural network prediction model to obtain an optimized neural network prediction model; inputting the highly correlated data obtained in step 2 into the optimized neural network prediction model to obtain an output value, i.e., a predicted value; the model prediction parameters are identified as normal parameters of the battery;

[0017] Step 5, calculate the residual of the predicted value and the true value of the neural network prediction model obtained in step 4. When the difference between the true value and the predicted value is greater than the abnormal threshold, it is considered abnormal, otherwise it is normal.

[0018] Furthermore, in step 1, the preprocessing principle is as follows:

[0019] 1) Sort vehicle operation data by time and perform redundancy removal operations;

[0020] 2) If there is a null value in the current data frame, the average of the two frames of data adjacent to the current data frame is used as the value of the current frame of data;

[0021] 3) The Z-score method was used to achieve data standardization.

[0022] Furthermore, in step 2, the Pearson correlation coefficient is used to evaluate the correlation between the data.

[0023] Furthermore, in step 5, the box plot method is used to set the abnormal threshold, and the calculation formula is as follows:

[0024] Thr = Q3 + 1.5 * IQR

[0025] Where Q3 is the third quartile of the residual between the predicted value and the true value, and IQR is the interquartile range.

[0026] Compared with the prior art, the present invention focuses on the abnormality identification method of electric vehicle power batteries. Based on the real vehicle operation big data uploaded by electric vehicles, the present invention aims at the instability of model efficiency and performance caused by the selection of hyperparameters in the neural network model in the vehicle big data, and proposes a power battery abnormality identification method based on greedy-relaxation hyperparameter optimization. Greedy ensures that the search for hyperparameters is carried out in the direction of reducing losses, and relaxation reduces the probability of falling into the local optimum. Battery abnormality identification is performed through the difference between the model prediction value and the true value, which can provide early warning for vehicles with thermal runaway and potential failure risks, thereby improving driving safety. The method of the present invention is based on the vehicle cloud platform monitoring data, seeks a balance between model performance and computational efficiency, and realizes the rapid and accurate construction of a battery abnormality identification neural network model, thereby improving the efficiency of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is the overall flow chart of the method of the present invention; DETAILED DESCRIPTION

[0028] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] like Figure 1 As shown, the battery anomaly identification method based on greedy-relaxed hyperparameter optimization provided by the present invention specifically includes the following steps:

[0030] Step 1: preprocess the vehicle operation data to obtain preprocessed vehicle operation data to ensure that the requirements of the model are met.

[0031] In this embodiment, according to the national standard GB / T 32960 "Technical Specifications for Electric Vehicle Remote Service and Management System", the historical data uploaded by the electric vehicle needs to be preprocessed to meet the model requirements. The specific preprocessing principles are as follows:

[0032] 1) Sort the vehicle operation data by time and perform redundancy removal operations to overcome the problems of time disorder and data duplication of vehicle operation data.

[0033] 2) If there is a null value in the current data frame, the average of the two frames of data adjacent to the current data frame is used as the value of the current frame of data;

[0034] 3) Since different data items in the national standard GB / T 32960 have different valid ranges, the Z-score method is used in this embodiment to achieve data standardization to ensure that each feature has a similar scale and range. The calculation formula is as follows:

[0035]

[0036] In the formula, x is the data to be standardized, μ is the average value of the data under the feature item corresponding to x, σ is the standard deviation of all data under the data item corresponding to x, and Z is the standardized score.

[0037] Step 2, in order to obtain better model performance, it is necessary to perform correlation analysis on different feature items and output features. Select data with strong correlation as the input item of the model. In this embodiment, the Pearson correlation coefficient is used to evaluate the correlation between data. Strong correlation represents greater contribution to model performance. Generally, when the correlation coefficient reaches 0.6 or above, it can be identified as a strong correlation, so it can be used as an input item of the model.

[0038] The Pearson correlation coefficient calculation formula is as follows:

[0039]

[0040] In the formula, n is the number of samples contained in the measured data item, x i and i Represent the i-th value of the measured data item and the target data item respectively, and are the average values ​​of the model candidate input items and target feature item data, respectively, and coef is the Pearson correlation coefficient.

[0041] Step 3, determine the hyperparameter categories that need to be optimized, and use the greedy-relaxed hyperparameter optimization method to optimize the hyperparameters to obtain the comprehensive optimal hyperparameters. Specifically, the hyperparameters that need to be optimized include four categories: learning rate lr, batch size batch_size, number of neural network layers, and number of neurons. At the same time, in order to evaluate the performance of the model, the loss function needs to be selected as the evaluation indicator. The specific steps are as follows:

[0042] Step 31, hyperparameter initialization: determine multiple hyperparameter categories that need to be optimized according to task requirements, list possible hyperparameter values ​​for each category of hyperparameters, and randomly select a group as the initial hyperparameters;

[0043] Step 32, hyperparameter pre-tuning: for each hyperparameter category in the initial hyperparameters, iterate all possible values ​​thereof while keeping the hyperparameter values ​​of other categories unchanged, train the model on such hyperparameter combination and calculate the loss (i.e., the value of the loss function) obtained in each iteration, and obtain the loss of the model under all possible values ​​of each hyperparameter category;

[0044] Step 33, hyperparameter category importance assessment: Calculate the range of the model loss (the difference between the maximum and minimum loss) of the same hyperparameter category under all possible values, which is used to characterize the importance of the hyperparameter category. The larger the range, the more significant the impact of the hyperparameter category on the model performance, and the higher the importance of the hyperparameter category.

[0045] The range of all hyperparameter categories is calculated according to the above method to quantitatively evaluate the importance of hyperparameter categories.

[0046] Step 34, search for local optimal hyperparameters: Based on the initial hyperparameters obtained in step 31, first optimize the hyperparameter category with the highest importance (i.e., the largest range) obtained in step 33, specifically: keep the remaining hyperparameter categories other than the optimized hyperparameter category unchanged, iterate all possible values ​​of the optimized hyperparameter category, and obtain the model loss of different hyperparameter combinations under the hyperparameter category. Assuming that there are n different hyperparameter values ​​under the current hyperparameter category (i.e., the optimized category), retain the round(n / ln(n)) corresponding hyperparameter values ​​with the smallest loss and use them as the local optimal candidate hyperparameters for the current hyperparameter category, where round(·) means rounding and n means the number of samples.

[0047] Step 35, search for the optimal comprehensive hyperparameters:

[0048] According to the local optimal candidate hyperparameters of the current hyperparameter category, the parameter values ​​under the optimized parameter category are fixed unchanged, and the local optimal hyperparameters of the most important category among the remaining hyperparameter categories are searched; when the hyperparameter category with the lowest importance is optimized, the hyperparameter with the smallest loss is selected as the comprehensive optimal hyperparameter.

[0049] Step 4, applying the comprehensive optimal hyperparameters to the neural network prediction model to obtain an optimized neural network prediction model; inputting the highly correlated data obtained in step 2 into the optimized neural network prediction model to obtain an output value, i.e., a predicted value.

[0050] Since the prediction results of the neural network prediction model tend to favor the majority class samples, and normal data accounts for the vast majority of vehicle operation data, the model prediction parameters can be considered as normal parameters of the battery.

[0051] Step 5, calculate the residual of the predicted value and the true value of the neural network prediction model obtained in step 4. When the difference between the true value and the predicted value is greater than the abnormal threshold, it is considered abnormal, otherwise it is normal.

[0052] Since the model prediction parameters are considered normal battery parameters, when there is a large deviation between the actual value and the predicted value, this is undoubtedly an abnormal performance. In order to accurately distinguish between normal and abnormal battery parameters, the box plot method is used to set the abnormal threshold based on vehicle historical data. The calculation formula is as follows:

[0053] Thr = Q3 + 1.5 * IQR

[0054] Where Q3 is the third quartile of the residual between the predicted value and the true value, and IQR is the interquartile range.

[0055] With the increasing number of electric vehicles and the gradual maturity of cloud computing technology, safety monitoring of electric vehicles through cloud platforms has gradually become a mainstream trend. In view of the important role of neural network models in cloud computing, the present invention proposes a greedy-relaxed hyperparameter optimization method, which effectively balances the efficiency and accuracy of the model. At the same time, according to the characteristics of vehicle big data, a power battery anomaly identification method based on greedy-relaxed hyperparameter optimization is proposed. It accurately predicts the normal parameters of the battery to detect battery anomalies, providing support for real-time safety risk monitoring of running vehicles.

Claims

1. A battery anomaly identification method based on greedy-relaxed hyperparameter optimization, characterized in that: The specific steps include: Step 1, preprocessing the vehicle operation data to obtain preprocessed vehicle operation data; Step 2: Perform correlation analysis on different feature items and output features, and select data with strong correlation as input items of the model; Step 3: Determine the hyperparameter categories that need to be optimized, and use the greedy-relaxed hyperparameter optimization method to optimize the hyperparameters to obtain the comprehensive optimal hyperparameters. The hyperparameters that need to be optimized include four categories: learning rate lr, batch size batch_size, number of neural network layers, and number of neurons. The loss function needs to be selected as the evaluation indicator. The specific steps are as follows: Step 31, hyperparameter initialization: determine multiple hyperparameter categories that need to be optimized according to task requirements, list possible hyperparameter values ​​for each category of hyperparameters, and randomly select a group as the initial hyperparameters; Step 32, hyperparameter pre-tuning: for each hyperparameter category in the initial hyperparameters, iterate all possible values ​​thereof while keeping the hyperparameter values ​​of other categories unchanged, train the model on such hyperparameter combination and calculate the loss obtained in each iteration, and obtain the loss of the model under all possible values ​​of each hyperparameter category; Step 33, hyperparameter category importance evaluation: calculate the range of the model loss obtained under all possible values ​​of the same hyperparameter category to characterize the importance of the hyperparameter category; Step 34, search for local optimal hyperparameters: based on the initial hyperparameters obtained in step 31, first optimize the hyperparameter category with the highest importance obtained in step 33; Step 35, searching for the overall optimal hyperparameter: according to the local optimal candidate hyperparameter of the current hyperparameter category, the parameter value under the optimized parameter category is fixed unchanged, and the local optimal hyperparameter of the most important category among the remaining hyperparameter categories is searched; when the hyperparameter category with the lowest importance is optimized, the hyperparameter with the smallest loss is selected as the overall optimal hyperparameter; Step 4, applying the comprehensive optimal hyperparameters to the neural network prediction model to obtain an optimized neural network prediction model; inputting the highly correlated data obtained in step 2 into the optimized neural network prediction model to obtain an output value, i.e., a predicted value; the model prediction parameters are identified as normal parameters of the battery; Step 5, calculate the residual of the predicted value and the true value of the neural network prediction model obtained in step 4. When the difference between the true value and the predicted value is greater than the abnormal threshold, it is considered abnormal, otherwise it is normal.

2. The battery anomaly identification method based on greedy-relaxed hyperparameter optimization according to claim 1, characterized in that: In step 1, the preprocessing principles are as follows: 1) Sort vehicle operation data by time and perform redundancy removal operations; 2) If there is a null value in the current data frame, the average of the two frames of data adjacent to the current data frame is used as the value of the current frame of data; 3) The Z-score method was used to achieve data standardization.

3. The battery anomaly identification method based on greedy-relaxed hyperparameter optimization according to claim 1, characterized in that: In step 2, the Pearson correlation coefficient was used to evaluate the correlation between the data.

4. The battery anomaly identification method based on greedy-relaxed hyperparameter optimization according to claim 1, characterized in that: In step 5, the box plot method is used to set the anomaly threshold, and the calculation formula is as follows: Thr = Q3 + 1.5 * IQR Where Q3 is the third quartile of the residual between the predicted value and the true value, and IQR is the interquartile range.

Images