Battery thermal runaway early warning method based on deep convolutional long short-term memory network

Abstract

The invention discloses a battery thermal runaway early warning method based on a deep convolutional long short-term memory network. The method comprises the following steps: S1, acquiring historical battery data field data of a power battery from a data platform and preprocessing the historical battery data field data; s2, selecting battery monomer voltage and probe temperature data from the preprocessed data, respectively extracting statistical characteristics of the battery monomer voltage and the probe temperature data, and respectively extracting time-frequency characteristics of the battery monomer voltage and the probe temperature data by adopting a time-frequency analysis technology; s3, using CLSTM to construct and train a prediction model, wherein the training input is the results of S1 and S2; s4, acquiring real-time data of the power battery and inputting the real-time data into the prediction model to obtain a single battery voltage and a probe temperature prediction value in a prediction time window; and S5, calculating the improved variance and the average extremum entropy of the two groups of predicted values, comparing the result with a set threshold value, and if the numerical value of any one item is greater than the set threshold value, giving out a thermal runaway early warning. The model used in the method can capture non-stationary features of battery data, and the precision of battery data prediction is improved.

Description

Technical Field

[0001] The present invention relates to the field of power battery safety, and in particular to a battery thermal runaway early warning method based on a deep convolutional long short-term memory network (Convolution based Long Short-Term Memory, hereinafter referred to as CLSTM). Background Art

[0002] Lithium-ion power batteries are widely used in the field of electric vehicles due to their advantages such as high specific energy and long cycle life. However, the safe and efficient management of lithium-ion power batteries is still a major technical challenge for the high-quality growth of electric vehicles. Batteries must be strictly managed to prevent overcharging, over-discharging, and overheating. In addition, due to its complex internal structure, the charging and discharging process of lithium batteries is a highly coupled physical and chemical reaction. Different charging and discharging rates, ambient temperature and other factors will have a great impact on the attenuation degree and service life of the battery pack, and the electrochemical parameters inside the battery will gradually change with the use of the battery. This requires the battery management system (Battery Management System, hereinafter referred to as BMS) to be able to identify parameter changes and adopt appropriate control strategies to ensure its safety. When the battery is about to cross the safety boundary and enter a fault state, the fault diagnosis system should give a timely warning, judge the fault based on the existing engineering application experience, and give reasonable countermeasures.

[0003] Recurrent neural networks are used for time series prediction because they can capture long dependencies over time periods. However, when using a long short-term memory (LSTM) network for battery data prediction, the statistical features extracted from the battery signal cannot represent non-stationary information well. In practical applications, non-stationary characteristic information is crucial for battery data prediction. Therefore, it is necessary to add the extraction of non-stationary information features to the traditional prediction algorithm to improve the accuracy of battery data prediction, which is also more conducive to extracting thermal runaway information of the battery. Summary of the invention

[0004] In view of the deficiency that the existing time series algorithm cannot extract the non-stationary information of the battery well, the present invention proposes a battery thermal runaway warning method based on a deep convolutional long short-term memory network, which adds the extraction of non-stationary information features and uses the CLSTM algorithm to predict the voltage and temperature of the battery in the future. At the same time, the improved variance and average extreme entropy of the predicted voltage and temperature are used to judge whether the battery is abnormal, which improves the accuracy of battery data prediction and is also more conducive to extracting the thermal runaway information of the battery.

[0005] The specific technical solutions are as follows:

[0006] A battery thermal runaway warning method based on a deep convolutional long short-term memory network comprises the following steps:

[0007] S1: Acquire historical data of the power battery from the data platform and preprocess it; the historical data includes battery data field data;

[0008] S2: selecting battery cell voltage and probe temperature data from the preprocessed data, extracting statistical features of the two selected data sets respectively, and extracting time-frequency features of the two selected data sets respectively by using time-frequency analysis technology;

[0009] S3: A prediction model is constructed and trained using a deep convolutional long short-term memory network. The input data for training is the preprocessed battery data field data obtained in S1, as well as the statistical characteristics and time-frequency characteristics of the battery cell voltage and probe temperature data obtained in S2. The trained prediction model is used to predict the battery cell voltage and probe temperature.

[0010] S4: acquiring real-time data of the power battery, the real-time data including battery data field data; inputting the real-time data into the trained prediction model to obtain a single cell voltage prediction value and a probe temperature prediction value within a prediction time window;

[0011] S5: Calculate the improved variance and average extreme entropy of the predicted value of the single cell voltage and the predicted value of the probe temperature respectively, and compare the obtained improved variance of the single cell voltage, the average extreme entropy of the single cell voltage, the improved variance of the probe temperature, and the average extreme entropy of the probe temperature with their set thresholds respectively. If the value of any one of them is greater than the set threshold, it is determined that the battery has a thermal runaway risk and a thermal runaway warning is issued; the set threshold is obtained by statistical analysis of historical data.

[0012] Furthermore, in S5, the improved variance S 2 The calculation expression is as follows:

[0013]

[0014] Where M is the total number of battery cells in the battery pack, avg is the average of the M battery cell data sequences K, the battery cell data sequence K represents the battery cell voltage E or the probe temperature T, where K(1) represents the first battery cell data sequence; K(i,j) represents the i-th battery cell data sequence at time j, i = 1, 2, ..., M, j = 1, 2, ..., N, and N is the prediction duration corresponding to the prediction time window.

[0015] Furthermore, in S5, the average extreme entropy The calculation expression is as follows:

[0016]

[0017] Where N is the prediction duration corresponding to the prediction time window, K max Indicates the maximum value in the battery cell data sequence at time t, K min Indicates the minimum value in the battery cell data sequence at time t.

[0018] Furthermore, in S5, the threshold value is set by calculating the improved variance and average extreme entropy of the battery cell voltage and the improved variance and average extreme entropy of the probe temperature in the historical data, analyzing the normal battery data and the thermal runaway data through a box plot, and converting the improved variance S of the battery cell voltage into the improved variance S of the battery cell voltage. E The upper quartile value of the box plot analysis is taken as S E The setting threshold, the average extreme entropy of the single battery voltage The upper quartile value of the box plot analysis is given as The set threshold value, the probe temperature improvement variance S T The upper quartile value of the box plot analysis is taken as S T The set threshold value, the probe temperature average extreme entropy The upper quartile value of the box plot analysis is given as The setting threshold.

[0019] Furthermore, in S1, the historical data of the power battery includes: battery cell voltage, probe temperature, battery pack voltage, current, SOC, SOH, and mileage.

[0020] Furthermore, in S1, the preprocessing method is specifically as follows: first, data cleaning is performed, then data normalization is performed, and finally, standardization is performed.

[0021] Furthermore, in S2, the statistical characteristics of the preprocessed battery cell voltage include the mean, standard deviation, variance, maximum value, and minimum value of the voltage; the statistical characteristics of the preprocessed temperature data include the mean, standard deviation, variance, maximum value, and minimum value of the temperature.

[0022] Furthermore, in S2, the time-frequency analysis technology includes: short-time Fourier transform, wavelet transform, empirical mode decomposition, and ensemble empirical mode decomposition.

[0023] Furthermore, in S3, the CLSTM model includes: a convolutional layer, an LSTM layer, and a fully connected layer; the convolutional layer is used to capture local dependencies, and the LSTM layer is used to perform sequence prediction.

[0024] The beneficial effects of the present invention are:

[0025] (1) The present invention uses time-frequency analysis technology to extract time-frequency features, and uses a convolutional neural network algorithm to improve the accuracy of battery data prediction; the CLSTM algorithm used can simultaneously process the time domain and frequency domain features of time series data through convolution operations. For non-stationary characteristics of signals such as battery voltage and temperature, CLSTM can well capture signal changes; moreover, CLSTM can be constructed into a deep network structure, and each layer can learn different levels of abstract features. This enables CLSTM to learn more complex patterns from raw data, making the algorithm more sensitive to subtle changes in battery performance.

[0026] (2) The present invention uses the improved variance and average extreme entropy of the power battery cell voltage and probe temperature in historical data to set the thermal runaway alarm threshold, and calculates the improved variance and average extreme entropy of the real-time predicted values ​​of the cell voltage and probe temperature, and compares them with the set threshold to determine whether the thermal runaway alarm condition is met. This method combining machine learning and mechanism analysis can accurately determine whether the power battery will have thermal runaway in the future for a long period of time. It is accurate and efficient and has certain application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a flow chart of a battery thermal runaway warning method based on a deep convolutional long short-term memory network proposed in an embodiment of the present invention.

[0028] Figure 2 This is a flow chart of obtaining a threshold value for determining whether a battery has a thermal runaway risk in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] The present invention will be described in detail below according to the accompanying drawings and preferred embodiments, and the purpose and effect of the present invention will become more clear. The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0030] like Figure 1 As shown, a battery thermal runaway warning method based on a deep convolutional long short-term memory network specifically includes the following steps:

[0031] S1: Obtain the historical data of the power battery from the data platform and preprocess it. The historical data of the power battery includes battery data fields such as battery cell voltage, probe temperature, battery pack voltage, current, SOC, SOH, mileage, etc. The preprocessing method is as follows: first clean the data, then normalize the data, and finally perform standardization.

[0032] S2: Select the battery cell voltage and probe temperature data from the preprocessed data, extract the statistical features of the two sets of data (i.e., the battery cell voltage and probe temperature data preprocessed by S1), and apply time-frequency analysis technology to extract the time-frequency features of the two sets of data, in order to better capture the non-stationary characteristics of the battery data. The data finally obtained by S2 are: the statistical features and time-frequency features of the battery cell voltage, and the statistical features and time-frequency features of the probe temperature data. Among them, the statistical features of the preprocessed battery cell voltage include the mean, standard deviation, variance, maximum value, and minimum value of the voltage; the statistical features of the preprocessed temperature data include the mean, standard deviation, variance, maximum value, and minimum value of the temperature. The time-frequency analysis technology includes: short-time Fourier transform (STFT), wavelet transform (WT), empirical mode decomposition (EMD), ensemble empirical mode decomposition (EEMD), and other algorithms. In this embodiment, the STFT algorithm is selected as the time-frequency analysis technology.

[0033] S3: Build a deep convolutional long short-term memory network (CLSTM) model. The CLSTM model includes: convolutional layer, LSTM layer, and fully connected layer. The convolutional layer is used to capture local dependencies, which is conducive to capturing local fluctuations in battery voltage and temperature; the LSTM layer is used to perform sequence prediction. The CLSTM model is trained using the battery data field data preprocessed by S1 (excluding battery cell voltage and probe temperature) and the data extracted by S2 (including the statistical characteristics and time-frequency characteristics of battery cell voltage, and the statistical characteristics and time-frequency characteristics of probe temperature data) to obtain a prediction model for predicting battery cell voltage and probe temperature.

[0034] S4: Acquire the real-time data of the power battery (including the battery data fields selected in S1, namely "battery cell voltage, probe temperature, battery pack voltage, current, SOC, SOH, mileage"), and pass these real-time data into the prediction model to obtain the power battery data prediction value within the prediction time window, namely the single cell voltage prediction value and the probe temperature prediction value.

[0035] S5: Calculate the improved variance and average extreme entropy of the predicted values ​​of the single cell voltage and the probe temperature obtained in S4, respectively, and use the improved variance of the single cell voltage, the average extreme entropy of the single cell voltage, the improved variance of the probe temperature, and the average extreme entropy of the probe temperature as the judgment indicators of thermal runaway. The specific calculation process is as follows:

[0036] 1) Improved variance S 2 :

[0037]

[0038]

[0039] Where, M is the total number of battery cells in the battery pack, avg represents the mean of the M battery cell data sequences K, and the battery cell data sequence K represents the battery cell voltage E or the probe temperature T (that is, when calculating the improved variance of the battery cell voltage, K in equations (1) and (2) represents the battery cell voltage E, and when calculating the improved variance of the probe temperature, K in equations (1) and (2) represents the probe temperature T), where K(1) represents the first battery cell data sequence. K(i,j) represents the i-th battery cell data sequence at time j, i = 1, 2, ..., M, j = 1, 2, ..., N, where N is the prediction duration corresponding to the prediction time window.

[0040] 2) Average extreme entropy:

[0041] The extreme entropy is defined as the ratio of the maximum value to the minimum value in the battery cell data sequence at a certain moment:

[0042] Average extreme entropy It is the average value of the extreme entropy at each moment in the prediction time window. Considering the possible fluctuation of abnormal values, the average extreme entropy is used. As a judgment indicator of thermal runaway, its calculation expression is as follows:

[0043]

[0044] The calculated single cell voltage improvement variance S E , Average extreme entropy of single cell voltage Probe temperature improvement variance S T , probe temperature average extreme entropy Compare them with their set thresholds respectively. If there is S E , S T , If a value in is greater than the set threshold, the battery is judged to have a thermal runaway risk and a thermal runaway warning is issued.

[0045] like Figure 2 As shown in Figure 1, the threshold is set by statistically analyzing the historical data. Specifically, first, the improved variance and average extreme entropy of the battery cell voltage and the improved variance and average extreme entropy of the probe temperature in the historical data are calculated according to equations (1) to (3). Then, the normal battery data and thermal runaway data are analyzed by box plots, and SE The upper quartile value of the box plot analysis is taken as S E The setting threshold value, The upper quartile value of the box plot analysis is given as The setting threshold, S T The upper quartile value of the box plot analysis is taken as S T The setting threshold value, The upper quartile value of the box plot analysis is given as The setting threshold.

[0046] Specifically, the improved variance and average extreme entropy of the cell voltage and temperature exceed the threshold in the following cases, and any one of them can trigger a thermal runaway warning:

[0047] a) Improved variance S of single cell voltage E In this embodiment, the improved variance S of the single cell voltage E The threshold value is set to 1.0, that is, S E >1.0, it is judged that the power battery has the risk of thermal runaway.

[0048] b) Improved variance S of probe temperature T In this embodiment, the improved variance S of the probe temperature T The threshold value is set to 0.5, that is, S T >0.5, it is judged that the power battery has the risk of thermal runaway.

[0049] c) Average extreme entropy of single cell voltage In this embodiment, the average extreme entropy of the single cell voltage The threshold value is set to 1.1, that is, at this time Determine whether the power battery has a risk of thermal runaway.

[0050] d) Average extreme entropy of probe temperature In this embodiment, the average extreme entropy of the probe temperature is The threshold value is set to 1.05, that is, at this time Determine whether the power battery has a risk of thermal runaway.

[0051] It should be noted that the above numerical values ​​such as 1.0, 0.5, 1.1, and 1.05 are all examples and are not intended to limit the protection scope of the present invention.

[0052] The present invention uses a convolutional neural network algorithm to improve the accuracy of battery data prediction. The CLSTM algorithm used can simultaneously process the time domain and frequency domain features of time series data through convolution operations. For the non-stationary characteristics of signals such as battery voltage and probe temperature, CLSTM can well capture signal changes. Moreover, CLSTM can be constructed into a deep network structure, and each layer can learn abstract features of different levels, which enables CLSTM to learn more complex patterns from the original data, making the algorithm more sensitive to subtle changes in battery performance. At the same time, the present invention uses the improved variance and average extreme entropy of the power battery cell voltage and probe temperature in historical data to set the threshold of thermal runaway warning, and calculates the real-time single cell voltage prediction value and probe temperature prediction value, and compares them with the set threshold value to determine whether the thermal runaway warning condition is met. This method combining machine learning and mechanism analysis can accurately determine whether the power battery will have thermal runaway in the future for a long period of time. It is accurate and efficient and has certain application prospects.

[0053] Those skilled in the art can understand that the above are only preferred examples of the invention and are not intended to limit the invention. Although the invention is described in detail with reference to the above examples, those skilled in the art can still modify the technical solutions recorded in the above examples or replace some of the technical features therein with equivalents. Any modification, equivalent replacement, etc. made within the spirit and principle of the invention shall be included in the protection scope of the invention.

Claims

1. A battery thermal runaway warning method based on deep convolutional long short-term memory network, characterized in that: The following steps are involved: S1: Acquire historical data of the power battery from the data platform and preprocess it; the historical data includes battery data field data; S2: selecting battery cell voltage and probe temperature data from the preprocessed data, extracting statistical features of the two selected data sets respectively, and extracting time-frequency features of the two selected data sets respectively by using time-frequency analysis technology; S3: A prediction model is constructed and trained using a deep convolutional long short-term memory network. The input data for training is the preprocessed battery data field data obtained in S1, as well as the statistical characteristics and time-frequency characteristics of the battery cell voltage and probe temperature data obtained in S2. The trained prediction model is used to predict the battery cell voltage and probe temperature. S4: acquiring real-time data of the power battery, the real-time data including battery data field data; inputting the real-time data into the trained prediction model to obtain a single cell voltage prediction value and a probe temperature prediction value within a prediction time window; S5: Calculate the improved variance and average extreme entropy of the predicted value of the single cell voltage and the predicted value of the probe temperature respectively, and compare the obtained improved variance of the single cell voltage, the average extreme entropy of the single cell voltage, the improved variance of the probe temperature, and the average extreme entropy of the probe temperature with their set thresholds respectively. If the value of any one of them is greater than the set threshold, it is determined that the battery has a thermal runaway risk and a thermal runaway warning is issued; the set threshold is obtained by statistical analysis of historical data.

2. The battery thermal runaway early warning method based on deep convolutional long short-term memory network according to claim 1 is characterized in that: In S5, the improved variance S 2 The calculation expression is as follows: Where M is the total number of battery cells in the battery pack, avg is the average of the M battery cell data sequences K, the battery cell data sequence K represents the battery cell voltage E or the probe temperature T, where K(1) represents the first battery cell data sequence; K(i,j) represents the i-th battery cell data sequence at time j, i = 1, 2, ..., M, j = 1, 2, ..., N, and N is the prediction duration corresponding to the prediction time window.

3. The battery thermal runaway early warning method based on deep convolutional long short-term memory network according to claim 1 is characterized in that: In S5, the average extreme entropy The calculation expression is as follows: Where N is the prediction duration corresponding to the prediction time window, K max Indicates the maximum value in the battery cell data sequence at time t, K min Indicates the minimum value in the battery cell data sequence at time t.

4. The battery thermal runaway early warning method based on deep convolutional long short-term memory network according to claim 1 is characterized in that: In S5, the threshold is set by calculating the improved variance and average extreme entropy of the battery cell voltage and the improved variance and average extreme entropy of the probe temperature in the historical data, analyzing the normal battery data and thermal runaway data through a box plot, and converting the improved variance S of the battery cell voltage into E The upper quartile value of the box plot analysis is taken as S E The setting threshold, the average extreme entropy of the single battery voltage The upper quartile value of the box plot analysis is given as The set threshold value, the probe temperature improvement variance S T The upper quartile value of the box plot analysis is taken as S T The set threshold value, the probe temperature average extreme entropy The upper quartile value of the box plot analysis is given as The setting threshold.

5. The battery thermal runaway early warning method based on deep convolutional long short-term memory network according to claim 1 is characterized in that: In S1, the historical data of the power battery includes: battery cell voltage, probe temperature, battery pack voltage, current, SOC, SOH, and mileage.

6. The battery thermal runaway early warning method based on deep convolutional long short-term memory network according to claim 1 is characterized in that: In S1, the preprocessing method is specifically as follows: first, data cleaning is performed, then data normalization is performed, and finally, standardization is performed.

7. The battery thermal runaway early warning method based on deep convolutional long short-term memory network according to claim 1 is characterized in that: In S2, the statistical characteristics of the pre-processed battery cell voltage include the mean, standard deviation, variance, maximum value, and minimum value of the voltage; The statistical characteristics of the preprocessed temperature data include the mean, standard deviation, variance, maximum value, and minimum value of the temperature.

8. The battery thermal runaway early warning method based on deep convolutional long short-term memory network according to claim 1 is characterized in that: In S2, the time-frequency analysis techniques include: short-time Fourier transform, wavelet transform, empirical mode decomposition, and ensemble empirical mode decomposition.

9. The battery thermal runaway early warning method based on deep convolutional long short-term memory network according to claim 1 is characterized in that: In S3, the CLSTM model includes: a convolutional layer, an LSTM layer, and a fully connected layer; the convolutional layer is used to capture local dependencies, and the LSTM layer is used to perform sequence prediction.

Images