Power battery high-temperature fault degree detection method based on ALSTM-CNN

Abstract

The invention discloses a power battery high-temperature fault degree detection method based on an ALSTM-CNN (Long Short-Term Memory-Convolutional Neural Network) (Attention Mechanism). A certain vehicle enterprise monitoring platform is adopted to obtain original data, data preprocessing is carried out, probe temperature is converted into battery average temperature, and a battery high-temperature fault correlation coefficient is obtained. According to the method, feature extraction is completed by performing noise reduction processing on data, then an ALSTM-CNN model is trained, parameter adjustment is performed based on verification set accuracy, and optimal network parameters are selected after multiple rounds of training. And real-time data are collected, fault degree analysis is carried out on the real-time data, and the fault level of the vehicle is output. The fault levels are totally four levels of 0-3, and 0 represents no fault; '1' represents a first-level fault and refers to a fault which does not affect normal driving of the vehicle; '2' represents a secondary fault and refers to a fault which affects the vehicle performance and needs a driver to limit driving; 3 represents a third-level fault, is a highest-level fault, and refers to a fault that a driver should immediately stop the vehicle for processing or request for rescue.

Description

Technical Field

[0001] The patent of this invention relates to the field of power battery safety control technology, and in particular to a method for detecting the degree of high-temperature fault of a power battery based on ALSTM-CNN. Background Art

[0002] The power battery system is usually composed of actuators, power batteries and components, and sensors. During its production and use, each component may fail, including actuator failure, battery failure, component failure, and sensor failure. Battery failure is the most common fault among all faults, mainly including internal short circuit, external short circuit, overcharge, over discharge, and inhomogeneity failure. As the battery continues to work, these faults eventually manifest as battery high temperature failures until thermal runaway occurs. Compared with other faults, battery high temperature failures are very harmful to battery performance, causing battery capacity loss, accelerating the aging process, increasing battery impedance, and may cause battery overheating, and even cause serious safety problems such as thermal runaway, combustion, and explosion. Therefore, the study of battery high temperature failures is crucial. There are many potential causes of battery high temperature failures. When a failure occurs, multiple parameters such as battery temperature, voltage, and resistance will change. Therefore, the diagnosis of battery high temperature failures requires coupling of multi-dimensional characteristic parameters. When the heat generated by the exothermic reaction inside the battery cannot be released to the environment in a timely and effective manner, once the heat reaches the "critical state", thermal runaway is inevitable. Before the "critical state", various thermal abuses and internal latent defects will cause high-temperature failures in the battery. Therefore, detecting and handling faults at high battery temperatures is one of the effective ways to ensure the safe operation of power batteries. Summary of the invention

[0003] The purpose of the present invention is to provide a method for detecting the degree of high temperature fault of a power battery based on ALSTM-CNN to solve the problems existing in the above-mentioned prior art.

[0004] To achieve the above object, the present invention provides the following solution: The present invention provides a method for detecting the degree of high temperature fault of a power battery based on ALSTM-CNN, and the method steps are as follows: S1. Preprocess the raw data obtained from a certain automobile enterprise monitoring platform, including cleaning duplicate data, processing outliers and missing values; S2. Use the probe temperature to estimate the average temperature of the battery online, and use the average temperature of the battery, the list of cell voltages, temperature consistency, voltage consistency and insulation resistance as the input of the ALSTM-CNN model to predict the battery high temperature fault level; S3. Use Sparrow Search Algorithm (SSA) to optimize the variational mode decomposition (VMD) and determine the modal decomposition number. and penalty factor VMD decomposes the voltage signal into a low-frequency static component and a high-frequency dynamic component. The envelope entropy is selected as the fitness function of SSA optimization. Each SSA calculates the envelope entropy once and selects the one with the lowest envelope entropy. and As the modal component number and penalty factor of VMD decomposition. Multiscale Permutation Entropy (MPE) is introduced as the evaluation basis of fault components and signal components, and the scale factor is selected. =12, embedding dimension m 6. Delay time The value of MPE is 1, and the mean of the permutation entropy of each IMF component is calculated as the final MPE value. The threshold of MPE is 0.6. The IMF components greater than 0.6 can be regarded as useful signal components, and those less than 0.6 are low-frequency signal components. S4. Establish an LSTM neural network, introduce the attention mechanism into the LSTM neural network, and combine it with the convolutional neural network to build a parallel ALSTM-CNN model based on the attention mechanism. This model combines two consecutive LSTM layers and an Attention layer to form an ALSTM module, and two one-dimensional convolutional layers and two pooling layers to form a CNN module. The activation function is Relu, and the features output by the ALSTM layer and the CNN layer are fused in the fully connected layer, and the final fault degree classification is output through the Softmax layer. The noise component screened out in S3, that is, the useful signal component, is used to train the ALSTM-CNN model, and the parameters are adjusted based on the accuracy of the validation set. After multiple rounds of training, the optimal network parameters are selected, among which the parameters that need to be determined include the number of convolutional layers, the number and size of convolution kernels, the number of LSTM units, the number of LSTM hidden layers, etc. The basis for adjusting the parameters is first the accuracy of the validation set training, followed by the time and resource occupation of the model training. Finally, the optimal ALSTM-CNN is selected as the power battery high temperature fault state detection model; S5. By collecting real-time data and inputting it into the ALSTM-CNN power battery high temperature fault state detection model, the power battery high temperature fault state can be predicted, and the fault degree, i.e., the fault level, can be output. Output "0" represents no fault, output "1" represents a first-level fault, output "2" represents a second-level fault, and output "3" represents a third-level fault.

[0005] Preferably, the S1 comprises the following steps: S1-1. Cleaning duplicate data: If there are two adjacent rows with the same data, it is likely due to data collection errors. They are considered duplicate records and the latter row is deleted. S1-2. Abnormal value processing: Abnormal values ​​refer to the obvious abnormal values ​​of battery voltage, current and temperature. For example, the temperature suddenly rises from the normal 25℃ to 214℃, and the single cell voltage suddenly drops from 3.7V to 0V. These may be due to sensor failure, resulting in individual measurement values ​​that are not in line with common sense, so they can be deleted directly. S1-3. Missing value processing: a) Continuous 1 minute sampling data is missing: Since the sampling time is 20 s, when data of 3 or more consecutive sampling points are missing, all data within the missing time can be directly deleted; b) Discontinuous data missing: If data of a certain sampling point is missing, or the continuous time does not exceed 2 sampling points, the adjacent cell average interpolation method is used to interpolate the missing values.

[0006] Preferably, S2 comprises the following steps: S2-1 generates a fixed number of sampling points The voltage and current sequence and ; S2-2 respectively for the voltage sequence , current sequence and Morlet wavelet time domain series Perform a fast Fourier transform; S2-3 Obtaining the wavelet coefficients of voltage and current series and ; S2-4 obtains the time at different times within the time length t At different frequencies Battery impedance under and the corresponding impedance angle; S2-5 queries the offline established table of the relationship between battery impedance and average temperature to obtain the average temperature of the battery at different times ; S2-6 uses a smoothing filter method to smooth and remove outliers on the estimated temperatures at different times; S2-7 repeats steps (S2-1)-(S2-6) to complete the online estimation of the average battery temperature.

[0007] Preferably, S3 comprises the following steps: S3-1. Initialize SSA parameters, such as population size, maximum number of iterations, etc. Set the value range of the modal decomposition number K to [3, 10], and the value range of the penalty factor to [100, 4500] to prevent the parameters from being set too small and affecting the fidelity of signal reconstruction in a noisy environment; S3-2, using VMD to decompose the voltage, temperature, current and other input signals of the power battery to obtain multiple IMF components; S3-3. Calculate each iteration The envelope entropy function value at the corresponding position updates the global optimal solution according to the envelope entropy size. The envelope entropy calculation formula is: In the formula, is the number of sampling points of the signal, is the first IMF components, Indicates The value of the IMF component after Hilbert transformation, is the envelope amplitude after Hilbert transformation, for The normalized form of is the envelope entropy; S3-4, iterate to the maximum number of times and output the optimal combination; S3-5, based on the best Combine the input signal to perform VMD decomposition and get IMF components; S3-6. Calculate the MPE of the IMF component and select the scale factor is 12, the embedding dimension m is 6, and the delay time The MPE threshold is 0.6, and the IMF components greater than 0.6 can be regarded as useful signal components, while those less than 0.6 are low-frequency signal components.

Claims

1. A method for detecting high temperature fault degree of power battery based on ALSTM-CNN, characterized in that: It includes the following steps: S1. Preprocess the raw data obtained from a certain automobile enterprise monitoring platform, including cleaning duplicate data, processing outliers and missing values; S2. Use the probe temperature to estimate the average temperature of the battery online, and use the average temperature of the battery, the list of cell voltages, temperature consistency, voltage consistency and insulation resistance as the input of the ALSTM-CNN model to predict the battery high temperature fault level; S3. Use Sparrow Search Algorithm (SSA) to optimize the variational mode decomposition (VMD) and determine the modal decomposition number. and penalty factor , VMD decomposes the voltage signal into low-frequency static components and high-frequency dynamic components, and selects envelope entropy as the fitness function of SSA optimization. The envelope entropy calculation formula is: In the formula, is the number of sampling points of the signal, is the first IMF components, Indicates The value of the IMF component after Hilbert transformation, is the envelope amplitude after Hilbert transformation, for The normalized form of is the envelope entropy; Each SSA calculates the envelope entropy once and selects the one with the lowest envelope entropy. and As the modal component number and penalty factor of VMD decomposition. Multiscale Permutation Entropy (MPE) is introduced as the evaluation basis of fault components and signal components, and the scale factor is selected. =12, embedding dimension m 6. Delay time The value of MPE is 1, and the mean of the permutation entropy of each IMF component is calculated as the final MPE value. The threshold of MPE is 0.

6. The IMF components greater than 0.6 can be regarded as useful signal components, and those less than 0.6 are low-frequency signal components. S4. Establish an LSTM neural network, introduce the attention mechanism into the LSTM neural network, and combine it with the convolutional neural network to build a parallel ALSTM-CNN model based on the attention mechanism. This model combines two consecutive LSTM layers and an Attention layer to form an ALSTM module, and two one-dimensional convolutional layers and two pooling layers to form a CNN module. The activation function is Relu, and the features output by the ALSTM layer and the CNN layer are fused in the fully connected layer, and the final fault degree classification is output through the Softmax layer. The noise component screened out in S3, that is, the useful signal component, is used to train the ALSTM-CNN model, and the parameters are adjusted based on the accuracy of the validation set. After multiple rounds of training, the optimal network parameters are selected, among which the parameters that need to be determined include the number of convolutional layers, the number and size of convolution kernels, the number of LSTM units, the number of LSTM hidden layers, etc. The basis for adjusting the parameters is first the accuracy of the validation set training, followed by the time and resource occupation of the model training. Finally, the optimal ALSTM-CNN is selected as the power battery high temperature fault state detection model; S5. By collecting real-time data and inputting it into the ALSTM-CNN power battery high temperature fault state detection model, the power battery high temperature fault state can be predicted, and the fault degree, i.e., the fault level, can be output. Output "0" represents no fault, output "1" represents a first-level fault, output "2" represents a second-level fault, and output "3" represents a third-level fault.

2. The method for detecting the degree of high temperature fault of a power battery based on ALSTM-CNN according to claim 1 is characterized in that: The S1 comprises the following steps: S1-1. Cleaning duplicate data: If there are two adjacent rows with the same data, it is likely due to data collection errors. They are considered duplicate records and the latter row is deleted. S1-2. Abnormal value processing: Abnormal values ​​refer to the obvious abnormal values ​​of battery voltage, current and temperature. For example, the temperature suddenly rises from the normal 25℃ to 214℃, and the single cell voltage suddenly drops from 3.7V to 0V. These may be due to sensor failure, resulting in individual measurement values ​​that are not in line with common sense, so they can be deleted directly. S1-3. Missing value processing: a) Continuous 1 minute sampling data is missing: Since the sampling time is 20 s, when data of 3 or more consecutive sampling points are missing, all data within the missing time can be directly deleted; b) Discontinuous data missing: If data of a certain sampling point is missing, or the continuous time does not exceed 2 sampling points, the adjacent cell average interpolation method is used to interpolate the missing values.

3. The method for detecting the degree of high temperature fault of a power battery based on ALSTM-CNN according to claim 1, characterized in that: The S2 comprises the following steps: S2-1 generates a fixed number of sampling points The voltage and current sequence and ; S2-2 respectively for the voltage sequence , current sequence and Morlet wavelet time domain series Perform a fast Fourier transform; S2-3 Obtaining the wavelet coefficients of voltage and current series and ; S2-4 obtains the time at different times within the time length t At different frequencies Battery impedance under and the corresponding impedance angle; S2-5 queries the offline established table of the relationship between battery impedance and average temperature to obtain the average temperature of the battery at different times ; S2-6 uses a smoothing filter method to smooth and remove outliers on the estimated temperatures at different times; S2-7 repeats steps (S2-1)-(S2-6) to complete the online estimation of the average battery temperature.

4. The method for detecting the degree of high temperature fault of a power battery based on ALSTM-CNN according to claim 1, characterized in that: The S3 comprises the following steps: S3-1. Initialize SSA parameters, such as population size, maximum number of iterations, etc. Set the modal decomposition number The value range is [3, 10], setting the penalty factor The value range of is [100, 4500], to prevent the parameter from being set too small and affecting the fidelity of signal reconstruction in a noisy environment; S3-2, using VMD to decompose the voltage, temperature, current and other input signals of the power battery to obtain multiple IMF components; S3-3. Calculate each iteration The envelope entropy function value at the corresponding position, updates the global optimal solution according to the size of the envelope entropy; S3-4, iterate to the maximum number of times and output the optimal combination; S3-5, based on the best Combine and perform VMD decomposition on the input signal to obtain IMF components; S3-6. Calculate the MPE of the IMF component and select the scale factor =12, embedding dimension m 6. Delay time The MPE threshold is 0.6, and the IMF components greater than 0.6 can be regarded as useful signal components, while those less than 0.6 are low-frequency signal components.

5. The method for detecting the degree of high temperature fault of a power battery based on ALSTM-CNN according to claim 1, characterized in that: The S4 content is as follows: An LSTM neural network is established, and the attention mechanism is introduced into the LSTM neural network. Combined with the convolutional neural network, a parallel ALSTM-CNN model based on the attention mechanism is constructed. The model combines two consecutive LSTM layers and an Attention layer to form an ALSTM module, and two one-dimensional convolutional layers and two pooling layers to form a CNN module. The activation function is Relu, and the features output by the ALSTM layer and the CNN layer are fused in the fully connected layer, and the final fault degree classification is output through the Softmax layer. The noise component screened out by S3, that is, the useful signal component, is used to train the ALSTM-CNN model. At the same time, the parameters are adjusted based on the accuracy of the validation set. After multiple rounds of training, the optimal network parameters are selected. The parameters that need to be determined include the number of convolutional layers, the number and size of convolutional kernels, the number of LSTM units, and the number of LSTM hidden layers. The basis for adjusting the parameters is first the accuracy of the validation set training, followed by the time and resource occupation of the model training. Finally, the optimal ALSTM-CNN is selected as the power battery high temperature fault state detection model.

6. The method for detecting the degree of high temperature fault of a power battery based on ALSTM-CNN according to claim 1, characterized in that: The content of S5 is as follows: By collecting real-time data and inputting it into the ALSTM-CNN power battery high temperature fault state detection model, the power battery high temperature fault state can be predicted and the fault degree, i.e., the fault level, can be output. Output "0" represents no fault, output "1" represents a level 1 fault, output "2" represents a level 2 fault, and output "3" represents a level 3 fault.

Images