Online vehicle battery integrated learning anomaly recognition method and system based on autoencoder

Abstract

The present invention discloses an abnormality identification method and system based on online vehicle battery integrated learning of autoencoders, including: stratifying and sampling all vehicles according to different mileages, and dividing them into the first layer of vehicle groups, ..., the cth layer of vehicle groups; randomly selecting j% of vehicles from each layer of vehicle groups as training sample vehicles; selecting power battery data of sample vehicles and generating sample data; using the data of each sample vehicle to train a corresponding autoencoder; calculating a threshold for each trained autoencoder; obtaining the autoencoder group of each vehicle group; calculating a result for each vehicle in each layer of vehicle groups according to each autoencoder in the corresponding autoencoder group; re-ranking based on the calculation results; and performing abnormal analysis on the top-ranked vehicles. The present invention stratifies vehicles, and analyzes and judges the vehicles belonging to the layer based on multiple autoencoders, and ranks abnormal vehicles in the front and finds them out by sorting.

Description

Technical Field

[0001] The present invention belongs to the technical field of online vehicle clustering methods based on autoencoder technology, and in particular relates to an online vehicle battery integrated learning anomaly recognition method based on autoencoder. Background Art

[0002] With the rapid development of the new energy industry, the sales of new energy vehicles account for an increasingly higher proportion of the total sales of passenger vehicles, and the number of new energy vehicles in use is also growing rapidly. Generally speaking, the factory status of batteries of the same type and model of vehicles meets the consistency requirements, but in subsequent use, due to differences in user driving habits, usage environment, operating conditions, etc., the batteries of the same type and model of vehicles will have certain differences as the usage time and mileage vary. A full understanding of the status of the vehicle group and the differences between individual batteries of different vehicles can be better used to study user conditions and improve vehicle design indicators.

[0003] In the existing patent literature, there are methods for battery anomaly identification and fault classification based on autoencoders or other clustering technologies. For example: Patent document CN111027625A discloses a battery screening method based on SAE and K-means clustering algorithm, which extracts time and voltage on the voltage-time curve as features, establishes a similarity function and gives a certain threshold criterion to complete the screening of the voltage-time curve, and analyzes the consistency of the sorting results. This method compares the difference characteristics between cells within a single battery, and the role of the autoencoder is to extract the implicit compression features of the voltage data, and the clustering uses the K-means algorithm. Patent document CN109375116A discloses a battery system abnormal battery identification method based on an autoencoder, which proposes to use the autoencoder to mine the differences in the intrinsic information of the battery and judge the consistency of the cells within the battery. Both patents are aimed at judging the differences between different cells within a single battery. Another example is a battery fault classification diagnosis method based on cluster analysis disclosed in patent document CN111126449A, which is also based on the K-means clustering algorithm to find the characteristics of the battery's electrical performance data during operation, using the power battery's voltage difference, temperature, average current, and average vehicle speed as clustering variables to divide the data of different batteries. This method compares the differences between different batteries.

[0004] The above patent documents all use unsupervised learning methods. The first two are aimed at the consistency analysis of a single battery pack, and do not compare the differences between battery packs of the same type and model of vehicles. Although the third one compares the differences between different batteries, it uses the conventional K-means clustering algorithm, which has the defect that the number of cluster centers needs to be manually determined. At the same time, the above methods do not adopt the idea of ​​ensemble learning, but only use one model for calculation. Summary of the invention

[0005] The present invention provides an autoencoder-based online vehicle battery integrated learning anomaly recognition method, which can compare the differences in battery packs between vehicles of the same type and model, and rank and find out abnormal vehicles.

[0006] In a first aspect, the present invention provides an autoencoder-based online vehicle battery integrated learning anomaly recognition method, comprising the following steps:

[0007] Step 1: All vehicles of the same type and model in operation on the market are divided into c layers. In each layer, j% of vehicles are randomly selected as training sample vehicles. The total number of sample vehicles is h. The sample vehicles are numbered v1, v2, v3, v4, …, v h-1 ,v h , the number of sample vehicles in each layer is h1, h2, h3, h4, ..., and the total number of vehicles in each layer is H1, H2, H3, H4, ..., where H1 = h1 / j%, H2 = h2 / j%, and so on;

[0008] Step 2: Select the power battery data of each sample vehicle in the past t months, and generate a total of s sample data of a fixed time length. Each sample is an m×n matrix, where m is the time length and n is the number of features.

[0009] Step 3: Construct h autoencoders. Each autoencoder has the same structure, consisting of several one-dimensional convolutional layers, one-dimensional deconvolutional layers, and fully connected layers, and randomly initializes the weight matrix; the autoencoders are numbered AE1, AE2, AE3, AE4, …, AE h-1 ,AE h ,Corresponding to the car number, each autoencoder is bound to a sample vehicle;

[0010] Step 4: The sample data of each vehicle obtained in step 2 is used as training data and input into the corresponding autoencoder for training the autoencoder;

[0011] Step 5: Use back propagation to train each autoencoder and save each trained autoencoder;

[0012] Step 6: For the vehicle data numbered v1 and the corresponding trained autoencoder AE1, the input is s m×n matrices I, and the output is s m×n matrices O of the same size. Calculate the MAE of the s corresponding inputs s and outputs;

[0013] Step 7: The calculation method of MAE is where i ij ∈I,o ij ∈O;

[0014] Step 8. There are s MAE values ​​calculated in step 7. The largest one is taken as the threshold th1 of the autoencoder AE1.

[0015] Step 9: For the remaining vehicle data and the corresponding trained autoencoders, calculate the corresponding thresholds according to the methods of steps 6 to 8, which are th2, th3, th4, ..., th h-1 ,th h ;

[0016] Step 10: Perform abnormal identification on all vehicles in the first layer of vehicle group H1 in step 1;

[0017] Step 11: Select the trained autoencoders AE1, AE2, ..., AE that belong only to the first layer of vehicles. i-1 ,AE i and the corresponding thresholds th1,th2,…,th i-1 ,th i ;

[0018] Step 12: Sort all the cars by frame number, numbering them as vin1, vin2, ..., vin H1-1 ,vin H1 ;

[0019] Step 13, generating data samples of s m×n matrices I1 for the vehicle numbered vin1 in the manner of step 2;

[0020] Step 14: Send the generated vehicle data samples of vin1 to the autoencoder AE1 for calculation, and the output obtained is s m×n matrices O1;

[0021] Step 15: Calculate the MAE of I1 and O1. There are s MAE values ​​in total, of which a values ​​are greater than the threshold th1 corresponding to the autoencoder AE1, accounting for a percentage of

[0022] Step 16: For other vehicles, i.e., vehicles numbered vin2,…,vin H1-1 ,vin H1 According to the method of steps 13 to 15, the corresponding percentages are obtained, which are p2,…,p H1-1 ,p H1 ;

[0023] Step 17: Arrange all vehicles according to percentages p1, p2, ..., p H1-1 ,p H1 Sort them, the serial numbers are 1, 2, 3, ..., H1;

[0024] Step 18: All vehicles are trained on the remaining autoencoders AE2,…,AE according to the method of steps 12 to 17. i-1 ,AE i Perform calculations;

[0025] Step 19: After all vehicles are calculated, each vehicle has i percentages and ranking numbers. The i percentages are averaged and all vehicles are ranked according to the average.

[0026] Step 20: Focus on monitoring and analyzing the top 1% of vehicles;

[0027] Step 21: For the vehicles in the remaining layers in step 1, follow the method of steps 11 to 20 to complete the analysis and sorting of all online vehicles.

[0028] Optionally, in step 1, all vehicles of the same model in operation on the market are stratified according to different mileages.

[0029] Optionally, in step 2, the current, total voltage, single cell voltage and temperature data of the power battery charging and discharging stages in the past t months are selected for each sample vehicle.

[0030] In the second aspect, the present invention discloses an autoencoder-based online vehicle battery integrated learning anomaly identification system, comprising a memory and a controller, wherein the memory stores a computer-readable program, and when the computer-readable program is called by the controller, it can execute the steps of the autoencoder-based online vehicle battery integrated learning anomaly identification method as described in the present invention.

[0031] The present invention has the following advantages: the present invention divides vehicles into layers, and then analyzes and judges the vehicles belonging to the layer based on multiple autoencoders, and puts abnormal vehicles in the front row and finds them out by sorting. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 Flow chart of the main steps of this embodiment. DETAILED DESCRIPTION

[0033] The present invention will be further described below in conjunction with the accompanying drawings.

[0034] like Figure 1As shown, in this embodiment, a method for identifying anomalies of vehicle battery ensemble learning based on autoencoders online includes: stratifying and sampling all vehicles according to different mileages to divide them into a first-layer vehicle group, ..., a c-th-layer vehicle group; randomly selecting j% of vehicles from each layer of the vehicle group as training sample vehicles; selecting power battery data of the sample vehicles and generating sample data; using the data of each sample vehicle to train a corresponding autoencoder; calculating a threshold for each trained autoencoder; obtaining an autoencoder group for each vehicle group; calculating a result for each vehicle in each layer of the vehicle group according to each autoencoder in the corresponding autoencoder group; re-ranking based on the calculation results; and performing anomaly analysis on the top-ranked vehicles.

[0035] In this embodiment, a method for identifying abnormalities of vehicle battery integrated learning based on an autoencoder online specifically includes the following steps:

[0036] The first part is to build the overall model:

[0037] Step 1: All vehicles of the same type and model in operation on the market are divided into c layers according to different mileage lengths. In each layer, j% of vehicles are randomly selected as training sample vehicles. The total number of sample vehicles is h, and the sample vehicle numbers are v1, v2, v3, v4, …, v h-1 ,v h , the number of sample vehicles in each layer is h1, h2, h3, h4, …, and the total number of vehicles in each layer is H1, H2, H3, H4, …, where H1 = h1 / j%, H2 = h2 / j%, and so on.

[0038] Step 2: Select the power battery data of each sample vehicle in the past t months, the current, total voltage, single cell voltage and temperature data of the charging and discharging segments, and generate a total of s sample data of a fixed time length through sliding windows or other methods. Each sample is an m×n matrix, where m is the time length and n is the feature number. Because the driving habits of each vehicle are different, the number of samples may be inconsistent, but the size of the matrix is ​​consistent.

[0039] Step 3: Construct h autoencoders. Each autoencoder has the same structure, consisting of several one-dimensional convolutional layers, one-dimensional deconvolutional layers, and fully connected layers, and randomly initializes the weight matrix; the autoencoders are numbered AE1, AE2, AE3, AE4, …, AE h-1 ,AE h ,Corresponding to the car number, each autoencoder is bound to a sample vehicle.

[0040] Step 4: The sample data of each vehicle obtained in step 2 is used as training data and input into the corresponding autoencoder for training the autoencoder.

[0041] Step 5: Use back propagation to train each autoencoder and save each trained autoencoder.

[0042] Step 6: For the vehicle data numbered v1 and the corresponding trained autoencoder AE1, the input is s m×n matrices I, and the output is s m×n matrices o of the same size. Calculate the MAE of the s corresponding inputs s and outputs.

[0043] Step 7: The calculation method of MAE is where i ij ∈I,o ij ∈O.

[0044] Step 8. There are s MAE values ​​calculated in step 7. The largest one is taken as the threshold th1 of the autoencoder AE1.

[0045] Step 9: For the remaining vehicle data and the corresponding trained autoencoders, calculate the corresponding thresholds according to the methods of steps 6 to 8, which are th2, th3, th4, ..., th h-1 ,th h .

[0046] For example: Assume that there are 50,000 vehicles in operation on the market, 20,000 vehicles with mileage between 0 and 30,000, and 30,000 vehicles with mileage above 30,000, then they can be divided into two layers, 20,000 vehicles in the first layer and 30,000 vehicles in the second layer. 2% of the vehicles in each layer are selected as sample vehicles, so there are 40 sample vehicles in the first layer and 60 sample vehicles in the second layer. Because the calculation method of each layer is the same, only the vehicles in the first layer are described in this embodiment. Sample data is generated for the 40 vehicles in the first layer according to the method in step 2. Then, an autoencoder is trained based on the sample data of each vehicle, so a total of 40 autoencoders need to be trained. Among the 40 sample vehicles and 40 autoencoders in the first layer, taking the first vehicle as an example, assume that the sample data of the first vehicle is a 1000 m×n matrix I, and the output is also a 1000 m×n matrix O. Calculate 1000 corresponding MAE values, and take the largest MAE value as the threshold of the first autoencoder. The thresholds of the remaining 39 vehicles are calculated in the same way. This completes the construction of the overall model.

[0047] The following is the second part, which is mainly for abnormal identification. The specific steps are as follows:

[0048] Step 10: Perform abnormal identification on all vehicles in the first layer of vehicle group H1 in step 1.

[0049] Step 11: Select the trained autoencoders AE1, AE2, ..., AE that belong only to the first layer of vehicles. i-1 ,AE i and the corresponding thresholds th1,th2,…,th i-1 ,th i .

[0050] Step 12: Sort all the cars by frame number, numbering them as vin1, vin2, ..., vin H1-1 ,vin H1 .

[0051] Step 13: Generate data samples of s m×n matrices I1 for the vehicle numbered vin1 in the manner of step 2.

[0052] Step 14: Send the generated vehicle data samples of vin1 to the autoencoder AE1 for calculation, and the output obtained is s m×n matrices O1.

[0053] Step 15: Calculate the MAE of I1 and O1. There are s MAE values ​​in total, of which a values ​​are greater than the threshold th1 corresponding to the autoencoder AE1, accounting for a percentage of

[0054] Step 16: For other vehicles, i.e., vehicles numbered vin2,…,vin H1-1 ,vin H1 Repeat steps 13 to 15 for the vehicles to obtain the corresponding percentages, which are p2,…,p H1-1 ,p H1 .

[0055] Step 17: Arrange all vehicles according to percentages p1, p2, ..., p H1-1 ,p H1 Sort them and the serial numbers are 1, 2, 3, …, H1.

[0056] Step 18: All vehicle pairs are assigned to the remaining autoencoders AE2,…,AE i-1 ,AE i Perform calculations, i.e. repeat steps 12 to 17.

[0057] Step 19. After all calculations are completed, each vehicle has i percentages and ranking numbers. The i percentages are averaged, and all vehicles are comprehensively ranked based on the average to form a result table.

[0058] Step 20: Focus on monitoring and analyzing the top 1% of vehicles.

[0059] Step 21: For the vehicles in the remaining layers in step 1, follow the method of steps 11 to 20 to complete the analysis and sorting of all online vehicles.

[0060] For example: Let's take the first layer in the above example as an example, 20,000 vehicles, 40 sample vehicles, and 40 trained autoencoders and their thresholds. The 40 autoencoders are numbered AE1, AE2, ..., AE 40 and the corresponding thresholds th1, th2, ...th 40 . Sort the 20,000 vehicles by their chassis numbers, numbered vin1, vin2, …, vin 20000 . Generate data samples for these 20,000 vehicles in the same way as in step 2. Here, we take the vehicle numbered vin1 as an example, assuming that there are 2000 m×n matrices I1. First, send these data samples to the autoencoder AE1 for calculation, and get the output as 2000 m×n matrices O1. Calculate the MAE of I1 and O1, and there are a total of 2000 MAE values. Assume that there are 1800 of them that are greater than the threshold th1 corresponding to the autoencoder AE1, and the percentage is Perform the same calculation on the remaining 19,999 vehicles to get the corresponding percentage. Sort all the vehicles by percentage, and the results are shown in the following table:

[0061]

[0062]

[0063] For all vehicles, the remaining 39 autoencoders AE2,…,AE 40 After all the calculations are completed, each vehicle has 40 percentages and ranking numbers. The average of these 40 percentages is calculated, and all vehicles are ranked comprehensively based on the average, forming the following table:

[0064]

[0065] From this table, we can conclude that the top 20 vehicles are outliers, and we need to focus on monitoring the top 20 vehicles in the future.

Claims

1. A method for identifying abnormalities of vehicle battery ensemble learning based on online autoencoder, characterized in that: The following steps are involved: Step 1: All vehicles of the same type and model in operation on the market are divided into c layers. In each layer, j% of vehicles are randomly selected as training sample vehicles. The total number of sample vehicles is h. The sample vehicles are numbered v1, v2, v3, v4, …, v h-1 ,v h , the number of sample vehicles in each layer is h1, h2, h3, h4, ..., and the total number of vehicles in each layer is H1, H2, H3, H4, ..., where H1 = h1 / j%, H2 = h2 / j%, and so on; Step 2: Select the power battery data of each sample vehicle in the past t months, and generate a total of s sample data of a fixed time length. Each sample is an m×n matrix, where m is the time length and n is the number of features. Step 3: Construct h autoencoders. Each autoencoder has the same structure, consisting of several one-dimensional convolutional layers, one-dimensional deconvolutional layers, and fully connected layers, and randomly initializes the weight matrix; the autoencoders are numbered AE1, AE2, AE3, AE4, …, AE h-1 ,AE h ,Corresponding to the car number, each autoencoder is bound to a sample vehicle; Step 4: The sample data of each vehicle obtained in step 2 is used as training data and input into the corresponding autoencoder for training the autoencoder; Step 5: Use back propagation to train each autoencoder and save each trained autoencoder; Step 6: For the vehicle data numbered v1 and the corresponding trained autoencoder AE1, the input is s m×n matrices I, and the output is s m×n matrices O of the same size. Calculate the MAE of the s corresponding inputs s and outputs; Step 7: The calculation method of MAE is where i ij ∈I,o ij ∈O; Step 8. There are s MAE values ​​calculated in step 7. The largest one is taken as the threshold th1 of the autoencoder AE1. Step 9: For the remaining vehicle data and the corresponding trained autoencoders, calculate the corresponding thresholds according to the methods of steps 6 to 8, which are th2, th3, th4, ..., th h-1 ,th h ; Step 10: Perform abnormal identification on all vehicles in the first layer of vehicle group H1 in step 1; Step 11: Select the trained autoencoders AE1, AE2, ..., AE that belong only to the first layer of vehicles. i-1 ,AE i and the corresponding thresholds th1,th2,…,th i-1 ,th i ; Step 12: Sort all the cars by frame number, numbering them as vin1, vin2, ..., vin H1-1 ,vin H1 ; Step 13, generating data samples of s m×n matrices I1 for the vehicle numbered vin1 in the manner of step 2; Step 14: Send the generated vehicle data samples of vin1 to the autoencoder AE1 for calculation, and the output obtained is s m×n matrices O1; Step 15: Calculate the MAE of I1 and O1. There are s MAE values ​​in total, of which a values ​​are greater than the threshold th1 corresponding to the autoencoder AE1, accounting for a percentage of Step 16: For other vehicles, i.e., vehicles numbered vin2,…,vin H1-1 ,vin H1 According to the method of steps 13 to 15, the corresponding percentages are obtained, which are p2,…,p H1-1 ,p H1 ; Step 17: Arrange all vehicles according to percentages p1, p2, ..., p H1-1 ,p H1 Sort them, the serial numbers are 1, 2, 3, ..., H1; Step 18: All vehicles are trained on the remaining autoencoders AE2,…,AE according to the method of steps 12 to 17. i-1 ,AE i Perform calculations; Step 19: After all vehicles are calculated, each vehicle has i percentages and ranking numbers. The i percentages are averaged and all vehicles are ranked according to the average. Step 20: Focus on monitoring and analyzing the top 1% of vehicles; Step 21: For the vehicles in the remaining layers in step 1, follow the method of steps 11 to 20 to complete the analysis and sorting of all online vehicles.

2. The method for identifying abnormalities of vehicle battery integrated learning based on online autoencoder according to claim 1 is characterized in that: In step 1, all vehicles of the same model in operation on the market are stratified according to different mileage lengths.

3. The method for identifying abnormalities of vehicle battery integrated learning based on online autoencoder according to claim 2 is characterized in that: In step 2, the current, total voltage, single cell voltage and temperature data of the power battery charging and discharging stages in the past t months are selected for each sample vehicle.

4. An autoencoder-based online vehicle battery integrated learning anomaly recognition system, characterized by: It comprises a memory and a controller, wherein the memory stores a computer-readable program, and when the computer-readable program is called by the controller, it can execute the steps of the autoencoder-based online vehicle battery integrated learning anomaly identification method as claimed in any one of claims 1 to 3.

Images