Method for improving rapid detection precision of health state of retired battery

Abstract

The invention discloses a method for improving the rapid detection precision of the health state of an out-of-service battery. The method comprises the following steps: testing a battery to obtain training verification data; processing data; a # imgabs0 # estimation model is pre-trained; training an abnormal data feature mode recognition module; identifying abnormal data; formally training a # imgabs1 # estimation model; estimating a battery # imgabs2 # to be detected; by introducing an abnormal data characteristic mode recognition and optimization model training strategy, the overall reliability and estimation precision of the battery health state estimation model in the face of abnormal data characteristic modes and unbalanced health state distribution caused by different aging paths are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery health status detection, and in particular relates to a method for improving the accuracy of rapid detection of the health status of retired batteries. Background Art

[0002] In recent years, the number of retired electric vehicle power batteries has continued to grow, and their environmentally friendly disposal and maximizing their residual value have become critical issues. Cascaded utilization is an effective approach to addressing this issue, with sorting being a particularly critical step. Currently, sorting primarily relies on capacity calibration testing to assess battery health, but this is inefficient. To improve sorting efficiency, the industry is beginning to develop battery health detection methods that incorporate machine learning technology to efficiently detect retired batteries.

[0003] The rapid detection method of battery health status combined with machine learning technology faces many challenges in practical applications. First, since different batteries are affected by factors such as user usage habits and environmental conditions during service, their aging paths are significantly different, resulting in data feature patterns of some batteries that are significantly different from other batteries. This difference makes it difficult for detection methods based on data-driven and machine learning to accurately identify, which in turn results in large detection errors. Secondly, the health status distribution of retired batteries usually shows a concentrated trend, that is, batteries with severe aging and mild aging account for a small proportion, while batteries with medium health account for a high proportion. This distribution characteristic causes the trained battery health status estimation model to have a high estimation accuracy for batteries in medium health, but a low estimation accuracy for batteries with severe or mild aging, significantly reducing the reliability and accuracy of the model.

[0004] The applicant previously applied for an invention patent entitled "A method for rapid detection of the health status of retired power lithium batteries", which discloses a method for rapid detection of the health status of retired batteries based on machine learning, which uses short-time charging test data as model input to achieve rapid estimation of the health status. However, in actual applications, it was found that this method still has limitations: for batteries with significantly different data feature patterns and batteries with severe or mild aging that account for a small proportion in the health status distribution, the estimation accuracy is low. In response to the above problems, the present invention proposes an improved method, which significantly improves the overall reliability and estimation accuracy of the battery health status estimation model when facing abnormal data feature patterns and uneven health status distribution caused by different aging paths by introducing abnormal data feature pattern recognition and optimizing model training strategies. Summary of the Invention

[0005] In response to the problems existing in the existing technology, the present invention provides a method for improving the accuracy of rapid detection of the health status of retired batteries, which significantly improves the overall reliability and estimation accuracy of the battery health status estimation model when facing abnormal data characteristic patterns and uneven health status distribution caused by different aging paths.

[0006] The technical solution of the present invention is achieved as follows:

[0007] A method for improving the accuracy of rapid detection of the health status of retired batteries includes the following steps:

[0008] S100: test battery to obtain training and verification data;

[0009] S200: data processing;

[0010] S300: Pre-training Estimating models;

[0011] S400: training an abnormal data feature pattern recognition module;

[0012] S500: abnormal data identification;

[0013] S600: Formal Training Estimating models;

[0014] S700: Battery to be tested Estimate, obtain The estimated value is used as a quick test result for the health status of the battery.

[0015] Furthermore, in S100, the test battery obtains training verification data, specifically including: using a battery capacity distribution cabinet to perform capacity calibration tests on 5-50% of the test batteries of the same batch and model and record the data;

[0016] Specifically, the capacity calibration test includes: using a battery capacity distribution cabinet, performing constant current discharge at a rate of 0.25C-1C, discharging to the lower limit of the battery voltage, standing for 10 minutes, performing constant current charging at a rate of 0.25C-1C, charging to the upper limit of the battery voltage, then charging at a constant voltage rate of 0.01C-0.05C to the cut-off current, standing for 10 minutes, and performing constant current discharge at a rate of 0.25C-1C, discharging to the lower limit of the battery voltage.

[0017] Specifically, the recorded data includes: changes in battery voltage, current, and capacity over time during the test.

[0018] Furthermore, in S200, the data processing specifically includes:

[0019] S210: Generate Estimate the model training and validation dataset labels. In the capacity calibration test, the final discharge capacity is used as the current discharge capacity of the battery. , according to the battery health status The capacity definition of the battery is used to calculate the , the formula is:

[0020] in is the current discharge capacity of the battery, is the rated capacity of the battery, calculated using As Estimate model training and validation dataset labels.

[0021] S220: Generate Estimate the input features of the model training and validation datasets and set the cutoff charging time , from the capacity calibration test data, intercept the time from the start of charging to The constant current charging stage data is used as the original data, and the original data includes voltage, current, capacity and time; is the minimum value of the voltage sequence in the intercepted data, The maximum value of the voltage sequence in the intercepted data; set the number of resampling points , based on the battery voltage upper limit and lower limit As the boundary, generate uniformly distributed resampled voltage values; according to the original data, , ] range, and calculate the corresponding current, capacity and time values ​​by interpolation. , ]and[ , ] range, the current, capacity and time values ​​are all set to 0, and the voltage range is [ , The resampled data in ] is a ternary sequence including current, capacity and time; the shape of the sequence is ( , as Estimate the input features of the model training and validation datasets; the input features and The labels of the estimation model correspond one to one; the above input features and their corresponding Estimate labels, divide them into training set: validation set = 4:1, and generate The training and validation datasets on which the model is estimated.

[0022] Furthermore, in S300, pre-training The estimation model specifically includes: using the training validation data set obtained in step S220 to The estimation model is pre-trained for multiple rounds; the error of each battery data sample in each round of pre-training is recorded; after the pre-training is completed, the absolute error (AE) of each sample is greater than a certain threshold. Frequency ;mark The data samples are samples of abnormal data characteristic patterns. 5%-8%, is 1 or 2.

[0023] Specifically, The estimation model is a regression model, which can be a convolutional neural network (CNN) model, a recurrent neural network (RNN) model, or other models for solving regression problems;

[0024] Specifically, the model input is the input feature obtained in step S200, and the output is Estimated value.

[0025] Specifically, the The estimation model is a regression model; specifically, it is an Inception-ResNet hybrid model based on causal dilated convolution. The specific structure includes: input layer, mask layer, multi-layer causal dilated convolution Inception-ResNet module and output layer;

[0026] The Inception-ResNet hybrid model combines the Inception module and the residual network (ResidualNetwork, ResNet);

[0027] Classification models and The estimation model structure is basically the same, only the output layer is different. The output layer of the classification model uses the sigmoid activation function to obtain a probability value between 0 and 1, and then outputs 0 or 1 according to the classification threshold of 0.5. The output layer of the estimation model uses the ReLu activation function to output a numerical value, namely estimated value.

[0028] Furthermore, in S400, the abnormal data feature pattern recognition module is a classification model with a dual-branch structure, specifically including:

[0029] Branch 1: The input data passes through a deep convolutional neural network, and through multi-layer convolution and pooling operations, six time series features are extracted. These features can reflect the dynamic changes and complex patterns of the data.

[0030] Branch 2: The input data passes through the curve feature recognition branch to extract the features of the three variable curves in the input data: the platform midpoint position and platform width of the current curve; the peak position and the slope on the left side of the peak of the capacity curve; the peak position and the slope on the left side of the peak of the time curve.

[0031] Feature merging: The six time series features extracted from branch one and the six curve features extracted from branch two are merged to form a merged feature vector containing 12 features, which is input into a deep neural network for further processing to capture higher-level abstract features and patterns.

[0032] Output layer: The features processed by the deep neural network are passed to the output layer. The output layer uses the sigmoid activation function to map the results to a probability value ranging from 0 to 1. According to the preset classification threshold of 0.5, the model converts the probability value into a binary classification result of 0 or 1, where 0 represents a normal data feature pattern sample and 1 represents an abnormal data feature pattern sample.

[0033] Specifically, during the training process, the input data processed in step S220 is used as a feature, and the labeling result of the sample corresponding to the input data in step S300 is used as a label, wherein the label value of the normal data feature pattern sample is 0, and the label value of the abnormal data feature pattern sample is 1.

[0034] Specifically, during the training process, binary cross entropy is used as the loss function.

[0035] Furthermore, in the above-mentioned S500, abnormal data is identified by using the abnormal data characteristic pattern recognition module trained in step S400 to filter the training verification data set generated after the data processing in step S200, and remove samples identified as abnormal data characteristic patterns.

[0036] Furthermore, in step S600, the training validation data set from which the abnormal data characteristic pattern samples identified in step S500 are removed is used to formally train Estimating models;

[0037] Specific, formal training Before estimating the model, perform data balancing and data enhancement on the training and validation datasets: , divide the data set into several equal intervals Interval; for a large number of samples interval, the undersampling method is used to randomly remove some samples; for the small number of samples In the interval, a random oversampling method is used to increase the number of samples by duplicating existing samples, and in the oversampling process, random Gaussian noise is added to the features and labels to enhance the diversity of the data.

[0038] Furthermore, in step S700, the test data of other batteries to be tested of the same batch and model are obtained and input into the data trained in step S600. Estimate the model and obtain the battery to be tested Estimated values, including:

[0039] Test the battery to be tested: Use a battery capacity distribution cabinet to perform constant current discharge at a rate of 0.25C-1C, discharge to the lower limit of the battery voltage, let it stand for 10 minutes, and then perform a short-term discharge at a rate of 0.25C-1C. Constant current charging; 30 seconds - 600 seconds;

[0040] Processing the test data of the battery to be tested: According to the step of generating the input features of the training and verification set in step S220, the test data of the battery to be tested is processed to obtain input data with the same form as the input features of the training and verification set;

[0041] Estimated battery to be tested : Input the processed input data into the trained data in step S600 Estimate the model and obtain the battery to be tested The estimated value is used as a quick test result for the health status of the battery.

[0042] Compared with the prior art, the present invention achieves the following beneficial effects:

[0043] (1) The present invention uses the frequency of samples with large errors to mark out the samples that are not easily identified through the pre-training process. The estimation model accurately estimates the characteristic pattern samples of abnormal data; this method can effectively identify subtle feature differences that are difficult to distinguish manually.

[0044] (2) The present invention adopts an abnormal data feature pattern recognition module with a dual-branch structure, extracts 6 feature values ​​respectively through deep neural network and curve feature recognition, and comprehensively distinguishes the feature patterns of normal data and abnormal data. Accurately identify and eliminate the battery to be tested that is difficult to accurately estimate samples, thus significantly improving Estimate the reliability and precision of the model.

[0045] (3) In the process of model training, the present invention effectively improves the adverse effects of the concentrated distribution of retired battery health status through data balancing and data enhancement. The estimation accuracy of the interval battery is further improved Estimate the reliability and precision of the overall estimation results of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 This is a flow chart of a method for improving the accuracy of rapid detection of the health status of retired batteries provided in Example 1 of the present invention;

[0047] Figure 2 This is a schematic diagram of the model structure of the abnormal data feature pattern recognition module in the method for improving the accuracy of rapid detection of the health status of retired batteries provided in Example 1 of the present invention;

[0048] Figure 3 The battery is provided in a method for improving the accuracy of rapid detection of the health status of retired batteries provided in Example 1 of the present invention. Schematic diagram of the estimation error situation;

[0049] Figure 4 The battery in the method for quickly detecting the health status of retired batteries provided in Comparative Example 1 of the present invention is Schematic diagram of the estimation error situation; DETAILED DESCRIPTION

[0050] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0051] Example 1

[0052] like Figures 1 to 3 A method for improving the accuracy of rapid detection of the health status of retired batteries includes the following steps:

[0053] S100: Test battery to obtain training and verification data, including: obtaining a batch of brand A lithium iron phosphate square batteries, battery rated capacity It is 86Ah.

[0054] Use a battery capacity classification cabinet to perform capacity calibration tests on 20% of the batteries in the same batch and model of brand A lithium iron phosphate square batteries and record the data; that is, the number of batteries undergoing capacity calibration tests accounts for 20% of the total number of batteries in the batch.

[0055] The capacity calibration test includes: using a battery capacity distribution cabinet, performing constant current discharge at a rate of 0.3C to the lower limit of the battery voltage, standing for 10 minutes, performing constant current charging at a rate of 0.3C to the upper limit of the battery voltage, then using a constant voltage charge at a rate of 0.01C to the cutoff current, standing for 10 minutes, and performing constant current discharge at a rate of 0.3C to the lower limit of the battery voltage;

[0056] The recorded data includes: the changes in battery voltage, current and capacity over time during the test.

[0057] S200: Data processing, specifically including:

[0058] S210: Generate Estimate the model training and validation dataset labels. In the capacity calibration test, the final discharge capacity is used as the current discharge capacity of the battery. , according to the battery health status The capacity definition of the battery is used to calculate the , the formula is:

[0059] in is the current discharge capacity of the battery, is the rated capacity of the battery, calculated using As Estimate the labels of the model training and validation datasets;

[0060] S220: Generate Estimate the input features of the model training and validation datasets and set the cutoff charging time , from the capacity calibration test data, intercept the time from the start of charging to The constant current charging stage data (voltage, current, capacity and time) is used as the original data; is the minimum value of the voltage sequence in the intercepted data, The maximum value of the voltage sequence in the intercepted data; set the number of resampling points , based on the battery voltage upper limit and lower limit As the boundary, generate uniformly distributed resampled voltage values; according to the original data, , ] range, and calculate the corresponding current, capacity and time values ​​by interpolation. , )and( , ] range, the current, capacity and time values ​​are all set to 0, and the voltage range is [ , The resampled data in ] is a ternary sequence including current, capacity and time; the shape of the sequence is ( , as Estimate the input features of the model training and validation datasets; the input features and The labels of the estimation model correspond one to one; the above input features and their corresponding Estimate labels, divide them into training set and validation set according to a certain ratio, and generate The training and validation datasets on which the model is estimated.

[0061] In this embodiment, the number of resampling points in step S220 is It is 271.

[0062] S300: Pre-training The estimation model specifically includes: using the training validation data set obtained in step S220 to The estimation model is pre-trained for multiple rounds; the error of each battery data sample in each round of pre-training is recorded; after the pre-training is completed, the absolute error (AE) of each sample is greater than the threshold. Frequency ;mark The data samples are samples of abnormal data characteristic patterns;

[0063] The estimation model is a regression model, which can be a Convolutional Neural Network (CNN) model, a Recurrent Neural Network (RNN) model, or other models for solving regression problems;

[0064] The model input is the input features processed in step S200, and the output is Estimated value.

[0065] In this embodiment, in step S300 is 5%, is 1; the used The estimation model is an Inception-ResNet hybrid model based on causal dilated convolution. The specific structure includes: input layer, mask layer, multi-layer causal dilated convolution Inception-ResNet module and output layer;

[0066] The Inception-ResNet hybrid model combines the Inception module and the residual network (ResidualNetwork, ResNet);

[0067] Classification models and The estimation model structure is basically the same, only the output layer is different. The output layer of the classification model uses the sigmoid activation function to obtain a probability value between 0 and 1, and then outputs 0 or 1 according to the classification threshold of 0.5. The output layer of the estimation model uses the ReLu activation function to output a numerical value, namely estimated value.

[0068] S400: training abnormal data feature pattern recognition module; Figure 2 As shown in Figure 2, the abnormal data feature pattern recognition module is a classification model with a dual-branch structure, specifically including:

[0069] Branch 1: The input data passes through a deep convolutional neural network, which extracts six time series features through multi-layer convolution and pooling operations. These features can reflect the dynamic changes and complex patterns of the data.

[0070] Branch 2: The input data passes through the curve feature recognition branch to extract the features of the three variable curves in the input data: the platform midpoint position and platform width of the current curve; the peak position and left slope of the peak of the capacity curve; the peak position and left slope of the peak of the time curve;

[0071] Feature merging: The six time series features extracted from branch one and the six curve features extracted from branch two are merged to form a joint feature vector containing 12 features. This vector is then input into a deep neural network for further processing to capture higher-level abstract features and patterns.

[0072] Output layer: The features processed by the deep neural network are passed to the output layer. The output layer uses the sigmoid activation function to map the results to a probability value ranging from 0 to 1. According to the preset classification threshold of 0.5, the model converts the probability value into a binary classification result of 0 or 1, where 0 represents a normal data feature pattern sample and 1 represents an abnormal data feature pattern sample.

[0073] During the training process, the input data processed in step S220 is used as a feature, and the labeling result of the sample corresponding to the input data in step S300 is used as a label, wherein the label value of the normal data feature pattern sample is 0, and the label value of the abnormal data feature pattern sample is 1;

[0074] During training, binary cross entropy is used as the loss function.

[0075] S500: Abnormal data identification, using the abnormal data feature pattern identification module trained in step S400 to filter the training and verification data set generated after the data processing in step S200, and eliminate samples identified as abnormal data feature patterns.

[0076] S600: Formal Training The estimation model is trained using the training validation data set from which the abnormal data feature pattern samples identified in step S500 are removed. Estimating models;

[0077] Formal training Before estimating the model, perform data balancing and data enhancement on the training and validation datasets: , divide the data set into several equal intervals Interval; for a large number of samples interval, the undersampling method is used to randomly remove some samples; for the small number of samples In the interval, a random oversampling method is used to increase the number of samples by duplicating existing samples, and in the oversampling process, random Gaussian noise is added to the features and labels to enhance the diversity of the data.

[0078] In this embodiment, in step S600, Divided into 5 equal intervals from 80% to 105% interval.

[0079] S700: Battery to be tested Estimate; obtain other battery test data to be tested for brand A lithium iron phosphate square battery, and input the trained data in step S600 Estimate the model and obtain the battery to be tested Estimated values, including:

[0080] Test the battery to be tested: Use a battery capacity distribution cabinet to perform constant current discharge at a rate of 0.3C until the battery voltage reaches the lower limit, let it stand for 10 minutes, and then perform a short-term charge at a rate of 0.3C. Constant current charging;

[0081] Processing the test data of the battery to be tested: According to the step of generating the input features of the training and verification set in step S220, the test data of the battery to be tested is processed to obtain input data with the same form as the input features of the training and verification set;

[0082] Estimated battery to be tested : Input the processed input data into the trained data in step S600 Estimate the model and obtain the battery to be tested Estimated value.

[0083] In this embodiment, in step S700, the short-time constant current charging time 600S.

[0084] In this embodiment, the final A brand lithium iron phosphate square battery is the battery used to obtain the training verification data in step S100. The estimated mean absolute error is 1.65%, and the mean relative error is 1.70%; Figure 3 , showing the error of each estimated sample in this embodiment.

[0085] The results show that the present invention provides a method for improving the accuracy of rapid detection of the health status of retired batteries, which significantly improves the overall reliability and estimation accuracy of the battery health status estimation model when facing abnormal data characteristic patterns and uneven health status distribution caused by different aging paths.

[0086] Example 2

[0087] In this embodiment, in step S100, a batch of B brand lithium iron phosphate square batteries are used, and the battery rated capacity is The ratio of batteries used for capacity calibration test and data recording is 5% of the total number of batteries; the constant current charge and constant current discharge rates in the capacity calibration test are both 0.25C, and the constant voltage charge cut-off current rate is 0.03C. is 271. In step S300 is 6%, In step S700, the short-time constant current charging time is The remaining steps are the same as those in Example 1, and the same steps will not be repeated here.

[0088] Finally, except for the battery used to obtain training verification data in step S100, the rest of the battery of brand B lithium iron phosphate square battery The estimated mean absolute error is 1.23% and the mean relative error is 1.62%.

[0089] Example 3

[0090] In this embodiment, in step S100, a batch of brand C ternary 18650 batteries are used, with a rated capacity of The ratio of batteries used for capacity calibration test and data recording is 50% of the total number of batteries; the constant current charge and constant current discharge rates in the capacity calibration test are both 1C, and the constant voltage charge cut-off current rate is 0.05C. is 241. In step S300 is 8%, In step S700, the short-time constant current charging time is The remaining steps are the same as those in Example 1, and the same steps will not be repeated here.

[0091] Finally, except for the battery used to obtain training verification data in step S100, the rest of the batteries of brand C are The estimated mean absolute error is 0.78% and the mean relative error is 0.81%.

[0092] Comparative Example 1

[0093] In this comparative example 1, steps S300, S400 and S500 are not performed; the remaining steps are the same as those in Example 1, and the same steps are not repeated here.

[0094] In this comparative example, the remaining batteries of the same batch and the same model, except for the batteries used to obtain training verification data in step S100 (accounting for 20% of the total number of batteries in the batch), The estimated mean absolute error is 2.10%, and the mean relative error is 2.15%; Figure 4 , showing the error of each estimated sample in this comparison.

[0095] It can be seen that compared with Comparative Example 1, Example 1 has higher overall reliability and accuracy. The estimation is less affected by the distribution characteristics.

[0096] Based on the disclosure and teachings of the above description, those skilled in the art may also make changes and modifications to the above embodiments. Therefore, the present invention is not limited to the specific embodiments disclosed and described above, and modifications and variations of the present invention should also fall within the scope of protection of the claims of the present invention. In addition, although certain specific terms are used in this description, these terms are only for convenience of description and do not constitute any limitation to the present invention.

Claims

1. A method for improving the accuracy of rapid detection of the health status of retired batteries, characterized in that: The following steps are involved: S100: test battery to obtain training and verification data; S200: data processing; S300: Pre-training The estimation model specifically includes: using the training validation data set obtained in step S200 to The estimation model is pre-trained for multiple rounds; the error of each battery data sample in each round of pre-training is recorded; after the pre-training is completed, the absolute error of each sample is counted to be greater than the predetermined threshold Frequency ;mark The data sample is a sample of abnormal data characteristic pattern. 5%-8%, is 1 or 2; The estimated model is a regression model; The model input is the input features processed in step S200, and the output is estimated value; S400: Training an abnormal data feature pattern recognition module. The abnormal data feature pattern recognition module is a classification model with a dual-branch structure, specifically including: Branch 1: The input data passes through a deep convolutional neural network, which extracts six time series features through multi-layer convolution and pooling operations. These features can reflect the dynamic changes and complex patterns of the data. Branch 2: The input data passes through the curve feature recognition branch to extract the features of the three variable curves in the input data: the platform midpoint position and platform width of the current curve; the peak position and left slope of the peak of the capacity curve; the peak position and left slope of the peak of the time curve; Feature merging: The six time series features extracted from branch one and the six curve features extracted from branch two are merged to form a merged feature vector containing 12 features. This vector is then input into a deep neural network for further processing to capture higher-level abstract features and patterns. Output layer: The features processed by the deep neural network are passed to the output layer. The output layer uses the sigmoid activation function to map the results to a probability value between 0 and 1. Based on the preset classification threshold of 0.5, the model converts the probability value into a binary classification result of 0 or 1, where 0 represents a normal data feature pattern sample and 1 represents an abnormal data feature pattern sample; During the training process, the input data processed in step S220 is used as features, and the labeling results of the samples corresponding to the input data in step S300 are used as labels, where the label value of the normal data feature pattern sample is 0, and the label value of the abnormal data feature pattern sample is 1; during the training process, binary cross entropy is used as the loss function; S500: abnormal data identification; S600: Formal Training Estimating models; S700: Battery to be tested Estimate, obtain The estimated value is used as a quick test result for the health status of the battery.

2. A method for improving the accuracy of rapid detection of the health status of retired batteries according to claim 1, characterized in that: In S100, the test battery obtains training and verification data, specifically including: using a battery capacity distribution cabinet to perform capacity calibration tests on 5-50% of the test batteries of the same batch and model and record the data; The capacity calibration test includes: using a battery capacity distribution cabinet, performing constant current discharge at a rate of 0.25C-1C to discharge to the lower limit of the battery voltage, standing for 10 minutes, performing constant current charging at a rate of 0.25C-1C to charge to the upper limit of the battery voltage, then charging at a rate of 0.01C-0.05C to the cutoff current, standing for 10 minutes, and performing constant current discharge at a rate of 0.25C-1C to discharge to the lower limit of the battery voltage; The recorded data includes: the changes in battery voltage, current and capacity over time during the test.

3. The method for improving the accuracy of rapid detection of the health status of retired batteries according to claim 1, characterized in that: In S200, the data processing specifically includes: S210: Generate Estimate the model training and validation dataset labels. In the capacity calibration test, the final discharge capacity is used as the current discharge capacity of the battery. , according to the battery health status The capacity definition of the battery is used to calculate the , the formula is: ,in is the current discharge capacity of the battery, is the rated capacity of the battery, calculated using As Estimate the labels of the model training and validation datasets; S220: Generate Estimate the input features of the model training and validation datasets and set the cutoff charging time , from the capacity calibration test data, intercept the time from the start of charging to The constant current charging stage data is used as the original data, and the original data includes voltage, current, capacity and time; is the minimum value of the voltage sequence in the intercepted data, The maximum value of the voltage sequence in the intercepted data; set the number of resampling points , based on the battery voltage upper limit and lower limit As the boundary, generate uniformly distributed resampled voltage values; according to the original data, , ] range, and calculate the corresponding current, capacity and time values ​​by interpolation. , ]and[ , ] range, the current, capacity and time values ​​are all set to 0, and the voltage range is [ , The resampled data in ] is a ternary sequence including current, capacity and time; the shape of the sequence is ( , as Estimate the input features of the model training and validation datasets; the input features and The labels of the estimation model correspond one to one; the above input features and their corresponding Estimate labels, divide them into training set: validation set = 4:1, and generate The training and validation datasets on which the model is estimated.

4. The method for improving the accuracy of rapid detection of the health status of retired batteries according to claim 1, characterized in that: In the S300, the The estimation model is a regression model; specifically, it is an Inception-ResNet hybrid model based on causal dilated convolution. The specific structure includes: input layer, mask layer, multi-layer causal dilated convolution Inception-ResNet module and output layer; The Inception-ResNet hybrid model combines the Inception module and the Residual Network (ResNet); Classification models and The estimation model structure is basically the same, only the output layer is different. The output layer of the classification model uses the sigmoid activation function to obtain a probability value between 0 and 1, and then outputs 0 or 1 according to the classification threshold of 0.

5. The output layer of the estimation model uses the ReLu activation function to output a numerical value, namely estimated value.

5. The method for improving the accuracy of rapid detection of the health status of retired batteries according to claim 1, characterized in that: In the above-mentioned S500, abnormal data is identified by using the abnormal data characteristic pattern recognition module trained in step S400 to filter the training and verification data set generated after the data processing in step S200, and remove samples identified as abnormal data characteristic patterns.

6. The method for improving the accuracy of rapid detection of the health status of retired batteries according to claim 1, characterized in that: In step S600, the training validation data set is used to remove the abnormal data feature pattern samples identified in step S500, and the formal training Estimating models; Formal training Before estimating the model, perform data balancing and data enhancement on the training and validation datasets: , divide the data set into several equal intervals Interval; for a large number of samples interval, the undersampling method is used to randomly remove some samples; for the small number of samples In the interval, a random oversampling method is used to increase the number of samples by duplicating existing samples, and in the oversampling process, random Gaussian noise is added to the features and labels to enhance the diversity of the data.

7. The method for improving the accuracy of rapid detection of the health status of retired batteries according to claim 1, characterized in that: In step S700, the test data of other batteries to be tested of the same batch and model are obtained and input into the data trained in step S600. Estimate the model and obtain the battery to be tested Estimated values, including: Test the battery to be tested: Use a battery capacity distribution cabinet to perform constant current discharge at a rate of 0.25C-1C, discharge to the lower limit of the battery voltage, let it stand for 10 minutes, and then perform a short-term discharge at a rate of 0.25C-1C. Constant current charging; 30 seconds - 600 seconds; Processing the test data of the battery to be tested: According to the step of generating the input features of the training and verification set in step S220, the test data of the battery to be tested is processed to obtain input data with the same form as the input features of the training and verification set; Estimated battery to be tested : Input the processed input data into the trained data in step S600 Estimate the model and obtain the battery to be tested The estimated value is used as a quick test result for the health status of the battery.

Images