SOH estimation and self-supervised learning-based SOH prediction method for power battery

Abstract

The invention discloses a power battery SOH estimation and self-supervised learning-based SOH prediction method. SOH estimation and prediction can be carried out according to vehicle condition data originally collected by an electric vehicle. The complete charging process is calibrated by accumulating the rise amount of the SOC; then, power battery health degree feature sampling is completed through statistics and calculation, and a power battery health degree data set is made; then designing a multi-layer encoder neural network architecture based on a frequency domain attention mechanism; performing pre-training by using the power battery health degree data set, learning common characteristics of battery aging, and performing fine adjustment by using independent data in the power battery health degree data set to learn individual characteristics of battery aging; and finally, completing SOH prediction of the power battery.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power batteries, and particularly relates to a power battery SOH estimation and SOH prediction method based on self-supervised learning. BACKGROUND

[0002] In actual use, the discharge and charge of the power battery of the electric vehicle are not completely reversible. The performance of the power battery tends to decline due to factors such as cycle aging and storage aging. The ratio of the current capacity of the power battery to the nominal capacity is often used as the SOH (State of Health) parameter in engineering to represent the health degree of the power battery. The SOH of a brand-new power battery is 100%. When the health degree of the power battery decreases to a certain extent, the power battery cannot provide sufficient power delivery for the electric vehicle, at which time a new power battery needs to be replaced for the electric vehicle, so that the electric vehicle can recover to a stable and efficient state. Therefore, in the current booming development of electric vehicles, it is of great significance to simply and accurately evaluate and predict the SOH of the power battery for the persistent use of the electric vehicle and the health management of the power battery.

[0003] The current mainstream power battery SOH estimation method is to count the charged capacity as the current battery capacity in a relatively stable charging process, and then calculate the ratio of the charged capacity to the rated capacity as the SOH of the battery. This method usually needs to fully charge and fully discharge the battery to calculate the charge in the complete SOC (State of Charge, i.e., current capacity percentage) interval (0~100%). In daily use scenarios, the power battery is rarely fully discharged, and the general charging process is a charging process in a non-complete SOC interval, i.e., a charging process from a non-complete discharge low capacity state to a non-complete charge high capacity state. This method cannot utilize the charging data in the non-complete SOC interval. In addition, excessive charging and discharging of the battery will cause certain damage to the battery. It is not worth the cost to estimate the current battery health degree at the expense of the health degree. Another power battery SOH estimation method is to count the charge in a fixed SOC interval (such as 40%~90%), and then estimate the charge in the complete SOC interval by scaling up, to calculate the SOH of the battery. This method avoids the damage to the battery caused by complete charging and discharging, but the error is amplified when the statistical interval is reduced, and the condition of the fixed SOC interval excludes many charging data, so there is also a problem of not being able to fully utilize the charging data.

[0004] Among the mainstream prediction algorithms, the methodological framework of supervised learning is often adopted, such as long short-term memory neural network (LSTM), Transformer, etc. These methods put all the collected electric vehicle data into the model for training, ignoring the differences between electric vehicles; in addition, these methods have high complexity, which is not conducive to deployment and implementation in on-board devices. SUMMARY

[0005] The main purpose of the present application is to overcome the shortcomings and deficiencies of the prior art, and to provide a power battery SOH estimation and SOH prediction method based on self-supervised learning.

[0006] In order to achieve the above purpose, the following technical solutions are adopted in the present application: A power battery SOH estimation and SOH prediction method based on self-supervised learning, the SOH prediction method comprising the following steps: S1, constructing SOH estimation and data set: extracting the charging segment from the electric vehicle power battery raw data, performing cumulative SOC rise amount calculation operation on the charging segment data, sampling the capacity value once every time the cumulative SOC rise amount reaches 100%, and obtaining the power battery SOH data set by cyclically traversing all the charging data; S2, constructing neural network: constructing a neural network based on frequency domain attention mechanism, the neural network comprising information embedding layer, encoder layer and feedforward neural network layer connected in sequence; S3, self-supervised learning training: adopting a two-stage framework of pre-training and fine-tuning, in the pre-training stage, inputting the whole power battery SOH data set into the neural network to perform the masked sample prediction task, and learning the aging common characteristics; in the fine-tuning stage, inputting the single electric vehicle power battery SOH data into the neural network to perform the recent sample prediction future sample task, and learning the aging individual characteristics; S4, SOH prediction: inputting the recent SOH data of the electric vehicle power battery into the neural network to perform the future SOH prediction operation.

[0007] Further, the step S1 performs battery capacity estimation by cumulative SOC increment reaching 100%, solving the problem of complete charging and discharging or relying on fixed SOC interval in the prior art, and realizing non-destructive capacity calculation by using charging segment data, avoiding damage to the battery by complete charging and discharging.

[0008] Specifically, the step S1 process is as follows: S1.1, performing charging segment screening operation on the electric vehicle power battery raw data, and eliminating non-charging data; S1.2, performing ampere-hour method operation on the charging data:

[0009] Among them, For current, The time interval is used to calculate the difference in charging amount between adjacent samples. ; S1.3, Perform cyclic accumulation operation: When the SOC rises to 100%, the amount of charge accumulated during the accumulation period... As current capacity Calculate the current capacity and the factory-shipped capacity. The ratio yields SOH; S1.4 Repeat the above operations to output the battery SOH dataset.

[0010] Furthermore, step S2 innovatively adopts a frequency domain attention mechanism to address the problem of low efficiency in processing time-series data using traditional time-domain attention mechanisms. By calculating attention weights in the frequency domain through Fourier transform, the computational complexity is reduced, and the low-frequency characteristics of battery aging are effectively captured, significantly improving the accuracy of long-term predictions. At the same time, it enhances the neural network's ability to resist high-frequency noise.

[0011] Specifically, step S2 is as follows: S2.1, Information embedding layer for input sample timestamps Perform time encoding operations to generate time feature vectors. ; Transform the time feature vector Embedded into sample features In this process, the temporal embedded sample features are obtained. ; Time-embedded sample features Perform linear projection operation , to obtain higher-dimensional representation ,in For linear projection matrix, the symbol " " indicates the dot product operation; S2.2. Connect multiple encoders in series. The execution process of each encoder is as follows: S2.2.1 Execute the multi-head frequency domain attention mechanism function operation:

[0012] in For Fourier transform, , , These are the query matrix, key matrix, and value matrix, respectively. It is the frequency domain attention mechanism operation function. For dimensions, " represents a matrix Conjugate transpose; S2.2.2 Perform the residual connection operation, adding the output of the multi-head frequency domain attention mechanism function to the input to obtain the residual result. ; S2.2.3, performing a nonlinear transformation operation on the residual result Inputting the residual result into a feedforward neural network for nonlinear transformation, the feedforward neural network is mathematically represented as follows:

[0013]

[0014] wherein, is calculated by the feedforward neural network, , is a weight matrix, , is a bias term, is a nonlinear activation function; S2.2.4, performing a residual connection operation to add the input and output of the feedforward neural network to obtain the residual result .

[0015] Further, the step S3 solves the problem of ignoring individual differences and requiring a large amount of labeled data in the traditional supervised learning method by using a self-supervised learning method. The pre-training operation learns the common knowledge of battery aging by using all vehicle data through a masking reconstruction task, and the fine-tuning operation performs a sequence prediction task for single vehicle data to learn the individual knowledge of battery aging. This design not only makes full use of unlabeled data, but also realizes precise adaptation to single vehicle characteristics.

[0016] Specifically, the step S3 process is as follows: S3.1, pre-training operation: first, performing a random masking operation on all data set samples to generate masked samples ; second, taking the generated masked samples as input, the complete all data set samples as label, training the neural network to perform the missing data reconstruction task; S3.2, fine-tuning operation: in the sample space of single vehicle power battery original data divided into windows of size T , taking the first k samples as input, the last T-k samples as label, training the neural network to perform the future SOH prediction task.

[0017] Further, the step S3 fine-tuning operation adopts a parameter freezing strategy, keeps the information embedding layer and encoder layer parameters unchanged after pre-training operation, and only adds a fully connected layer after the encoder layer to perform parameter update operation. This method protects the common knowledge learned by pre-training operation, effectively prevents overfitting caused by insufficient data of a single vehicle, significantly reduces the consumption of fine-tuning computing resources, and makes the model more easily deployed on vehicle-mounted devices.

[0018] Specifically, the mathematical expression of the fully connected layer in step S3 is as follows:

[0019] wherein, is the fully connected layer calculation, is the input, is the weight matrix, is the bias term.

[0020] Compared with the prior art, the present application has the following advantages and beneficial effects: 1. Efficient data utilization: the SOH is estimated by accumulating the SOC rise, which fully utilizes the data of the incomplete charging segment and avoids the damage to the battery caused by complete charging and discharging; 2. Novel neural network design: the frequency domain attention mechanism is adopted to more efficiently capture the long-term dependence relationship of the low-frequency dominant process of battery aging and reduce the computational complexity; 3. Advanced deep learning framework: the self-supervised learning framework of "pre-training + fine-tuning" is innovatively introduced, which first learns the aging commonality using all vehicle data, and then adapts the aging individuality using single vehicle data, effectively solving the problem of individual differences of vehicles and reducing the dependence of a single vehicle on a large amount of labeled data; 4. Deployment friendly: by freezing the pre-training layer and only fine-tuning the light-weight layer at the end, the model has small parameter update amount, fast speed, low resource consumption, strong anti-overfitting ability in the fine-tuning stage, and is more easily deployed and updated in resource-constrained environments such as vehicle-mounted devices; 5. High prediction accuracy: combining the advantages of commonality learning and individual adaptation, as well as frequency domain modeling, the accuracy of predicting the future SOH of the power battery of a single electric vehicle is significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0021] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0022] Figure 1is a SOH prediction neural network flowchart based on self-supervised learning in the present application, which shows the forward calculation process of pre-training, fine-tuning and prediction in detail; Figure 2 is a SOH prediction flowchart in the present application; Figure 3 is a training error curve diagram in the pre-training stage of the present application; Figure 4 is a training error curve diagram in the fine-tuning stage of the present application. DETAILED DESCRIPTION

[0023] In order for those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] In the present application, "embodiment" means that the specific features, structures or characteristics described in conjunction with the embodiment can be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily mean the same embodiment, nor is it an independent or alternative embodiment to other embodiments. Those skilled in the art explicitly and implicitly understand that the embodiments described in the present application can be combined with other embodiments.

[0025] Embodiment 1 This embodiment takes cloud data of twenty electric vehicles of the same model as the original data set, and provides a power battery SOH estimation and SOH prediction method based on self-supervised learning. The cloud data is the electric vehicle running data uploaded by twenty electric vehicles every 20 seconds, which contains data time, vehicle speed, charge state, cumulative mileage, total voltage, total current, SOC and other field information. Table 1 shows part of the information of three consecutive original data, in which the data time of the first

[0026] Table 1. Original data display table

[0027] The power battery SOH estimation and SOH prediction method based on self-supervised learning disclosed in this embodiment includes four steps: ​​​S1, constructing SOH estimation and dataset: extracting charging segments from the raw data of the power battery of an electric vehicle, performing cumulative SOC rise amount calculation operation on the charging segment data, sampling the capacity value once every time the cumulative SOC rise amount reaches 100%, and obtaining the power battery SOH dataset by cyclically traversing all the charging data; S2, constructing neural network: constructing a neural network based on a frequency domain attention mechanism, the neural network comprising an information embedding layer, an encoder layer and a feedforward neural network layer connected in sequence; S3, self-supervised learning training: adopting a two-stage framework of pre-training and fine-tuning, inputting the entire power battery SOH dataset into the neural network to perform a masking sample prediction task in the pre-training stage to learn aging common features, and inputting the SOH data of a single electric vehicle power battery into the neural network to perform a recent sample predicting future sample task in the fine-tuning stage to learn aging individual features; S4, SOH prediction: inputting the recent SOH data of the power battery of the electric vehicle into the neural network to perform future SOH prediction operation.

[0028] In step S1, constructing SOH estimation and dataset, the flow chart is shown in Figure 2 , and the specific steps are as follows: Data preprocessing and interval charging amount calculation: first, eliminate the non-charging data in the original dataset, then calculate the interval charging amount between adjacent data points for the remaining charging data points , the calculation formula is , wherein is the current in the time interval, and are the time stamps of adjacent data points. The result of this process is shown in Table 2.

[0029] Table 2. Interval charging amount calculation result table

[0030] S1.1, marking complete charging period: starting from a certain starting data point , accumulating the SOC rise amount of the subsequent data points , until the accumulated value reaches 100% when the corresponding data point . The charging data sequence from to records a complete charging process of the power battery.

[0031] S1.2, feature engineering: extracting additional statistics from the data of the same complete charging period to enrich the feature dimension, the extracted features include extreme value, mean value, range, variance, kurtosis, skewness, etc.

[0032] S1.3, estimating SOH: calculate the sum of all interval charge amount in the complete charging cycle marked in step 1.2 , the charge amount Q can represent the current actual capacity of the power battery . The state of health SOH of the power battery is defined as the ratio of the current capacity to the nominal capacity at factory , that is .

[0033] S1.4, data set making: generate a power battery health data record for the above extracted features, iterate the above process, and finally summarize all generated data records to form a power battery health data set. An example of a power battery health data record is shown in Table 3.

[0034] Table 3. A power battery health data example display table

[0035] Step S2: build a neural network, the neural network is shown in Figure 2 .

[0036] S2.1, information embedding layer: convert the timestamp information of the input sample into digital information that can be processed by the computer , and embed it into the features of the input sample , to get the time embedding sample features ; then perform linear projection operation to obtain high-dimensional representation , where is the linear projection matrix, and "·" represents dot product; S2.2, encoder layer: a multi-layer encoder structure is built, and the encoders are connected in series. Each encoder performs the following operations at a time: first, perform a multi-head frequency domain attention mechanism function operation, and add the result to the input to get the residual ; then input to the feedforward neural network for nonlinear transformation, and add the result to to get the residual , keeping the input and output dimensions of the encoders uniform. The mathematical representation of the frequency domain attention mechanism and the feedforward neural network is as follows:

[0037] where is the Fourier transform, , , are the query matrix, key matrix, and value matrix, is the frequency domain attention mechanism operation function, is the dimension, and ​” denotes a matrix conjugate transpose;

[0038]

[0039] wherein, is calculated for a feedforward neural network, , is a weight matrix, , is a bias term, is a nonlinear activation function; Step S3: self-supervised learning training, divided into pre-training and fine-tuning two steps, the specific steps are as follows: S3.1, neural network pre-training: the health data set of the whole twenty electric vehicles power battery is divided into training set, validation set and test set according to the proportion of 12:3:5 for pre-training. The pre-training task is set to mask data reconstruction. Specifically, 16 consecutive health data records are extracted from the data set, and random masking operation is performed on the 16 data records. The goal of pre-training neural network is to learn the internal structure and pattern of data and accurately fill in the masked part of input. This task setting can make the neural network learn the complex correlation between features and the time sequence dynamic characteristics of the battery aging process. The pre-training neural network flow chart is shown in Figure 1 The pre-training route of the pre-training route is as follows: S3.1.1, random 4 of the input 16 health data records are randomly masked, that is, the masking rate is 25%; S3.1.2, position embedding. The date information of each data record is converted into a dense vector representation through the position embedding layer. The embodiment of the application adopts sinusoidal position encoding to convert the date information into a high-dimensional vector, so that the neural network can effectively perceive the time sequence relationship between data points.

[0040] S3.1.3, dimension projection. The original dimensional features are mapped to a high-dimensional latent space with dimension by a linear projection layer. The purpose is to project low-dimensional features to high-dimensional space, which helps the model to learn more rich representations and complex nonlinear relationships.

[0041] S.3.1.4, encoder stack. The projected high-dimensional sequence is input to a module composed of encoders in series. In each encoder, the input will be simultaneously input to , , a multi-head frequency domain attention block with heads. The trainable content of the multi-head frequency domain attention block lies in the linear layer contained therein. The dimension of the frequency domain attention mechanism calculation result after splicing is ; then batch normalization is applied to reduce the effect of distribution shift; then input into a feedforward neural network with function as activation function, first expand the dimension to = 256, then reduce the dimension back to , and input into the next encoder for calculation; S3.1.5, linear layer output: after the deep feature extraction of the encoder stack, a linear layer is used to map the high-dimensional features back to the original dimension, and finally output the complete 16 health degree data records.

[0042] The pre-trained neural network performs well in the masking reconstruction task. When the evaluation index is selected as the mean square error, the training set error is 0.5581, and the validation set error is 0.4929. Table 4 shows the test effect of a masking reconstruction, and the pre-training error curve is shown in Figure 3 .

[0043] Table 4. Pre-training example effect display table

[0044] S3.2, neural network fine-tuning: the present application realizes that there are individual differences in the use mode and environment of the power battery of different vehicles. In order to accurately adapt to the degradation characteristics of a specific vehicle on the basis of a general pre-trained model, and realize personalized high-precision SOH prediction, the present application designs a lightweight and efficient fine-tuning strategy. The tenth electric vehicle is used as the research vehicle, and its own health degree data set is used for fine-tuning, which is constructed by step S1 and contains the feature sequence extracted from the complete charging period of the vehicle. The fine-tuning task is set to input 16 consecutive historical health degree data records, and the fine-tuning neural network directly predicts the future 8 health degree data records. The fine-tuning neural network flow chart is shown in Figure 1 The fine-tuning route is as follows: S3.2.1, freeze parameters. Keep the parameters of the pre-trained neural network unchanged, which makes the knowledge of the general mode of power battery degradation learned by the pre-training operation be completely preserved, avoids the problem of knowledge forgetting caused by fine-tuning on limited single vehicle data, and significantly improves the robustness and generalization of the model under few-shot fine-tuning; S3.2.2, fine-tuning. A lightweight feedforward neural network is introduced after the encoder layer as the only trainable module for fine-tuning, which completes the sequence-level mapping from 16 sequences to 8 sequences and the dimension-level mapping from high-dimensional features to original dimension features.

[0045] After fine-tuning, the neural network performs well in the health degree data set The performance is good, and when the evaluation index is selected as the mean square error, the training set error is 0.3460, the validation set error is 0.2372, and the fine-tuning process error curve is shown in Figure 4 .

[0046] Step S4: SOH prediction. Input the SOH historical data of the power battery in the recent period, and complete the prediction of the future SOH value.

[0047] In summary, the SOH prediction method disclosed in the embodiment fully utilizes the charging data, reduces the SOH estimation error, and avoids the damage to the battery caused by full charging and discharging. The SOH prediction method combines the advantages of common learning and individual adaptation, and the advantages of frequency domain modeling, and significantly improves the accuracy of the future SOH prediction of the power battery of a single electric vehicle.

[0048] The technical features of the above embodiments can be combined in any manner. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the present application.

[0049] The above embodiments are the preferred embodiments of the present application, but the embodiments of the present application are not limited to the above embodiments, and any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present application are equivalent replacement methods and are included in the protection scope of the present application.

Claims

1. A method for estimating SOH of a power battery and predicting SOH based on self-supervised learning, characterized in that, The SOH prediction method comprises the following steps: S1, constructing SOH estimation and data set: extracting the charging segment from the electric vehicle power battery original data, performing cumulative SOC rising amount calculation operation on the charging segment data, sampling the capacity value once every time the cumulative SOC rising amount reaches 100%, and obtaining the power battery SOH data set by circularly traversing all charging data; S2, constructing neural network: constructing a neural network based on frequency domain attention mechanism, the neural network comprising information embedding layer, encoder layer and feedforward neural network layer connected in turn; S3, self-supervised learning training: adopting two-stage framework of pre-training and fine-tuning, inputting the whole power battery SOH data set into the neural network to perform the masking sample prediction task in the pre-training stage, learning the aging common characteristics; in the fine-tuning stage, inputting the SOH data of single electric vehicle power battery into the neural network to perform the recent sample prediction future sample task, learning the aging individual characteristics; S4, SOH prediction: inputting the recent SOH data of electric vehicle power battery into the neural network to perform future SOH prediction operation.

2. The method of claim 1, wherein the method further comprises: The step S1 process is as follows: S1.1, performing charging segment screening operation on the electric vehicle power battery original data, and eliminating non-charging data; S1.2, performing an ampere-hour method operation on the charging data: wherein, is the current, is the time interval, the charge difference between adjacent samples is calculated ; S1.3, Perform a cycle accumulation operation: when the SOC increase reaches 100%, accumulate the charge amount during the period as the current capacity , calculate the SOH by the ratio of the current capacity to the factory capacity ​ S1.4, circularly performing the above operation, and outputting the power battery SOH data set.

3. The method of claim 1, wherein the method further comprises: The neural network working process in the step S2 is as follows: S2.1, information embedding layer, timestamp of input sample performing time encoding operation to generate time feature vector ; embedding time feature vector into sample feature , obtaining time embedding sample feature ; performing linear projection operation on time embedding sample feature , obtaining high-dimensional representation , where is linear projection matrix, and symbol ” represents dot product operation; S2.2, connecting multiple encoders in series, and each encoder performing the process as follows: S2.2.1, performing multi-head frequency domain attention mechanism function operation: wherein is the Fourier transform, , , are a query matrix, a key matrix, and a value matrix, respectively, is a frequency domain attention mechanism operation function, is a dimension, denotes a matrix transposition, and softmax() is a normalization exponential function for an input vector , the j-th output of the i-th row of the matrix is calculated as follows: S2.2.2, performing a residual connection operation to add the output of the multi-headed frequency domain attention mechanism function to the input to obtain a residual result ; S2.2.3, performing a non-linear transform operation on the residual result input to a feed-forward neural network for non-linear transformation, the feed-forward neural network being mathematically represented as follows: wherein, is calculated for the feedforward neural network, , is a weight matrix, , is a bias term, is a non-linear activation function; S2.2.4, performing a residual connection operation to add the input and output of the feedforward neural network to obtain a residual result .

4. The method of claim 1, wherein the method further comprises: The step S3 comprises: S3.1, pre-training operation: firstly, the whole data set samples performing a random masking operation to generate masked samples ; secondly, taking the generated masked samples as input, the complete whole data set samples as labels, training the neural network to perform the missing data reconstruction task; S3.2, fine-tuning operation: in the original data of a single vehicle power battery, divide the window size T sample space The previous k samples As input, the last T-k samples As label, train neural network to perform future SOH prediction task.

5. The method of claim 4, wherein the method further comprises: In the fine-tuning operation of the step S3, the parameters of the information embedding layer and the encoder layer are kept unchanged, a fully connected layer is added after the encoder layer, and only the parameter update operation of the fully connected layer is performed, and the fully connected layer is mathematically expressed as follows: wherein, is a fully connected layer computation function, is a nonlinear activation function, is an input, is a weight matrix, is a bias term.

Images