Lithium battery capacity degradation knee interval early prediction method based on deep learning

Abstract

The invention discloses a lithium battery capacity degradation knee interval early prediction method based on deep learning, and the method comprises the following steps: S1, obtaining two early discharge capacity-voltage curves at different intervals in the charge-discharge cycle of a lithium battery, carrying out the feature extraction of the discharge capacity-voltage curves through an auto-encoder, and carrying out the feature extraction of the discharge capacity-voltage curves; two discharge capacity-voltage curve compression characteristics are obtained; and S2, inputting the compression characteristics of the two discharge capacity-voltage curves into a feedback confrontation AE-FCNN structure to obtain a predicted knee interval. According to the method, the knee interval is predicted by adopting the feedback confrontation AE-FCNN structure, the knee interval is predicted based on the early charge-discharge cycle data of the lithium battery through a deep learning-based method, the capacity degradation track and trend of the battery are reflected at the initial stage when the capacity degradation is not obvious, and the early monitoring of the state of health of the battery is realized.

Description

Technical Field

[0001] The present invention belongs to the research field of lithium battery health status, and specifically relates to an early prediction method for the lithium battery capacity degradation knee interval based on deep learning. Background Art

[0002] At present, there are some studies in the prior art dedicated to predicting and identifying the knee point on the battery capacity degradation curve. Kim et al. proposed a gradient-knee curve method to identify the knee point without using a complex differential equation battery model. This method uses the minimum and maximum change rates of the slope of the capacity curve, finds the points corresponding to each change rate, draws the tangents of the corresponding period, and defines the intersection of these tangents as the knee point. However, this method can only be used offline, requires knowledge of the complete capacity degradation curve, and requires estimation of the second-order derivative, which may amplify noise in discrete battery data sets and affect detection accuracy. Heze et al. used a sliding window stacked long short-term memory neural network (SLSTM) for knee point prediction. This method combines sliding windows to construct a data set and uses SLSTM for prediction. Although SLSTM improves prediction accuracy by stacking a large number of LSTM networks, its computational complexity increases dramatically, making online prediction difficult. Compared with LSTM, SLSTM models are more complex and require a large amount of data for training. This specificity may limit the applicability of this method in situations where data is scarce or data diversity is insufficient. Greenbank and Tao et al. proposed a data-driven approach to predict the entire capacity degradation trajectory, including the knee point and the end of life. They used Gaussian process regression to regress automatically selected and processed features, directly outputting the knee point and the end of life. While automatic feature selection can improve efficiency, it may not always capture all relevant factors that affect battery degradation. There is a risk of ignoring important variables or the model tending to favor the most prevalent features in the training data. LI et al. proposed a deep learning-based approach to predict the entire capacity curve, including the knee point and the end of life. Their model is a sequence-to-sequence (S2S) model based on a recurrent neural network (RNN) / long short-term memory network (LSTM) that continuously adapts to growing sequences. It uses two multi-layer LSTM architectures: the encoder encodes the input sequence into a static embedding vector, and the decoder processes these vectors to provide the desired output sequence. The model can continuously correct the prediction using newly generated data, enabling online use on electric vehicles through cloud servers. As more data is collected, the output curve becomes more accurate. However, the model has poor prediction performance when early training data is limited, making it unsuitable for early verification of batteries. Deng et al. explored the statistical properties of various battery charging parameters such as current, battery pack voltage, state of charge, and temperature differences. They used methods such as Pearson correlation coefficient and grey relational analysis to relate these factors to battery capacity. Their results highlight the importance of selecting the optimal set of features for predicting battery capacity and emphasize the key role of feature engineering in data-driven models.

[0003] Current research on lithium battery state of health (SOH) estimation methods focuses on the final life, capacity degradation trajectory and knee point, and relies on offline data, with limited early prediction accuracy. Current research has not clearly defined the knee interval on the lithium battery capacity degradation curve, and lacks accurate positioning and early prediction methods. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides a deep learning-based early prediction method for the knee interval of lithium battery capacity degradation, which solves the problem that the prior art lacks accurate positioning and early prediction of the knee interval on the lithium battery capacity degradation curve.

[0005] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: a method for early prediction of the knee interval of lithium battery capacity degradation based on deep learning, comprising the following steps:

[0006] S1. Obtain two early discharge capacity-voltage curves with different intervals in the charge and discharge cycle of the lithium battery, extract features of the discharge capacity-voltage curves through an autoencoder, and obtain compression features of the two discharge capacity-voltage curves;

[0007] S2. The compressed features of the two discharge capacity-voltage curves are fed back into the adversarial AE-FCNN structure to obtain the predicted knee interval.

[0008] Further: In S1, the self-encoder outputs the compression feature of the discharge capacity-voltage curve The specific expression is:

[0009]

[0010] h=f θ (x) = σ(W e x+b e )

[0011] In the formula, is the decoder function, f θ (·) is the encoder function, h is the hidden representation, W e and W d is the weight matrix, b e and b d is the bias vector and σ(·) is the activation function.

[0012] The beneficial effects of the above further scheme are as follows: the present invention reduces the dimensionality of the data through an encoder-decoder structure, and then reconstructs it to output the compressed features of the discharge capacity-voltage curve. This method forces the model to capture a deeper level of representation that can be mapped back to the original sequence features, thereby achieving automatic dimensionality reduction. In the encoding layer, these features carry the most important information of the original data in a compressed form. By using these proposed features to predict the inflection point interval, this method not only reduces the number of network parameters, but also improves the calculation speed.

[0013] Further: In S2, the feedback adversarial AE-FCNN structure includes a fully connected neural network and a feedback fully connected neural network connected to each other, and S2 includes the following sub-steps:

[0014] S21, inputting the compression features of the two discharge capacity-voltage curves into a fully connected neural network to obtain three key points in the knee interval including the knee point, including a starting point, an end point and a knee point;

[0015] S22, combining the three key points of the knee interval in pairs, and splicing them with the compression features of the two discharge capacity-voltage curves, and inputting the spliced ​​results into the feedback fully connected neural network;

[0016] S23, outputting feedback correction factors of the three key points of the knee interval through a feedback fully connected neural network, and correcting the three key points of the knee interval through the feedback correction factors to obtain a predicted knee interval.

[0017] The beneficial effect of the above further scheme is: the present invention proposes an AE-FCNN model, which reduces training parameters and compresses the network size, while also preventing the overfitting problem.

[0018] Further: In S2, the expression of the feedback adversarial AE-FCNN structure loss function J(θ) is specifically:

[0019] J(θ)=αJ prim +(1-α)J feedback

[0020] J prim is the prediction loss of the fully connected neural network, J feedback is the prediction loss of the feedback fully connected neural network, and α is the weighting coefficient.

[0021] The beneficial effects of the above further scheme are as follows: the present invention improves the prediction accuracy of the AE-FCNN structure by adjusting the hyperparameters and optimizing the weights. The structure ensures the independence of the prediction points of each network branch, while combining the coupling relationship between different prediction points to maximize performance. Through the feedback mechanism, the structure enhances the mutual coupling between the three outputs and allows the reuse of intermediate features extracted by the autoencoder, thereby reducing feature loss.

[0022] The beneficial effects of the present invention are:

[0023] (1) The present invention provides a method for early prediction of the knee interval of lithium battery capacity degradation based on deep learning, which can determine the position of the knee interval on the capacity degradation curve online and in real time without relying on offline data. Compared with the existing methods, it avoids the problem of difficulty in solving curvature due to noise.

[0024] (2) The present invention adopts a feedback adversarial AE-FCNN structure to predict the knee interval. Through a deep learning-based method, the knee interval is predicted based on the early charge and discharge cycle data of the lithium battery. In the early stage when the capacity degradation is not obvious, the capacity degradation trajectory and trend of the battery are reflected, thereby realizing early monitoring of the battery health status. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 This is a flow chart of a method for early prediction of the lithium battery capacity degradation knee interval based on deep learning of the present invention.

[0026] Figure 2 It is the third-order fitting effect diagram of the present invention.

[0027] Figure 3 It is the 4th order fitting effect diagram of the present invention.

[0028] Figure 4 It is the fifth-order fitting effect diagram of the present invention.

[0029] Figure 5 The knee interval is determined for the curvature method of the present invention.

[0030] Figure 6 It is the overall structure diagram of the prediction method of the present invention. DETAILED DESCRIPTION

[0031] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0032] like Figure 1 As shown, in one embodiment of the present invention, a method for early prediction of a lithium battery capacity degradation knee interval based on deep learning comprises the following steps:

[0033] S1. Obtain two early discharge capacity-voltage curves with different intervals in the charge and discharge cycle of the lithium battery, extract features of the discharge capacity-voltage curves through an autoencoder, and obtain compression features of the two discharge capacity-voltage curves;

[0034] S2. The compressed features of the two discharge capacity-voltage curves are fed back into the adversarial AE-FCNN structure to obtain the predicted knee interval.

[0035] Predicting the knee interval based on early cycle data is of great significance for rapid verification of battery energy management and manufacturing processes. For users, the knee interval can serve as an early warning for adjusting duty cycles, scheduling battery maintenance, and preventing inconveniences caused by battery health issues. It is very difficult to predict the knee interval in the early stage of capacity degradation, as the aging of lithium-ion batteries is a highly complex and nonlinear process with diverse internal mechanisms and degradation dynamic characteristics that are difficult to accurately model. This makes it very difficult to predict the knee interval in the early stage of capacity degradation. The present invention accurately locates the knee interval on the capacity degradation curve through a curvature method, and predicts the knee interval using a deep learning method.

[0036] In this embodiment, the present invention also applies the least square method to fit the lithium battery capacity degradation curve, and tests polynomial fitting of different orders. The formula of curvature k is specifically:

[0037]

[0038] Wherein, f”(x) is the second-order derivative of the original capacity degradation curve f(x), and f'(x) is the first-order derivative of the original capacity degradation curve f(x).

[0039] After the fifth-order fitting, the mean absolute percentage error (MAPE) dropped to 0.068 and the root mean square percentage error (RMSE) dropped to 0.00086. The fitting curve almost coincides with the original capacity degradation curve, as shown in Figure 2 to Figure 4 shown.

[0040] This fitting process eliminates the noise interference in the original data and obtains an analytical solution for the curvature. Finally, the derived curvature and its derivative are used to simultaneously mark the location of the knee interval, as shown in Figure 5 As shown. The knee point is usually the point with the largest curvature on the capacity decay curve, which can be identified by the extreme value of the curvature. The starting point and the descending point of the knee interval represent the position where the curvature changes fastest, and the derivative of the curvature can be solved, and the position where the curvature changes fastest can be determined by its extreme value. The method for determining the knee interval is only used as a real labeling method for the prediction model data set in the present invention, and does not participate in the training and testing of the prediction model.

[0041] In S1, the self-encoder outputs the compression feature of the discharge capacity-voltage curve The specific expression is:

[0042]

[0043] h=f θ (x) = σ(W e x+b e )

[0044] In the formula, is the decoder function, f θ (·) is the encoder function, h is the hidden representation, W e and W d is the weight matrix, b e and b d is the bias vector and σ(·) is the activation function.

[0045] In this embodiment, the autoencoder used in the present invention is an unsupervised learning algorithm with powerful feature extraction capabilities. It reduces the dimensionality of the data through an encoder-decoder structure, and then reconstructs it to output the compressed features of the discharge capacity-voltage curve. This method forces the model to capture a deeper level of representation that can be mapped back to the original sequence features, thereby achieving automatic dimensionality reduction. In the encoding layer, these features carry the most important information of the original data in a compressed form. By using these proposed features to predict the inflection point interval, this method not only reduces the number of network parameters, but also improves the calculation speed.

[0046] like Figure 6 As shown, the model proposed in the present invention is used to predict the inflection point interval, and the prediction effect is enhanced by considering factors such as parameter scale, test error and data set scale. When considering the parameters of the neural network, a balance must be achieved with the scale of the data set. Initially, a fully connected neural network model (FCNN) was used to extract and train features directly on the original discharge capacity curve without additional feature engineering. For this method, the number of input neurons has reached thousands, and the parameters will increase sharply as the network deepens. This causes the network parameter scale to be much larger than the data set scale, causing overfitting. The fully connected neural network (FCNN) method not only has high training costs due to its large number of parameters, but also performs poorly on the test set. In the data set, there are only about one hundred capacity degradation data points, while the discharge capacity cycle data is relatively sufficient. Based on these characteristics, a solution to the overfitting problem caused by too many parameters is provided. Therefore, an AE-FCNN model is proposed, which reduces training parameters and compresses the network scale while preventing overfitting problems.

[0047] In S2, the feedback adversarial AE-FCNN structure includes a fully connected neural network and a feedback fully connected neural network connected to each other, and S2 includes the following sub-steps:

[0048] S21, inputting the compression features of the two discharge capacity-voltage curves into a fully connected neural network to obtain three key points in the knee interval including the knee point, including a starting point, an end point and a knee point;

[0049] S22, combining the three key points of the knee interval in pairs, and splicing them with the compression features of the two discharge capacity-voltage curves, and inputting the spliced ​​results into the feedback fully connected neural network;

[0050] S23, outputting feedback correction factors of the three key points of the knee interval through a feedback fully connected neural network, and correcting the three key points of the knee interval through the feedback correction factors to obtain a predicted knee interval.

[0051] In this embodiment, the present invention also considers the interaction between the three points in the inflection point interval. The inflection point interval consists of three parts: the starting point, the decline point and the inflection point. The predictions of these three points will affect each other. The advantage is that the three points have a certain coupling relationship on the capacity curve and can be correlated with each other. The disadvantage is that predicting these three points at the same time will weaken the network's prediction effect on any single point. In this method, the model predicts each point separately through three independent branches. However, the three points predicted by this method lose the advantage of their mutual correlation. In order to further enhance the correlation between the output results, the present invention introduces a feedback adversarial AE-FCNN (AE-FCNN+) structure, such as Figure 6 As shown in the figure, during the prediction process, two of the three prediction points are combined with the compressed features of the discharge capacity-voltage curve output by the autoencoder to predict the remaining points. The final model produces two sets of results: the prediction results from the fully connected neural network and the prediction results from the feedback fully connected neural network. The loss function contains these two sets of prediction errors and represents the importance ratio between different prediction errors by weighting. The expression of the feedback adversarial AE-FCNN structure loss function J(θ) is specifically:

[0052] J(θ)=αJ prim +(1-α)J feedback

[0053] J prim is the prediction loss of the fully connected neural network, J feedback is the prediction loss of the feedback fully connected neural network, and α is the weighting coefficient.

[0054] In this embodiment, the present invention adjusts the hyperparameters and optimizes the weights to improve the prediction accuracy of the AE-FCNN structure. The structure ensures the independence of the prediction points of each network branch, while combining the coupling relationship between different prediction points to maximize performance. Through the feedback mechanism, the structure enhances the mutual coupling between the three outputs and allows the reuse of intermediate features extracted by the autoencoder, thereby reducing feature loss.

[0055] The beneficial effects of the present invention are as follows: the present invention provides a method for early prediction of the knee interval of lithium battery capacity degradation based on deep learning, which can determine the position of the knee interval on the capacity degradation curve online and in real time without relying on offline data. Compared with the existing methods, the problem of difficulty in solving curvature due to noise is avoided.

[0056] The present invention adopts a feedback adversarial AE-FCNN structure to predict the knee interval. Through a deep learning-based method, the knee interval is predicted based on the early charge and discharge cycle data of the lithium battery. In the early stage when the capacity degradation is not obvious, the capacity degradation trajectory and trend of the battery are reflected, thereby realizing early monitoring of the battery health status.

[0057] In the description of the present invention, it is necessary to understand that the orientation or positional relationship indicated by the terms "center", "thickness", "upper", "lower", "horizontal", "top", "bottom", "inner", "outer", "radial", etc. is based on the orientation or positional relationship shown in the drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second", and "third" are used only for descriptive purposes, and cannot be understood as indicating or implying the relative importance or the number of implicitly specified technical features. Therefore, the features defined by "first", "second", and "third" may explicitly or implicitly include one or more of the features.

Claims

1. A method for early prediction of lithium battery capacity degradation knee interval based on deep learning, characterized in that: The following steps are involved: S1. Obtain two early discharge capacity-voltage curves with different intervals in the charge and discharge cycle of the lithium battery, extract features of the discharge capacity-voltage curves through an autoencoder, and obtain compression features of the two discharge capacity-voltage curves; S2. The compressed features of the two discharge capacity-voltage curves are fed back into the adversarial AE-FCNN structure to obtain the predicted knee interval.

2. The method for early prediction of lithium battery capacity degradation knee interval based on deep learning according to claim 1 is characterized in that: In S1, the self-encoder outputs the compression feature of the discharge capacity-voltage curve The specific expression is: h=f θ (x)=σ(W e x+b e ) In the formula, is the decoder function, f θ (·) is the encoder function, h is the hidden representation, W e and W d is the weight matrix, b e and b d is the bias vector and σ(·) is the activation function.

3. The method for early prediction of lithium battery capacity degradation knee interval based on deep learning according to claim 1 is characterized in that: In S2, the feedback adversarial AE-FCNN structure includes a fully connected neural network and a feedback fully connected neural network connected to each other, and S2 includes the following sub-steps: S21, inputting the compression features of the two discharge capacity-voltage curves into a fully connected neural network to obtain three key points in the knee interval including the knee point, including a starting point, an end point and a knee point; S22, combining the three key points of the knee interval in pairs, and splicing them with the compression features of the two discharge capacity-voltage curves, and inputting the spliced ​​results into the feedback fully connected neural network; S23, outputting feedback correction factors of the three key points of the knee interval through a feedback fully connected neural network, and correcting the three key points of the knee interval through the feedback correction factors to obtain a predicted knee interval.

4. The method for early prediction of lithium battery capacity degradation knee interval based on deep learning according to claim 3 is characterized in that: In S2, the expression of the feedback adversarial AE-FCNN structure loss function J(θ) is specifically: J(θ)=αJ prim +(1-a)J feedback J prim is the prediction loss of the fully connected neural network, J feedback is the prediction loss of the feedback fully connected neural network, and α is the weighting coefficient.

Images