Lithium battery cycle life and capacity degradation curve inflection point prediction method

Abstract

The invention discloses a lithium battery cycle life and capacity degradation curve inflection point prediction method, which comprises the following steps: acquiring key performance parameters such as voltage, temperature, internal resistance, charging time and discharging capacity, extracting the cycle index when the initial capacity is reduced to 80% as a life prediction label, identifying the capacity degradation curve inflection point, and marking the corresponding cycle index; by adopting a multi-time scale feature extraction and sliding window strategy, extracting weekly and intra-period features from previous 100 cycle data; performing feature selection in combination with a Spearman correlation coefficient method and a recursive feature elimination method; combining a multi-head attention mechanism and a multi-layer perceptron to construct a hybrid deep learning network, and predicting cycle life and inflection points; through training and testing, the cycle life and the capacity degradation curve inflection point of the lithium battery can be accurately predicted under different working conditions. According to the method, the utilization efficiency of the aging information is effectively improved, and the prediction precision and the generalization ability of the model are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of health management of lithium-ion batteries, and in particular to a method for predicting the inflection point of a lithium battery cycle life and capacity degradation curve. Background Art

[0002] Lithium-ion batteries are widely used in electric vehicles, portable electronic devices, electric vertical take-off and landing aircraft (eVTOL), satellites and other fields due to their high energy density, low self-discharge rate and long life. However, lithium-ion batteries will experience capacity decay during use, especially after a certain use cycle, the capacity decay rate is significantly accelerated. In order to ensure the long-term safe and reliable operation of lithium-ion batteries, it is crucial to carry out health management and cycle life prediction of lithium-ion batteries.

[0003] At present, the methods for predicting the cycle life of lithium-ion batteries are mainly divided into model-based methods and data-driven methods. Model-based methods describe the battery degradation process through empirical models or physical models, but this method relies on precise physical and chemical parameters and is computationally complex. With the rapid development of artificial intelligence technology, data-driven methods are gaining more and more attention. This method uses intelligent algorithms to extract health indicators and predict battery cycle life by analyzing the degradation data in actual use.

[0004] However, most existing data-driven methods focus on extracting features from cross-cycle data, ignoring the aging information within the cycle. In addition, existing studies usually only predict the cycle life of a single battery or batteries from the same production batch, failing to fully consider the differences between battery batches, resulting in poor generalization performance of the model.

[0005] To address these problems, the present invention proposes a lithium-ion battery cycle life and capacity inflection point prediction method based on a multi-time scale feature extraction method and a hybrid deep learning network to improve the prediction accuracy and generalization ability.

[0006] The above information disclosed in this Background section is only for enhancement of understanding of the background of the invention and therefore it may contain information that does not form the prior art that is already known in this country to a person of ordinary skill in the art. Summary of the invention

[0007] In view of the deficiencies in the prior art, the purpose of the present invention is to provide a method for predicting the inflection point of the cycle life and capacity degradation curve of a lithium battery. The method achieves high-precision early prediction of lithium-ion batteries through multi-time scale feature extraction and a hybrid deep learning network.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A method for predicting the inflection point of a lithium battery cycle life and capacity degradation curve comprises the following steps:

[0010] S100: Test and collect process quantity signals of the lithium battery charging and discharging process under different working conditions through the battery health management system;

[0011] S200: extracting the number of cycles corresponding to when the initial capacity of the lithium battery drops to 80% as a task label for predicting the cycle life of the lithium battery; identifying the inflection point of the capacity degradation curve, and using the number of cycles corresponding to the inflection point as a task label for predicting the inflection point of the capacity degradation curve;

[0012] S300: extracting inter-cycle features and intra-cycle features according to the process quantity signal;

[0013] S400: removing redundant features in the inter-cycle features;

[0014] S500: Building a hybrid deep learning network;

[0015] S600: training the hybrid deep learning network;

[0016] S700: After the training is completed, the inter-cycle features and intra-cycle features are simultaneously input into the trained hybrid deep learning network for forward propagation. The trained hybrid deep learning network outputs cycle life or inflection point prediction results to achieve the prediction of the inflection point of the lithium battery cycle life and capacity degradation curve under different working conditions.

[0017] Preferably, in step S100, the process quantity signal includes an internal resistance , Charging time , discharge voltage , discharge capacity and temperature .

[0018] Preferably, except for the internal resistance and charging time In addition, the discharge voltage , discharge capacity and temperature It is only necessary to collect data during the discharge phase of each cycle.

[0019] Preferably, for the capacity degradation curve of the lithium battery, two straight lines are used to fit the decay trajectory near the beginning and end of the life respectively, and the intersection of the two straight lines is the inflection point.

[0020] Preferably, in step S300, a multi-time scale feature extraction method is used to extract the inter-cycle features and intra-cycle features.

[0021] Preferably, in step S300, a sliding window strategy is used to extract the inter-cycle features and intra-cycle features.

[0022] Preferably, in step S500, constructing the hybrid deep learning network is combined with a multi-head attention mechanism.

[0023] Preferably, in step S500, constructing the hybrid deep learning network is performed in combination with a multilayer perceptron.

[0024] Preferably, in step S500, the hybrid deep learning network includes a periodic feature extractor.

[0025] Preferably, in step S500, the hybrid deep learning network includes an intra-cycle feature extractor.

[0026] Preferably, in step S200, the capacity degradation curve of the lithium battery is fitted with two straight lines to respectively fit the decay trajectory near the beginning and end of the life, and the intersection of the two straight lines is the inflection point. The fitting formula is expressed as:

[0027]

[0028] in, Indicates the discharge capacity, Indicates the number of cycles, Represents the turning point, Indicates an offset, set to 0, parameter It is used to control the slope and bias of the two straight lines. The optimal value of the parameter can be obtained through nonlinear least squares fitting, and the number of cycles corresponding to the inflection point is used as the task label for predicting the inflection point of the capacity degradation curve.

[0029] Preferably, in step S300, the multi-time scale feature extraction method includes the following sub-steps:

[0030] S301: Discharge capacity based on the first 100 cycle data With discharge voltage The sliding window strategy is adopted, the window size is set to 50 cycles, the sliding window starts at the 10th cycle, and slides 10 cycles each time. The discharge capacity of the last cycle of the window is subtracted from the discharge capacity of the first cycle of the window to obtain , a total of 5 different , , , and Calculate the values ​​of these five values ​​separately , , , , , , , , , , , ,in In , thus a total of 75 statistical values ​​are obtained, recorded as

[0031] ;

[0032] S302: Calculate the slope value of the linear fit of the capacity degradation curve from the 2nd to the 100th cycle based on the cycle number and capacity degradation curve of the first 100 cycle data and the intercept value , the slope value of the linear fit of the capacity degradation curve from the 91st to the 100th cycle and the intercept value , discharge capacity values ​​at the 2nd cycle and the 100th cycle and , the difference between the maximum discharge capacity value and the second discharge capacity value , so a total of 7 values ​​are obtained, recorded as:

[0033] ;

[0034] S303: Calculate the internal resistance value at the second cycle based on the cycle number and internal resistance curve of the first 100 cycle data , Minimum internal resistance between the 2nd cycle and the 100th cycle , the difference between the internal resistance of the 100th cycle and the 2nd cycle , thus we get three values ​​in total, recorded as:

[0035] ;

[0036] S304: Calculate the integral value of the temperature with the cycle number from the 2nd cycle to the 100th cycle according to the cycle number and discharge temperature curve of the first 100 cycle data , the maximum temperature from the 2nd cycle to the 100th cycle and minimum temperature , thus we get three values ​​in total, recorded as:

[0037] ;

[0038] S305: Calculate the average charging time value of the first five cycles ;

[0039] S306: The 89 values ​​calculated in steps S301, S302, S303, S304 and S305 are , respectively, and perform square root , cube root , original value , logarithmic transformation with base 10 And logarithm transformation with natural number e as base , thus a total of 445 cycle features were extracted, recorded as:

[0040] ;

[0041] S307: Calculate eight statistical values, including the maximum value, minimum value, average value, variance, median, kurtosis, skewness, and area under the curve, in each cycle of the first 100 cycle data according to the incremental capacity curve, and record them as:

[0042] ;

[0043] S308: Calculate eight statistical values, including the maximum value, minimum value, average value, variance, median, kurtosis, skewness, and area under the curve, in each cycle of the first 100 cycle data according to the discharge temperature and discharge voltage curve, and record them as:

[0044]

[0045] S309: Calculate eight statistical values, including the maximum value, minimum value, average value, variance, median, kurtosis, skewness, and area under the curve, in each cycle of the first 100 cycle data according to the discharge capacity and discharge voltage curve, and record them as:

[0046]

[0047] S310: According to steps S307, S308 and S309, 24 statistical values ​​can be calculated from each cycle data, and the first 100 cycles Combined together, the intra-period characteristics are obtained, expressed as:

[0048] ;

[0049] Preferably, in step 400, removing redundant features in the inter-cycle features includes the following sub-steps:

[0050] S401: Calculate inter-cycle features based on cross-cycle data extraction The absolute value of the Spearman correlation coefficient with the cycle life or inflection point is calculated as follows:

[0051]

[0052] in represents the periodic characteristic value, represents the prediction target, n is the number of samples, and the periodic features whose values ​​are greater than 0.4 are retained and recorded as ;

[0053] S402: Use recursive feature elimination search strategy, select XGBRegressor as evaluation model, and calculate the periodic features with Spearman correlation coefficient greater than 0.4 Importance Rating , select the top 320 period features and obtain the subset after fusion feature selection .

[0054] Preferably, in step 500, the multi-head attention mechanism and the multi-layer perceptron are combined to construct a periodic feature extractor and intra-cycle feature extractor The hybrid deep learning network includes the following steps:

[0055] S501: The cycle feature extractor It consists of two fully connected layers;

[0056] S502: The intra-cycle feature extractor It consists of two Transformer encoding layers, one one-dimensional convolutional layer, and one flattening layer;

[0057] S503: The and A Dropout layer is connected to each of them, and finally the outputs of the two feature extractors are connected in series and mapped to the predicted value by a fully connected layer.

[0058] Preferably, in step 600, the learning rate adopted by the snapshot ensemble learning strategy can be expressed as:

[0059]

[0060] in Indicates the current training cycle, represents the learning rate of the current cycle, and represents the maximum and minimum learning rate values, Indicates the total number of cycles in a snapshot cycle. Indicates modular operation. In a training process, the last ten snapshot models are saved.

[0061] Preferably, in step 600, after obtaining the predicted value of the cycle life or inflection point during the training process, the loss of the model can be calculated based on the true label. , the calculation formula is:

[0062]

[0063] Where N is the number of samples, is an estimate, is the true label.

[0064] The present invention also discloses a lithium battery cycle life and capacity degradation curve inflection point prediction system, the system comprising:

[0065] The acquisition module is used to collect process quantity signals of the lithium battery charging and discharging process under different working conditions;

[0066] An identification module, used to identify task labels for cycle life prediction of lithium batteries and task labels for capacity degradation curve inflection point prediction;

[0067] An extraction module, used to extract inter-cycle features and intra-cycle features;

[0068] Building blocks for constructing hybrid deep learning networks;

[0069] Training module, used to train hybrid deep learning networks;

[0070] The prediction module is used to predict the cycle life of lithium batteries and the inflection point of the capacity degradation curve under different working conditions.

[0071] The present invention also discloses a computer storage medium, wherein the storage medium includes computer instructions, and when the computer is run on the computer, the computer executes any of the methods described above.

[0072] The present invention also discloses an electronic device, wherein the electronic device comprises: a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the methods described above when executing the program.

[0073] Compared with the prior art, the beneficial effects brought by the present invention are: for the first time, it realizes the simultaneous extraction of features from inter-cycle and intra-cycle data, and maximizes the use of aging information through a sliding window strategy; a fusion feature selection method is used to effectively screen redundant features, while ensuring the screening rate and taking into account the quality of the features; the designed hybrid model is combined with an integrated snapshot learning strategy, which significantly improves the generalization ability and robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 It is a flow chart of a method for predicting the inflection point of a lithium battery cycle life and capacity degradation curve provided by an embodiment of the present invention;

[0075] Figure 2It is a schematic diagram of inflection point identification in a method for predicting inflection point of a lithium battery cycle life and capacity degradation curve provided by an embodiment of the present invention;

[0076] Figure 3 It is a schematic diagram of a sliding window strategy in a method for predicting an inflection point of a lithium battery cycle life and capacity degradation curve provided by an embodiment of the present invention;

[0077] Figure 4 It is a schematic diagram of a fusion feature selection method in a method for predicting an inflection point of a lithium battery cycle life and capacity degradation curve provided by an embodiment of the present invention;

[0078] Figure 5 This is a detailed diagram of a hybrid deep learning network for predicting an inflection point of a lithium battery cycle life and capacity degradation curve provided by an embodiment of the present invention;

[0079] Figure 6 It is a schematic diagram of a snapshot integrated learning strategy for a method for predicting the inflection point of a lithium battery cycle life and capacity degradation curve provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0080] The following will refer to the attached Figures 1 to 6 Specific embodiments of the present invention are described in detail. Although specific embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0081] It should be noted that certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different nouns to refer to the same component. This specification and claims do not use the difference in nouns as a way to distinguish components, but use the difference in the functions of the components as the criterion for distinction. As mentioned throughout the specification and claims, "including" or "comprising" is an open term, so it should be interpreted as "including but not limited to". The subsequent description of the specification is a preferred embodiment of the present invention, but the description is based on the general principles of the specification and is not intended to limit the scope of the present invention. The scope of protection of the present invention shall be determined by the attached claims.

[0082] To facilitate understanding of the embodiments of the present invention, further explanation will be given below by taking specific embodiments as examples in conjunction with the accompanying drawings, and each of the accompanying drawings does not constitute a limitation on the embodiments of the present invention.

[0083] like Figure 1 As shown, the present invention provides a method for predicting the inflection point of a lithium battery cycle life and capacity degradation curve, comprising:

[0084] S100: The battery health management system tests and collects process quantity signals of the lithium battery charging and discharging process under different working conditions, including: voltage, temperature, internal resistance, charging time and discharge capacity;

[0085] S200: extracting the number of cycles corresponding to when the initial capacity of the lithium battery drops to 80% as a task label for predicting the cycle life of the lithium battery; identifying the inflection point of the capacity degradation curve, and using the number of cycles corresponding to the inflection point as a task label for predicting the inflection point of the capacity degradation curve;

[0086] Furthermore, for the capacity degradation curve of the lithium battery, two straight lines are used to fit the decay trajectory near the beginning and end of the life, and the intersection of the two straight lines is the inflection point.

[0087] It should be noted that the "task label for cycle life prediction" can directly regard the "number of cycles corresponding to the initial capacity of the lithium battery dropping to 80%" as the label; while the "task label for predicting the inflection point of the capacity degradation curve" requires an "identification" operation, "using two straight lines to fit the decay trajectory near the beginning and end of the life, respectively, and the intersection of the two straight lines is the inflection point", and then regard the number of cycles corresponding to the inflection point as the label.

[0088] S300: A multi-time scale feature extraction method is used, and a sliding window strategy is used to extract inter-cycle features from the cross-cycle data of the first 100 cycles of the lithium battery, and intra-cycle features are extracted from each single cycle of the first 100 cycle data;

[0089] S400: performing feature selection to remove redundant features in the periodic features;

[0090] S500: Combining multi-head attention mechanism and multi-layer perceptron to build a feature extractor including periodicity and intra-cycle feature extractor Hybrid deep learning network;

[0091] S600: Input the inter-cycle features and intra-cycle features in the training set into the inter-cycle feature extractor respectively and intra-cycle feature extractor , and input any task label, and use the snapshot ensemble learning strategy to train the hybrid deep learning network;

[0092] It should be noted that in a prediction task, you can enter any one of the two labels here. If you enter the "task label for cycle life prediction", the prediction target is "cycle life"; if you enter the "task label for capacity degradation curve inflection point prediction", the prediction target is "capacity degradation curve inflection point".

[0093] S700: After the training is completed, the inter-cycle features and intra-cycle features of the test set are simultaneously input into the trained hybrid deep learning network for forward propagation. The trained hybrid deep learning network outputs cycle life or inflection point prediction results to achieve the prediction of the inflection point of the lithium battery cycle life and capacity degradation curve under different working conditions.

[0094] In a preferred embodiment of the method for predicting the inflection point of the cycle life and capacity degradation curve of a lithium battery, in step S100, the process quantity signal includes an internal resistance , Charging time , discharge voltage , discharge capacity and temperature , except for the internal resistance and charging time In addition, the discharge voltage , discharge capacity and temperature It is only necessary to collect data during the discharge phase of each cycle.

[0095] In a preferred embodiment of the method for predicting the inflection point of the cycle life and capacity degradation curve of a lithium battery, in step S200, two straight lines are used to fit the decay trajectory near the beginning and end of the capacity degradation curve of the lithium battery, respectively, such as Figure 2 As shown, the intersection of the two straight lines is the inflection point, and the fitting formula is expressed as:

[0096]

[0097] in, Indicates the discharge capacity, Indicates the number of cycles, Represents the turning point, Indicates an offset, set to 0, parameter It is used to control the slope and bias of the two straight lines. The optimal value of the parameter can be obtained through nonlinear least squares fitting, and the number of cycles corresponding to the inflection point is used as the task label for predicting the inflection point of the capacity degradation curve. Figure 2 A schematic diagram of inflection point identification is shown, with parameter values ​​noted in the figure That is the optimal value for identifying the inflection point of the battery capacity degradation curve.

[0098] In a preferred embodiment of the method for predicting the inflection point of the cycle life and capacity degradation curve of a lithium battery, Figure 3 The schematic diagram of the sliding window strategy is shown. As shown in the figure, during the inter-cycle feature extraction process, the sliding window strategy is used to calculate 5 different The five values ​​are then transformed through 15 statistical transformations and 5 nonlinear transformations. This strategy makes full use of the aging information in the cycle-to-cycle data, so that even when only a small amount of early cycle data is used, a high cycle life or inflection point prediction accuracy can be obtained.

[0099] Specifically, in step S300, the multi-time scale feature extraction method includes the following sub-steps:

[0100] S301: Discharge capacity based on the first 100 cycle data With discharge voltage The sliding window strategy is adopted, the window size is set to 50 cycles, the sliding window starts at the 10th cycle, and slides 10 cycles each time. The discharge capacity of the last cycle of the window is subtracted from the discharge capacity of the first cycle of the window to obtain , a total of 5 different , , , and Calculate the values ​​of these five values ​​separately , , , , , , , , , , , ,in In , thus a total of 75 statistical values ​​are obtained, recorded as

[0101] ; It should be noted that the discharge capacity With discharge voltage The curve is composed of 1000 sampling points connected smoothly. i That is the horizontal coordinate of each point on the curve, V i It is the ordinate of each point on the curve. That is, the sum of the products of the horizontal and vertical coordinates of all points.

[0102] S302: Calculate the slope value of the linear fitting of the capacity decay curve from the 2nd to the 100th cycle based on the cycle number and the discharge capacity degradation curve of the first 100 cycle data and the intercept value , the slope value of the linear fit of the capacity decay curve from 91 to 100 cycles and the intercept value , discharge capacity values ​​at the 2nd cycle and the 100th cycle and , the difference between the maximum discharge capacity value and the second discharge capacity value , so a total of 7 values ​​are obtained, recorded as:

[0103] ;

[0104] S303: Calculate the internal resistance value at the second cycle based on the cycle number and internal resistance curve of the first 100 cycle data , Minimum internal resistance between the 2nd cycle and the 100th cycle , the difference between the internal resistance of the 100th cycle and the 2nd cycle , thus we get three values ​​in total, recorded as:

[0105] ;

[0106] S304: Calculate the integral value of the temperature with the cycle number from the 2nd cycle to the 100th cycle according to the cycle number and discharge temperature curve of the first 100 cycle data , the maximum temperature from the 2nd cycle to the 100th cycle and minimum temperature , thus we get three values ​​in total, recorded as:

[0107] ;

[0108] S305: Calculate the average charging time value of the first five cycles ;

[0109] S306: The 89 values ​​calculated in steps S301, S302, S303, S304 and S305 are , respectively, and perform square root , cube root , original value , logarithmic transformation with base 10 And logarithm transformation with natural number e as base , thus a total of 445 cycle features were extracted, recorded as:

[0110] ;

[0111] S307: Calculate eight statistical values, including the maximum value, minimum value, average value, variance, median, kurtosis, skewness, and area under the curve, in each cycle of the first 100 cycle data according to the incremental capacity curve, and record them as:

[0112] ;

[0113] S308: Calculate eight statistical values, including the maximum value, minimum value, average value, variance, median, kurtosis, skewness, and area under the curve, in each cycle of the first 100 cycle data according to the discharge temperature and discharge voltage curve, and record them as:

[0114]

[0115] S309: Calculate eight statistical values, including the maximum value, minimum value, average value, variance, median, kurtosis, skewness, and area under the curve, in each cycle of the first 100 cycle data according to the discharge capacity and discharge voltage curve, and record them as:

[0116]

[0117] S310: According to steps S307, S308 and S309, 24 statistical values ​​can be calculated from each cycle data, and the first 100 cycles Combined together, the intra-period characteristics are obtained, expressed as:

[0118] .

[0119] For the above steps, combined with what is mentioned above, the proposed multi-time scale feature extraction method can not only maximize the feature extraction from the inter-cycle data through the sliding window strategy, but also effectively extract features from the intra-cycle data, fully mining the aging information contained in the data, thereby achieving high-precision battery cycle life and inflection point prediction.

[0120] In a preferred embodiment of the method for predicting the inflection point of the cycle life and capacity degradation curve of a lithium battery, Figure 4 The schematic diagram of the fusion feature selection method is shown. As shown in the figure, the Spearman correlation coefficient is first used to quickly screen out features with low correlation, and then the features with low importance scores are removed one by one through recursive feature elimination. This method is both efficient and can retain the features with the best prediction performance.

[0121] Specifically, in step 400, the feature selection includes the following sub-steps:

[0122] S401: Calculate inter-cycle features based on cross-cycle data extraction The absolute value of the Spearman correlation coefficient with the cycle life or inflection point is calculated as follows:

[0123]

[0124] in represents the periodic characteristic value, represents the prediction target, n is the number of samples, and the periodic features whose values ​​are greater than 0.4 are retained and recorded as ;

[0125] S402: Use recursive feature elimination search strategy, select XGBRegressor as evaluation model, and calculate the periodic features with Spearman correlation coefficient greater than 0.4 Importance Rating , select the top 320 period features and obtain the subset after fusion feature selection .

[0126] With respect to the above steps, in combination with what has been mentioned above, the proposed method performs feature selection by combining the Spearman correlation coefficient method and the recursive feature elimination method, giving full play to the advantages of both of fast calculation speed and excellent screening effect. The method of this embodiment can not only quickly screen out features with low correlation, but also retain features with the best prediction performance.

[0127] In a preferred embodiment of the method for predicting the inflection point of the cycle life and capacity degradation curve of a lithium battery, Figure 5 The hybrid deep learning network diagram is shown. As shown in the figure, the hybrid deep learning network structure consists of and It consists of two parts. In , the sequence features are first extracted through two Transformer encoder layers (each with 3 heads and a feedforward network dimension of 128), followed by a one-dimensional convolutional layer (Conv1D, with a kernel size of 16, a stride and padding of 1), and the output is flattened through a Flatten layer, followed by further processing through a fully connected layer (FC-16) with Dropout (0.5). At the same time, The first layer uses Sigmoid activation function, and the next layer uses Dropout (0.1) to prevent overfitting. Finally, the outputs of the two encoders are concatenated through a Concatenate operation and output through a fully connected layer (FC-1). , feature input within the cycle Since the inter-cycle features already contain a lot of physical information, a shallow neural network is used for training, while the intra-cycle features require a deep neural network to further extract aging information. The network can simultaneously process inter-cycle and intra-cycle features of different dimensions and flexibly adjust the depth of the encoder according to the complexity of the features. This design not only makes full use of the feature information, but also significantly improves the generalization ability of the model, thereby improving the prediction accuracy.

[0128] Specifically, in step S500, the multi-head attention mechanism and the multi-layer perceptron are combined to construct a period feature extractor and intra-cycle feature extractor The hybrid deep learning network includes the following steps:

[0129] S501: Using two fully connected layers to form the inter-cycle feature extractor ;

[0130] S502: Use two layers of Transformer encoding layers, one layer of one-dimensional convolution layer and one layer of flattening layer to form the intra-cycle feature extractor ;

[0131] S503: The cycle feature extractor and intra-cycle feature extractor A Dropout layer is connected to each of them, and finally the outputs of the two feature extractors are connected in series and mapped to the predicted value by a fully connected layer.

[0132] For the above steps, combined with what was mentioned above, the designed model architecture can simultaneously process inter-cycle features and intra-cycle features of different dimensions, and flexibly adjust the complexity of the encoder according to the complexity of the features. This design not only makes full use of multi-time scale feature information, but also effectively improves the generalization ability of the model, thereby significantly improving the prediction accuracy of this method.

[0133] In a preferred embodiment of the method for predicting the inflection point of the cycle life and capacity degradation curve of a lithium battery, Figure 6 The learning rate diagram of the snapshot ensemble learning strategy is shown. As shown in the figure, the snapshot ensemble learning strategy enables the model to jump out of the local optimal solution and further explore the local optimal solution by periodically increasing and decreasing the learning rate, thereby obtaining a model that converges to multiple local optimal solutions in one training process.

[0134] Specifically, the learning rate adopted by the snapshot ensemble learning strategy can be expressed as:

[0135]

[0136] in Indicates the current training cycle, represents the learning rate of the current cycle, and represents the maximum and minimum learning rate values, Indicates the total number of rounds in a snapshot cycle. Indicates modular operation. In a training process, the last ten snapshot models are saved.

[0137] In a preferred embodiment of the method for predicting the inflection point of the cycle life and capacity degradation curve of a lithium battery, after obtaining the predicted value of the cycle life or inflection point during the training process, the loss of the model can be calculated based on the real label. , the calculation formula is:

[0138]

[0139] Where N is the number of samples, is an estimate, is the true label.

[0140] In another embodiment, a method for predicting the inflection point of a lithium battery cycle life and capacity degradation curve comprises the following steps:

[0141] S10: Test and collect process quantity signals of the lithium battery charging and discharging process under different working conditions through the battery health management system, including: voltage, temperature, internal resistance, charging time and discharge capacity;

[0142] In this step, the process quantity signal includes the internal resistance , Charging time , discharge voltage , discharge capacity and temperature , except for the internal resistance and charging time In addition, the discharge voltage , discharge capacity and temperature It is only necessary to collect data during the discharge phase of each cycle.

[0143] It can be understood that in the process of collecting signals, the process quantity signals collected by the battery under different working conditions are different, and the corresponding cycle life and capacity degradation curve inflection points are also different.

[0144] S20: extracting the number of cycles corresponding to when the initial capacity of the lithium battery drops to 80% as a task label for predicting the cycle life of the lithium battery; by identifying the inflection point of the capacity degradation curve, the number of cycles corresponding to the inflection point is used as a task label for predicting the inflection point of the capacity degradation curve;

[0145] In this step, the capacity degradation curve of the lithium battery is fitted with two straight lines to fit the decay trajectory near the beginning and end of the life respectively. The intersection of the two straight lines is the inflection point. The fitting formula is expressed as:

[0146]

[0147] in, Indicates the discharge capacity, Indicates the number of cycles, Represents the turning point, Indicates an offset, set to 0, parameter It is used to control the slope and bias of the two straight lines. The optimal value of the parameter can be obtained through nonlinear least squares fitting, and the number of cycles corresponding to the inflection point is used as the task label for predicting the inflection point of the capacity degradation curve.

[0148] S30: Using a multi-time scale feature extraction method, a sliding window strategy is used to extract inter-cycle features from the cross-cycle data of the first 100 cycles of the lithium battery, and intra-cycle features are extracted from each single cycle of the first 100 cycle data;

[0149] S40: performing feature selection to remove redundant features in the inter-cycle features;

[0150] S50: Combining multi-head attention mechanism and multi-layer perceptron to build a feature extractor including period and intra-cycle feature extractor Hybrid deep learning network;

[0151] S60: Input the inter-cycle features and intra-cycle features in the training set into the inter-cycle feature extractor respectively and intra-cycle feature extractor , and input any task label, and use the snapshot ensemble learning strategy to train the hybrid deep learning network; among them, when the input label is "task label for cycle life prediction", the training of cycle life prediction is completed; after completing the corresponding training, input "task label for capacity degradation curve inflection point prediction" to complete the training of capacity degradation curve inflection point prediction.

[0152] In this step, the learning rate adopted by the snapshot ensemble learning strategy can be expressed as:

[0153]

[0154] in Indicates the current training cycle, represents the learning rate of the current cycle, and represents the maximum and minimum learning rate values, Indicates the total number of cycles in a snapshot cycle. Indicates modular operation. In a training process, the last ten snapshot models are saved;

[0155] After obtaining the predicted value of cycle life or inflection point during training, the loss of the model can be calculated based on the true label. , the calculation formula is:

[0156]

[0157] Where N is the number of samples, is an estimate, is the true label;

[0158] S70: After the initial training is completed, the inter-cycle features and intra-cycle features of the test set are simultaneously input into the trained hybrid deep learning network for forward propagation. The trained hybrid deep learning network outputs the cycle life or inflection point prediction results to achieve the prediction of the cycle life of lithium batteries and the inflection point prediction of the capacity degradation curve under different working conditions. It should be noted that the initial training can set the training rounds (epochs), and the present embodiment sets 4000 epochs. After 4000 epochs of training, the model converges and can be used for prediction, that is, it can be verified on the first test set and the second test set.

[0159] In this embodiment, the cycle life and capacity degradation curve inflection point of lithium batteries are predicted by combining multi-time scale feature extraction and a hybrid model based on a multi-head attention mechanism and a multi-layer perceptron. First, features are extracted from both the inter-cycle data and the intra-cycle data to maximize the use of the aging information hidden in the data. Feature selection is then performed to ensure the efficiency of the screening process while retaining high-quality features. In order to enhance the predictive generalization ability of the model, a hybrid neural network model based on a multi-head attention mechanism and a multi-layer perceptron is constructed, and a snapshot integration strategy is used for training, thereby significantly improving the generalization of the method.

[0160] Next, the technical solution of the present invention is further illustrated by conducting comparative experiments with methods A and B (comparative methods A and B are both representative prediction methods that are widely used and well-known in the current field).

[0161] Specifically, the main evaluation indicators used in this experiment are the root mean square error (RMSE) and the mean relative error (MAPE). The smaller these two values ​​are, the more stable the prediction is.

[0162] The data used in the experiment were collected in three time periods. The data collection time of the first and second batches was relatively close, so their initial states were similar. The battery data of the third batch was collected after another year. Due to calendar aging and other reasons, the battery data of the third batch was quite different from the data of the first two batches. The odd-numbered batteries of the first two batches were used as the training set, the even-numbered batteries were used as the first test set, and the batteries of the third batch were used as the second test set.

[0163] To ensure the fairness of the experiment, all methods used the same hyperparameter settings. The prediction accuracy of the three methods on the first test set and the second test set is shown in Table 1.

[0164] Table 1

[0165]

[0166] It can be seen from Table 1 that the prediction accuracy of the method of the present invention is always higher than that of the prior art, that is, comparison methods A and B.

[0167] The present invention also discloses a lithium battery cycle life and capacity degradation curve inflection point prediction system, the system comprising:

[0168] The acquisition module is used to collect process quantity signals of the lithium battery charging and discharging process under different working conditions;

[0169] An identification module, used to identify task labels for cycle life prediction of lithium batteries and task labels for capacity degradation curve inflection point prediction;

[0170] An extraction module, used to extract inter-cycle features and intra-cycle features;

[0171] Building blocks for constructing hybrid deep learning networks;

[0172] Training module, used to train hybrid deep learning networks;

[0173] The prediction module is used to predict the cycle life of lithium batteries and the inflection point of the capacity degradation curve under different working conditions.

[0174] The present invention also discloses a computer storage medium, wherein the storage medium includes computer instructions, and when the computer is run on the computer, the computer executes any of the methods described above.

[0175] The present invention also discloses an electronic device, wherein the electronic device comprises: a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements any of the methods described above when executing the program.

[0176] The above specific embodiments are only illustrative and instructive, and are not restrictive. A person skilled in the art can make many forms under the guidance of this specification and without departing from the scope of protection of the claims of the present invention, all of which belong to the scope of protection of the present invention.

Claims

1. A method for predicting the inflection point of a lithium battery cycle life and capacity degradation curve, characterized in that: The method comprises the following steps: S100: Test and collect process quantity signals of the lithium battery charging and discharging process under different working conditions through the battery health management system; S200: extracting the number of cycles corresponding to when the initial capacity of the lithium battery drops to 80% as a task label for predicting the cycle life of the lithium battery; identifying the inflection point of the capacity degradation curve, and using the number of cycles corresponding to the inflection point as a task label for predicting the inflection point of the capacity degradation curve; S300: extracting inter-cycle features and intra-cycle features according to the process quantity signal; S400: removing redundant features in the inter-cycle features; S500: Building a hybrid deep learning network; S600: training the hybrid deep learning network; S700: After the training is completed, the inter-cycle features and intra-cycle features are simultaneously input into the trained hybrid deep learning network for forward propagation. The trained hybrid deep learning network outputs cycle life or inflection point prediction results to achieve the prediction of the inflection point of the lithium battery cycle life and capacity degradation curve under different working conditions.

2. The method according to claim 1, wherein: Preferably, in step S100, the process quantity signal includes an internal resistance , Charging time , discharge voltage , discharge capacity and temperature .

3. The method according to claim 2, wherein: Remove internal resistance and charging time In addition, the discharge voltage , discharge capacity and temperature It is only necessary to collect data during the discharge phase of each cycle.

4. The method according to claim 1, wherein: For the capacity degradation curve of lithium batteries, two straight lines are used to fit the decay trajectory near the beginning and end of life respectively, and the intersection of the two straight lines is the inflection point.

5. The method according to claim 1, wherein: In step S300, a multi-time scale feature extraction method is used to extract the inter-cycle features and intra-cycle features.

6. The method according to claim 1, wherein: In step S300, a sliding window strategy is used to extract the inter-cycle features and intra-cycle features.

7. The method according to claim 1, wherein: In step S500, constructing the hybrid deep learning network requires combining a multi-head attention mechanism.

8. A lithium battery cycle life and capacity degradation curve inflection point prediction system, characterized in that: The system comprises: The acquisition module is used to collect process quantity signals of the lithium battery charging and discharging process under different working conditions; An identification module, used to identify task labels for cycle life prediction of lithium batteries and task labels for capacity degradation curve inflection point prediction; An extraction module, used to extract inter-cycle features and intra-cycle features; Building blocks for constructing hybrid deep learning networks; Training module, used to train hybrid deep learning networks; The prediction module is used to predict the cycle life of lithium batteries and the inflection point of the capacity degradation curve under different working conditions.

9. A computer storage medium, wherein: The storage medium includes computer instructions, which, when executed on a computer, enable the computer to execute the method according to any one of claims 1 to 7.

10. An electronic device, wherein: The electronic device comprises: a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the method according to any one of claims 1 to 7 when executing the program.

Images