A method for evaluating the full life cycle life of an electric vehicle battery based on transfer learning

Abstract

The present application relates to the technical field of electric vehicle battery full life cycle life evaluation, and in particular to an electric vehicle battery full life cycle life evaluation method based on transfer learning. The present application is based on high correlation charging segment features, and source training set and target training set are respectively constructed according to source data set of source domain battery and target data set of target battery. The SOH evaluation model is pre-trained using the source training set, and the pre-trained SOH evaluation model is fine-tuned using the target training set. Finally, the SOH evaluation model after fine-tuning is used to evaluate the SOH of the target battery in the late service period and the utilization period. The present application realizes the full life cycle evaluation of the SOH value of the power battery under the reasonable training data scale. The effective evaluation interval of the present application runs through the service period and the utilization period after retirement of the power battery, and the early prediction of the battery SOH and the evaluation of the secondary service ability of the battery are considered, which effectively helps the full life cycle fine management of the power battery.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of electric vehicle battery full life cycle life evaluation, and in particular to an electric vehicle battery full life cycle life evaluation method based on transfer learning. BACKGROUND

[0002] The improvement of the environmental cost and performance demand of power batteries drives the battery management to be refined and extended to the full life cycle. Under complex working conditions, extending the effective interval of the aging state evaluation method from the service period to the post-retirement utilization period is the key to achieving the above goal.

[0003] As a complex closed electrochemical system, power batteries often exhibit obvious nonlinear time-varying characteristics. Therefore, the accuracy of the battery degradation path prediction depends significantly on the size of the training data. This leads to a huge challenge in prospectively evaluating the aging state of power batteries only relying on a small amount of early data. For this reason, the prior art proposes, for example: a multi-output Gaussian process regression (MOGPR) based on the initial cycle of the battery pack and the full life cycle data of the cell to extract features and achieve the effect of battery pack SOH prediction. Or use a deep learning network based on attention information to combine the health features extracted in the charging segment to reduce the size of the SOH prediction model training set. However, these schemes have obvious edge distortion near the battery retirement point, the effective prediction window is small, and it is difficult to achieve SOH evaluation of the full life cycle of power batteries.

[0004] Therefore, the prior art urgently needs an electric vehicle battery full life cycle life evaluation method that can solve the edge distortion caused by the nonlinear time-varying characteristics of the battery, and balance the early prediction of the battery SOH and the evaluation of the secondary service ability of the battery under a reasonable training data size. SUMMARY

[0005] The purpose of the present application is to provide an electric vehicle battery full life cycle life evaluation method based on transfer learning to solve the above technical problems.

[0006] The present application provides an electric vehicle battery full life cycle life evaluation method based on transfer learning, which comprises:

[0007] According to the cycle number and SOH value correlation curve of the multiple source domain batteries in the service period, a source data set is constructed; according to the cycle number and SOH value correlation curve of the target battery in the early service period of the target battery in need of SOH evaluation, a target data set is constructed;

[0008] According to the SOC distribution interval of the electric vehicle charging starting time, the battery voltage interval for extracting the charging segment features in each cycle is determined;

[0009] performing correlation analysis on the pre-selected candidate features and the battery SOH values to determine the charging segment features for battery SOH value prediction;

[0010] taking the charging segment features in the battery voltage interval of each cycle as input features and the SOH value corresponding to the cycle as output label, and constructing a source training set and a target training set according to the source data set and the target data set respectively;

[0011] pre-training the SOH evaluation model constructed in advance by using the source training set, and fine-tuning the pre-trained SOH evaluation model by using the target training set;

[0012] using the fine-tuned SOH evaluation model to evaluate the SOH of the target battery in the late service period and the utilization period.

[0013] Preferably, the cycle number-SOH value correlation curves of the plurality of source domain batteries are obtained under different preset operating condition parameters, and the operating condition parameters include environmental temperature, charging current, discharging current and SOH value end-of-life.

[0014] Preferably, the cycle number-SOH value correlation curve of the target battery in the early service period represents the cycle number-SOH value correlation curve from 100% SOH to 90% SOH.

[0015] Preferably, the correlation analysis on the pre-selected candidate features and the battery SOH values is based on Kendall correlation analysis.

[0016] Preferably, according to the SOC distribution interval at the starting time of electric vehicle charging, the battery voltage interval for extracting the charging segment features in each cycle is determined, and the determined battery voltage interval is 3.5V-4.2V.

[0017] The charging segment features include:

[0018] a VIT feature representing the time required for the charging voltage of the battery to change from 3.8V to 4.0V in the constant current charging link;

[0019] a ST feature representing the time from the start of battery charging to the constant current charging-constant voltage charging conversion node;

[0020] a CIT feature representing the time required for the charging current of the battery to change from 2.5A to 0.8A in the constant voltage charging link.

[0021] Preferably, the SOH evaluation model includes a Transformer-BiLSTM model structure.

[0022] Preferably, the Transformer-BiLSTM model structure comprises a BiLSTM module, a Transformer encoder, a Transformer decoder and an output module connected in sequence; wherein,

[0023] The BiLSTM module is configured to process the time series data of the input features to obtain enhanced time series data; the BiLSTM module comprises a feedforward LSTM unit layer, a feedback LSTM unit layer and a linearization unit layer connected in sequence, the feedforward LSTM unit layer and the feedback LSTM unit layer each comprise an LSTM unit, the LSTM unit comprises an input gate, a forget gate, an output gate and a cell state, and the enhanced time series data output by the BiLSTM module is represented as:

[0024]

[0025] wherein, represents the hidden state of the feedforward LSTM unit layer at time step t, represents the hidden state of the feedback LSTM unit layer at time step t.

[0026] The Transformer encoder comprises a multi-head attention layer and a feedforward network layer connected in sequence, the outputs of the multi-head attention layer and the feedforward network layer are each connected with a residual connection and a layer normalization unit, the multi-head attention layer comprises a plurality of parallel single-head attention mechanism heads, the single-head attention mechanism head is configured to take the enhanced time series data output by the BiLSTM module as an input sequence and map the input sequence into an embedding vector, add a position encoding to the embedding vector to form a fusion matrix wherein, is a sequence length, is an embedding dimension, the fusion matrix is multiplied by a query weight matrix , a key weight matrix and a value weight matrix to obtain a query vector , a key vector and a value vector , and the expression is:

[0027]

[0028] The output of the multi-head attention layer is represented as:

[0029]

[0030]

[0031] ​

[0032] in This represents the output of the multi-head attention layer. This represents the output of the i-th single-head attention mechanism head. Indicates the number of heads of attention. , and Let be the query weight matrix, key weight matrix, and value weight matrix of the i-th single-head attention mechanism head, respectively. This represents the output transformation matrix. This represents the transpose of the key vector. Represents the similarity matrix. This represents the scaling factor; Softmax is used to normalize similarity values ​​into probabilities.

[0033] The output of the multi-head attention layer is then passed through a feedforward network layer to output the encoder-generated sequence. ;

[0034] The Transformer decoder includes a hidden attention layer, a cross attention layer, and a feedforward network layer connected in series; the outputs of the hidden attention layer, cross attention layer, and feedforward network layer are all connected to residual connections and layer normalization units; the Transformer decoder is used to generate sequences based on the encoder. The target input Y, including the true SOH value, is processed through a hidden attention layer, a cross attention layer, and a feedforward network layer to generate a prediction result. Prediction results Represented as:

[0035]

[0036] The output module includes the feedforward network layer and the prediction results output by the Transformer decoder. The feedforward network layer of the output module finally outputs the SOH prediction value of the Transformer-BiLSTM model structure.

[0037] Preferably, all neural network layers of the Transformer-BiLSTM model structure are opened when pre-training the pre-built SOH evaluation model using the source training set;

[0038] When fine-tuning the pre-trained SOH evaluation model using the target training set, the linearization unit of the BiLSTM module, the cross-attention layer in the Transformer decoder, and the feedforward network layer of the output layer are enabled, while the remaining neural network layers are frozen.

[0039] Preferably, the expression for the forget gate of the LSTM unit is:

[0040]

[0041] wherein, denotes the time step, denotes the input at the current time step, denotes the hidden state at the previous time step, denotes the weight matrix of the forget gate, denotes the Sigmoid activation function, denotes the bias term, denotes the output of the forget gate;

[0042] The input gate of the LSTM unit is expressed as:

[0043]

[0044]

[0045] wherein, , are the weight matrices of the input gate and the candidate neuron state, respectively, and are the bias terms of the input gate and the candidate neuron state, respectively, denotes the activation function, denotes the scaling parameter, denotes the alternative cell state;

[0046] The cell state of the LSTM unit is updated by merging the outputs of the forget gate and the input gate, and is expressed as:

[0047]

[0048] wherein, is the cell state at the previous time step, is the cell state at the current time step;

[0049] The output gate of the LSTM unit is expressed as:

[0050]

[0051]

[0052] wherein, is the weight matrix of the output gate, is the bias term of the output gate, is the output of the output gate, is the output of the LSTM unit.

[0053] Compared to existing technologies, this invention offers the following advantages: Based on highly correlated charging segment features, this invention constructs source and target training sets according to the source dataset of the source battery and the target dataset of the target battery, respectively. The source training set is used to pre-train a pre-built SOH evaluation model, and the target training set is used to fine-tune the pre-trained SOH evaluation model. Finally, the fine-tuned SOH evaluation model is used to evaluate the SOH of the target battery during its later service life and its subsequent secondary use. This invention achieves full lifecycle evaluation of the SOH value of power batteries with a reasonable training data scale. The effective evaluation range of this invention spans the service life and the secondary use period after retirement of the power battery, taking into account both early prediction of battery SOH and evaluation of battery secondary service capability, effectively contributing to the refined management of the entire lifecycle of power batteries. Attached Figure Description

[0054] Figure 1 A flowchart illustrating a method for assessing the lifespan of an electric vehicle battery based on transfer learning, provided in an embodiment of the present invention.

[0055] Figure 2 The curves showing the correlation between the number of cycles and the SOH value for datasets 1 and 2 collected in this embodiment of the invention;

[0056] Figure 3 This is a statistical diagram of the SOC distribution of the power battery at the charging start time in an embodiment of the present invention.

[0057] Figure 4 This is a schematic diagram of the charging segment features according to an embodiment of the present invention;

[0058] Figure 5 This is a graph showing the correlation coefficients between the VIT and ST features and the SOH of the battery in an embodiment of the present invention.

[0059] Figure 6 The correlation coefficients between CIT and CVT characteristics and battery SOH in embodiments of the present invention;

[0060] Figure 7 This is a structural diagram of the Transformer-BiLSTM model according to an embodiment of the present invention;

[0061] Figure 8 This is a diagram showing the SOH prediction effect of transfer learning throughout the entire lifecycle of an embodiment of the present invention. Detailed Implementation

[0062] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0063] Combination Figure 1This invention provides a method for assessing the lifespan of electric vehicle batteries based on transfer learning, comprising:

[0064] Step S1: Construct a source dataset based on the correlation curves of cycle count and SOH value of multiple source domain batteries during their service life; construct a target dataset based on the correlation curves of cycle count and SOH value of the target battery for which SOH evaluation is required in the early stage of its service life; wherein, the correlation curves of cycle count and SOH value of the multiple source domain batteries are obtained under different preset operating conditions, including ambient temperature, charging current, discharging current and SOH value life end point.

[0065] Specifically, this embodiment conducted multiple battery cycle aging experiments to obtain a dataset of fast charging under various operating conditions. The experimental battery information is detailed in Table 1, corresponding to Dataset 1 (corresponding to the source dataset). Battery samples were placed at three ambient temperatures: -10 °C, 0 °C, and 25 °C, and fast-charged at 3C to 5C. Furthermore, this embodiment also incorporates the publicly available CALCE dataset from the University of Maryland, i.e., Dataset 2 (corresponding to the target dataset), to verify the generalization of the proposed method. The battery number and its corresponding operating condition are shown in Table 2. Figure 2 The correlation curves between the number of cycles and the SOH value for the two datasets are shown respectively.

[0066] Table 1: Battery Dataset

[0067]

[0068] Table 2: Battery Operating Conditions

[0069]

[0070] Step S2: Based on the SOC distribution range at the start of electric vehicle charging, determine the battery voltage range for extracting charging segment features in each cycle.

[0071] Because the initial SOC of electric vehicles is uncertain and does not start from 0%, this example illustrates the distribution of initial SOC of power batteries in three battery swapping stations. The survey data spans one year. Based on the statistics of 27,110 battery swapping transactions, the results are as follows: Figure 3As shown in Dataset 1, the constant current / constant voltage charging conversion point of batteries charged at 3C (15A) at 25℃ is around 80% SOC. According to statistics from 27,110 battery swaps, the scenario where the power battery energy is consumed to below 60% SOC before recharging accounts for a large proportion, reaching 76%. In Dataset 1, the voltage corresponding to 60% SOC in the static state is 3.75V. This means that the characteristic obtained when the battery voltage is above 3.7V has a high probability of occurring in the application scenario of vehicle power batteries, which is suitable for practical applications of SOH prediction. In this embodiment, the battery voltage range finally determined based on the distribution of battery charging voltage is 3.5V-4.2V.

[0072] Step S3: Perform correlation analysis on the pre-selected candidate features and the battery SOH value to determine the charging segment features used for battery SOH value prediction; specifically, the candidate features include the voltage interval time (VIT) of the constant current charging stage, the current interval time (CIT) of the constant voltage charging stage, the constant current-constant voltage transition time (ST), and the constant voltage charging duration (CVT).

[0073] The voltage interval time (VIT) of the constant current charging stage CC is explained in detail, within a relatively small voltage range. Within V, the electrochemical reaction within the lithium-ion battery LIB is stable, and the lithium-ion battery LIB can be approximated as an equivalent capacitance. But be careful. It is not a constant value and varies with SOC. Therefore, within a defined SOC range, when When V is sufficiently small, the amount of electricity obtained is [amount]. Q and voltage range V-relation:

[0074]

[0075] For the constant current charging (CC) phase of the battery, the current For a constant value, V corresponds to Within time t, the following relationship can be obtained:

[0076]

[0077]

[0078] This refers to the voltage interval time, hereinafter referred to as the VIT (Voltage Interval Time) characteristic, such as... Figure 4 As shown in (a). If The choice of V is definite, and I is constant, so we can obtain... and Corresponding relationship:

[0079]

[0080] while The value represents the power performance of the battery, that is, the time of continuously receiving a large current rate, which is related to the SOH of the battery. When the battery ages, its power density decreases, t will shorten. By the change of t, the aging behavior of the battery can be traced, as shown in Figure 4 (c).

[0081] The VIT feature reflects the battery's ability to receive electric charge. Its principle and expression formula are approximately similar to ICA (Incremental Capacity Analysis), but it changes from differential calculation to integral calculation, reducing the influence of sensor sampling noise on the feature. Since ICA processes the entire constant current charging curve of the battery to find peaks to extract SOH-related features. Therefore, it needs to be clear before using VIT in which V interval, taking the VIT feature is more conducive to the prediction of the battery SOH.

[0082] The charging curve of the constant current charging link CC in each cycle is used to extract the VIT feature. A VIT of V is taken every 0.1V between 3.5V - 4.2V, and the interval widths of V are 0.1V, 0.2V, 0.3V, and 0.4V respectively. Obtain the VIT features of the 1 - n cycles at

[0083]

[0084]

[0085]

[0086]

[0087] Specifically describe the current interval time (CIT) of the constant voltage charging link, as shown in Figure 4 (b). Similarly, in a relatively small current interval I, the lithium battery can be approximated as an equivalent capacitor . At a certain SOC, when I is small enough and the voltage V is constant, we can obtain the corresponding relationship between t and :

[0088]

[0089] Here t represents the voltage interval time CIT (Current Interval Time). The value is related to the flatness of the graphite in the battery's negative electrode. As the battery's negative electrode becomes rougher with use, t will be prolonged, especially the duration of low-current charging at the end of constant-voltage charging. This can then be achieved through... Changes in t can track the aging behavior of the battery, such as Figure 4 As shown in (d).

[0090] Similar to VIT characteristics, it is necessary to know which current range is being used. Extracting CIT features from the lower current limit (I) is more beneficial for predicting battery SOH. The constant voltage (CV) cycle of each charge cycle is used to extract CIT features; a specific lower current limit is selected between 0.8-15A to determine the CIT. The lower limits of current I are 0.8A, 1.0A, 2.5A, ..., 10A, 15A. Obtain the corresponding... Under condition I, the CIT features of cycles 1 to N are used for subsequent correlation analysis.

[0091]

[0092]

[0093]

[0094]

[0095]

[0096] This section provides a detailed explanation of the Constant Current-Constant Voltage (CC-CV) Switch Time (ST) and Constant Voltage Charging Time (CVT). The CC-CV Switch Time (ST) is the time from the start of battery charging to the point where the battery transitions from the CC (Constant Current) stage to the CV (Constant Voltage) stage. Similar to the Vitality Indicator (VIT), it reflects the battery's power performance. At the ST moment, the battery has the maximum charging current and the highest voltage, representing the peak charging power under CC-CV charging conditions. As the battery ages, the ST shifts earlier, appearing earlier in the charging phase. When the initial battery charge is relatively fixed, the ST characteristic can be used to predict the battery's State of Harshness (SOH).

[0097] Constant Voltage Time (CVT) is the duration of constant voltage charging for a battery, belonging to the category of CIT (Constant Voltage Injection) characteristics, and is used to reflect the flatness of the battery electrodes. As the battery ages, the surface roughness of the electrodes increases, and the CVT increases. However, the CVT characteristic requires recording the complete constant voltage charging curve, resulting in a relatively long data segment. When battery charging data is relatively complete, the CVT characteristic can be used to predict the battery's State of Harshness (SOH) value.

[0098] To analyze the effectiveness of candidate features, this embodiment analyzes the correlation between candidate features and battery SOH values. Kendall's Tau is a rank-based nonparametric correlation measure used to measure the monotonic relationship (not limited to linear relationship) between two variables. Its core idea is to assess correlation by comparing the order consistency of observation pairs. For cases with duplicate data, Tau-b is generally used, as shown in the following formula:

[0099]

[0100] C: Number of concordant pairs

[0101] D: Number of Discordant Pairs

[0102] T X : The logarithm of repeated rank only on variable X

[0103] T Y : The logarithm of repeated rank only on variable Y

[0104] The results of Kendall correlation coefficient calculation for the candidate features are as follows: Figure 5 and Figure 6 As shown, the closer the correlation coefficient is to 1, the better the feature performs in predicting SOH (corresponding to red in the figure). The closer the correlation coefficient is to 0, the worse the feature performs in predicting SOH (corresponding to blue in the figure). Features with a correlation coefficient below 0.6 are not considered as training features for neural network models (corresponding to gray in the figure). The blank areas in the figure indicate that VIT features cannot be obtained under the corresponding operating conditions.

[0105] Figure 5 The results of the correlation analysis between VIT and ST show that the quality of the VIT characteristics is highly dependent on the selection of the voltage range. (Observation) The feature correlation coefficient is high at V=0.1 V, and high at 3.8 V-3.9 V and 3.9 V-4.0 V. The correlation decreases at lower or higher voltage ranges, especially near the CC-CV transition point, such as 4.1 V-4.2 V, where the feature correlation drops significantly. Therefore, analysis shows that the VIT feature is unsuitable as a model training feature near the battery charging power limit (the CC-CV transition point has the maximum charging power). As the width of V increases, the correlation coefficient between VIT features and SOH also increases. This is also why ST features have a high correlation; ST features can be considered as VIT features with a width of 2.8 V-4.2 V. Furthermore, with... As V increases, the stability of the feature is significantly improved, for example... When V=0.4 V, the correlation coefficient does not fluctuate with the selection of the voltage range. From the perspective of correlation, A larger V value is more suitable as a feature for predicting SOH, but as V increases... As V increases, the time scale of the features also increases. Features from 3.8 V to 4.1 V require 600 seconds, or 10 minutes. In the constant current charging stage, CC cannot guarantee that such long VIT features can be obtained. Therefore, while ensuring the effectiveness of the features, this embodiment selects the VIT features with the shortest possible time as the charging segment features. Given that ST features also have good correlation, ST is added as one of the charging segment features for SOH prediction, which is marked with an orange box in the figure.

[0106] Because the constant current charging stage (CC) is incomplete, the VIT feature may not be obtainable. However, the constant voltage charging stage (CV) occurs at the end of charging and has a higher degree of completeness; therefore, the CIT feature is used as an alternative in this embodiment. Figure 6 The results of the CIT correlation analysis are shown. The overall correlation of CIT features is higher than that of VIT features, and it is significantly correlated with battery temperature conditions. The CIT feature of the lithium-ion battery LIB operating at 25℃ on the right is significantly higher than that of the lithium-ion battery LIB operating at -10℃ on the left. This is because CIT predicts battery aging based on the non-uniformity of the battery electrodes. For the battery, high electrode non-uniformity leads to a longer trickle charging time. However, at low temperatures, poor electrolyte fluidity also causes a longer trickle charging time, which is an interfering factor. Since the CIT feature cannot decouple the two phenomena of electrode aging and decreased electrolyte fluidity, the effectiveness of the CIT feature in predicting the SOH of batteries operating at low temperatures is reduced. The closer the I-interval is to the charging cutoff current, the significantly higher the correlation coefficient of the CIT characteristic. Taking battery #5 as an example, when When the lower bound of interval I is 2.5A (0.5C), the correlation coefficient is 0.9. When the lower bound of interval I is 1 A (0.2C), the correlation coefficient is 0.94. The lower bound of the I interval decreased further from 1A (0.2C) to 0.8A (0.16C), but this did not significantly improve the CIT correlation. However, The upper bound of the I interval has no significant impact on the correlation of CIT features. In order to minimize the time scale of charging segment features, this embodiment selects CIT features of 0.8A-2.5A, which is marked with an orange box in the figure.

[0107] Step S4: Use the charging segment features within the battery voltage range of each cycle as input features, and the SOH value corresponding to that cycle as output label. Construct source training sets and target training sets based on the source dataset and target dataset, respectively. The charging segment features determined after correlation analysis include VIT features, ST features, and CIT features. VIT features represent the time required for the battery charging voltage to increase from 3.8V to 4.0V. ST features represent the time from the start of battery charging to the transition node between constant current charging and constant voltage charging. CIT features represent the time required for the battery charging current to increase from 2.5A to 0.8A.

[0108] Step S5: Pre-train the pre-built SOH evaluation model using the source training set, and fine-tune the pre-trained SOH evaluation model using the target training set; for example... Figure 7 As shown, the SOH evaluation model includes a Transformer-BiLSTM model structure.

[0109] The Transformer-BiLSTM model structure includes a BiLSTM module, a Transformer encoder, a Transformer decoder, and an output module connected in series.

[0110] Among them, LSTM (Long Short-Term Memory) networks perform exceptionally well in processing time-series data, especially in predicting battery capacity evolution over time. LSTM can capture long-term dependencies, which is crucial for predicting battery state-of-the-art (SOH) degradation. A BiLSTM module is used to process the time-series data of the input features to obtain enhanced time-series data; the BiLSTM module includes a feedforward LSTM unit layer, a feedback LSTM unit layer, and a linearization unit layer connected in series. Both the feedforward and feedback LSTM unit layers include LSTM units, and each LSTM unit includes an input gate, a forget gate, an output gate, and a unit state.

[0111] The forget gate controls which information is deleted from the previous memory unit (cell state). The expression for the forget gate is:

[0112]

[0113] in, Indicates a time step. This represents the input at the current time step. This represents the hidden state of the previous time step. The weight matrix represents the forget gate. This represents the Sigmoid activation function. Indicates the bias term. Indicates the output of the forget gate;

[0114] The input gate is used to determine which new information should be stored in the candidate memory cells. The expression for the input gate is:

[0115]

[0116]

[0117] in, , These are the weight matrices for the input gate and the candidate neuron states, respectively. and These represent the bias terms for the input gate and the candidate neuron states, respectively. This represents the activation function. Indicates the proportional parameter. Indicates the status of the candidate cell;

[0118] The cell state is the core of an LSTM cell, determining its long-term memory capacity. The cell state of an LSTM cell is updated by merging the outputs of the forget gate and the input gate, as shown below:

[0119]

[0120] in, It is the cell state of the previous time step. It represents the cell state at the current time step;

[0121] The expression for the output gate of the LSTM unit is:

[0122]

[0123]

[0124] in, It is the weight matrix of the output gate. It is the bias term of the output gate. It is the output of the output gate. It is the output of the LSTM unit.

[0125] The BiLSTM module is a bidirectional architecture based on LSTM. It processes the sequence bidirectionally through two LSTM layers, one forward and one backward, thereby capturing richer contextual information. The BiLSTM module outputs enhanced time-series data. Represented as:

[0126]

[0127] in, This indicates that the feedforward LSTM cell layer is at time step t The hidden state, This indicates that the feedback LSTM cell layer is at time step t The hidden state.

[0128] BiLSTM offers improvements over traditional LSTM networks in capturing long-term dependencies and enhancing accuracy. For example, in predicting battery aging, the battery's capacity decay process is related not only to the current state but also to multiple past time steps; BiLSTM can simultaneously consider these past and future dependencies. In this embodiment, we add a BiLSTM layer before the Transformer encoder to process the initial training features and generate more context-aware representations, which then provide a richer input for subsequent Transformer modules.

[0129] Transformer is a deep learning model based on attention mechanisms, originally used for natural language processing tasks. Its core feature is efficient modeling through self-attention and parallelized encoding and decoding architecture. Self-attention allows the model to focus on words in other positions in the sequence while encoding a word, thereby capturing global dependencies.

[0130] The Transformer encoder comprises a multi-head attention layer and a feedforward network layer connected in series. The outputs of both the multi-head attention layer and the feedforward network layer are connected to residual connections and layer normalization units. Residual connections help prevent gradient vanishing, while layer normalization accelerates training and improves stability. The expression is:

[0131]

[0132] in, This represents the input to the residual connection and layer normalization unit. The layer normalization function is represented. This indicates a residual connection.

[0133] The multi-head attention layer includes multiple parallel single-head attention mechanisms. Each single-head attention mechanism is used to take the enhanced time-series data output by the BiLSTM module as an input sequence and map it into an embedding vector. The embedding vector is then combined with position encoding to form a fusion matrix. ,in It is the sequence length. It is the embedding dimension, the fusion matrix Multiply by the query weight matrix respectively Key weight matrix Sum weight matrix Get the query vector Key vector Sum value vector The expression is:

[0134]

[0135] The output of the multi-head attention layer is represented as follows:

[0136]

[0137]

[0138]

[0139] in This represents the output of the multi-head attention layer. This represents the output of the i-th single-head attention mechanism head. Indicates the number of heads of attention. , and Let be the query weight matrix, key weight matrix, and value weight matrix of the i-th single-head attention mechanism head, respectively. This represents the output transformation matrix. This represents the transpose of the key vector. Represents the similarity matrix. This represents the scaling factor; Softmax is used to normalize similarity values ​​into probabilities.

[0140] The output of the multi-head attention layer is then passed through a feedforward network layer to output the encoder-generated sequence. ;

[0141] The Transformer decoder consists of a hidden attention layer, a cross attention layer, and a feedforward network layer connected in series. The cross attention layer in the Transformer decoder is the core of SOH prediction and is responsible for fusing the battery features of the Transformer encoder with the SOH estimation results of the decoder. The differences in aging patterns of different battery datasets require relearning the attention weights.

[0142] The outputs of the hidden attention layer, cross attention layer, and feedforward network layer are all connected to residual connections and layer normalization units; the Transformer decoder is used to generate sequences based on the encoder. The target input Y, including the true SOH value, is processed through a hidden attention layer, a cross attention layer, and a feedforward network layer to generate a prediction result. Prediction results Represented as:

[0143]

[0144] The output module includes the feedforward network layer and the prediction results output by the Transformer decoder. The feedforward network layer of the output module ultimately outputs the SOH prediction value of the Transformer-BiLSTM model structure.

[0145] The feedforward network layer includes two fully connected layers:

[0146]

[0147] in, This represents the input to the fully connected layer. For the output of the fully connected layer, , These are the first weight matrix and the second weight matrix, respectively. , These are the first output bias and the second output bias, respectively.

[0148] This embodiment employs Bayesian optimization to optimize the model parameters of the Transformer-BiLSTM model structure. Bayesian optimization utilizes a Gaussian process to establish a probabilistic model of the objective function, enabling the prediction of potential optimization regions within a relatively small number of iterations, thereby avoiding wasting computational resources on suboptimal parameter settings. This method not only effectively adjusts multiple sets of hyperparameters in the neural network but also simultaneously reduces the risk of overfitting and improves the model's generalization ability.

[0149] For a detailed explanation of the training process, this embodiment uses transfer learning as a solution to address the problem of limited data in the target domain. Transfer learning allows knowledge to be transferred from a data-rich domain (called the source domain).

[0150] In this embodiment, when pre-training the pre-built SOH evaluation model using the source training set, all neural network layers of the Transformer-BiLSTM model structure are enabled; when fine-tuning the pre-trained SOH evaluation model using the target training set, the linearization unit of the BiLSTM module, the cross-attention layer in the Transformer decoder, and the feedforward network layer of the output layer are enabled, while the remaining neural network layers are frozen.

[0151] First, dataset 1 (as the source training set) was used to verify the generalization ability of the proposed Transformer-BiLSTM model structure to battery operating rate conditions and ambient temperature. Capacity decay curves of each battery sample from 100% SOH to 80% SOH were obtained from dataset 1 (to control experimental variables, only 100% SOH to 80% SOH were used for batteries #7 to #9), and the training and test sets were divided in a 2:3 ratio.

[0152] Table 3: SOH Prediction Results

[0153]

[0154] Table 3 shows the prediction results of the Transformer-BiLSTM model structure for battery SOH, including the minimum root mean square error (RMSE), the RMSE after three rounds of Bayesian optimization, and the SOH prediction error under three temperature conditions. Overall, except for battery #1, the SOH prediction error of the other batteries remains below 2%, which preliminarily proves that the Transformer-BiLSTM model structure proposed in this embodiment has good generalization ability.

[0155] Step S6: Use the fine-tuned SOH evaluation model to evaluate the SOH of the target battery during the later stages of its service life and the secondary utilization period.

[0156] In transfer learning, the SOH evaluation model was pre-trained using dataset 1 and then fine-tuned using early data from dataset 2 (the CALCE public dataset is selected as an application example; in actual applications, the battery data to be evaluated can be directly replaced with dataset 2) representing 103% to 93% of the SOH. (Since the initial battery capacity in dataset 2 is higher than the rated capacity, the calculated SOH is higher than 100%). During fine-tuning, the CIT feature in dataset 2 was adjusted accordingly due to the smaller battery capacity, ranging from 0.05A to 0.25A.

[0157] During model fine-tuning, only the linearization unit of the BiLSTM module, the cross-attention layer in the Transformer decoder, and the feedforward network layer of the output layer are enabled. BiLSTM is used to learn battery time-series aging patterns, but the charging segment duration features vary significantly across different battery models and operating conditions, requiring adjustment of the feature projection dimension. The cross-attention layer in the Transformer decoder, as the core of the framework's time-series prediction, is responsible for fusing the battery features from the encoder with the SOH estimation results from the decoder. Differences in aging patterns across different battery datasets necessitate relearning the attention weights. Finally, the fully connected layer of the feedforward network layer in the output layer must be fine-tuned. The pre-trained SOH output dimension (100%~60% SOH) differs from the prediction target (100%~20% SOH), requiring fine-tuning to adapt to the non-linearity of SOH in different datasets.

[0158] The SOH prediction performance of transfer learning is as follows Figure 8 As shown, using transfer learning techniques, model fine-tuning can be performed using less than 15% of the data (less than 150 cycles) to achieve SOH prediction for the remaining lifetime. The prediction performance shows a MAE of less than 2.3% (optimal result 2.07%) and an RMSE of less than 3.34% (optimal result 2.57%). Figure 8 As shown in (a) to (c), the method proposed in this embodiment can accurately capture the accelerated aging drop point of the battery at 70% SOH, and has good prediction accuracy for the degradation curve near the drop point. It is worth noting that the proposed method also has sufficient identification capability for extremely low SOH ranges, such as... Figure 8 (c) shows the 20% SOH region.

[0159] Figure 8 (d) shows the prediction performance of battery samples in Dataset 2 (target dataset). It is evident that this method possesses extremely high prediction accuracy when the battery SOH > 40%, and its effective prediction window lower limit is significantly better than the existing level of SOH = 60%. The region below 40% SOH contributes the majority of the prediction error for the entire lifecycle SOH. This embodiment, under the premise of reasonably available data scale and quality, extends the effective range of the evaluation method at both ends of the entire lifecycle. Its effective range extends from 93% SOH to 20% SOH, spanning the battery's service life and secondary utilization period, enabling this method to possess both early prediction and secondary service capability assessment capabilities.

[0160] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0161] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0162] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0163] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0164] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A method for assessing the lifespan of electric vehicle batteries based on transfer learning, characterized in that, include: A source dataset was constructed based on the correlation curves between the number of cycles and the state of harm (SOH) value of batteries from multiple source domains during their service life. Construct a target dataset by using the correlation curve between the number of cycles and the SOH value of the target battery in the early stage of its service life, as needed for SOH evaluation; Based on the SOC distribution range at the start of electric vehicle charging, determine the battery voltage range for extracting charging segment features in each cycle. Correlation analysis was performed on pre-selected candidate features and battery SOH values ​​to determine the charging segment features used for battery SOH value prediction; The charging segment features within the battery voltage range of each cycle are used as input features, and the SOH value corresponding to that cycle is used as the output label. Source training set and target training set are constructed based on the source dataset and target dataset, respectively. The pre-built SOH evaluation model is pre-trained using the source training set, and the pre-trained SOH evaluation model is fine-tuned using the target training set. The SOH assessment of the target battery was carried out in the late service life and secondary utilization period using a fine-tuned SOH assessment model. Based on the SOC distribution range at the start of electric vehicle charging, the battery voltage range for extracting charging segment features in each cycle is determined to be 3.5V-4.2V. The charging segment features include: VIT characteristic, representing the time required for the battery charging voltage to increase from 3.8V to 4.0V during constant current charging; ST characteristic, representing the time from the start of battery charging to the transition node between constant current charging and constant voltage charging; CIT characteristic represents the time required for the battery charging current to decrease from 2.5A to 0.8A during the constant voltage charging stage.

2. The method for assessing the full life cycle of electric vehicle batteries based on transfer learning according to claim 1, characterized in that, The cycle number and SOH value correlation curves of the multiple source domain batteries were obtained under different preset operating conditions, including ambient temperature, charging current, discharging current and SOH value life end point.

3. The method for assessing the full life cycle of electric vehicle batteries based on transfer learning according to claim 1, characterized in that, The correlation curve between the number of cycles and the SOH value of the target battery in the early stage of its service life represents the correlation curve between the number of cycles and the SOH value of the target battery from 100% SOH to 90% SOH.

4. The method for assessing the full life cycle of electric vehicle batteries based on transfer learning according to claim 1, characterized in that, Correlation analysis was performed on pre-selected candidate features and battery SOH values ​​based on Kendall correlation analysis.

5. The method for assessing the full life cycle of electric vehicle batteries based on transfer learning according to claim 1, characterized in that, The SOH evaluation model includes the Transformer-BiLSTM model structure.

6. The method for assessing the full life cycle of electric vehicle batteries based on transfer learning according to claim 5, characterized in that, The Transformer-BiLSTM model structure includes a BiLSTM module, a Transformer encoder, a Transformer decoder, and an output module connected in series; wherein, The BiLSTM module processes input feature-rich temporal data to obtain enhanced temporal data. The BiLSTM module comprises a feedforward LSTM layer, a feedback LSTM layer, and a linearization layer connected in series. Both the feedforward and feedback LSTM layers include LSTM units, each containing an input gate, a forget gate, an output gate, and a unit state. The enhanced temporal data output by the BiLSTM module is... Represented as: in, This represents the hidden state of the feedforward LSTM cell layer at time step t. This represents the hidden state of the feedback LSTM cell layer at time step t; The Transformer encoder includes a multi-head attention layer and a feedforward network layer connected in series. The outputs of both the multi-head attention layer and the feedforward network layer are connected to residual connections and layer normalization units. The multi-head attention layer includes multiple parallel single-head attention mechanisms. Each single-head attention mechanism takes the enhanced temporal data output from the BiLSTM module as an input sequence and maps it to an embedding vector. The embedding vector is then augmented with positional encoding to form a fusion matrix. ,in It is the sequence length. It is the embedding dimension, the fusion matrix Multiply by the query weight matrix respectively Key weight matrix Sum weight matrix Get the query vector Key vector Sum value vector The expression is: The output of the multi-head attention layer is represented as follows: in This represents the output of the multi-head attention layer. This represents the output of the i-th single-head attention mechanism head. Indicates the number of heads of attention. , and Let be the query weight matrix, key weight matrix, and value weight matrix of the i-th single-head attention mechanism head, respectively. This represents the output transformation matrix. This represents the transpose of the key vector. Represents the similarity matrix. This represents the scaling factor; Softmax is used to normalize similarity values ​​into probabilities. The output of the multi-head attention layer is then passed through a feedforward network layer to output the encoder-generated sequence. ; The Transformer decoder includes a hidden attention layer, a cross attention layer, and a feedforward network layer connected in series; the outputs of the hidden attention layer, cross attention layer, and feedforward network layer are all connected to residual connections and layer normalization units; the Transformer decoder is used to generate sequences based on the encoder. The target input Y, including the true SOH value, is processed through a hidden attention layer, a cross attention layer, and a feedforward network layer to generate a prediction result. Prediction results Represented as: The output module includes the feedforward network layer and the prediction results output by the Transformer decoder. The feedforward network layer of the output module finally outputs the SOH prediction value of the Transformer-BiLSTM model structure.

7. The method for assessing the full life cycle of electric vehicle batteries based on transfer learning according to claim 6, characterized in that, When pre-training the pre-built SOH evaluation model using the source training set, all neural network layers of the Transformer-BiLSTM model structure are opened. When fine-tuning the pre-trained SOH evaluation model using the target training set, the linearization unit of the BiLSTM module, the cross-attention layer in the Transformer decoder, and the feedforward network layer of the output layer are enabled, while the remaining neural network layers are frozen.

8. The method for assessing the full life cycle of electric vehicle batteries based on transfer learning according to claim 6, characterized in that, The expression for the forget gate of the LSTM unit is: in, Indicates a time step. This represents the input at the current time step. This represents the hidden state of the previous time step. The weight matrix represents the forget gate. This represents the Sigmoid activation function. Indicates the bias term. Indicates the output of the forget gate; The expression for the input gate of the LSTM unit is: in, , These are the weight matrices for the input gate and the candidate neuron states, respectively. and These represent the bias terms for the input gate and the candidate neuron states, respectively. This represents the activation function. Indicates the proportional parameter. Indicates the status of the candidate cell; The cell state of the LSTM unit is updated by merging the outputs of the forget gate and the input gate, as follows: in, It is the cell state of the previous time step. It represents the cell state at the current time step; The output gate expression of the LSTM unit is: in, It is the weight matrix of the output gate. It is the bias term of the output gate. It is the output of the output gate. It is the output of the LSTM unit.

Images