A training and prediction method for battery health status

Abstract

This invention relates to the field of battery technology, and more particularly to a training method and a prediction method for battery state of health (SOH). The training method includes: acquiring training feature vectors of the battery; iteratively training an initial prediction model for the battery's SOH using these training feature vectors until the current total loss function value obtained from the total loss function corresponding to the iterative prediction model is not greater than a total loss function threshold, at which point the current prediction model is used as the target prediction model; during each iteration of training, inputting each training feature vector into an SOH mapping network module to obtain the current predicted SOH value and a hidden state vector; inputting each training feature vector, the current predicted SOH value, and the hidden state vector into a degradation dynamics network module to obtain the physical degradation rate, and obtaining the total loss function value based on the physical degradation rate and the total loss function. The target prediction model obtained through this training method improves the prediction accuracy of the battery's SOH.

Description

Technical Field

[0001] This invention relates to the field of battery technology, and in particular to a method for training and predicting battery health status. Background Technology

[0002] In the pursuit of carbon neutrality, the full electrification of the transportation industry has become an inevitable global trend. The greening process of transportation is no longer limited to the ground but is expanding into all dimensions of sea, land, and air. For example, the electrification of transportation vehicles powered by batteries (such as lithium batteries), including new energy vehicles, electric aircraft, and electric vertical takeoff and landing (eVTOL) aircraft, is rapidly accelerating. However, the operating conditions of air transportation vehicles like eVTOLs are complex: eVTOLs require extremely high power density output during vertical takeoff and landing and face severe load fluctuations when switching between hovering and cruise. These special operating conditions easily accelerate the physicochemical degradation of the battery, resulting in significant internal changes and increased heterogeneity. This means that batteries face more stringent high-energy, high-power, and high-safety standards and requirements than ground-based electric vehicles, and their lifespan also needs to meet higher controllability, reliability, and cycle conditions.

[0003] Battery State of Health (SOH) is a key indicator for measuring battery aging and remaining capacity. Accurate SOH prediction is crucial for extending battery life, optimizing the Battery Management System (BMS), and ensuring driving safety. When a battery's SOH falls below a critical threshold, irreversible internal physicochemical changes are likely to occur. For example, lithium dendrite formation and electrolyte decomposition significantly increase the probability of catastrophic events such as thermal runaway. If the SOH prediction is inaccurate, the BMS may misjudge the battery as healthy even when it is nearing dangerous conditions such as lithium plating, electrolyte decomposition, or micro-short circuits. In aviation scenarios, such misjudgments could directly lead to irreversible thermal runaway during flight, or even cause a crash. Therefore, to meet the higher demands of aviation-grade batteries, there is an urgent need for a prediction method that is both highly accurate and physically explainable and safe.

[0004] Existing methods for predicting battery state of harm (SOH) mainly fall into two categories. The first category is based on physical models, such as equivalent circuit models and electrochemical models. These methods focus on the internal mechanisms of the battery, combining parameters such as internal resistance, voltage, current, and ion concentration to construct complex physicochemical models. These models are used to predict changes within the battery, thereby achieving SOH prediction. While these methods have clear physical meaning and interpretability, their computational complexity is high, with numerous parameters that are prone to dynamic drift. Parameter identification is difficult, they are hard to obtain in practical applications, and they rely heavily on idealized assumptions, making them prone to mismatch with real-world operating conditions. Under complex and variable real-world conditions (such as changes in temperature, state of charge, and drastic fluctuations in charge / discharge rates), the prediction accuracy of SOH significantly decreases, making them unsuitable for eVTOL applications. The second category is based on data-driven machine learning methods (such as CNNs and LSTMs). These methods do not rely on the battery itself to build models but instead predict the battery's SOH using a large amount of battery operating data. Machine learning methods, under conditions of sufficient data, can demonstrate strong nonlinear fitting capabilities and relatively accurate prediction results. However, machine learning models are black-box models, lacking physical constraints and exhibiting poor interpretability. They also require large amounts of high-quality labeled data, limiting their generalization ability and consequently limiting the accuracy of battery SOH prediction. Therefore, existing battery SOH prediction methods still suffer from insufficient prediction accuracy and generalization, making it difficult to meet the higher SOH requirements of eVTOL. Summary of the Invention

[0005] This application provides a training and prediction method for battery health status, which solves the technical problems of insufficient prediction accuracy and generalization in the prior art. It improves the prediction accuracy, stability and robustness of the prediction model, and the prediction model has high physical interpretability and generalization, which can reduce prediction error and improve the generalization ability and credibility of the prediction model under complex actual working conditions. This meets the engineering application requirements of high safety standards and low fault tolerance fields such as electric aircraft.

[0006] In a first aspect, embodiments of the present invention provide a method for training battery health status, comprising:

[0007] Obtain the training feature vector for each battery, which includes: global features, voltage range features, higher-order statistical features, and time features;

[0008] The initial prediction model for the state of health (SOH) of the battery is iteratively trained using each of the training feature vectors until the current total loss function value obtained by the total loss function corresponding to the iterative prediction model is not greater than the total loss function threshold. Then, the current prediction model is taken as the target prediction model. The prediction model includes: an SOH mapping network module and a degradation dynamics network module.

[0009] In each iteration of training, each training feature vector is input into the SOH mapping network module to obtain the current SOH prediction value and the hidden state vector. The SOH mapping network module includes a feature segmenter, a Transformer module and a first LSTM module connected in sequence.

[0010] Each training feature vector, the current predicted value of SOH, and the hidden state vector are input into the degradation dynamics network module to obtain the physical degradation rate. Based on the physical degradation rate and the total loss function, the total loss function value is obtained. The degradation dynamics network module includes a second LSTM module.

[0011] Optionally, obtaining the training feature vector for each battery includes:

[0012] Obtain the raw data for each of the batteries;

[0013] Each of the original data is subjected to rest time constraint processing, state of charge interval constraint processing, monotonicity constraint processing, and outlier constraint processing to obtain the effective charging segment for each battery. The rest time constraint processing selects the rest segment where the rest time of the battery before charging exceeds a preset time. The state of charge interval constraint processing selects the charging segment within a set state of charge interval of the battery during the charging process. The set state of charge interval is a state of charge interval that satisfies the following conditions: the initial state of charge is not greater than a first state of charge threshold, the final state of charge reaches a second state of charge threshold, and the charging state of the battery is marked as having met the marking conditions. The monotonicity constraint processing selects the charging segment where the state of charge increases monotonically. The outlier constraint processing removes the charging segment where the state of charge of the battery is abnormal during the charging process.

[0014] The SOH tag of the battery is obtained based on the effective charging segment;

[0015] The global features, voltage range features, and higher-order statistical features are extracted from the effective charging segments of each battery to obtain the training feature vector of each battery. The global features include: the sequence number, maximum voltage value, peak current value, average current value, peak temperature, average temperature, and minimum temperature in the charging segment. The voltage range features include: the voltage value and capacity within a set stable voltage range. The higher-order statistical features include: the kurtosis, skewness, and slope of the voltage and current relationship curve in the charging segment.

[0016] Optionally, the step of extracting the global features, voltage range features, and higher-order statistical features from the effective charging segments of each battery to obtain the training feature vector for each battery includes:

[0017] Map the timestamps of the effective charging segments to time features;

[0018] The global features, voltage range features, higher-order statistical features, and time features are concatenated along the feature dimension to obtain the training feature vector.

[0019] Optionally, the step of inputting each of the trained feature vectors into the SOH mapping network module to obtain the current predicted value of SOH and the hidden state vector includes:

[0020] The training feature vector is segmented and embedded using the feature segmenter to obtain a learnable feature vector.

[0021] The Transformer module performs global feature interaction processing on the learnable feature vector to obtain an updated feature vector. The Transformer module includes: The Transformer encoder consists of two layers, with adjacent Transformer encoders connected by a residual layer. Each Transformer encoder layer includes a multi-head attention mechanism and a feedforward neural network, as well as a normalization layer connected to the multi-head attention mechanism and a normalization layer connected to the feedforward neural network.

[0022] The updated feature vector is predicted using the first LSTM module to obtain the current predicted value of SOH and the hidden state vector.

[0023] Optionally, the step of performing word segmentation and embedding processing on the trained feature vector through the feature segmenter to obtain a learnable feature vector includes:

[0024] Assign independent weight vectors and bias vectors to each feature in the training feature vector to obtain the feature embedding vector, as shown in the following formula:

[0025] ;

[0026] in, For the first The feature embedding vectors are, and the training feature vectors are... , for One characteristic, The first of the training feature vectors Features , It is the first The weight vector of each feature. It is the first The bias vectors corresponding to each feature, where R is the real number field;

[0027] A learnable classification token is appended to the beginning of the feature embedding vector. The learnable feature vector is obtained. .

[0028] Optionally, the step of performing prediction processing on the updated feature vector through the first LSTM module to obtain the current predicted value of SOH and the hidden state vector includes:

[0029] Extract the updated classification token from the updated feature vector. d represents the d-dimensional dimension;

[0030] Based on the updated classification token, the current predicted value of SOH is obtained as shown in the following formula:

[0031] ;

[0032] in, The current predicted value of SOH is given by MLP, where MLP is a multilayer perceptron and LayerNorm is a normalization layer.

[0033] The updated classification token This serves as the implicit state vector.

[0034] Optionally, the step of inputting each of the training feature vectors, the current predicted value of SOH, and the hidden state vector into the degradation dynamics network module to obtain the physical degradation rate includes:

[0035] Through formula The physical degradation rate is obtained. Where G is the physical evolution function, Let be the hidden state vector, and t be the time feature. These are the internal weights of the LSTM model of the degenerate dynamics network.

[0036] Optionally, obtaining the total loss function value based on the physical degradation rate and the total loss function includes:

[0037] Through formula The total loss function value is obtained. ,in, For data loss values, This represents the physical and dynamic loss value. The loss value is the monotonicity constraint value of SOH;

[0038] Through formula The data loss value is obtained. ;

[0039] in, This represents the total number of training feature vectors input to the SOH mapping network module each time. Let be the SOH prediction value of the j-th training feature vector out of the total number of training feature vectors input each time. This is the true SOH value of the j-th training feature vector out of the total number of training feature vectors input each time.

[0040] Optionally, via formula The physical and dynamic loss value is obtained. ;in, Let be the derivative of the SOH prediction value of the j-th training feature vector out of the total number of training feature vectors input each time. The physical degradation rate is the value of the j-th training feature vector out of the total number of training feature vectors input each time.

[0041] Through formula The monotonicity constraint loss value of SOH is obtained. ; where ReLU is the activation function.

[0042] Based on the same inventive concept, in a second aspect, the present invention also provides a method for predicting battery health status, comprising:

[0043] Acquire the target prediction model and the test data of the battery, wherein the target prediction model is a model trained by the battery health state training method as described in the first aspect.

[0044] The test data is subjected to resting time constraint processing, state of charge interval constraint processing, monotonicity constraint processing and outlier constraint processing respectively to obtain the effective charging segment of the battery;

[0045] Global features, voltage range features, and higher-order statistical features are extracted from the effective charging segment to obtain a feature vector;

[0046] The feature vector is input into the target prediction model to obtain the predicted SOH value of the battery.

[0047] One or more technical solutions in the embodiments of the present invention have at least the following technical effects or advantages:

[0048] This embodiment employs the FT-Transformer module to construct the SOH mapping network module. The FT-Transformer module introduces a Feature Tokenizer (FT) to independently map the heterogeneous physical parameters of the battery into high-dimensional semantic tokens. It utilizes an attention mechanism to deeply learn the nonlinear interaction weights between battery features, thereby significantly enhancing the battery health prediction model's ability to express complex heterogeneous data features. Simultaneously, the SOH mapping network module directly outputs the current predicted SOH value and the deep hidden state vector. The current predicted SOH value and the hidden state vector are then input into a degradation dynamics network module based on LSTM (Long Short-Term Memory). This network module does not directly predict the SOH value but utilizes the LSTM's long short-term memory gating mechanism to accurately capture the long-term temporal dependencies during battery aging, specifically for fitting the physical degradation rate (i.e., the time derivative) of the battery's SOH over time.

[0049] Then, a composite loss function, namely the total loss function, is constructed, which includes the residuals of the physical partial differential equations (PDEs) and monotonicity constraints. This total loss function enforces a strict consistency between the output derivative of the preceding SOH mapping network module and the physical degradation rate of the subsequent degradation dynamics network module. Unlike models that simply stack networks and directly output SOH predictions, this model mechanism, through feedback adjustment of the total loss function, allows the deep learning prediction model to be explicitly guided by the physical dynamics equations during gradient descent. This enhances the ability to deeply analyze the heterogeneous characteristics of the battery while endowing the prediction model with the ability to obey physical evolution laws. This enables the prediction model to learn efficiently from degradation data, capture time decay dynamics, and comply with physical laws, thereby significantly improving the accuracy, stability, and robustness of the battery's SOH prediction values. Thus, the battery health state prediction model trained using the method of this embodiment, i.e., the target prediction model, improves the prediction accuracy, stability, and robustness of the battery health state, thereby enhancing the accuracy, precision, and reliability of the battery health state prediction results. Meanwhile, the prediction model for battery health status has high physical interpretability and generalization ability, which can reduce prediction errors and improve the generalization ability and reliability of the prediction model under complex actual working conditions. This enables the SOH prediction value of the prediction model to meet the engineering application requirements of fields with high safety standards and low fault tolerance, such as electric aircraft. Attached Figure Description

[0050] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0051] Figure 1 A flowchart illustrating the steps of a battery health status training method according to an embodiment of the present invention is shown.

[0052] Figure 2 A schematic diagram of the structure of the battery health state training method in an embodiment of the present invention is shown;

[0053] Figure 3 The loss curves of the traditional PINN model and the PINN model in this embodiment of the invention are shown.

[0054] Figure 4 The figure shows the fitting curves of the predicted SOH value and the actual SOH value of the conventional PINN model in an embodiment of the present invention.

[0055] Figure 5The figure shows the fitting curve of the SOH predicted value and the actual SOH value of the PINN model in the embodiment of the present invention.

[0056] Figure 6 A comparison chart of various evaluation metrics between the traditional PINN model and the PINN model in this embodiment of the invention is shown.

[0057] Figure 7 A flowchart illustrating the steps of a battery health state prediction method according to an embodiment of the present invention is shown. Detailed Implementation

[0058] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0059] Example 1

[0060] The first embodiment of the present invention provides a training method for battery health status, such as... Figure 1 and Figure 2 As shown, it includes:

[0061] S101, Obtain the training feature vector for each battery. The training feature vector includes: global features, voltage range features, higher-order statistical features, and time features.

[0062] S102, through each training feature vector, the initial prediction model of the battery health state SOH is iteratively trained until the current total loss function value obtained by the total loss function corresponding to the iterative prediction model is not greater than the total loss function threshold. Then, the current prediction model is used as the target prediction model. The prediction model includes: SOH mapping network module and degradation dynamics network module.

[0063] S103, in each iteration of training, each training feature vector is input into the SOH mapping network module to obtain the current SOH prediction value and hidden state vector. The SOH mapping network module includes: a feature segmenter, a Transformer module and a first LSTM module connected in sequence.

[0064] S104, each training feature vector, the current predicted value of SOH, and the hidden state vector are input into the degradation dynamics network module to obtain the physical degradation rate, and the total loss function value is obtained based on the physical degradation rate and the total loss function. The degradation dynamics network module includes: a second LSTM module.

[0065] It should be noted that the battery's State of Health (SOH) stands for State of Health. The training method for determining the battery's state of health can run on personal computers and other electronic devices, neural network processing units (NPUs), or chips. The batteries used in this embodiment include, but are not limited to, lithium batteries, lead-acid batteries, nickel-metal hydride batteries, sodium-ion batteries, and zinc-manganese batteries.

[0066] This embodiment employs the FT-Transformer module to construct the SOH mapping network module. The FT-Transformer module introduces a Feature Tokenizer (FT) to independently map the heterogeneous physical parameters of the battery into high-dimensional semantic tokens. It utilizes an attention mechanism to deeply learn the nonlinear interaction weights between battery features, thereby significantly enhancing the battery health prediction model's ability to express complex heterogeneous data features. Simultaneously, the SOH mapping network module directly outputs the current predicted SOH value and the deep hidden state vector. The current predicted SOH value and the hidden state vector are then input into a degradation dynamics network module based on LSTM (Long Short-Term Memory). This network module does not directly predict the SOH value but utilizes the LSTM's long short-term memory gating mechanism to accurately capture the long-term temporal dependencies during battery aging, specifically for fitting the physical degradation rate (i.e., the time derivative) of the battery's SOH over time.

[0067] Then, a composite loss function, namely the total loss function, is constructed, which includes the residuals of the physical partial differential equations (PDEs) and monotonicity constraints. This total loss function enforces a strict consistency between the output derivative of the preceding SOH mapping network module and the physical degradation rate of the subsequent degradation dynamics network module. Unlike models that simply stack networks and directly output SOH predictions, this model mechanism, through feedback adjustment of the total loss function, allows the deep learning prediction model to be explicitly guided by the physical dynamics equations during gradient descent. This enhances the ability to deeply analyze the heterogeneous characteristics of the battery while endowing the prediction model with the ability to obey physical evolution laws. This enables the prediction model to learn efficiently from degradation data, capture time decay dynamics, and comply with physical laws, thereby significantly improving the accuracy, stability, and robustness of the battery's SOH prediction values. Thus, the battery health state prediction model trained using the method of this embodiment, i.e., the target prediction model, improves the prediction accuracy, stability, and robustness of the battery health state, thereby enhancing the accuracy, precision, and reliability of the battery health state prediction results. Meanwhile, the prediction model for battery health status has high physical interpretability and generalization ability, which can reduce prediction errors and improve the generalization ability and reliability of the prediction model under complex actual working conditions. This enables the SOH prediction value of the prediction model to meet the engineering application requirements of fields with high safety standards and low fault tolerance, such as electric aircraft.

[0068] Below, in conjunction with Figure 1 and Figure 2 The following details the specific implementation steps of the battery health status training method provided in this embodiment:

[0069] First, step S101 is executed to obtain the training feature vector for each battery. The training feature vector includes: global features, voltage range features, higher-order statistical features, and time features.

[0070] Specifically, step S1011 is executed first to obtain the raw data for each battery. The raw data for the battery is the battery charge and discharge dataset. The raw data stream includes mixed charge and discharge data under various operating conditions such as resting, intermittent charging, and short-distance driving. The charge and discharge data includes, but is not limited to: charging voltage, charging current, charging temperature, state of charge (SOC), discharging voltage, discharging current, discharging temperature, and resting time.

[0071] Next, step S1012 is executed, where resting time constraint processing, state of charge interval constraint processing, monotonicity constraint processing, and outlier constraint processing are applied to each piece of raw data to obtain the effective charging segments for each battery. The number of effective segments for a battery is several, and the specific number can be set according to actual needs. Specifically, the resting time constraint processing selects resting segments where the battery's resting time before charging exceeds a preset time. The state of charge interval constraint processing selects charging segments within a set state of charge interval during the battery's charging process. The set state of charge interval is defined as a range where the initial state of charge is no greater than a first state of charge threshold, the final state of charge reaches a second state of charge threshold, and the battery's charging state is marked as having met the marking conditions. The monotonicity constraint processing selects charging segments where the state of charge increases monotonically, and the outlier constraint processing removes charging segments where the battery's state of charge is abnormal during charging.

[0072] Specifically, in step S1012, the raw data undergoes rest time constraint processing: rest segments with a rest time exceeding a preset time (e.g., 1 hour) before battery charging begins are selected. The preset time can be set according to actual needs. The principle of rest time constraint processing is that, due to concentration polarization and electrochemical polarization phenomena inside the battery when it has just finished discharging or when the load changes drastically, the measured voltage of the battery deviates from the true open circuit voltage (OCV). Setting a long rest time preset aims to ensure that the diffusion of chemical substances inside the battery reaches a balanced state, eliminate the falsely high or low voltage caused by polarization effects, thereby ensuring that the values ​​of the battery charging segments are as close as possible to the actual situation of the battery, improve the accuracy of the battery's initial state data, and thus accurately extract the various parameter characteristics of the battery.

[0073] The raw data undergoes state of charge (SOC) range constraint processing: During battery charging, a charging segment within a defined SOC range is selected. The SOC range is defined as a range that satisfies the following conditions: the initial SOC is no greater than a first SOC threshold, the final SOC reaches a second SOC threshold, and the battery's charging status is marked as complete. The first SOC threshold, second SOC threshold, and charging status marker can be set according to actual needs. For example, during battery charging, a charging segment within this SOC range can be selected where the initial SOC is no greater than 60% of the first SOC threshold, the final SOC reaches 100% of the second SOC threshold, and the charging status is marked as complete. This charging segment is also considered a fully charged segment.

[0074] The principle behind the state-of-charge (SOC) range constraint processing is that the battery's OCV curve typically shows a plateau in the near-fully charged (high SOC) region, with a small voltage change rate, making it difficult to capture rich aging information. Longer data segments, including the transition from constant current (CC) to constant voltage (CV), can more effectively reflect changes in battery internal resistance and capacity decay characteristics. Therefore, selecting the portion of the battery with a small SOC change contains less information and is less meaningful. However, selecting a charging range with a larger span (covering at least 60%-100% of the charging process's SOC) ensures the completeness and richness of the battery's feature vectors while effectively eliminating data fluctuations caused by inconsistent charging depths. This significantly improves the information entropy of the battery's feature vectors, ensuring the accuracy and precision of various battery parameters, thereby improving the accuracy and precision of the prediction model.

[0075] Monotonicity constraints are applied to the raw data: During battery charging, the State of Charge (SOC) of the battery maintains a monotonically increasing relationship. For example, if the SOC of a battery initially increases and then decreases, or decreases or remains unchanged during charging, it indicates that the charging segment is an incorrect charging segment and needs to be filtered out. Introducing a monotonically increasing SOC constraint into the selection of battery charging segments improves the quality of the battery's feature vectors during the training of the prediction model, ensuring that there are no erroneous features in the feature vectors, preventing the prediction model from learning incorrect features, and thus improving the accuracy, stability, and robustness of the prediction model.

[0076] Outlier constraint processing is applied to the raw data: charging segments with abnormal state of charge (SOC) during the charging process are identified. For example, if the SOC of the battery is detected to be null, outside the 0-100% range, or stagnant for an extended period in a non-fully charged state, it indicates a sensor malfunction or an estimation error in the BMS (Battery Management System). Such charging segments are then treated as invalid and directly discarded to ensure the physical consistency of the battery's training data, guarantee the accuracy and precision of various battery parameters, and ultimately improve the accuracy and precision of the prediction model.

[0077] By processing the original battery data under the above four constraints, the effective charging segments of the battery are selected, ensuring the physical consistency of the battery training data and guaranteeing the accuracy and precision of the various parameters and features of the battery. This enables the prediction model to learn based on high-quality battery training data, thereby improving the accuracy, precision, stability, and robustness of the prediction model.

[0078] Next, step S1013 is executed to obtain the SOH tag of the battery based on the valid charging segment.

[0079] Specifically, in step S1013, the SOH tag of the battery is obtained based on the effective charging segment of the battery. Specifically, the effective charging segment of the battery will automatically have an SOH tag. For example, the SOH tag of the effective charging segment of the battery is generated by the BMS. The BMS acquires data such as battery voltage, current, and SOC, and automatically compares this data with pre-set OCV (Open Circuit Voltage) curve data to generate the SOH tag. If the effective charging segment of the battery lacks an SOH tag, the actual charging capacity of the effective charging segment is used as the current maximum available capacity using the ampere-hour integration method, and then the SOH tag is calculated.

[0080] For example, a single data point about a battery (i.e., a charging segment) contains multiple feature data such as charging voltage, current, and temperature at a certain moment, where the SOH tag is defined as the current maximum available capacity. With factory rated capacity The ratio. For the data corresponding to time A in an effective full-charge segment, the current maximum usable capacity of the battery at time A. By controlling the charging current In time The integral is obtained from the integral, and the calculation is shown in formula (1).

[0081] (1);

[0082] in, and These are the start and end times of the effective charging segment, respectively. The charging time for this charging segment. Afterwards, the SOH tag of the charging segment at time A. The calculation formula is:

[0083] (2);

[0084] As can be seen, the multi-feature data of the battery at each time point needs to be labeled using the ampere-hour integral method based on the time and current features in the data. The ampere-hour integral method can quantify the current capacity retention rate of the battery at the physical level, providing high-precision labels for subsequent model learning and training.

[0085] Then, step S1014 is executed to extract global features, voltage range features, and higher-order statistical features from the effective charging segments of each battery, obtaining the training feature vector for each battery. The global features include: the sequence number, maximum voltage value, peak current, average current, peak temperature, average temperature, minimum temperature, and cumulative mileage (mileage data is taken from the batteries of electric vehicles) within the charging segment. Voltage range features include: voltage values ​​and capacities within a set stable voltage range, such as the mean, median, standard deviation, and range of cumulative capacitance within the set stable voltage range; the mean, median, standard deviation, and range of the maximum and minimum single-cell voltage difference sequence (for battery packs) or the voltage value (for non-battery packs). Higher-order statistical features include: the kurtosis, skewness, and slope of the voltage-current relationship curve within the charging segment. For example, the voltage kurtosis, current kurtosis, voltage skewness, current skewness, voltage entropy, current entropy, voltage slope, and current slope of the curve.

[0086] Specifically, in step S1014, the timestamps of effective charging segments are mapped to time features. Global features, voltage range features, higher-order statistical features, and time features are concatenated along the feature dimension to obtain a training feature vector. It should be noted that one training feature vector can be extracted from one effective charging segment of a battery.

[0087] To fully exploit the multi-source heterogeneous feature information during battery degradation, this embodiment extracts three types of features with strong health indicators from each effective charging segment and constructs a training feature vector with standardized time characteristics. The three types of indicative features are global features, voltage range features, and high-order statistical features. Global features are key charging parameters used to capture the macroscopic aging performance of a single charging process. Global features include: maximum voltage value, current value, peak temperature, etc. To learn the battery's charging characteristics in a fine-grained manner, voltage range features select parameters from a specific stable voltage range (e.g., a voltage range of 3.8V to 4.2V). The set stable voltage range can be set according to actual needs. Key indicators within this range, such as voltage value and capacity, are calculated. High-order statistical features are high-level statistics of the voltage / current relationship curve, such as the kurtosis, skewness, and slope of the curve.

[0088] To accommodate the input requirements of subsequent time-series prediction models and preserve evolutionary information over time, this embodiment constructs the extracted features into a fixed-length sequence as input to the subsequent prediction model. First, the original absolute timestamps in each charging segment are mapped to relative time series, thereby introducing a discrete time feature variable that monotonically increases with the charging process. , time variable As an independent feature dimension besides physical quantities such as voltage and current, it explicitly informs the prediction model of the temporal position of each sampling point in the charging process. Then, equal-step resampling is performed, with a fixed sequence length set. The three types of features extracted above are then compared with discrete time variables. The features are combined to form a shape. The standardized feature tensor (where, The feature tensor (i.e., the total dimension of the features) serves as the training feature vector, which is then used as input to the subsequent prediction model. Each dimension of this feature tensor includes time features, as well as global features, voltage range features, and higher-order statistical features corresponding to the time features (i.e., time points). This process ensures the consistency of the input dimensions of the prediction model and helps the prediction model accurately capture the long-term dependencies of battery parameters over time. This allows the prediction model to efficiently learn the physical characteristics of the battery, accurately predict the SOH (State of Health) value, and improve the accuracy, stability, robustness, and generalization of the SOH prediction value.

[0089] For the battery charging and discharging dataset acquired in real-world scenarios, this embodiment employs a unique feature extraction and preprocessing method based on the battery's operating characteristics. Given that the original data is mixed data from various operating conditions and contains sensor noise interference, direct use would severely impact the accuracy of the prediction model. Therefore, this embodiment proposes a set of screening criteria to filter the original dataset, extracting effective charging segments from the massive dataset. The training feature vectors of the battery are then extracted from these effective segments, ensuring the accuracy and precision of the battery's training feature vectors. This allows the prediction model to learn based on high-quality battery training data, improving the prediction model's accuracy, precision, stability, robustness, and generalization ability.

[0090] Next, step S102 is executed, using each training feature vector to iteratively train the initial prediction model for the battery's state of health (SOH) until the current total loss function value obtained by the iterative prediction model is not greater than the total loss function threshold. At this point, the current prediction model is used as the target prediction model. For example, Figure 2 As shown, the prediction model includes: SOH mapping network module and degradation dynamics network module The training process of each iteration of the prediction model is described in detail in steps S103 and S104.

[0091] This embodiment's prediction model abandons the traditional unidirectional feature stacking mode and constructs a cascaded gray-box prediction architecture of "state mapping-dynamic constraints" coupled with physical information. This embodiment's prediction model not only solves the technical problems of insufficient prediction accuracy and generalization in existing technologies for battery health state, but also further addresses the issues of low interpretability and lack of physical constraints in complex and high-safety application scenarios such as eVTOL. Furthermore, to address the shortcomings of existing technologies in modeling complex feature interactions and long-term temporal dependencies in SOH prediction, an enhanced PINN prediction model is proposed.

[0092] Unlike existing methods that commonly involve simply stacking or directly cascading multiple different networks, this embodiment uses two network modules (i.e. and (Through the first network module) The output SOH prediction value and the hidden state are connected together. Furthermore, this is achieved through two network modules (i.e....) and The joint loss function (i.e., the total loss function) is fed back. and Further optimization of the output of these two network modules. The first network module... and the second network module Both will output SOH prediction values, the first network module The output SOH prediction value is mainly used to solve for the degradation rate, which guides the second network module. The learning process. Ultimately, the SOH prediction value output by the prediction module is in the first network module. Building upon this foundation, the study further explored the dynamic laws of degradation, improving the prediction accuracy, stability, and robustness of the battery health status prediction model. Simultaneously, the battery health status prediction model exhibits high physical interpretability and generalization ability, reducing prediction errors and enhancing its generalization capability and reliability under complex real-world conditions. This meets the engineering application requirements of high-safety-standard, low-fault-tolerance fields such as electric aircraft.

[0093] During each iteration of training, such as Figure 2 As shown, in step S103, each training feature vector is input into the SOH mapping network module to obtain the current predicted value of SOH and the hidden state vector. The SOH mapping network module includes a feature segmenter, a Transformer module and a first LSTM module connected in sequence.

[0094] Specifically, As a solution function, it is responsible for learning from the input feature sequence. (i.e., training feature vectors) and time To SOH prediction value This implementation employs a complex nonlinear mapping to address the feature semantic space mismatch problem caused by the simple concatenation of heterogeneous features in traditional models. It utilizes a Feature Tokenizer Transformer (FT-Transformer) architecture to independently project continuous heterogeneous physical quantities into a high-dimensional manifold space. A multi-head attention mechanism is then used to uncover the nonlinear coupling relationships between features, ultimately achieving adaptive weighted fusion of heterogeneous features to map strongly nonlinear SOH prediction values. Through this network module, the prediction model can learn the nonlinear mapping relationship between heterogeneous features and SOH prediction values, automatically identify the importance of features, and adaptively adjust the weights.

[0095] like Figure 2 As shown, step S1031 is executed first, and the training feature vector is segmented and embedded by the feature segmenter FT to obtain a learnable feature vector, so as to perform heterogeneous feature segmentation and high-dimensional embedding processing on the training feature vector.

[0096] Assume the training feature vector is R is the real number field, and the training feature vectors include This embodiment introduces a heterogeneous physical feature segmenter, FT, to input the training feature vector X directly into the fully connected layer, unlike existing technologies that directly input the training feature vector X into the fully connected layer. FT provides scalar features for each input feature vector. Assign independent learning weights and biases. For the th element in the input training feature vector... scalar of features It is mapped to a higher dimension through a specific linear transformation. The feature embedding vector of dimension 1, the linear projection formula is shown in formula (3):

[0097] (3);

[0098] in, For the first There are feature embedding vectors, and the training feature vectors are . , for One characteristic, To train the th feature vector Features , It is the first The weight vector of each feature. It is the first The bias vector corresponding to each feature.

[0099] This step transforms feature scalars with completely different physical meanings and dimensions into those within the same high-dimensional semantic space. dimensional vector This enables the predictive model to learn a unique representation for each feature and effectively handle differences in scale and distribution.

[0100] To aggregate global information for subsequent regression prediction, a learnable classification token (CLS Token) is appended to the beginning of the feature embedding vector, denoted as... The learnable feature vector is obtained, as shown in formula (4). Here, the classification token is... Used as an aggregator for global information.

[0101] (4);

[0102] in, These are learnable feature vectors. Learnable feature vectors contain... Each token has the following dimensions: .

[0103] Next, step S1032 is executed, where the learnable feature vectors undergo global feature interaction processing via the Transformer module to obtain updated feature vectors. The Transformer module includes: The Transformer encoder consists of layers, with adjacent Transformer encoder layers connected by residual layers. Each Transformer encoder layer includes a multi-head attention mechanism and a feedforward neural network, as well as a normalization layer connected to the multi-head attention mechanism and a normalization layer connected to the feedforward neural network.

[0104] Specifically, learnable feature vectors The input is fed into the Transformer module. The Transformer module uses a multi-head attention mechanism to... Perform global feature interaction. The input consists of a feature interaction layer composed of L+1 stacked Transformer encoder layers. Each encoder layer includes a connected multi-head self-attention (MSA) mechanism and a feed-forward neural network (FFN), optimized through residual connections and normalization layers.

[0105] In each layer of the Transformer encoder, such as the first... In the L+1 layer Transformer encoder, a multi-head attention mechanism is used to calculate the correlation strength between different physical features. For the L+1 layer... Each attention point calculates the query. Matrix, Key Matrix, Value The matrix is ​​shown in formula (5):

[0106] (5);

[0107] in, For the first The L+1 layer Transformer encoder The weights of the query matrix for each attention head. No. The first layer of the Transformer encoder The weights of the key matrix of each attention head. No. The first layer of the Transformer encoder The weights of the value matrix of each attention head.

[0108] Subsequently, the attention score matrix is ​​obtained, and feature interaction calculation is performed to quantify the features using the matrix. Features The influence weights are then normalized using the softmax function to obtain the attention weight matrix, as shown in formula (6):

[0109] (6);

[0110] Where, d a Indicates the first A scaling factor for the attention head is used to prevent gradient vanishing and ensure training stability; softmax is the activation function.

[0111] The output of this layer is obtained through concatenation and linear mapping using a multi-head attention mechanism, and then a nonlinear transformation is performed through a feedforward network to obtain the updated feature vector. .

[0112] Through the feature fusion and update process of the Transformer module, classification tokens are made possible. The Transformer encoder gradually aggregates global information most relevant to battery aging from all physical features. Internally, it captures complex interactions and nonlinear dependencies between features through a multi-head attention mechanism and a feedforward neural network. The multi-head attention mechanism allows the prediction model to dynamically weigh the importance of all other features when generating a representation for each feature, thus achieving robust global information fusion. Through this mechanism, the prediction model can dynamically identify and assign higher weights to key features (such as voltage changes during constant current charging) and suppress interference from irrelevant features. Compared to the simple feature stitching of traditional PINN fully connected layers, this mechanism significantly improves feature utilization, enabling the prediction model to maintain accurate predictions even with slight fluctuations in some sensor data. This significantly improves the accuracy, precision, stability, robustness, and generalization of the prediction model, ensuring the accuracy, precision, and reliability of the SOH prediction values.

[0113] Then, step S1033 is executed, and the updated feature vector is predicted through the first LSTM module to obtain the current predicted value of SOH and the hidden state vector.

[0114] Specifically, in each Transformer encoder layer, the updated classification token is extracted from the updated feature vector. This allows us to obtain the updated classification token output by the final Transformer encoder layer. Specifically, after... After processing by the Transformer encoder layers, the updated feature vector is output from the final Transformer encoder layer. Extract the updated classification token As a deep state representation of the entire battery.

[0115] Based on the updated classification token, the current predicted value of SOH is obtained. Specifically, the value is... Input the SOH predictor, which is a simple LSTM (Long Short-Term Memory) network (i.e., the first LSTM module), and output the SOH estimate at the current time step, i.e., the current SOH prediction value:

[0116] (7);

[0117] in, is the current predicted value of SOH, MLP is a multilayer perceptron, and LayerNorm is a normalization layer.

[0118] Updated category tokens As a hidden state vector.

[0119] In the process of obtaining the current predicted value and hidden state vector of SOH, not only is the following utilized... Predicting the current predicted value of SOH will also involve this high-dimensional hidden state vector. and the current predicted value of SOH This information is then passed to the subsequent degradation dynamics network module. Unlike traditional methods that directly discard intermediate features, this design in this embodiment ensures that the subsequent LSTM network (i.e., the second LSTM network module) not only uses the SOH prediction value from the SOH mapping network module at time t, but also... Also through By understanding the underlying physical characteristics that lead to the current predicted SOH value, the degradation dynamics network module can deduce the degradation rate based on these underlying physical characteristics.

[0120] This embodiment introduces a feature segmenter (FT) into the SOH mapping network module. Compared to methods that directly concatenate multi-dimensional features and input them into an MLP or CNN, this offers advantages in format alignment and dynamic interaction. Battery operating data is inherently multi-source and heterogeneous (e.g., voltage is a continuous variable, temperature is a slow variable, time is a monotonically increasing variable, and driving range and other features have different physical meanings). Existing technologies simply normalize these data and concatenate them into a long vector before inputting it into an MLP or CNN, which can lead to confusion in physical meaning.

[0121] The feature segmenter FT in this embodiment assigns independent embedding weights to each scalar feature. This approach projects data with different physical properties onto a unified high-dimensional space, assigning an independent token to each physical feature. This allows the prediction model to process heterogeneous information in mathematically equivalent dimensions, significantly reducing the model's sensitivity to differences in data distribution and solving the problem of feature space mismatch. Traditional MLPs or CNNs use static weight parameters when processing concatenated features; after training, the feature fusion method for all samples is fixed. This embodiment uses the FT-Transformer. After feature extraction, the data directly enters the Transformer module. Utilizing a multi-head attention mechanism, it can dynamically adjust and calculate the interaction weights between features based on the independent tokens generated by FT, exhibiting scene adaptability. Furthermore, while FT maintains the independence of each feature, the Transformer module provides a global receptive field, enabling the model to accurately capture cross-modal nonlinear coupling relationships.

[0122] like Figure 2As shown, step S104 is executed, where each training feature vector, the current predicted value of SOH, and the hidden state vector are input into the degradation dynamics network module to obtain the physical degradation rate, and the total loss function value is obtained based on the physical degradation rate and the total loss function. The degradation dynamics network module includes a second LSTM module.

[0123] Specifically, Modeling battery degradation kinetics and introducing physical constraints is crucial for achieving these constraints. This network module does not directly calculate the SOH (State of Health) prediction value; instead, it acts as a physical equation fitter, learning the kinetics of the SOH prediction value's evolution over time by capturing its decay rate. Since battery degradation is a time-dependent process, this embodiment integrates an LSTM module, i.e., a second LSTM module, into this network module to receive the input sequence. (i.e., training feature vectors), time And SOH predictions from the SOH mapping network module and hidden state vector .in This provides the current SOH estimate, i.e., the current predicted SOH value. Hidden State Vector The input sequence provides features and high-dimensional aggregation information between them, with time serving as the independent variable for rate calculation. The gating mechanism (input gate, forget gate, output gate) within the second LSTM module enables it to selectively retain relevant historical information and discard noise, thereby effectively learning long-term degradation trends and short-term fluctuations.

[0124] The forget gate determines the discarding of historical aging information, as shown in formula (8):

[0125] (8);

[0126] in, For the first The forgetting factor at each time step determines the degree to which memories from the previous moment need to be retained; Here is the weight matrix for the forget gate; For the corresponding bias term; It was the previous step The hidden states of the degradation dynamics network module; It is the first SOH prediction values ​​at each time step This represents the hidden state vector of the FT-Transformer module, containing high-dimensional physical features after interaction through a multi-head attention mechanism. As a time feature, This represents the discrete time step.

[0127] The input gate determines the update of the current physical state information, as shown in formula (9):

[0128] (9);

[0129] in, For the first The input gate weights at the i-th time step determine the i-th time step. The extent to which information from each time step needs to be written into long-term memory; This is the weight matrix of the input gate; This is the corresponding bias term.

[0130] Cell state update Record the aging memory of the battery throughout its entire life cycle, as shown in formula (10):

[0131] (10);

[0132] in, Representing the The cell state at each time step, which in practice represents the long-term aging evolution trajectory of the battery; Forgetting factor; is the input weight; tanh is the activation function; This is the weight matrix; This is the corresponding bias term.

[0133] The output gate expressions are shown in equations (11) and (12):

[0134] (11);

[0135] (12)

[0136] in, Output gating controls the degree to which information is extracted from the cell state; This is the weight matrix for output gating; This is the corresponding bias term.

[0137] in, It is the output hidden state of the degradation dynamics network module, which contains the current dynamic characteristics combined with historical memory, and is used for the next step of fitting the physical degradation rate.

[0138] The hidden state of the output of the first LSTM module The input is fed into a multilayer perceptron (MLP) and mapped to a scalar output. In this case, unlike traditional methods, the second LSTM module in this embodiment no longer fits the SOH prediction value itself, but rather fits an unknown physical evolution function. , so that:

[0139] (13);

[0140] in, The rate of physical degradation. For physical evolution function, These are the weights of the LSTM network modules.

[0141] at this time The time derivative of the predicted SOH value represents the physical degradation rate. It is not used directly to provide SOH prediction results, but is specifically used to participate in the physical consistency loss (PDELoss, i.e., physical dynamic loss value) in the subsequent composite loss function (i.e., the total loss function). The calculation of ). Through During training, gradient backpropagation forces the output trajectory of the preceding SOH mapping network module to conform to this physical degradation dynamic rate, thereby achieving a closed-loop fusion of data-driven approaches and physical laws. Finally, after feedback correction using the total loss function, the final SOH prediction value is output based on the SOH prediction value output by the first SOH mapping network module.

[0142] This embodiment uses a composite loss function (i.e., the total loss function). This guides the training of the enhanced PINN prediction module. The total loss function is defined as a weighted sum of the data loss, physics and dynamics loss, and SOH monotonicity constraint loss, as shown in the formula:

[0143] (14);

[0144] in, For data loss values, This represents the physical and dynamic loss value. This represents the loss value due to the monotonicity constraint of SOH.

[0145] Data loss The mean squared error (MSE) is used to quantify the learning level of the training data for the prediction model. It represents the difference between the predicted SOH value and the true SOH value (i.e., the SOH label), calculated using the following formula:

[0146] (15);

[0147] in, This represents the total number of training feature vectors input to the SOH mapping network module each time. Let be the SOH prediction value of the j-th training feature vector out of the total number of training feature vectors input each time. This is the true SOH value of the j-th training feature vector out of the total number of training feature vectors input each time.

[0148] Physical and dynamic losses Constrained Degeneracy Dynamics Network Module Predicted physical degradation rate Mapping network module with SOH Output The export rate matches. Specifically, the export rate... It is approximated by the finite difference method. The formula is:

[0149] (16).

[0150] in, Let be the derivative of the SOH prediction value of the j-th training feature vector out of the total number of training feature vectors input each time. This represents the physical degradation rate of the j-th training feature vector out of the total number of training feature vectors input each time.

[0151] SOH monotonic constraint loss We introduce the physical prior knowledge that SOH must decrease monotonically. This loss term penalizes predictions that are contrary to the actual degradation trend (i.e., the predicted SOH should decrease over time).

[0152] (17).

[0153] ReLU is the activation function.

[0154] By minimizing the composite loss The predictive model is driven to fit the observed data while adhering to the physical laws and monotonic trends of battery degradation, thereby enhancing the accuracy, robustness and physical reliability of the prediction.

[0155] How the prediction model in this embodiment works:

[0156] Existing technologies typically only concatenate neural network layers for end-to-end numerical fitting, learning only from data without imposing physical constraints on predictions. This not only fails to adequately explore the interactions between heterogeneous features such as voltage and temperature but also cannot effectively learn the dynamic laws governing the long-term aging process of batteries. Consequently, existing models lack physical interpretability and are prone to non-monotonic prediction errors, low reliability, low accuracy, and poor stability under complex operating conditions. To address this issue, the enhanced PINN prediction model in this embodiment abandons the traditional unidirectional feature stacking mode and constructs a cascaded gray-box prediction architecture of "state mapping-dynamic constraints" coupled with physical information. It consists of two networks: a SOH mapping network module and a degradation dynamics network module. Its core lies in using the FT-Transformer as the SOH mapping network module, utilizing its multi-head attention mechanism to achieve adaptive weighted fusion of heterogeneous features. Furthermore, a Long Short-Term Memory (LSTM) network is integrated into the physical dynamics network, utilizing its gating mechanism to effectively capture the long-term temporal features of battery degradation and enhance the learning of the long-term degradation laws of batteries.

[0157] First, an SOH mapping network module is constructed using FT-Transformer. A feature segmenter is introduced to independently map the heterogeneous physical parameters of the battery into high-dimensional semantic tokens. An attention mechanism is used to learn the nonlinear interaction weights between features, significantly enhancing the model's ability to represent complex heterogeneous data features while directly outputting the predicted SOH value and deep hidden states. Then, the output of the SOH mapping network module is input into an LSTM-based degradation dynamics network module. The degradation dynamics network module does not directly calculate the predicted SOH value; instead, it utilizes the long short-term memory gating mechanism of LSTM to accurately capture the long-term temporal dependencies in the battery aging process, specifically for fitting the physical degradation rate (i.e., the time derivative) of SOH over time. Finally, a composite loss function (i.e., the total loss function) is constructed, incorporating physical partial differential equation (PDE) residuals and monotonicity constraints, forcing the output derivative of the preceding SOH mapping network module to maintain strict consistency with the physical degradation rate of the subsequent degradation dynamics network module. Unlike models and methods that simply stack networks and directly output SOH predictions, this embodiment uses feedback adjustment of the total loss function to explicitly guide the deep learning prediction model during gradient descent using physical dynamics equations. This enhances the model's ability to deeply analyze the heterogeneous characteristics of batteries while also enabling it to obey physical evolution laws. The model can learn efficiently from degraded data, capture time decay dynamics, and comply with physical laws, thereby significantly improving the accuracy, stability, and robustness of SOH predictions.

[0158] In principle, traditional electrochemical models rely on precise modeling of the complex electrochemical reaction mechanisms within the battery, resulting in significant challenges in parameter identification and extremely high computational complexity. This embodiment eliminates the need for separate modeling of the complex internal mechanisms of different battery systems. Instead, it uses a neural network to fit the physical dynamics, significantly reducing computational complexity and parameter identification difficulty. This makes the prediction model more universal and generalizable across different battery types. Traditional pure machine learning methods are "black box" models, directly outputting SOH predictions without considering the physical logic. This embodiment constructs a "gray box" model with deep coupling of physical information, outputting not only SOH predictions but also explicitly outputting the physical degradation rate and hidden states. By embedding a composite loss function with PDE residuals and monotonicity constraints, the model output is forced to conform to the physical evolution of battery aging. Compared to pure black box models, this embodiment has stronger physical interpretability, reliability, and robustness, making it more suitable for high-safety-requirement scenarios such as eVTOL, which this embodiment targets. Traditional PINN modules typically use simple MLPs to process inputs, making it difficult to capture the complex heterogeneous interactions between battery operating data. This embodiment introduces a feature segmenter into the SOH mapping network module, then utilizes the multi-head attention mechanism of the Transformer module to achieve adaptive weighted fusion of features, and uses LSTM to capture temporal features, thereby improving the model's fitting accuracy for complex operating data and its ability to learn temporal features.

[0159] Therefore, the battery health state prediction model trained by the training method in this embodiment, i.e., the target prediction model, improves the prediction accuracy, stability, and robustness of the battery health state, thereby enhancing the accuracy, precision, and reliability of the prediction results. Simultaneously, the battery health state prediction model possesses high physical interpretability and generalization ability, reducing prediction errors and improving the generalization ability and reliability of the prediction model under complex real-world operating conditions. This ensures that the SOH prediction value generated by the model meets the engineering application requirements of high safety standards and low fault tolerance fields such as electric aircraft.

[0160] The training method and target prediction model in this embodiment have the following advantages:

[0161] 1. Fast convergence speed

[0162] Thanks to the FT-Transformer network's ability to quickly extract features and the effective physical constraints on the solution space, the training efficiency of this embodiment is significantly better than that of the existing traditional PINN. Compared with traditional PINN, the target prediction module obtained by the training method of this embodiment achieves lower total loss and lower sub-loss (especially data loss). Both exhibit significantly faster convergence speeds. For example... Figure 3As shown, the horizontal axis represents the training epochs of the model, and the vertical axis represents the various losses of the model. In the early stages of training, the slope of the loss curve in this embodiment is significantly greater than that of the traditional PINN loss curve, indicating that the target prediction model in this embodiment can capture the core features of the data more quickly. The overall loss can quickly stabilize at a lower loss level, indicating that the target prediction model has higher learning efficiency, and the final convergence loss is significantly lower, proving that the target prediction model performs well in fitting training data and satisfying physical constraints.

[0163] 2. Good fitting effect

[0164] To visually demonstrate the fitting effect, scatter plots of the predicted SOH values ​​versus the actual SOH values ​​from a single experiment were generated for both the traditional PINN model and the enhanced PINN prediction model of this embodiment. For example... Figure 4 and Figure 5 As shown, compared to the traditional PINN model, the target prediction model in this embodiment produces predictions that are more closely clustered around the diagonal, demonstrating a better fit.

[0165] 3. High prediction accuracy and high stability

[0166] This embodiment addresses the problem that traditional PINN models struggle to effectively capture heterogeneous feature interactions and long-term temporal dependencies, thereby achieving a significant improvement in accuracy across key error metrics. A comparison of the average results from ten independent, repeated experiments shows that... Figure 6 As shown, the Mean Absolute Error (MAE) is significantly reduced, decreasing by 20.97% compared to the traditional PINN model. This embodiment greatly improves the accuracy of SOH prediction. Simultaneously, the enhanced PINN model in this embodiment outperforms the traditional PINN model in terms of Mean Square Error (MSE), Normalized Mean Square Error (NMSE), and Mean Absolute Percentage Error (MAPE), demonstrating a significant improvement. This proves that the prediction model in this embodiment has substantially enhanced its ability to extract SOH-related degradation information from heterogeneous and time-series feature sequences, meaning that the model in this embodiment has stronger learning ability and stronger SOH prediction capability for battery characteristics.

[0167] Example 2

[0168] Based on the same inventive concept, the second embodiment of the present invention also provides a method for predicting battery health status, such as... Figure 7 As shown, it includes:

[0169] S201, Obtain the target prediction model and the battery test data, wherein the target prediction model is a model trained by the battery health status training method shown in Example 1.

[0170] S202, the test data is subjected to resting time constraint processing, state of charge interval constraint processing, monotonicity constraint processing and outlier constraint processing respectively to obtain the effective charging segment of the battery;

[0171] S203, extract global features, voltage range features and high-order statistical features from the effective charging segment to obtain the feature vector;

[0172] S204. Input the feature vector into the target prediction model to obtain the predicted SOH value of the battery.

[0173] It should be noted that the method for predicting battery health can run on electronic devices such as personal computers, neural network processing units (NPUs), or chips.

[0174] In the battery health state prediction method, the target prediction model uses a feature segmenter, a Transformer module, and a first LSTM module; that is, the target prediction model is a SOH mapping network module model. Finally, the first LSTM module outputs the battery's predicted SOH value. The processing procedure of each module is consistent with that of the modules in the training method, and will not be repeated here.

[0175] The prediction method and target prediction model in this embodiment can accurately predict the SOH (State of Health) value of the battery, improving the prediction accuracy, stability, and robustness of battery health status. Simultaneously, the predicted SOH value exhibits high physical interpretability and generalization, possessing extremely high reliability, meeting the engineering application requirements of high-safety-standard, low-fault-tolerance fields such as electric aircraft. In this embodiment, by introducing physical dynamics loss and monotonicity constraint loss, the model can effectively identify and correct irrational prediction fluctuations caused by sensor noise or extreme load fluctuations during the inference phase, ensuring that the SOH output trajectory always conforms to the physical evolution law of irreversible battery aging, thus improving the reliability and credibility of the prediction results. Furthermore, the FT-Transformer architecture's adaptive processing capability for heterogeneous features significantly reduces the migration and adaptation costs of the algorithm. During real-time monitoring, this scheme can not only accurately output the current predicted SOH value but also predict the time point when the battery reaches the failure threshold based on the physical degradation rate, upgrading traditional passive maintenance to predictive maintenance, providing technical support for preventing thermal runaway risks in high-safety fields such as electric aircraft. In addition, the high-precision prediction capability provides a scientific basis for battery residual value assessment and cascade utilization, which can effectively improve the economic benefits and safety assurance level of the entire battery life cycle.

[0176] 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.

[0177] 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.

[0178] 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.

[0179] 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.

[0180] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0181] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for training battery health status, characterized in that, include: Obtain the training feature vector for each battery, which includes: global features, voltage range features, higher-order statistical features, and time features; The initial prediction model for the state of health (SOH) of the battery is iteratively trained using each of the training feature vectors until the current total loss function value obtained by the total loss function corresponding to the iterative prediction model is not greater than the total loss function threshold. Then, the current prediction model is taken as the target prediction model. The prediction model includes: an SOH mapping network module and a degradation dynamics network module. In each iteration of training, each training feature vector is input into the SOH mapping network module to obtain the current SOH prediction value and the hidden state vector. The SOH mapping network module includes a feature segmenter, a Transformer module and a first LSTM module connected in sequence. Each training feature vector, the current predicted value of SOH, and the hidden state vector are input into the degradation dynamics network module to obtain the physical degradation rate. Based on the physical degradation rate and the total loss function, the total loss function value is obtained. The degradation dynamics network module includes a second LSTM module.

2. The battery health state training method of claim 1, wherein, The process of obtaining the training feature vector for each battery includes: Obtain the raw data for each of the batteries; Each of the original data is subjected to rest time constraint processing, state of charge interval constraint processing, monotonicity constraint processing, and outlier constraint processing to obtain the effective charging segment for each battery. The rest time constraint processing selects the rest segment where the rest time of the battery before charging exceeds a preset time. The state of charge interval constraint processing selects the charging segment within a set state of charge interval of the battery during the charging process. The set state of charge interval is a state of charge interval that satisfies the following conditions: the initial state of charge is not greater than a first state of charge threshold, the final state of charge reaches a second state of charge threshold, and the charging state of the battery is marked as having met the marking conditions. The monotonicity constraint processing selects the charging segment where the state of charge increases monotonically. The outlier constraint processing removes the charging segment where the state of charge of the battery is abnormal during the charging process. The SOH tag of the battery is obtained based on the effective charging segment; The global features, voltage range features, and higher-order statistical features are extracted from the effective charging segments of each battery to obtain the training feature vector of each battery. The global features include: the sequence number, maximum voltage value, peak current value, average current value, peak temperature, average temperature, and minimum temperature in the charging segment. The voltage range features include: the voltage value and capacity within a set stable voltage range. The higher-order statistical features include: the kurtosis, skewness, and slope of the voltage and current relationship curve in the charging segment.

3. The battery health status training method as described in claim 2, characterized in that, The step of extracting the global features, voltage range features, and higher-order statistical features from the effective charging segment of each battery to obtain the training feature vector for each battery includes: Map the timestamps of the effective charging segments to time features; The global features, voltage range features, higher-order statistical features, and time features are concatenated along the feature dimension to obtain the training feature vector.

4. The battery health status training method as described in claim 3, characterized in that, The step of inputting each of the trained feature vectors into the SOH mapping network module to obtain the current predicted value of SOH and the hidden state vector includes: The training feature vector is segmented and embedded using the feature segmenter to obtain a learnable feature vector. The Transformer module performs global feature interaction processing on the learnable feature vector to obtain an updated feature vector. The Transformer module includes: The Transformer encoder consists of two layers, with adjacent Transformer encoders connected by a residual layer. Each Transformer encoder layer includes a multi-head attention mechanism and a feedforward neural network, as well as a normalization layer connected to the multi-head attention mechanism and a normalization layer connected to the feedforward neural network. The updated feature vector is predicted using the first LSTM module to obtain the current predicted value of SOH and the hidden state vector.

5. The battery health status training method as described in claim 4, characterized in that, The step of performing word segmentation and embedding processing on the trained feature vector through the feature segmenter to obtain a learnable feature vector includes: Assign independent weight vectors and bias vectors to each feature in the training feature vector to obtain the feature embedding vector, as shown in the following formula: ； in, For the first The feature embedding vectors are, and the training feature vectors are... , for One characteristic, The first of the training feature vectors Features , It is the first The weight vector of each feature. It is the first The bias vectors corresponding to each feature, where R is the real number field; A learnable classification token is appended to the beginning of the feature embedding vector. The learnable feature vector is obtained. .

6. The battery health status training method as described in claim 5, characterized in that, The step of performing prediction processing on the updated feature vector through the first LSTM module to obtain the current predicted value of SOH and the hidden state vector includes: Extract the updated classification token from the updated feature vector. d represents the d-dimensional dimension; Based on the updated classification token, the current predicted value of SOH is obtained as shown in the following formula: ； in, The current predicted value of SOH is given by MLP, where MLP is a multilayer perceptron and LayerNorm is a normalization layer. The updated classification token This serves as the implicit state vector.

7. The battery health status training method as described in claim 6, characterized in that, The step of inputting each of the training feature vectors, the current predicted value of SOH, and the hidden state vector into the degradation dynamics network module to obtain the physical degradation rate includes: Through formula The physical degradation rate is obtained. Where G is the physical evolution function, Let be the hidden state vector, and t be the time feature. These are the internal weights of the LSTM model of the degenerate dynamics network.

8. The battery health status training method as described in claim 7, characterized in that, The step of obtaining the total loss function value based on the physical degradation rate and the total loss function includes: Through formula The total loss function value is obtained. ,in, For data loss values, This represents the physical and dynamic loss value. The loss value is the monotonicity constraint value of SOH; Through formula The data loss value is obtained. ; in, This represents the total number of training feature vectors input to the SOH mapping network module each time. Let be the SOH prediction value of the j-th training feature vector out of the total number of training feature vectors input each time. This is the true SOH value of the j-th training feature vector out of the total number of training feature vectors input each time.

9. The battery health status training method as described in claim 8, characterized in that, Through formula The physical and dynamic loss value is obtained. ;in, Let be the derivative of the SOH prediction value of the j-th training feature vector out of the total number of training feature vectors input each time. The physical degradation rate is the value of the j-th training feature vector out of the total number of training feature vectors input each time. Through formula The monotonicity constraint loss value of SOH is obtained. ; where ReLU is the activation function.

10. A method for predicting battery health status, characterized in that, include: Obtain the target prediction model and the test data of the battery, wherein the target prediction model is a model trained by the battery health status training method as described in any one of claims 1 to 9; The test data is subjected to resting time constraint processing, state of charge interval constraint processing, monotonicity constraint processing and outlier constraint processing respectively to obtain the effective charging segment of the battery; Global features, voltage range features, and higher-order statistical features are extracted from the effective charging segment to obtain a feature vector; The feature vector is input into the target prediction model to obtain the predicted SOH value of the battery.

Images