A lightweight detection method for the health status of lithium-ion batteries based on physical information neural networks.

Abstract

This invention provides a lightweight method for detecting the health status of lithium-ion batteries based on a physical information neural network. The method includes collecting time, voltage, current, temperature, and state of charge data during the battery's discharge process in the take-off and landing phases, processing this data into discharge time-series data, designing feature factors related to battery aging based on this data, and using Pearson correlation coefficient and grey relational analysis algorithms to select the optimal feature sequence data. This optimal feature sequence data is then input into a physical information neural network model constructed from two cascaded neural networks for training. During training, a Bayesian optimization algorithm is used to fine-tune the hyperparameters of the two neural networks, followed by a hierarchical transfer learning strategy to fine-tune the second neural network. Finally, the trained physical information neural network model is used for battery health status detection. This invention improves the quality of feature data, reduces the computational complexity of features and the model, and achieves accuracy and reliability in battery health status detection under airborne conditions.

Description

Technical Field

[0001] This invention relates to the field of lithium battery technology, and in particular to a lightweight detection method for the health status of lithium-ion batteries based on a physical information neural network. Background Technology

[0002] With the rapid development of the low-altitude economy, electric vertical take-off and landing (eVTOL) aircraft are widely used in logistics, emergency rescue, and urban transportation. Lithium-ion batteries, with their advantages of high energy density, long cycle life, and environmental friendliness, have become the primary power source for eVTOL aircraft. However, during critical flight phases such as take-off and landing, eVTOL aircraft not only place stringent requirements on the high-rate discharge of the power battery but also impose higher standards on the precision and reliability of battery management. As a complex electrochemical system, the internal state evolution mechanism of lithium-ion batteries is highly nonlinear and difficult to observe directly, making it difficult to accurately estimate their State of Health (SOH). Furthermore, different aircraft operate under different temperature conditions, different operating modes, and different charge / discharge rates, creating complex operating conditions for the battery. These issues constrain the optimized design and implementation of battery management systems and seriously affect the operational reliability and flight safety of eVTOL aircraft. Therefore, there is an urgent need for a rapid, accurate, and lightweight detection technology for the battery health status of eVTOL aircraft.

[0003] Existing methods for assessing battery health mainly include quantitative calculation-based, model-based, and data-driven approaches. However, the high coupling of heat-electric-mass transfer within the battery makes the degradation mechanism difficult to calculate and quantify directly. Therefore, quantitative calculation-based methods require significant computational resources to accurately obtain aging parameters, hindering their widespread adoption. Physical model-based methods require extensive prior knowledge of battery mechanisms and familiarity with various physical models, but these methods still lead to accumulated biases across temperature and rate scenarios, placing extremely high demands on modelers. Furthermore, as battery aging progresses, model parameters may change, weakening the model's ability to identify parameters and causing significant errors in the estimation results. In contrast, data-driven methods do not require explicit physical modeling. They only need to obtain features characterizing battery aging (such as voltage, current, and temperature) and then use artificial intelligence algorithms to apply these aging features to accurately assess the battery's state of health (SOH). However, data-driven methods generally lack interpretability at the feature level and in the model inference process, making it difficult to establish an interpretable and stable mapping with the aging mechanism. Furthermore, the historical battery data spanning different types and operating conditions results in heterogeneous battery operation data and difficulties in obtaining battery capacity labels, which exacerbates the data inconsistency problem and increases the difficulty of model training and feature extraction.

[0004] Furthermore, existing battery state of health estimation models are mainly based on historical data from vehicle-mounted batteries, with only a few studies specifically addressing the state of health (SOH) estimation of eVTOL aircraft batteries. Unlike traditional vehicle-mounted lithium battery operating data, eVTOL aircraft require high-rate discharge during takeoff and landing, resulting in time-series data exhibiting non-stationary, highly transient, and multi-scale coupled characteristics. Based on these characteristics, current research on the local mutations, recovery processes, and common patterns across temperature and rate during eVTOL aircraft battery operation remains insufficient. Existing technologies mainly suffer from the following problems:

[0005] Existing battery health status detection methods are mainly applied to vehicle lithium batteries, which are suitable for lower discharge rates, but are difficult to apply to high-rate discharge, strong transient and non-steady operating conditions such as take-off and landing of electric vertical take-off and landing aircraft.

[0006] Traditional data-driven methods typically involve complex models and a large number of parameters, requiring significant computational resources and making them unsuitable for deployment on aircraft.

[0007] Due to manufacturing differences and varying operating conditions, battery cells exhibit inconsistent aging trajectories, making it difficult to apply the model to batteries under all operating conditions. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes a lightweight detection method for the health status of lithium-ion batteries based on a physical information neural network. By collecting data during the battery's discharge process in takeoff and landing phases, discharge time-series data for these phases is designed. Based on this time-series data, feature factors related to battery aging are designed, and the Pearson correlation coefficient and grey relational analysis algorithms are used to filter these feature factors, obtaining the optimal feature sequence data. This optimal feature sequence data is then input into a physical information neural network model constructed from two cascaded neural networks for training. During training, a Bayesian optimization algorithm is used to fine-tune the hyperparameters of the two neural networks. After the first neural network is trained, a hierarchical transfer learning strategy is used to fine-tune the second neural network. Finally, the trained physical information neural network model is used for battery health status detection, which can improve the accuracy and reliability of battery health status detection under airborne conditions.

[0009] A lightweight detection method for the health status of lithium-ion batteries based on physical information neural networks, the method comprising the following steps:

[0010] Step S1, Feature Segment Extraction and Label Definition: Analyze the lithium battery dataset, extract the current sequence I and voltage sequence U from the data, and introduce the current change rate dI / dt and voltage change rate dV / dt as transient features. Then calculate the real-time power P=UI and its change rate dP / dt. Combining power with current and voltage change rates, construct an adaptive boundary recognition algorithm, and use this algorithm to extract the segment data of the takeoff and landing phases from the dataset. Then use the ampere-hour integration method to calculate the battery capacity Q in the battery data, and use the battery capacity as the label value of the battery health status. Finally, use the linear interpolation method to repair the missing values ​​in the takeoff and landing phase data, and use the neighbor mean method to smooth the abnormal data points to complete the data preprocessing.

[0011] Step S2, calculate the optimal feature sequence: Based on the preprocessed takeoff and landing segment data, design feature factors related to battery aging, and use Pearson correlation coefficient to verify the correlation between these features and battery health status. Then, use the grey relational algorithm to remove features with high autocorrelation to obtain the optimal feature sequence.

[0012] Step S3: Construct the Physical Information Neural Network Model (PINN): PINN consists of two cascaded neural networks. The first neural network is a solution function network, which captures local features during the battery aging process to learn the model's data loss and consistency loss. The second neural network is a dynamics network, which learns the physical constraint loss of partial differential equations to simulate the temporal evolution of battery aging characteristics. Then, the optimal feature sequence data is divided into a training set and a test set, and the training set data is input into the model to train the parameter sets of the two neural networks.

[0013] Step S4: Use a TPE-based Bayesian algorithm to optimize the hyperparameters learning rate and number of network layers of PINN;

[0014] Step S5: Fine-tune PINN using a hierarchical transfer learning strategy, and use the fine-tuned PINN to detect the health status of battery data.

[0015] Furthermore, the construction of the adaptive boundary recognition algorithm includes:

[0016] Step S11, Load the battery cycling dataset Set the takeoff clip list and landing clip list And initialized to 0, where, Indicates the first One sampling point, k ∈ [ 1 , 2 , ⋯ , N ] , Indicates the first Sampling time of each sampling point This represents the total number of sampling points. for Voltages collected at all times for The current collected at all times, for Temperatures collected at all times for The state of charge collected at any time;

[0017] Step S12: The voltage and current data are filtered using the sliding window method;

[0018] Step S13: Calculate the power P=UI, the power change rate dP / dt, the voltage change rate dU / dt, and the current change rate dI / dt based on the filtered voltage U and current I.

[0019] Step S14, if |dP / dt|> And dU / dt dI / dt Then the value of counter_high is incremented by 1. To reach the high magnification threshold, It is a strong transient value, and counter_high is a counter that indicates entering a high-multiplication rate;

[0020] Step S15: If counter_high ≥ N1, it indicates that this is the start point of the takeoff phase, and the starting point position is saved. N1 represents the number of times to confirm entering high-magnification mode, which is set during initialization. The number of sampling points at the start of takeoff;

[0021] Step S16, if |dP / dt| < And dU / dt < and dI / dt< Then the value of counter_low is incremented by 1. The threshold for exiting high-magnification mode is defined by counter_low, which is a counter representing the exit from high-magnification mode.

[0022] Step S17, if counter_low N2 indicates that this is the end point of the takeoff phase, and the end point position is saved. N2 represents the number of times to confirm exiting high-multiplier mode, which is set during initialization. This represents the number of sampling points at the end of takeoff.

[0023] Step S18, extract the battery cycle data from the start time. By the end time The discrete fragment data is saved to the takeoff fragment list. S takeoff = { ( t k , U ( t k ), I ( t k ), T ( t k ), SOC ( t k )) ， | t k ∈ [ t R 1 , t R 2 ] } And reset counter_high=0, counter_low=0;

[0024] Step S19: During the descent phase, repeat steps S12 to S17. When counter_high≥N1 occurs again, it indicates that this is the start point of the descent phase, and the start point position is saved. When counter_low occurs again When N2 is reached, it indicates the end point of the descent phase; the end point position is saved. Extract battery cycle data from the start time By the end time The discrete fragment data is saved to the landing fragment list. S landing = { ( t k , U ( t k ), I ( t k ), T ( t k ), SOC ( t k )) ， | t k ∈ [ t R 3 , t R 4 ] } ;

[0025] Step S20: Repeat steps S12 to S19 on the battery cycle dataset until all takeoff and landing segment data are extracted, and then end the program.

[0026] Furthermore, the battery capacity value for each cycle is calculated using the ampere-hour integration method. And obtain the battery capacity sequence. Q = [ Q 1 , Q 2 , ⋯ Q n ] , For the first The next loop The total number of cycles is calculated as follows:

[0027] ;

[0028] in, For the first The battery capacity of the next cycle, Δt is the sampling interval, I(t) is the current data, t1 and t2 are the start and end times of the discharge process in this cycle, respectively, and SOC. t1 This is the initial state of charge (SOC) for this discharge cycle. t2 This terminates the SOC (State of Charge) for this discharge cycle.

[0029] Further, step S2 specifically includes:

[0030] Step S21: Based on the preprocessed takeoff and landing segment data, statistical features are designed for voltage, current, and temperature. These features include the maximum, minimum, and average voltage values, as well as the voltage difference at the start and end times. After incorporating all statistical features into the candidate feature pool, the Pearson correlation coefficient between each candidate feature and the battery capacity Q is calculated. Candidate features with Pearson correlation coefficients higher than a preset threshold are retained and sorted in descending order by Pearson correlation coefficient;

[0031] Step S22: For the feature sequence after descending order, take the first ranked feature as the reference feature, and use the grey relational algorithm to calculate the similarity between the remaining features and the reference feature in turn. Delete redundant features with similarity higher than a set threshold, and finally obtain the optimal feature sequence X, which is used as the input for the next step.

[0032] Furthermore, the Pearson correlation coefficient The calculation method is as follows:

[0033] ;

[0034] in, Features In the The value taken in the next loop Features of The average value of the next cycle. This represents the average value of the battery capacity sequence. For the first Battery capacity per cycle This is for the cumulative summation operation;

[0035] The gray relational degree is calculated as follows:

[0036] ;

[0037] ;

[0038] in, Let the gray relational degree be the m-th feature. For the m-th feature in the th... The correlation coefficient of the next cycle, The total number of cycles, For the reference feature sequence, the first The statistical characteristic value of the next cycle, For the m-th feature The statistical characteristic value of the next cycle, The resolution coefficient is defined by min and max, which represent the minimum and maximum values, respectively.

[0039] Furthermore, step S3 specifically includes:

[0040] Constructing the optimal features of the battery through a solution function network The mapping relationship to battery capacity is set as follows. Then we have: ,in, In order to be in The predicted battery capacity is as follows. For time, The input feature vector sequence is the optimal feature sequence. For the parameter set of the solution function network;

[0041] Parameter set This includes the weight parameters and bias terms of each layer of the solution function neural network, where the first... The learning process for the layer's weight parameters and bias terms is represented as follows:

[0042] h j F = { σ ( W j F ⋅ [ X , t ] + b j F ) , j = 1 σ ( W j F ⋅ [ h j − 1 F , t ] + b j F ) , j > 1 ;

[0043] in, To solve the function network F(·) of the th The output of the layer, Given the input feature vector sequence, where t is time. For the first Layer weight parameters, For the first The layer's bias term, where σ is the activation function;

[0044] Then, the parameters of the learned solution function network are used to apply the input. Make a prediction to obtain the predicted battery capacity. This predicted value is then used as one of the inputs to the dynamics network to continue learning the network's parameter set. Parameter set Including weight parameters and bias terms, the learning dynamics network is in its first stage. The learning process for the layer's weight parameters and bias terms is as follows:

[0045] h j G = { σ ( W j G ⋅ [ t , X , u , ∂ u ∂ t , ∂ u ∂ X ] + b j G ) , j = 1 σ ( W j G ⋅ h j − 1 G + b j G ) , j > 1 ;

[0046] in, For the dynamic network The layer's output, where t is time. Given the input feature vector sequence, Let be the battery capacity predicted by the solution function network, ∂u / ∂t be the rate of change of the predicted battery capacity, and ∂u / ∂X be the partial derivative of the predicted battery capacity with respect to the eigenvector. For the dynamic network Layer weight parameters, For the dynamic network The layer's bias term, where σ is the activation function;

[0047] The physical constraint loss of the two networks is expressed through the residual function. accomplish:

[0048] ;

[0049] Where F(t, X; Φ) is the output of the solution function neural network. The output of the dynamic neural network, and These are the parameter sets for the two networks, respectively.

[0050] The total loss function of the PINN network is composed of the data loss function. Consistency loss function and physical constraint loss function The composition and calculation method are as follows:

[0051] ;

[0052] in, This represents the total number of battery cycles. For the first Battery capacity per cycle For the first Predicted battery capacity for the next cycle For the first Predicted battery capacity for the next cycle For the first The physical constraint loss for the next iteration uses ReLU as the activation function and α and β as weighting coefficients. Let this be the total loss function, when the total loss value The model is optimal when the preset minimum value is reached.

[0053] Further, step S4 specifically includes: randomly setting multiple initial learning rates and obtaining the predicted battery capacity corresponding to each learning rate; comparing with the actual battery capacity to obtain the one among these multiple predicted battery capacities that is closest to the actual battery capacity, which is denoted as the current optimal observation value; constructing two density functions based on this optimal observation value to model the conditional probability distribution of battery capacity at different learning rates; subsequently, calculating the expected improvement EI based on the conditional probability distribution, and obtaining the optimal learning rate by maximizing EI; learning another hyperparameter network layer in the same way to obtain the optimal model.

[0054] Further, step S5 specifically includes: keeping the parameters of the solution function network unchanged, and then fine-tuning the parameters in the dynamic network layer by layer on the target domain battery data, setting the parameters in the dynamic network at the first... The parameters of the layer are The loss function of this layer is expressed as follows:

[0055] ;

[0056] in, For the first The loss function of the layer, Let MSE be the total number of samples, and MSE be the mean squared error loss function. For dynamic networks to input features and the Layer parameters The predicted output, For the first The battery capacity in the next cycle is determined by the loss function. Layer-by-layer fine-tuning optimizes the parameters of each layer in the dynamics network, resulting in the optimal dynamics network. Finally, a physical information neural network model constructed from two optimized neural networks is used to detect the battery health status.

[0057] This invention effectively solves the limitations of existing health estimation methods, such as difficulty in applying them to high-rate discharge scenarios, large data requirements, poor interpretability, and low estimation accuracy. This invention has the following beneficial technical effects:

[0058] Intelligent High-Rate Operating Condition Feature Extraction and Screening: This invention addresses the high-rate, strong transient operating conditions of low-altitude aircraft during takeoff and landing. It proposes an adaptive boundary recognition algorithm based on the power density change rate. By monitoring the change rate of power density P=UI in real time, when |dP / dt| exceeds a preset threshold τ, it automatically marks it as a segment boundary, accurately capturing the high-rate discharge characteristic interval. Simultaneously, it introduces the current change rate dI / dt and voltage change rate dV / dt as transient features, constructing a preliminary feature factor set that can map the battery aging mechanism. Linear interpolation is used to optimize the time resolution, and through dual calculations using Pearson correlation coefficient and grey relational algorithm, a high-quality feature set highly correlated with battery capacity and exhibiting low autocorrelation between features is screened from the preliminary features. This method effectively solves the problem of insufficient feature extraction in non-stationary, multi-scale coupled operating conditions of traditional methods, avoids subjectivity in the data acquisition process, greatly reduces data volume requirements, and achieves a balance between high estimation accuracy and low data acquisition difficulty.

[0059] Physical Information Constraint Neural Network Modeling: To address the lack of physical interpretability and generalization ability in traditional data-driven methods, this invention introduces the Physical Information Neural Network (PINN) architecture. The PINN consists of two cascaded neural networks. The first network captures local features during battery aging from the optimal feature sequence, while the second network learns the temporal evolution of battery aging features from both local features and the optimal feature sequence data. A constraint relationship is constructed between the two based on the physical mechanism of battery aging. The loss function combines data fitting error and physical consistency residuals to ensure that the model learning results conform to the physical laws of battery aging. A composite loss function is constructed, and the Adam optimizer and cosine annealing learning rate scheduling strategy are used for model training.

[0060] Hyperparameter optimization: A Bayesian algorithm based on TPE is used to optimize the hyperparameter learning rate and number of network layers of PINN. TPE establishes a nonparametric probabilistic model to divide historical evaluation results into two categories, good and bad, corresponding to different density functions. This modeling method avoids the limitations of the Gaussian process assumption, making TPE more adaptable to noise, non-smooth, and non-convex objective functions, and less susceptible to the influence of complex function shapes. In addition, the next hyperparameter point selected by TPE each time is determined by maximizing the expected improvement EI, which allows it to intelligently jump to potentially better regions. Therefore, under the same computational budget, TPE can usually significantly reduce the number of trials and achieve fast convergence. Compared with traditional neural network hyperparameter optimization methods, TPE has higher efficiency, faster convergence speed, and stronger search capabilities, and is especially suitable for complex hyperparameter spaces.

[0061] Layered transfer learning strategy for fine-tuning the pre-trained model: In the Physical Information Neural Network (PINN), the layered transfer learning strategy is used to fine-tune the pre-trained model, which can effectively improve the convergence efficiency and prediction accuracy of the model in new tasks. By fine-tuning the network parameters layer by layer, a progressive adaptation from low-level features to high-level structures is achieved. Layered transfer learning can fully retain the general physical features learned in the original task while avoiding knowledge forgetting or training instability caused by one-time global fine-tuning. Low-level networks are usually responsible for learning general physical laws, partial differential equation structure features, and basic function representation capabilities. These parts are often universal across scenarios. High-level parameters focus more on the boundary conditions, parameter offsets, or field distribution patterns of specific problems. Therefore, layer-by-layer fine-tuning can make the model more stably adapt to new physical scenarios. Compared with training PINN from scratch or directly fine-tuning the overall parameters, the layered transfer strategy can significantly reduce training costs, improve optimization efficiency, and perform more robustly in small sample or complex physical constraint scenarios. It is an effective method to improve the generalization ability and transferability of PINN.

[0062] In summary, this invention targets the unique operating mode of electric vertical takeoff and landing aircraft, extracting segment data from the takeoff and landing phases, and using this segment data to select a set of features strongly correlated with battery aging for modeling. This improves the model's relevance and accuracy. Furthermore, the model framework designed in this invention requires less feature computation, occupies less storage space, has high model detection accuracy, and features a simple network structure. These characteristics enable the proposed method to be quickly deployed on aircraft for real-time SOH detection, achieving accurate and reliable battery health status detection under airborne conditions. Attached Figure Description

[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart illustrating a lightweight detection method for the health status of lithium-ion batteries based on a physical information neural network, provided in an embodiment of the present invention.

[0065] Figure 2 This is a schematic diagram of different tasks in one cycle of an eVTOL battery provided in an embodiment of the present invention;

[0066] Figure 3 This is a schematic diagram of the process for extracting feature segments during the takeoff and landing phases of an eVTOL battery according to an embodiment of the present invention;

[0067] Figure 4 This is a schematic diagram of the feature filtering process provided in an embodiment of the present invention;

[0068] Figure 5 This is a schematic diagram of the PINN model structure and training strategy provided in an embodiment of the present invention;

[0069] Figure 6 This is a schematic diagram of the SOH estimation results of the PINN model provided in the embodiments of the present invention. Detailed Implementation

[0070] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] This invention provides a lightweight detection method for the health status of lithium-ion batteries based on a physical information neural network. This method, tailored to the unique operating mode of batteries in electric vertical takeoff and landing (EVTOL) aircraft, analyzes operational data to extract feature factors highly correlated with battery aging, constructs a neural network mapping model from these features to battery capacity, and finally utilizes the learned model to achieve rapid detection of battery health status. The overall process is as follows: Figure 1 As shown, the method includes:

[0072] Step S1, Feature fragment extraction and label definition

[0073] Step S11: Analyze the lithium battery dataset, extract the current sequence I and voltage sequence U from the data, and introduce the current change rate dI / dt and voltage change rate dV / dt as transient features. Subsequently, calculate the real-time power P=UI and its change rate dP / dt. Combining the above power, current, and voltage change rates, construct an adaptive boundary recognition algorithm, and use this algorithm to extract fragment data of the takeoff and landing phases from the data. The lithium battery dataset is obtained from the battery management system.

[0074] In order to accurately extract the characteristics that characterize battery aging, it is necessary to first analyze the battery operation data and design features accordingly. Figure 2 This paper illustrates the changes in voltage, current, and temperature of an electric vertical takeoff and landing (EVTOL) aircraft battery over time during a complete cycle. As shown in the figure, significant fluctuations in voltage, current, and temperature occur during takeoff and landing, reflecting substantial changes in the battery's internal mechanisms during this process. Therefore, the characteristic data from these two stages are more effective in characterizing the battery's aging state. In contrast, data from other stages show more gradual changes and are less effective in reflecting battery aging information.

[0075] Therefore, this invention first extracts segmented data from the takeoff and landing phases of the complete cycle data of the aircraft battery, such as... Figure 3 As shown, this method is implemented by constructing an adaptive boundary recognition algorithm based on voltage, current, and power change rate. The main steps are as follows:

[0076] Step 1, Load the battery cycling dataset Set the takeoff clip list and landing clip list And initialize it to 0; Indicates the first One sampling point, k ∈ [ 1 , 2 , ⋯ , N ] , Indicates the first Sampling time of each sampling point This represents the total number of sampling points. They are The voltage, current, temperature, and state of charge are collected in real time.

[0077] Step 2: Use the sliding window method to filter the voltage and current data, with the window width set to 5 sampling points;

[0078] Step 3: Based on the filtered voltage U and current I, calculate the power P=UI, the power change rate dP / dt, the voltage change rate dU / dt, and the current change rate dI / dt;

[0079] Step 4, if |dP / dt|>τ_high, and dU / dt dI / dt Then the value of counter_high is incremented by 1, and τ_high is the threshold for entering high magnification. For strong transient values, counter_high is a counter representing entering a high-magnification range, and τ_hig and The value is set based on experience;

[0080] Step 5: If counter_high ≥ N1, it indicates that this is the start point of the takeoff phase, and the starting point position is saved. N1 is the number of times to confirm entering high magnification, which is set during initialization; This represents the number of sampling points at the start of takeoff.

[0081] Step 6, if |dP / dt| < τ_low and dU / dt < and dI / dt< If τ_low is incremented by 1, then the value of counter_low is incremented by 1. τ_low is the threshold for exiting high magnification, and counter_low is a counter that represents the exit from high magnification.

[0082] Step 7, if counter_low N2 indicates that this is the end point of the takeoff phase, and the end point position is saved. N2 is the number of times to confirm exiting high-multiplier mode, which is set during initialization; Number of sampling points at the end of takeoff

[0083] Step 8, extract the data from the beginning of the loop. By the end time The discrete fragment data is saved to the takeoff fragment list. S takeoff = { ( t k , U ( t k ), I ( t k ), T ( t k ), SOC ( t k )) ， | t k ∈ [ t R 1 , t R 2 ] } Reset counter_high=0 and counter_low=0.

[0084] Step 9: During the descent phase, the battery operating mode remains characterized by strong transient high power. Therefore, repeat steps 2 to 7. When counter_high ≥ N1 appears again, it indicates that this is the start point of the descent phase. Save the start point position. When counter_low occurs again When N2 is reached, it indicates the end point of the descent phase; the end point position is saved. Extract the data from the beginning of the loop. By the end time The discrete fragment data is saved to the landing fragment list. S landing = { ( t k , U ( t k ), I ( t k ), T ( t k ), SOC ( t k )) ， | t k ∈ [ t R 3 , t R 4 ] } .

[0085] Step 10: Repeat steps 2 through 9 on the loop dataset until all takeoff and landing data segments are extracted, then end the program.

[0086] Step S12: Calculate the battery capacity Q in the battery data using the ampere-hour integration method, and use it as the label value of the battery health status.

[0087] From the battery dataset, a sample is constructed for each cycle of data, and the battery capacity of that cycle is used as the ground truth label for the state of battery health (SOH). Then, all "sample-label" pairs are divided into training and test sets, with one flight mission corresponding to one battery cycle.

[0088] In the battery field, a generally accepted rule is that a lithium-ion battery is considered to have reached the end of its lifespan when its capacity decays to 80% of its initial capacity. Therefore, the core of detecting the State of Health (SOH) of a lithium-ion battery lies in accurately estimating its current capacity. To achieve this goal, it is necessary to construct a precise mapping relationship from features characterizing battery aging to battery capacity. This process mainly includes the following three steps: First, obtain the battery capacity for each cycle in the battery's historical data and use it as a label for the training data; second, design features that are strongly correlated with battery aging; and finally, establish a mapping model between features and capacity.

[0089] In model training, battery capacity serves as the label value for samples and is a benchmark for evaluating the model's prediction accuracy: the closer the model's prediction is to the label value, the better its performance. Therefore, accurately obtaining battery capacity data has a decisive impact on the model training effect.

[0090] In a Battery Management System (BMS), the data acquisition module can acquire key parameters such as time, voltage, current, temperature, and state of charge (SOC) during battery operation in real time. To accurately construct the sample labels required for model training, an improved ampere-hour integration method is used to calculate the battery capacity value for each cycle. And obtain the battery capacity sequence. Q = [ Q 1 , Q 2 , ⋯ Q n ] , For the first The next loop The total number of cycles is calculated as follows:

[0091] ;

[0092] in, For the first The battery capacity of the next cycle, Δt is the sampling interval, I(t) is the current data, t1 and t2 are the start and end times of the discharge process in this cycle, respectively, and SOC. t1 This is the initial state of charge (SOC) for this discharge cycle. t2 This terminates the SOC (State of Charge) for this discharge cycle.

[0093] Step S13: Use linear interpolation to repair missing values ​​in the takeoff and landing data, and use the nearest neighbor mean method to smooth out abnormal data points to complete data preprocessing.

[0094] Step S2: Calculate the optimal feature sequence. Based on the extracted takeoff and landing data segments, firstly, design feature factors related to battery aging, and use the Pearson correlation coefficient to verify the correlation between these features and battery health status. Then, use the grey relational analysis algorithm to remove features with high autocorrelation, including:

[0095] Step S21: Based on the preprocessed takeoff and landing segment data, design statistical features for voltage, current, and temperature. The statistical features include the maximum, minimum, and average voltage values, as well as the voltage difference at the start and end times. After including all statistical features in the candidate feature pool, calculate the Pearson correlation coefficient between each candidate feature and the battery capacity Q. Candidate features with Pearson correlation coefficients higher than a preset threshold are retained and sorted in descending order of Pearson correlation coefficient.

[0096] Step S22: For the feature sequence after descending order, take the first ranked feature as the reference feature, and use the grey relational algorithm to calculate the similarity between the remaining features and the reference feature in turn. Delete redundant features with similarity higher than a set threshold, and finally obtain the optimal feature sequence X, which is used as the input for the next step of the model.

[0097] Based on Figure 2 Based on the analysis of voltage, current, and temperature variation characteristics during the take-off and landing phases, and combined with industry experience, this embodiment designed 28 candidate features related to voltage, current, temperature, internal resistance, and capacity, as shown in Table 1. These features can reflect the battery aging state to varying degrees, therefore, it is necessary to select the key features with the greatest characterizing ability.

[0098] Table 1. Candidate Battery Features for Takeoff and Landing Phases of Vertical Electric Aircraft

[0099]

[0100] The specific process of optimal feature selection is as follows: Figure 4 As shown:

[0101] First, on the takeoff or landing segment data, calculate each candidate feature and the battery capacity sequence respectively. Q = [ Q 1 , Q 2 , ⋯ Q n ] The Pearson correlation coefficient, with a certain feature set as , K = [ k 1 , k 2 , ⋯ k n ] , Features In the The value taken in the next loop;

[0102] Next, calculate and Pearson correlation coefficient The calculation method is as follows:

[0103] ;

[0104] in, Features of The average value of the next cycle. It is the average value of the battery capacity sequence. For the first Battery capacity per cycle This is for the cumulative summation operation.

[0105] After calculation, candidate features with Pearson correlation coefficients higher than a preset threshold are retained and sorted in descending order of correlation coefficient. In this embodiment, the preset threshold is set to 0.6, resulting in Pearson correlation coefficients for 16 candidate features. If the value is greater than 0.6, save these features in the optimal feature sequence. X = [ f 1 , f 2 , ⋯ f 16 ] In the middle, then... f 1 = [ f 1 ( 1 ), f 1 ( 2 ), ⋯ , f 1 ( n )] As a reference feature sequence, calculate middle f m = [ f m ( 1 ), f m ( 2 ), ⋯ , f m ( n )] and Grey relational degree , m = [ 2 , 3 , ⋯ 16 ] The gray relational degree is calculated as follows:

[0106] ;

[0107] ;

[0108] in, Let the gray relational degree be the m-th feature. For the m-th feature in the th... The correlation coefficient of the next cycle, The total number of cycles, For the reference feature sequence, the first The statistical characteristic value of the next cycle, For the m-th feature The statistical characteristic value of the next cycle, The resolution coefficient is set to 0.5 in this embodiment, and min and max are used to find the minimum and maximum values, respectively.

[0109] Then, redundant features with a gray correlation degree higher than a set threshold are deleted. In this embodiment, the threshold is set to 0.9, meaning that if the gray correlation degree of a certain feature is higher than a set threshold, then the redundant features are deleted. If the similarity is greater than 0.9, it is considered highly similar to the reference feature, and the feature is removed. A total of eight optimal features are retained, forming the optimal feature sequence, denoted as... X = [ f 1 , f 2 , ⋯ f 8 ] This serves as the input for the next step of the model.

[0110] Step S3: Construct a physical information neural network model

[0111] The physical information neural network model (hereinafter referred to as the model) in this invention consists of two cascaded neural networks. The first neural network is a solution function network, which captures local features in the battery aging process to learn the model's data loss and consistency loss. The second neural network is a dynamics network, which learns the physical constraint loss of the partial differential equation to simulate the temporal evolution of battery aging characteristics. The total loss of the model consists of data loss, consistency loss, and physical constraint loss. When the total loss of the model reaches its minimum, the model reaches its optimum. The minimum value is preset based on experience.

[0112] Then, the optimal feature sequence data is divided into a training set and a test set, and the training set data is input into the model to train the parameter sets of the two neural networks.

[0113] In this embodiment, the optimal feature sequence X = [ f 1 , f 2 , ⋯ f 8 ] The dataset is divided into training and testing sets in a 7:3 ratio. The training set data is then input into the model for training, and a learning rate is set. The number of network layers is a hyperparameter of the model, and the initial learning rate is set. =0.001, initial layer=3, model training is performed using Adam optimizer and cosine annealing strategy.

[0114] The structure of a physical information neural network model is as follows: Figure 5 As shown, its detailed explanation is as follows:

[0115] The physical information neural network model comprises two core components: a solution function network F(·) and a dynamics network G(·). The solution function network is responsible for establishing the dynamics network from the input features. The nonlinear mapping relationship to battery capacity is represented by an input layer, a fully connected hidden layer, and an output layer. The structure of the dynamic network is consistent with that of the solution function network, except for the weight parameters. The dynamic network is used to simulate the dynamic behavior of the battery during degradation to capture the change of battery capacity over time. The two networks are connected through physical constraint loss and together form a unified physical information neural network model.

[0116] First, the solution function network F(·) is a neural network used to simulate a nonlinear function, thereby constructing the optimal features of the battery. The mapping relationship to battery capacity, assuming this mapping relationship is as follows: Then, the process can be expressed mathematically as follows:

[0117] ;

[0118] in, In order to be in The predicted battery capacity is as follows. For time, The input feature vector sequence is the optimal feature sequence. This is the set of parameters for the solution function network.

[0119] Obtain the mapping relationship The process of training a neural network is the same as the process of learning a set of parameters. The weight parameters and bias terms for each layer.

[0120] For the solution function network F(·), its first... The learning process for the layer's weight parameters and bias terms is represented as follows:

[0121] h j F = { σ ( W j F ⋅ [ X , t ] + b j F ) , j = 1 σ ( W j F ⋅ [ h j − 1 F , t ] + b j F ) , j > 1 ;

[0122] in, To solve the function network F(·) of the th The output of the layer, Given the input feature vector sequence, where t is time. For the first Layer weight parameters, For the first The layer's bias term, σ, is the activation function.

[0123] Once all the weight parameters and bias terms in the solution function network F(·) have been learned, the solution function network F(·) becomes a definite mapping relationship, which can be used to map the input... Make a prediction to obtain the predicted battery capacity. This predicted value is then used as one of the inputs to the dynamic network G(·), which continues to learn the parameter set of the dynamic network G(·). Parameter set Including weight parameters and bias terms, since the structure of the dynamic network G(·) is the same as that of the solution function network F(·), its first... The learning process for the layer's weight parameters and bias terms is represented as follows:

[0124] h j G = { σ ( W j G ⋅ [ t , X , u , ∂ u ∂ t , ∂ u ∂ X ] + b j G ) , j = 1 σ ( W j G ⋅ h j − 1 G + b j G ) , j > 1 ;

[0125] in, For the dynamic network G(·), the th The layer's output, where t is time. Given the input feature vector sequence, Let F(·) represent the battery capacity predicted by F(·), ∂u / ∂t represent the rate of change of the predicted battery capacity, and ∂u / ∂X represent the partial derivative of the predicted battery capacity with respect to the eigenvector. For the dynamic network G(·), the th Layer weight parameters, For the dynamic network G(·), the th The bias term of the layer, σ is the activation function, and the dynamic network captures the dynamic characteristics and inherent laws of the battery degradation process by integrating information on time, features, current battery capacity state and its partial derivatives.

[0126] The physical constraint loss of the two networks is expressed through the residual function. accomplish:

[0127] ;

[0128] Where F(t, X; Φ) is the output of the solution function neural network F(·). The output of the dynamic neural network G(·) and These are the parameter sets for the two networks, respectively.

[0129] Finally, the PINN network employs a multinomial composite loss function to optimize the model's prediction accuracy, derived from the data loss function. Consistency loss function and physical constraint loss function The composition and calculation method are as follows:

[0130] ;

[0131] in, This represents the total number of battery cycles. For the first Battery capacity per cycle For the first Predicted battery capacity for the next cycle For the first Predicted battery capacity for the next cycle For the first The physical constraint loss for the next iteration uses ReLU as the activation function and α and β as weighting coefficients. Let this be the total loss function, when the total loss value The model is optimal when it reaches its minimum value.

[0132] Step S4: The hyperparameter learning rate and number of network layers of PINN are optimized using a Bayesian algorithm based on TPE. Multiple initial learning rates are randomly set, and the predicted battery capacity corresponding to each learning rate is obtained. By comparing with the actual battery capacity, the one that is closest to the actual battery capacity among these multiple predicted battery capacities is obtained and recorded as the current optimal observation value. Based on this optimal observation value, two density functions are constructed to model the conditional probability distribution of battery capacity at different learning rates. Subsequently, the expected improvement EI is calculated based on the conditional probability distribution, and the optimal learning rate is obtained by maximizing EI. The other hyperparameter network layer number is learned in the same way to obtain the optimal model, and it is used to estimate the battery capacity on the test set, thereby performing battery health status detection.

[0133] In step S3, during the training process, the PINN network model primarily learns the weight parameters and bias terms, while the learning rate... The number of network layers, as hyperparameters, remains crucial to the model, but the model itself cannot learn them. Therefore, this invention employs a Bayesian optimization algorithm based on a tree-based Parzen estimator (TPE) to search for the optimal hyperparameters.

[0134] The generalized description of the TPE algorithm is as follows:

[0135] The TPE algorithm is a parameter optimization method based on a probabilistic model. Assuming the parameter to be optimized is x, and the objective function is y = f(x), the algorithm aims to find the optimal parameter x* that maximizes the objective function value y. Since the probability distribution of the objective function y cannot be directly obtained, the TPE algorithm uses two density functions to model the conditional probability distribution of the objective function. The specific implementation process is as follows: First, several points are randomly sampled in the value space of parameter x, and these sampled points are substituted into the objective function y = f(x). Through observation and comparison, the optimal value among these random sampled points can be determined and denoted as the current optimal observation value y*. Then, based on this optimal observation value, the following two density functions are constructed:

[0136] Assume It is all The parameter configuration set contains a total of One point, It is all The parameter configuration set contains a total of If there are 1 point, then:

[0137] ;

[0138] in, Let these be the assumed parameters to be optimized. The objective function value, This is the current optimal observation value. It is the conditional probability density function. and For two different density functions, It is a parameter Dimensions It is a balance parameter. To calculate Euclidean distance.

[0139] The optimization objective of the TPE algorithm is to find the optimal parameters. , making and To get as close as possible, we use expected improvement. To find the optimal parameters When the expected improvement reaches its maximum value, that value corresponds to the optimal parameter. The expression is:

[0140] ;

[0141] In order to be in under conditions Expected improvements This indicates the range from negative infinity to... Integral range, and Both are conditional probability density functions. , Let be the probability density function. ,in, The weighting coefficients are used for the expected improvement function. Simplify as follows:

[0142] ;

[0143] The final simplified expression is:

[0144] ;

[0145] in, It indicates a direct proportion.

[0146] Based on the generalized formula of the TPE algorithm, the parameters are... Replace with learning rate Then construct the learning rate respectively. The two density functions are used to calculate the maximum expected improvement, thus obtaining the optimal learning rate. And learn the optimal number of hyperparameter network layers in the same way.

[0147] Step S5: Improve the physical information neural network model using a hierarchical transfer learning strategy: Fine-tune the network model optimized by the Bayesian algorithm based on TPE using a hierarchical transfer learning strategy to obtain the optimal dynamic network. Finally, the physical information neural network model constructed by the two optimized neural networks is used to detect the battery health status.

[0148] For battery data of different battery types and operating conditions, there is an inconsistency between the source domain (battery data under one operating condition) and the target domain (battery data under another operating condition). To address this, this invention employs a hierarchical transfer learning strategy to fine-tune the pre-trained PINN model. The main implementation process involves freezing the parameters of the solution function neural network F(·), i.e., keeping the parameters in the solution function neural network F(·) unchanged, and then fine-tuning the parameters in the dynamic network G(·) layer by layer on the target domain battery data. Finally, the learned new features are used for the final target domain battery capacity prediction.

[0149] Assume that in the dynamic network G(·), the first... The parameters of the layer are The loss function of this layer is expressed by the following formula:

[0150] ;

[0151] in, For the first The loss function of the layer, Let MSE be the total number of samples, and MSE be the mean squared error loss function. For the dynamic network model of input features and the Layer parameters The predicted output, For the first Battery capacity per cycle.

[0152] Through the loss function Layer-by-layer fine-tuning brings the parameters of each layer in the dynamic network G(·) to their optimal values, thus obtaining the optimal dynamic network. Finally, the physical information neural network model constructed by the two optimized neural networks is used to detect the battery health status.

[0153] Finally, the method proposed in this invention was validated on a publicly available battery dataset for eVTOL aircraft, and its battery health state estimation results are as follows: Figure 6As shown in the figure, the X-axis represents the actual SOH value (discharge capacity of the battery) of the predicted sample, and the Y-axis represents the predicted value of the sample by the method proposed in this invention. The blue curve indicates that the predicted value is consistent with the actual value. The red scatter points are generally distributed along the blue line, indicating that the predicted value of the method of this invention is highly consistent with the actual value. Some red scatter points deviate from the blue solid line, indicating that there is a certain error between some predicted values ​​and the actual value. This is due to the poor quality of some individual data in the predicted sample. However, the overall prediction trend fits the actual value very well, indicating that the method has high accuracy and stability in predicting battery SOH.

[0154] This invention constructs an adaptive boundary recognition algorithm based on power change rate to accurately capture key feature intervals of batteries under high-rate and strong transient conditions during takeoff and landing. It also introduces a Physical Information Neural Network (PINN) to jointly incorporate data fitting error, physical loss function error, and consistency error into the loss function, overcoming the shortcomings of traditional data-driven methods that lack physical mechanism mapping and interpretability. This ensures that the estimation results conform to the physical laws of battery aging and significantly improves the model's generalization ability. Furthermore, by using a Bayesian optimization algorithm based on a TPE structure to quickly optimize the hyperparameters in the PINN network model, it achieves synergistic optimization of estimation accuracy, computational efficiency, and physical consistency. This allows the model to complete high-precision battery capacity estimation using only key short-term discharge data from takeoff and landing, enabling rapid detection of battery health status. The model in this invention uses two cascaded neural networks for feature learning, resulting in fewer model parameters and a lightweight design that meets the stringent requirements of real-time airborne applications. Meanwhile, feature sequence calculation is relatively simple, significantly reducing data requirements and computational resource consumption compared to traditional methods. The adoption of a layer-by-layer fine-tuning transfer learning strategy effectively solves the problem of inconsistency between target domain and source domain data, improving the model's robustness across operating conditions. The method proposed in this invention not only efficiently solves the problem of aircraft battery health detection but also provides a scientific basis for optimizing battery health management strategies, comprehensively enhancing the intelligence level and safety assurance capabilities of low-altitude aircraft battery management systems.

[0155] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for lightweight detection of lithium-ion battery state-of-health based on physical information neural network, characterized in that, The method comprises: Step S1, feature fragment extraction and label definition: analyze the lithium battery data set, extract the current sequence I and voltage sequence U in the data, and introduce the current change rate dI / dt and voltage change rate dV / dt as transient characteristics, then calculate the real-time power P=UI and its change rate dP / dt, integrate the power and the current and voltage change rates, construct an adaptive boundary recognition algorithm, and use the algorithm to extract the takeoff and landing stage fragment data from the data set, then use the ampere-hour integration method to calculate the battery capacity Q of the battery in the battery data, and take the battery capacity as the label value of the battery health state, finally use the linear interpolation method to repair the missing values in the takeoff and landing stage data, and use the neighbor point mean method to smooth the abnormal data points, and complete the data preprocessing; Step S2, calculate the optimal feature sequence: based on the preprocessed takeoff and landing fragment data, design the feature factors related to battery aging, and use the Pearson correlation coefficient to verify the correlation of these features with the battery health state, and then use the grey correlation degree algorithm to remove the features with high autocorrelation, to obtain the optimal feature sequence; Step S3, constructing a physical information neural network model PINN: PINN is composed of two series-connected neural networks, the first neural network is a solution function network, which captures local features in the battery aging process and is used to learn the data loss and consistency loss of the model, the second neural network is a dynamics network, which is used to learn the physical constraint loss of the partial differential equation, to realize the simulation of the time evolution law of the battery aging characteristics, then divide the optimal feature sequence data into training set and test set, and input the training set data into the model to train the parameter set of the two neural networks; Step S4, use the TPE-based Bayesian algorithm to optimize the learning rate and network layer of PINN; Step S5, use the hierarchical transfer learning strategy to fine-tune PINN, and use the fine-tuned PINN to detect the health state of the battery data; Wherein, the adaptive boundary recognition algorithm comprises: Step S11, loading battery cycle dataset , setting take-off segment list and landing segment list and initializing to 0, wherein, denotes the th sampling point, , denotes the sampling time of the th sampling point, is the total number of sampling points, is the time collected voltage, is the time collected current, is the time collected temperature, is the time collected state of charge; Step S12, filter the voltage and current data using the sliding window method; Step S13, according to the filtered voltage U and current I, calculate the power P=UI, the power change rate dP / dt, the voltage change rate dU / dt, and the current change rate dI / dt; Step S14, if |dP / dt| > 0.5 , and dU / dt , dI / dt , counter_high is added by 1, for entering high-multiplying threshold, for strong transient value, counter_high is a counter representing entering high-multiplying. Step S15, if counter_high≥N1, it indicates that this is the start point of the take-off phase, and the starting point position is saved N1 is the number of entries into high-multiplying confirmation, which is set during initialization, is the sampling point number at the start of take-off; Step S16, if |dP / dt| < 0.5 and dU / dt < 0.5 and dI / dt < 0.5 then counter_low is incremented by 1, for exit high rate threshold, counter_low is a counter representing exit high rate. Step S17, if counter_low N2, indicates the end point of the take-off phase at this time, and saves the end point position N2 is the number of times of exiting the high magnification confirmation, which is set at the initialization, is the number of sampling points at the end of take-off Step S18, extract discrete segment data from start time to end time in the battery cycle data, save to takeoff segment list and reset counter_high = 0, counter_low = 0; Step S19, the landing phase repeats steps S12 to S17, when counter_high≥N1 appears again, it indicates that this time is the starting point of the landing phase, save the starting point position , when counter_low N2 appears again, it indicates that this time is the ending point of the landing phase, save the ending point position , extract the discrete segment data from the starting time to the ending time in the battery cycle data, and save to the landing segment list ; Step S20, repeat steps S12 to S19 on the battery cycle data set until all takeoff and landing fragment data are extracted, and the program ends.

2. The method of claim 1, wherein the method is a lightweight method for detecting the state of health of a lithium-ion battery based on a physical information neural network. The battery capacity value of each cycle is calculated by using the ampere-hour integration method , and the battery capacity sequence is obtained , for the first cycle, for the nth cycle, for the total number of cycles, and the calculation method is: ； in, For the first The battery capacity of the next cycle, Δt is the sampling interval, I(t) is the current data, t1 and t2 are the start and end times of the discharge process in this cycle, respectively, and SOC. t1 This is the initial state of charge (SOC) for this discharge cycle. t2 This terminates the SOC (State of Charge) for this discharge cycle. 3.The method of claim 1, wherein, The step S2 further comprises: Step S21, based on the pre-processed take-off and landing segment data, statistical features are designed for voltage, current and temperature respectively, the statistical features include the maximum value, the minimum value, the average value of the voltage and the voltage difference of the start and end time, after all the statistical features are included in the candidate feature pool, the Pearson correlation coefficient of each candidate feature and the battery capacity Q is calculated , the candidate features with Pearson correlation coefficient higher than the preset threshold are retained, and are arranged in descending order according to the Pearson correlation coefficient; Step S22, for the feature sequence arranged in descending order, taking the first sorted feature as the reference feature, calculating the similarity between the remaining features and the reference feature in turn using the grey correlation degree algorithm, deleting the redundant features with similarity higher than the set threshold, and finally obtaining the optimal feature sequence X composed of the optimal features as the input of the next step.

4. The method of claim 3, wherein the method is characterized by: Pearson correlation coefficient is calculated as: ； wherein, is characteristic the value on the nth cycle, the value on the nth cycle, is characteristic the average value of the n cycles, is the average value of the battery capacity sequence, is the average value of the battery capacity sequence, is the battery capacity on the nth cycle, is the battery capacity on the nth cycle, is the cumulative summation operation; The calculation method of the grey correlation degree is: ； ； wherein, is the grey correlation degree of the mth feature, is the correlation coefficient of the mth feature in the nth cycle, is the total number of cycles, is the statistical feature value of the nth cycle in the reference feature sequence, is the statistical feature value of the nth cycle of the mth feature, is the resolution coefficient, min and max are the minimum and maximum values, respectively.​​​ 5.The method of claim 1, wherein, The step S3 further comprises: Constructing the optimal features of the battery through a solution function network The mapping relationship to battery capacity is set as follows. Then we have: ,in, In order to be in The predicted battery capacity is as follows. For time, The input feature vector sequence is the optimal feature sequence. For the parameter set of the solution function network; Parameter set comprising weight parameters and bias terms of each layer of the inverse function neural network, wherein the weight parameters and bias terms of the first layer are denoted as The learning process of the weight parameters and bias terms of the layer is represented as ； in, To solve the function network F(·) of the th The output of the layer, Given the input feature vector sequence, where t is time. For the first Layer weight parameters, For the first The layer's bias term, where σ is the activation function; Then, the parameters of the learned solution function network are used to apply the input. Make a prediction to obtain the predicted battery capacity. This predicted value is then used as one of the inputs to the dynamics network to continue learning the network's parameter set. Parameter set Including weight parameters and bias terms, the learning dynamics network is in its first stage. The learning process for the layer's weight parameters and bias terms is as follows: ； wherein, is the weight parameter of the jth layer of the dynamics network, is the output of the jth layer of the dynamics network, t is time, is the sequence of input feature vectors, is the predicted battery capacity by the solution function network, ∂u / ∂t is the rate of change of the predicted battery capacity, and ∂u / ∂X is the partial derivative of the predicted battery capacity with respect to the feature vector, is the weight parameter of the jth layer of the dynamics network, is the output of the jth layer of the dynamics network, t is time, is the sequence of input feature vectors, is the bias term of the jth layer of the dynamics network, and σ is the activation function. Physical constraint loss of two networks by residual function Implementation: ； where F(t, X; Φ) is the output of the solution function neural network, is the output of the dynamics neural network, and are the parameter sets of the two networks, respectively. The total loss function of the PINN network is composed of a data loss function , a consistency loss function , and a physical constraint loss function , and the calculation method is as follows: ； wherein, is the total number of battery cycles, is the battery capacity of the cycle, is the predicted battery capacity of the cycle, is the predicted battery capacity of the cycle, is the physical constraint loss of the cycle, ReLU is an activation function, and a and b are weighting coefficients, is the total loss function, and when the total loss value reaches a preset minimum value, the model is optimal. 6.The physical information neural network-based lithium-ion battery state of health lightweight detection method according to claim 1, characterized in that, The step S4 further comprises: randomly setting multiple initial learning rates, and obtaining corresponding predicted battery capacities under each learning rate; comparing the predicted battery capacities with the real battery capacity to obtain one of the multiple predicted battery capacities that is closest to the real battery capacity, denoted as a current optimal observation; constructing two density functions based on the optimal observation to model conditional probability distributions of the battery capacity when corresponding to different learning rates; subsequently, calculating an expected improvement EI based on the conditional probability distributions, and obtaining an optimal learning rate by maximizing the EI; and learning another hyperparameter network layer in the same manner to obtain an optimal model.

7. The method of claim 1, wherein the method is a lightweight method for detecting the state of health of a lithium-ion battery based on a physical information neural network. The step S5 further includes: keeping the parameters of the inverse function network unchanged, and performing layer-by-layer fine tuning on the parameters in the kinetic network on the target domain battery data, setting the parameters in the kinetic network in the target domain as the parameters of the i-th layer are , and the loss function of the i-th layer is represented by ； in, For the first The loss function of the layer, Let MSE be the total number of samples, and MSE be the mean squared error loss function. For dynamic networks to input features and the Layer parameters The predicted output, For the first The battery capacity in the next cycle is determined by the loss function. Layer-by-layer fine-tuning optimizes the parameters of each layer in the dynamics network, resulting in the optimal dynamics network. Finally, a physical information neural network model constructed from two optimized neural networks is used to detect the battery health status.

Images