Storage battery residual life prediction system and method based on physical information neural network algorithm model

Abstract

The invention relates to the technical field of battery residual life prediction, and discloses a storage battery residual life prediction system and method based on a physical information neural network algorithm model. The method comprises the following steps: applying multi-frequency-point alternating current excitation and pulse current to a storage battery, and collecting electrochemical impedance spectroscopy data, polarized internal resistance data and voltage and current temperature data; creating a feature vector based on the electrochemical impedance spectroscopy data, the polarized internal resistance data and the voltage and current temperature data; and inputting the feature vector into a physical information neural network model, and outputting a health state prediction value and a residual life prediction value. According to the method, the physical information neural network algorithm model can be guided to learn the mapping relation according with the battery degradation mechanism, the generalization ability of the model and the physical interpretability of the prediction result are improved, and the accuracy of the battery remaining life prediction of the storage battery is further improved.

Description

Technical Field

[0001] This invention relates to the field of battery remaining life prediction technology, and in particular to a battery remaining life prediction system and method based on a physical information neural network algorithm model. Background Technology

[0002] In the power industry, batteries are a core component of AC / DC power systems, and their health status directly affects power supply safety. Traditional battery monitoring methods mainly collect surface parameters such as voltage, temperature, and internal resistance at a single frequency point. The data is superficial and cannot effectively assess the aging and degradation status of the battery, making it difficult to provide accurate data support for early battery replacement.

[0003] While existing pure data-driven neural network prediction methods can learn statistical patterns from historical data, they lack the constraints of the physical mechanisms of battery degradation. This results in insufficient generalization ability of the model when the training samples are limited or when faced with unseen operating conditions. The prediction results lack physical interpretability and are prone to unreasonable predictions that violate electrochemical laws. Summary of the Invention

[0004] The main objective of this invention is to provide a battery remaining life prediction system and method based on a physical information neural network algorithm model. This invention can guide the physical information neural network algorithm model to learn a mapping relationship that conforms to the battery degradation mechanism, improve the generalization ability of the model and the physical interpretability of the prediction results, and thus improve the accuracy of battery remaining life prediction.

[0005] To achieve the above objectives, this invention provides a method for predicting the remaining life of a battery based on a physical information neural network algorithm model, comprising the following steps: Multi-frequency AC excitation and pulsed current were applied to the battery, and electrochemical impedance spectroscopy data, polarization internal resistance data, and voltage, current and temperature data were collected. A feature vector is created based on the electrochemical impedance spectroscopy data, the polarization internal resistance data, and the voltage, current, and temperature data. The feature vector is input into the physical information neural network model, which outputs the predicted health status and the predicted remaining lifespan.

[0006] Optionally, in a first implementation of the first aspect of the present invention, multi-frequency AC excitation and pulsed current are applied to the battery, and electrochemical impedance spectroscopy data, polarization internal resistance data, and voltage, current, and temperature data are collected, including: Multi-frequency AC excitation and pulse current are applied to the battery, and impedance response, pulse voltage response and voltage, current and temperature data are collected; The impedance response is subjected to frequency domain transformation to obtain electrochemical impedance spectroscopy data, and the pulse voltage response is subjected to state estimation to obtain polarization internal resistance data.

[0007] Optionally, in a second implementation of the first aspect of the present invention, creating a feature vector based on the electrochemical impedance spectroscopy data, the polarization resistance data, and the voltage, current, and temperature data includes: The complex impedance values ​​at each frequency point of the electrochemical impedance spectroscopy data are decomposed into real and imaginary parts to extract the ohmic internal resistance, charge transfer impedance and diffusion impedance parameters, and the ohmic internal resistance, charge transfer impedance and diffusion impedance parameters are used as the frequency domain impedance feature set. The steady-state value of polarization internal resistance, the polarization time constant, and the peak value of polarization voltage are extracted from the polarization internal resistance data, and the steady-state value of polarization internal resistance, the polarization time constant, and the peak value of polarization voltage are used as a time-varying feature set. The charging segment is extracted from the voltage, current and temperature data, and the mean voltage, standard deviation of voltage, rate of change of voltage gradient, peak current, current decay rate and temperature rise amplitude are calculated and used as the statistical feature set of the charging segment. A feature vector is constructed based on the frequency domain impedance feature set, the time-varying feature set, and the charging segment statistical feature set.

[0008] Optionally, in a third implementation of the first aspect of the present invention, before inputting the feature vector into the physical information neural network model, the battery remaining life prediction method based on the physical information neural network algorithm model further includes: The battery is characterized as a multi-level equivalent circuit with negative electrode impedance, electrolyte impedance and positive electrode impedance connected in series, and a state space equation is constructed. The state variables of the state space equation include solid electrolyte interface membrane impedance, charge transfer impedance and diffusion impedance. The predicted output features of the target neural network are used as the state variable estimates of the state space equation, and the physical residuals after substituting the state variable estimates into the state space equation are calculated to obtain the physical constraint loss term. The target neural network includes a convolutional neural network, a long short-term memory network, and a multilayer perceptron. A total loss function is constructed, which includes a data-driven loss term, a physical constraint loss term, and a boundary condition loss term. The target neural network is trained using the total loss function to obtain a physical information neural network model.

[0009] Optionally, in a fourth implementation of the first aspect of the present invention, a total loss function comprising a data-driven loss term, the physical constraint loss term, and a boundary condition loss term is constructed, and the target neural network is trained with the total loss function to obtain a physical information neural network model, comprising: Construct a total loss function that includes the data-driven loss term, the physical constraint loss term, and the boundary condition loss term, based on the physical constraint weight coefficient and the boundary condition weight coefficient; Randomly extract a batch of samples and their true labels from the training set, input the sample vectors into the target neural network, calculate the predicted output features, and calculate the total loss function value based on the predicted output features and the true labels; The gradient of the weight parameters of each layer is calculated by backpropagation based on the total loss function value, the weight parameters are updated using an optimizer, and the learning rate is adjusted exponentially. The total loss function value is monitored on the validation set. When the total loss function value does not decrease for N consecutive rounds, the iteration is terminated, and the physical information neural network model is obtained.

[0010] Optionally, in a fifth implementation of the first aspect of the present invention, the feature vector is input into a physical information neural network model, and the predicted health status and remaining life expectancy are output, including: The feature vectors are divided into impedance spectrum feature vectors, time series feature vectors, and statistical feature vectors. The impedance spectrum feature vector is input into the convolutional neural network in the physical information neural network model to extract frequency domain spatial features. The time series feature vector is input into the long short-term memory network in the physical information neural network model to extract degradation trend features. The statistical feature vector is input into the multilayer perceptron in the physical information neural network model to extract nonlinear mapping features. The frequency domain spatial features, the degradation trend features, and the nonlinear mapping features are input into the fully connected output layer of the physical information neural network model for nonlinear activation transformation, and the predicted health status and remaining lifespan are output.

[0011] Optionally, in a sixth implementation of the first aspect of the present invention, the frequency domain spatial features, the degradation trend features, and the nonlinear mapping features are input into the fully connected output layer of the physical information neural network model for nonlinear activation transformation, outputting health status prediction values ​​and remaining lifetime prediction values, including: The frequency domain spatial features, the degradation trend features, and the nonlinear mapping features are concatenated into a comprehensive feature vector; The comprehensive feature vector is input into the fully connected output layer of the physical information neural network model for weighted summation and nonlinear activation transformation. The health status prediction value is calculated through the first output node, and the remaining lifespan prediction value is calculated through the second output node.

[0012] Optionally, in the seventh implementation of the first aspect of the present invention, after outputting the predicted health status value and the predicted remaining life value, the battery remaining life prediction method based on the physical information neural network algorithm model further includes: The battery is controlled to perform constant current discharge by remote capacity control, and the discharge current and discharge time are collected in real time. The actual discharge capacity is obtained by integrating the product of the discharge current and the discharge time. Divide the actual discharge capacity by the rated capacity of the battery and multiply by 100% to obtain the actual capacity measurement value; Calculate the prediction error between the predicted health status value and the actual capacity measurement value; When the prediction error value exceeds a set threshold, a fine-tuning dataset is constructed to perform transfer learning updates on the physical information neural network model, resulting in an updated physical information neural network model.

[0013] Optionally, in an eighth implementation of the first aspect of the present invention, when the prediction error value exceeds a set threshold, a fine-tuning dataset is constructed to perform transfer learning updates on the physical information neural network model to obtain an updated physical information neural network model, including: The newly acquired electrochemical impedance spectroscopy data, polarization internal resistance data, voltage, current and temperature data, and corresponding actual capacity measurements are merged with historical training data to obtain a fine-tuning dataset. The weight parameters of the physical constraint branch in the physical information neural network model are frozen. The first bottom layer weights of the convolutional neural network and the second bottom layer weights of the long short-term memory network are set as the first learning rate. The first weights of the multilayer perceptron and the second weights of the fully connected output layer are set as the second learning rate. The second learning rate is greater than the first learning rate. A batch of samples is extracted from the fine-tuned dataset and input into the physical information neural network model to calculate the total loss function. The weight parameters are updated by backpropagation to obtain the updated physical information neural network model.

[0014] This invention also provides a battery remaining life prediction system based on a physical information neural network algorithm model, comprising: The acquisition module is used to apply multi-frequency AC excitation and pulse current to the battery and acquire electrochemical impedance spectroscopy data, polarization internal resistance data, and voltage, current and temperature data. A module is created to generate feature vectors based on the electrochemical impedance spectroscopy data, the polarization internal resistance data, and the voltage, current, and temperature data. The output module is used to input the feature vector into the physical information neural network model and output the health status prediction value and the remaining lifespan prediction value.

[0015] In summary, this invention comprehensively characterizes the internal degradation state of a battery from frequency, time, and statistical dimensions by acquiring electrochemical impedance spectroscopy data, polarization resistance data, and voltage, current, and temperature data through multi-frequency AC excitation and pulsed current acquisition. Compared to traditional methods that only acquire a single parameter, this approach can more accurately capture battery aging characteristics. By constructing a state-space equation based on an electrochemical equivalent circuit model and transforming it into physical constraint terms embedded in the neural network loss function, the neural network must not only fit the observed data during training but also satisfy the physical laws of battery degradation. Thus, even with limited training samples, physical constraints can guide the network to learn mapping relationships that conform to the battery degradation mechanism, improving the model's generalization ability and the physical interpretability of the prediction results. A multi-network collaborative architecture is employed, utilizing convolutional neural networks to extract frequency-domain spatial features of impedance spectra, long short-term memory networks to extract temporal degradation trend features, and multilayer perceptrons to extract statistical features, to achieve comprehensive multi-scale capture of battery degradation characteristics. By comparing the actual capacity measurement value with the predicted value through core capacity operation, when the prediction error exceeds the threshold, the transfer learning update mechanism is activated. The newly collected data and historical training data form a fine-tuning dataset, and the weight parameters of the model are updated with differentiated learning rate, so that the model can gradually learn the individual differences of different batches of batteries and the operating characteristics of different working conditions. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the steps of a battery remaining life prediction method based on a physical information neural network algorithm model in one embodiment of the present invention; Figure 2 This is a block diagram of the battery remaining life prediction system based on the physical information neural network algorithm model in an embodiment of the present invention.

[0017] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] Reference Figure 1 This embodiment provides a method for predicting the remaining life of a battery based on a physical information neural network algorithm model, including the following steps: S1 applies multi-frequency AC excitation and pulse current to the battery and collects electrochemical impedance spectroscopy data, polarization internal resistance data, and voltage, current and temperature data. S2, a feature vector is created based on electrochemical impedance spectroscopy data, polarization internal resistance data, and voltage, current, and temperature data; S3 inputs the feature vector into the physical information neural network model and outputs the predicted health status and the predicted remaining lifespan.

[0020] In one example, multi-frequency AC excitation and pulsed current are applied to the battery, and electrochemical impedance spectroscopy data, polarization resistance data, and voltage, current, and temperature data are collected, including: Multi-frequency AC excitation and pulse current are applied to the battery, and impedance response, pulse voltage response and voltage, current and temperature data are collected; The impedance response is frequency-domain transformed to obtain electrochemical impedance spectroscopy data, and the pulse voltage response is state-estimated to obtain polarization internal resistance data.

[0021] In this example, under the condition that the battery is in float charging or constant current charging state, the digital signal generator and the main control microcontroller (MCU) integrated in the single-cell monitoring module are used to control the output of a set of continuous multi-frequency AC sinusoidal signals as excitation current applied to the positive and negative terminals of the battery. The voltage amplitude of the excitation signal is set to 10 mV to 50 mV, the frequency starts from 0.5 Hz and increases logarithmically up to 7.5 kHz, and is maintained at each frequency point for a duration of not less than 3 to 10 cycles to ensure that the system fully responds to the steady state. At the same time, the instantaneous response curves of the battery terminal voltage and current corresponding to the applied signal are acquired with high precision, and the original analog signal is digitized by a high-speed sampling ADC to form a sampling set of not less than 20 discrete frequency points. The fast Fourier transform module built into the MCU performs frequency domain transformation on the acquired time-domain voltage and current signals to obtain the complex impedance value Z(ω) corresponding to each frequency point. The complex impedance consists of a real part and an imaginary part, which are used to characterize the ohmic component and the capacitive reactance and inductive reactance components, respectively. By plotting the complex impedance at all frequency points on the complex plane, a Nyquist spectrum can be obtained. The real intercept in the high-frequency region is extracted as the ohmic internal resistance, the semicircle diameter in the mid-frequency region is used to estimate the charge transfer impedance, and the slope in the low-frequency region reflects the diffusion impedance Warburg coefficient, thus obtaining impedance spectrum data characterizing the battery's intrinsic electrochemical processes. A constant discharge pulse current with an amplitude of 20 amps and a pulse width of 100 to 500 milliseconds is applied to the battery by the control cell monitoring module, with the current not exceeding 0.1C of the battery's rated capacity. During the pulse application period and the transition phase after its removal, the battery terminal voltage response curve V(t) is recorded in real time at a high sampling rate of 10 kHz to 100 kHz. This curve exhibits multi-level voltage drop and recovery characteristics caused by electrochemical polarization effects. Based on the known discharge current input and sampled voltage output, an extended state-space model is constructed, where state variables include ohmic voltage drop, charge transfer voltage drop, and concentration voltage drop. The input is the pulse current I(t), and the output is the terminal voltage V(t). After establishing the state equation and output equation, a Kalman filter or Luneburger observer is used to estimate the unmeasurable internal state, identifying the dynamically changing polarization internal resistance parameter. Its steady-state value represents the final internal resistance level at which the system tends to equilibrium after polarization. The entire polarization identification process is completed within three seconds after the pulse is removed, with an estimation accuracy better than 0.001 milliohms, and voltage, current, and temperature data under environmental and operating conditions are recorded simultaneously.

[0022] State estimation of the pulse voltage response yields polarization resistance data, including: establishing a second-order equivalent circuit model of the battery; decomposing the pulse voltage response into three components: instantaneous ohmic voltage drop, electrochemical polarization voltage drop, and concentration polarization voltage drop; constructing an extended state-space equation containing ohmic, electrochemical, and concentration polarization resistance state variables, with the pulse current as the input and the terminal voltage as the output; and using a Luneburg observer or Kalman filter to estimate the state of the extended state-space equation, directly calculating the voltage change amplitude to current ratio at the moment of pulse application. Ohmic internal resistance is used to extract electrochemical polarization internal resistance and its corresponding time constant by fitting an exponential decay curve based on the rapid recovery process of the voltage response after pulse removal. Concentration polarization internal resistance and its corresponding time constant are extracted by fitting the slow recovery process of the voltage response. Time series analysis is performed on the polarization internal resistance estimation results for multiple consecutive pulse cycles. The mean value of the polarization internal resistance is calculated as the steady-state value of the polarization internal resistance, the time constant of the rapid recovery process is calculated as the electrochemical polarization time constant, and the time constant of the slow recovery process is calculated as the concentration polarization time constant. The three together constitute the time-varying feature set of the polarization internal resistance data.

[0023] In one example, a feature vector is created based on electrochemical impedance spectroscopy data, polarization resistance data, and voltage, current, and temperature data, including: The real and imaginary parts of the complex impedance values ​​at each frequency point of the electrochemical impedance spectroscopy data are decomposed to extract the ohmic internal resistance, charge transfer impedance and diffusion impedance parameters, and the ohmic internal resistance, charge transfer impedance and diffusion impedance parameters are used as the frequency domain impedance feature set. The steady-state value of polarization internal resistance, polarization time constant, and peak value of polarization voltage are extracted from the polarization internal resistance data, and the steady-state value of polarization internal resistance, polarization time constant, and peak value of polarization voltage are used as a time-varying feature set. The charging segment is extracted from the voltage, current and temperature data, and the mean voltage, standard deviation of voltage, rate of change of voltage gradient, peak current, current decay rate and temperature rise amplitude are calculated and used as the statistical feature set of the charging segment. Feature vectors are constructed based on frequency domain impedance feature sets, time-varying feature sets, and charging segment statistical feature sets.

[0024] In this example, the complex impedance Z(ω) corresponding to each frequency point in the electrochemical impedance spectroscopy data is mathematically decomposed, and its real and imaginary parts are extracted. These parts are then plotted on the complex plane to form a Nyquist spectrum. Parameters are extracted by combining the different frequency band characteristics of this spectrum. The real intercept in the high-frequency region reflects the total ohmic internal resistance of the electrolyte-conductor path, the semicircle diameter in the mid-frequency region characterizes the charge transfer impedance, which reflects the reactivity of electrons and ions at the electrode interface, and the slope in the low-frequency region reflects the diffusion-limited process, corresponding to the Warburg diffusion impedance parameter, which is related to the evolution of the concentration gradient inside the electrode material. These three impedance parameters are combined to form a frequency domain impedance feature set. The polarization resistance change curve identified based on the pulse current response was analyzed. During the recovery phase after the discharge pulse ends, the monitoring terminal voltage gradually recovers to a stable value. During this process, the dynamic response of the polarization process was identified by fitting a state-space model. Three features were extracted: the steady-state value of the polarization resistance in the final stable phase, characterizing the residual internal resistance after polarization decay; the characteristic constant of the time required for the system to reach steady state, reflecting the time delay and inertia of the polarization process; and the voltage drop amplitude at the initial stage of the discharge pulse application, i.e., the peak polarization voltage, representing the nonlinear stress effect in the transient response. These three features together constitute a time-varying feature set reflecting the dynamic behavior of the battery. Representative charging process segments were selected as analysis windows from voltage, current, and temperature data collected during long-term operation monitoring. For example, the constant current charging range between 60% and 95% SOC was selected because the voltage curve rise rate is significant and the electrochemical reaction is relatively stable during this stage, making it suitable as a feature extraction range. Within this interval, the mean, standard deviation, and average value of the first-order difference sequence dV / dt of the voltage time series V(t) are calculated to reflect the changing trends of the average voltage level, voltage fluctuation, and charging slope during this period, respectively. The maximum current value and the rate of change of current over time dI / dt of the current time series I(t) are extracted to assess the battery's load capacity and output stability during the constant current phase. The temperature rise amplitude ΔT, i.e., the difference between the final and initial temperatures, is calculated for the temperature series T(t) to reflect the battery's thermal load response during charging. This set of statistics constitutes the statistical feature set of the charging segment. The frequency domain impedance feature set, time-varying feature set, and charging segment statistical feature set are then concatenated and integrated to form a feature vector.

[0025] The complex impedance values ​​at each frequency point of the electrochemical impedance spectroscopy data are decomposed into real and imaginary parts to extract ohmic resistance, charge transfer impedance, and diffusion impedance parameters. This includes: plotting the complex impedance values ​​at each frequency point on the complex plane to form an impedance spectrum; dividing the impedance spectrum into high-frequency, mid-frequency, and low-frequency regions according to frequency range; extracting the ohmic resistance value from the intercept in the high-frequency region; calculating the diameter of the semicircular arc from the semicircular arc fitting in the mid-frequency region to obtain the charge transfer impedance value; and calculating the diffusion impedance coefficient from the slope of the straight line segment in the low-frequency region. The center position and relaxation frequency of the semicircular arc in the mid-frequency region are extracted; the frequency dispersion index of the double-layer capacitance is calculated based on the degree of deviation of the center position from the real axis; and the time constant of the charge transfer process is calculated based on the relaxation frequency. The slope trend of the impedance curve in the low-frequency region is analyzed; when the slope is close to vertical, it is determined to be a semi-infinite diffusion characteristic; when the slope is a sloping straight line, it is determined to be a finite-length diffusion characteristic. The diffusion time constant and diffusion characteristic length are calculated according to different diffusion characteristic types.

[0026] After constructing feature vectors based on frequency domain impedance feature sets, time-varying feature sets, and charging segment statistical feature sets, the battery remaining life prediction method based on the physical information neural network algorithm model further includes: extracting charge transfer impedance values ​​at multiple frequency points from the frequency domain impedance feature set, calculating the gradient of charge transfer impedance values ​​as a function of frequency, determining that the battery is in a rapid degradation stage when the absolute value of the gradient is greater than a preset threshold, and determining that the battery is in a stable aging stage when the absolute value of the gradient is less than the preset threshold; establishing a mapping relationship model between charge transfer impedance and battery remaining capacity, substituting the charge transfer impedance values ​​into the mapping relationship model to calculate a preliminary capacity estimate, comparing the preliminary capacity estimate with the actual capacity measurement value obtained from the previous capacity verification operation, and calculating the capacity change rate; determining whether to immediately trigger a remote capacity verification operation based on the capacity change rate, prioritizing the triggering of a remote capacity verification operation to obtain an accurate capacity value before performing neural network prediction when the capacity change rate does not exceed the set degradation rate threshold, directly inputting the feature vector into the physical information neural network model for prediction.

[0027] In one example, before inputting the feature vector into the physical information neural network model, the battery remaining life prediction method based on the physical information neural network algorithm model also includes: The battery is characterized as a multi-level equivalent circuit with negative electrode impedance, electrolyte impedance and positive electrode impedance connected in series. A state-space equation is constructed. The state variables of the state-space equation include solid electrolyte interface membrane impedance, charge transfer impedance and diffusion impedance. The predicted output features of the target neural network are used as the state variable estimates of the state space equation, and the physical residuals after substituting the state variable estimates into the state space equation are calculated to obtain the physical constraint loss term. The target neural network includes convolutional neural networks, long short-term memory networks and multilayer perceptrons. Construct a total loss function that includes a data-driven loss term, a physical constraint loss term, and a boundary condition loss term. Train the target neural network with the total loss function to obtain a physical information neural network model.

[0028] In this example, at the physical modeling level, the battery is modeled as a segmented equivalent circuit network consisting of negative electrode impedance, electrolyte impedance, and positive electrode impedance connected in series. The negative electrode impedance includes, in turn, a parallel structure of the negative electrode interface film resistance and capacitance, a parallel structure of the negative electrode charge transfer resistance and double-layer capacitance, and a diffusion impedance term consisting of the negative electrode diffusion impedance coefficient. The electrolyte impedance is modeled as a single electrolyte ohmic resistance. Similarly, the positive electrode impedance includes a parallel structure of the positive electrode interface film resistance and capacitance, a parallel structure of the positive electrode charge transfer resistance and double-layer capacitance, and a diffusion impedance term corresponding to the positive electrode diffusion impedance coefficient. Based on this structure, each parameter constituting the electrochemical submodule is defined as a state variable in the state-space model, including: negative electrode interface film resistance, negative electrode interface film capacitance, negative electrode charge transfer resistance, negative electrode double-layer capacitance, negative electrode diffusion impedance coefficient, electrolyte ohmic resistance, positive electrode interface film resistance, positive electrode interface film capacitance, positive electrode charge transfer resistance, positive electrode double-layer capacitance, and positive electrode diffusion impedance coefficient, totaling 11 state variables. In the state-space equations, the temporal evolution of each set of state variables is controlled by external stimuli such as input variables, charging current, and operating temperature. Based on battery aging mechanisms such as SEI film growth, charge transfer reaction rate, or diffusion slowness, a recursive model is established in the form of discrete difference equations. At the neural network structure level, to enhance the multimodal perception capability of impedance spectrum, voltage and current time series, and statistical features, the target neural network adopts a hybrid architecture jointly constructed from convolutional neural networks, long short-term memory networks, and multilayer perceptrons to process frequency domain features, temporal features, and statistical features respectively. The intermediate layer outputs of the above three neural networks are integrated into a unified representation through feature concatenation and then fed into the main branch network. This branch output includes predicted battery health status and remaining lifetime, as well as estimated values ​​of 11 state variables used for electrochemical physics modeling, namely, estimated values ​​of various resistance, capacitance, and diffusion parameters. The estimated state variables predicted by the neural network are substituted into the physical state-space equations as input. The theoretical state prediction for the next time step is calculated using current input conditions such as charging current. This prediction is then compared with the output state variables predicted by the neural network at the next time step to obtain the physical consistency residual. The residual is squared and weighted to form the physical constraint loss term, which measures whether the neural network follows the basic laws of battery degradation evolution. To achieve joint learning under multi-objective constraints, the total loss function consists of three parts: a data-driven loss term, which measures the error between the neural network's predicted SOH and RUL values ​​and the actual values; a physical constraint loss term based on the consistency of state variable evolution, used to restrict the model's predicted path to conform to the electrochemical kinetic evolution trend; and a boundary condition loss term, used to constrain the physical feasibility of the output results, such as the battery health state must be between 0 and 100%, remaining lifetime cannot be negative, and any resistance or capacitance value must be positive.By backpropagating and optimizing the gradient of the total loss function, a physical information neural network model that conforms to physical laws, is highly interpretable, and has high-precision prediction capabilities is obtained.

[0029] The convolutional neural network layer is specifically designed to process electrochemical impedance spectroscopy (EIS) data. It receives a 2×20 two-dimensional matrix as input, where the first row represents the real impedance values ​​at 20 frequency points, and the second row represents the corresponding imaginary impedance values. The first convolutional layer has 16 convolutional kernels, each 1×3 in size (spanning three adjacent frequency points in the frequency dimension), with a stride of 1. The kernel weight matrix is ​​initialized using the Kaiming method before training to accommodate the subsequent ReLU activation function. The convolution operation is performed as follows: for each 1×3 local window of the input matrix, the kernel is multiplied element-wise with the elements within the window, summed, and then a bias term is added to obtain a scalar output value. The kernel slides across the frequency dimension with a stride of 1, moving from frequency point f1 to f18, generating 18 output positions. Therefore, the output feature map size of the first convolutional layer is 2×18×16 (2 channels, 18 positions, 16 feature maps). The output of the convolutional layer undergoes a non-linear transformation using the ReLU activation function, calculated as ReLU(x) = max(0,x). This truncates all negative values ​​to 0, preserving positive features and enhancing the network's non-linear expressive power. The activated feature map is then input into a max-pooling layer with a 1×2 pooling window and a stride of 2. The maximum value is taken from every two adjacent positions in the frequency dimension, downsampling the 18 positions to 9, resulting in an output feature map size of 2×9×16. The second convolutional layer, configured with 32 kernels (also 1×3), convolves the 2×9×16 feature map output from the first layer, resulting in an output feature map size of 2×7×32. After a second max-pooling (1×2 window), the feature map size is reduced to 2×3×32. Finally, the 3D feature map is flattened into a 1D vector with dimensions of 2×3×32 = 192, but the actual implementation may use 256 dimensions (possibly through an additional fully connected mapping layer). The working principle of CNN layers is to automatically learn the local patterns of impedance spectrum in different frequency ranges through local receptive fields and weight sharing mechanisms, such as the ohmic impedance plateau in the high frequency region, the charge transfer semicircle in the mid frequency region, and the diffusion impedance oblique line in the low frequency region. Compared with manually extracting feature parameters, convolutional neural networks can capture more subtle frequency domain texture changes and are more sensitive to the differences in impedance spectrum morphology at different stages of battery degradation.

[0030] The Long Short-Term Memory (LSTM) network layer specifically processes the time-series data of voltage, current, and temperature during battery charging and discharging. It receives a T×3 two-dimensional sequence matrix as input, where T is the time step size ranging from 50 to 200 (corresponding to a sampling duration of 5 to 20 seconds). The three channels correspond to the normalized values ​​of voltage, current, and temperature, respectively. The LSTM employs a two-layer stacked structure. The first LSTM layer contains 64 hidden units, and the second LSTM layer contains 32 hidden units. Each LSTM unit contains four core gating mechanisms, and the calculation process is as follows: At time t, the LSTM unit receives the current input x. t (3-dimensional vector) and the hidden state h from the previous time step {t-1} (The first layer is 64-dimensional, and the second layer is 32-dimensional.) First, calculate the forget gate f. t =σ(W f ·[h {t-1} , x t ]+b f ), where σ is the value between 0 and 1 of the Sigmoid activation function output, and W f The forget gate weight matrix (size 64×67 or 32×35), b f As the bias vector, the forget gate determines the state from cell state c. {t-1} How much historical information is forgotten; then calculate the input gate i. t =σ(W i ·[h {t-1} , x t ]+b i ) and candidate cell state C̃ t =tanh(W c ·[h {t-1} , x t ]+b c The input gate determines how much new information is accepted, and candidate states are mapped to the -1 to 1 interval using the tanh activation function; the cell state is updated to c. t =f t ⊙c {t-1} +i t ⊙C̃ t Where ⊙ denotes element-wise multiplication (Hadamard product), this formula reflects the balance between forgetting old information and accepting new information; finally, the output gate o is calculated. t =σ(W o ·[h {t-1} , x t ]+b o ) and hidden state h t =o t ⊙tanh(c tThe output gate controls how much cell state information is output to the hidden state. After the first LSTM layer processes all T time steps, it outputs the 64-dimensional hidden state h at each time step. t The input sequence to the second LSTM layer is used for gating calculation again, and finally the 32-dimensional hidden state h at the last time step is output. T The weight matrix W of the LSTM network serves as a feature representation of the entire time series. f W i W c W o Using an orthogonal initialization method, the bias b of the forget gate in the bias term... f Initialize to 1 (encouraging the retention of more historical information in the early stages), and initialize other biases to 0. The working principle of LSTM layers is to selectively remember or forget historical information through a gating mechanism. It is particularly suitable for capturing long-term dependencies such as battery capacity decay and internal resistance growth, such as the degradation trend of the battery over the past hundreds of charge-discharge cycles. Compared with traditional recurrent neural networks (RNNs), LSTM effectively alleviates the gradient vanishing problem through the linear propagation path of cell states and can learn long-term dependency patterns spanning tens or even hundreds of time steps. This is crucial for tasks that require historical trajectory information, such as predicting the remaining battery life (RUL).

[0031] The multilayer perceptron layer is responsible for processing the statistical feature vectors extracted from the charging segment, including 58-dimensional high-level statistics such as mean voltage, standard deviation, gradient rate of change, peak current, decay rate, and temperature rise. The MLP adopts a fully connected feedforward neural network structure, containing three hidden layers with 128, 64, and 32 neurons respectively. The calculation process of the first hidden layer is as follows: the 58-dimensional input feature vector x is multiplied by the weight matrix W1 (size 128×58), and the bias vector b1 (128-dimensional) is added to obtain a linear combination z1=W1·x+b1. Then, a nonlinear transformation h1=ReLU(z1)=max(0,z1) is performed through the ReLU activation function, and the 128-dimensional feature vector h1 is output. The calculation process of the second hidden layer is z2=W2·h1+b2, h2=ReLU(z2), where W2 has a size of 64×128, b2 is 64-dimensional, and the 64-dimensional feature vector h2 is output. The calculation process for the third hidden layer is z3 = W3·h2 + b3, where h3 = ReLU(z3), W3 has a size of 32×64, b3 is 32-dimensional, and the output 32-dimensional feature vector h3 is the final output of the MLP module. The weight matrices W1, W2, and W3 of each layer of the MLP are initialized using the Xavier method. The calculation formula is that the weights are randomly sampled from a uniform distribution. This initialization method ensures that the variance of the activation values ​​during forward propagation and the variance of the gradient during backpropagation remain relatively stable across layers. The bias vectors b1, b2, and b3 are all initialized to 0. The working principle of the MLP layers is to learn the high-order interaction relationships between input features through multi-layer nonlinear transformations. For example, the product effect of voltage standard deviation and temperature rise may indicate the risk of short circuit inside the battery, and the ratio of current peak to voltage gradient may reflect the degree of increase in battery internal resistance. These complex feature combination patterns are difficult to extract through manual design. By stacking multiple nonlinear transformation layers, the MLP can automatically learn the mapping function from the original statistical features to high-level abstract features, providing a highly discriminative feature representation for the final SOH and RUL predictions.

[0032] The feature fusion layer is responsible for integrating the 256-dimensional frequency domain spatial features extracted by CNN, the 32-dimensional temporal degradation features extracted by LSTM, and the 32-dimensional statistical features extracted by MLP. The fusion process employs a feature concatenation operation, specifically calculated as follows: defining three feature vectors f... cnn ∈R^256、f lstm ∈R^32、f mlp ∈R^32, the feature fusion operation is f concat =[f cnn ; f lstm ; f mlp The semicolon (] indicates vector concatenation, resulting in a 320-dimensional comprehensive feature vector f. concat∈R^320. The concatenation operation is computationally very simple, requiring only the elements of the three vectors to be arranged in order into a long vector, without involving any weight matrix or activation function, with a computational complexity of O(1). The concatenated 320-dimensional feature vector may have too high a dimension, leading to an explosion of parameters in the subsequent fully connected layer. Therefore, depending on the actual needs, a dimension-reducing fully connected layer can be added. This layer contains 58 neurons (consistent with the dimension of the original statistical features), and the calculation process is f reduced =W reduce ·f concat +b reduce W reduce The size is 58×320, outputting a 58-dimensional feature vector; or it can directly maintain 320 dimensions as input to the subsequent PINN main network. The working principle of the feature fusion layer is to achieve the complementarity of multimodal information. The frequency domain spatial patterns captured by CNN reflect the spatial distribution characteristics of the electrochemical processes inside the battery (such as impedance differences at different electrode locations), the temporal degradation trend captured by LSTM reflects the aging dynamics of the battery accumulated over time (such as capacity decaying exponentially), and the statistical features captured by MLP reflect the instantaneous state of the charging process (such as abnormal high temperature or current fluctuations during a charge). These three features describe the battery health status from different perspectives. By splicing and fusing them, more comprehensive information input can be provided, improving the robustness and accuracy of prediction.

[0033] The main network layer of the physical information neural network receives the feature vector (58-dimensional or 320-dimensional) output from the feature fusion layer, performs forward propagation through a multi-layer fully connected neural network, and embeds electrochemical physical constraints during training. The network structure contains three hidden layers and one output layer. The number of neurons in the hidden layers is 128, 64, and 32 respectively, and the output layer contains two neurons that output the predicted values ​​of SOH and RUL, respectively. The calculation process for the first hidden layer is z1 = W1·x input +b1, h1=tanh(z1), where x input The input feature vector, W1, has a size of 128×input. dim (input) dim The activation function is tanh(x) = (e^xe^(-x)) / (e^x+e^(-x)), with an output range of (-1, 1). Compared to ReLU, it maintains bounded output and avoids activation value explosion. The second hidden layer is calculated as z2 = W2·h1 + b2, h2 = tanh(z2), and W2 has a size of 64 × 128. The third hidden layer is calculated as z3 = W3·h2 + b3, h3 = tanh(z3), and W3 has a size of 32 × 64. In the output h2 of the second hidden layer, an 11-dimensional vector is separated through linear transformation as the neural network estimate of the electrochemical equivalent circuit parameters. The output layer is calculated as z out =W out·h3+b out The output layer applies a sigmoid activation function to the SOH and multiplies it by 100 to map it to the 0-100% range. The calculation formula is SOH = 100·σ(z out [0]), where σ(x) = 1 / (1 + e^(-x)); the ReLU activation function is used for RUL to ensure non-negativity, and the calculation formula is RUL = max(0, z). out [1]). The total number of network weight parameters is input. dim ×128+128×64+64×32+32×2+64×11≈15000 parameters (based on input) dim (For example, x = 58). The embedding of physical constraints is achieved through a loss function. During training, the total loss function is defined and calculated as follows: x n Substituting n into the state-space equation dx / dt=f(x,u), ∂x is calculated using automatic differentiation. n n / ∂t, compare ∂x n n / ∂t and f(x) n The difference between n and u). The working principle of the PINN main network layer is to use physical prior knowledge as a regularization term to guide the network learning, so that the network not only fits the training data, but also satisfies the physical laws of battery degradation (such as the SEI film impedance increasing according to the square root law and the capacity decreasing according to the exponential law). This physical information fusion method can significantly improve the model's generalization ability when the training samples are limited, avoid overfitting and unreasonable predictions, and at the same time give the model physical interpretability, and the prediction results can be traced back to the evolution of electrochemical parameters.

[0034] The model training employs the Adam (Adaptive Moment Estimation) optimizer, which combines the advantages of momentum and RMSprop methods, enabling adaptive adjustment of the learning rate for each parameter. The update rule of the Adam optimizer is as follows: first, calculate the loss function L... total The gradient g=∂L for each weight parameter w total / ∂w; then update the first moment estimate (moving average of the gradient) m=β1·m+(1-β1)·g, where β1=0.9 is the momentum coefficient; update the second moment estimate (moving average of the squared gradient) v=β2·v+(1-β2)·g², where β2=0.999; perform bias correction on the moment estimates m̂=m / (1-β1^t) and v̂=v / (1-β2^t), where t is the current iteration number; finally update the weights to w. new =w old-η·m̂ / (√v̂+ε), where η is the learning rate and ε=1e-8 is a minimal constant to prevent division by zero. The initial learning rate is set to η0=0.001, and is dynamically adjusted using an exponential decay strategy. The calculation formula is η(epoch)=η0·γ^(epoch / decaysteps), where γ=0.95 is the decay factor, and decay... steps =10 means that the learning rate decays to 95% of its original value every 10 epochs. During training, batches are randomly selected from the training set. size A batch consists of 32 samples; calculate the average loss L for that batch. batch =(1 / 32)·Σ i L total (i), for L batch Backpropagation is performed to compute the gradients of all parameters, and then the Adam optimizer is invoked to update the parameters. Training lasts for 500 to 1000 epochs, with each epoch iterating over the entire training set once. After each epoch, the model performance is evaluated on the validation set, and the validation loss L is calculated. val The training mechanism triggers an early stopping mechanism to terminate training when Lval (Learning Value) fails to decrease for 20 consecutive epochs, preventing overfitting, in accordance with prediction accuracy metrics (such as MAE and RMSE). The optimizer and learning rate scheduling strategy work to achieve efficient and stable parameter optimization. The Adam optimizer automatically adjusts the update step size for different parameters through adaptive learning rates. For parameters with sparse gradients (such as the weights of physical constraint branches), a larger step size is used to accelerate convergence, while for parameters with dense gradients (such as output layer weights), a smaller step size is used to avoid oscillations. The exponential decay learning rate strategy uses a larger learning rate in the early stages of training to quickly approach the optimal solution, and a smaller learning rate in the later stages for fine-tuning, improving the final convergence accuracy. Batch training and the early stopping mechanism ensure training efficiency while avoiding overfitting, ensuring the model has good generalization performance on the test set.

[0035] The output layer is the last layer of the network, responsible for mapping the 32-dimensional feature vector h3 of the third hidden layer to a 2-dimensional predicted output, corresponding to SOH and RUL respectively. The calculation of the output layer is divided into two independent neuron nodes. The first node calculates the SOH prediction value, and the second node calculates the RUL prediction value. The total loss function is a weighted sum of three terms. The working principle of the output layer and the loss function is to achieve end-to-end supervised learning. The output layer maps the network output to a physically meaningful range through an appropriate activation function. The loss function guides the network parameters to optimize in a direction that both fits the data and satisfies physical laws by penalizing prediction errors and physical violations. The design of the multi-objective loss function is the core innovation of PINN that distinguishes it from traditional neural networks, making the model not only a black-box data fitting tool, but also a physics-driven prediction model that integrates domain knowledge.

[0036] In one example, a total loss function is constructed, comprising a data-driven loss term, a physical constraint loss term, and a boundary condition loss term. The target neural network is trained using this total loss function to obtain a physical information neural network model, including: Construct a total loss function that includes data-driven loss terms, physical constraint loss terms, and boundary condition loss terms based on the physical constraint weight coefficients and boundary condition weight coefficients; Randomly select a batch of samples and their true labels from the training set, input the sample vectors into the target neural network, calculate the predicted output features, and calculate the total loss function value based on the predicted output features and the true labels; The gradient of the weight parameters of each layer is calculated by backpropagation based on the total loss function value. The weight parameters are updated using an optimizer, and the learning rate is adjusted exponentially. The total loss function value is monitored on the validation set. When the total loss function value does not decrease for N consecutive rounds, the iteration is terminated, and the physical information neural network model is obtained.

[0037] In this example, three core loss function components are defined during the network training initialization phase. The data-driven loss term measures the fitting error between the battery health status and remaining lifespan values ​​output by the neural network and the true labels, using a weighted mean square error to balance different dimensions. The physical constraint loss term calculates the physical residual based on the deviation between the estimated physical state variables output by the target neural network and the predicted values ​​from the state-space equations, measuring whether the model output conforms to electrochemical kinetics. The boundary condition loss term constrains the physical rationality of the model output, ensuring that the predicted health status value is within the range of 0 to 100%, the remaining lifespan is non-negative, and all resistance, capacitance, and diffusion parameters are positive. The positive domain constraint is achieved by setting an activation function or adding soft constraint penalties. Based on the above, physical constraint weight coefficients and boundary condition weight coefficients are introduced, constructing the overall objective function as: Total Loss = Data Loss + Weight 1 × Physical Loss + Weight 2 × Boundary Loss, serving as the overall optimization objective. During the training phase, a small batch of samples and their corresponding health status and lifespan labels are randomly selected from the training set. The input feature vector of this batch of samples is then fed into the target neural network for forward propagation calculation, obtaining a multi-dimensional prediction output including health status values, lifespan values, and a set of electrochemical and physical variables. Based on the prediction results, the three loss terms are calculated separately and then weighted and summed to obtain the total loss function value for the current batch. Backpropagation is then performed with the total loss function value as the objective. The gradient of each connection weight parameter in the network is calculated layer by layer using the chain rule of differentiation. Optimizers such as Adam, RMSProp, or SGD are used in combination with first-order and second-order moment estimation to perform gradient updates on each parameter. During the optimization process, the learning rate is dynamically adjusted according to a set exponential decay strategy, that is, the base learning rate is gradually reduced exponentially as the number of training rounds increases, achieving rapid convergence in the early stage and fine-tuning in the later stage. To avoid overfitting and improve generalization ability, the total loss function value is continuously monitored on an independent validation set during training. When the total loss function value on the validation set does not decrease further during N consecutive training rounds, it indicates that the model has converged or entered an oscillating state. At this point, the early stopping mechanism is triggered to terminate the training process and obtain the physical information neural network model.

[0038] In one example, the feature vector is input into a physical information neural network model, which outputs health status predictions and remaining life expectancy predictions, including: The eigenvectors are classified into impedance spectrum eigenvectors, time series eigenvectors, and statistical eigenvectors; Impedance spectrum feature vectors are input into the convolutional neural network in the physical information neural network model to extract frequency domain spatial features. Temporal feature vectors are input into the long short-term memory network in the physical information neural network model to extract degradation trend features. Statistical feature vectors are input into the multilayer perceptron in the physical information neural network model to extract nonlinear mapping features. Frequency domain spatial features, degradation trend features, and nonlinear mapping features are input into the fully connected output layer of the physical information neural network model for nonlinear activation transformation, and the predicted health status and remaining lifespan are output.

[0039] In this example, multi-dimensional feature engineering is performed based on the raw multi-source data collected from battery operation monitoring. The feature vectors are structured and grouped according to their physical source and information structure. Features obtained from electrochemical impedance spectroscopy analysis include the real and imaginary parts of complex impedance at different frequency points, as well as parameters such as charge transfer resistance, ohmic resistance, and diffusion impedance coefficient extracted by Nyquist spectrum fitting, which constitute the impedance spectrum feature vector. This part has obvious frequency domain structure regularity. At the same time, the time-series feature vector is formed by the voltage change rate, current decay slope, and temperature rise dynamics sequence data formed by preprocessing the constant current charging segment selected from the continuously monitored voltage, current, and temperature time series, which reflects the dynamic degradation trend of the battery during operation. Statistical feature vectors, including statistics such as voltage mean, current fluctuation variance, and temperature deviation, are used to reflect the central tendency and distribution characteristics of the overall operating status. These features are not time-dependent but have complex nonlinear interaction patterns. These three types of features are input into different sub-modules of the physical information neural network model for processing. The impedance spectrum feature vector is converted into a two-dimensional matrix form and input into the convolutional neural network substructure. Multiple convolutional kernels are used to perform sliding convolution operations on the frequency axis to extract local structural features and shape change patterns in the frequency domain. During the extraction process, activation functions and pooling operations are used to enhance the network's ability to recognize impedance spectrum features such as semi-circular arcs, slope changes, and pole structures layer by layer, forming compressed frequency domain spatial features. At the same time, the temporal feature vector is input into the long short-term memory network structure. This structure maintains the cumulative memory of key degradation features across multiple time steps through a gating mechanism, capturing the dynamic change trends of typical degradation signals such as gradual voltage increase, current decrease, or accelerated temperature rise during continuous charging of the battery. The hidden state of the last time step is output as the degradation trend feature vector. The statistical feature vector is input into a multilayer perceptron structure composed of multiple fully connected layers. Through nonlinear activation functions, high-dimensional feature interaction relationships are mapped layer by layer to extract complex correlation features and obtain nonlinear mapping feature expressions. Frequency domain spatial features, degradation trend features, and nonlinear mapping features are concatenated according to feature dimensions to form a fused feature vector, which is then input into the fully connected output layer of the physical information neural network. In the fully connected output layer, feature mapping and prediction decoding are completed through nonlinear transformation, and the current health status value and remaining life estimate of the battery are output. The health status prediction value reflects the relative level of the current capacity of the battery with respect to the rated capacity, while the remaining life prediction value is derived from the degradation rate to estimate the number of working cycles remaining to the failure threshold.

[0040] Before inputting frequency domain spatial features, degradation trend features, and nonlinear mapping features into the fully connected output layer of the physical information neural network model, the battery remaining life prediction method based on the physical information neural network algorithm model further includes: calculating the physical residual values ​​corresponding to the frequency domain spatial features, degradation trend features, and nonlinear mapping features respectively; substituting the frequency domain spatial features into the electrochemical equivalent circuit model to calculate the theoretical impedance spectrum and comparing it with the measured impedance spectrum to obtain the frequency domain physical residual; substituting the degradation trend features into the capacity decay empirical equation to calculate the theoretical capacity decay trajectory and comparing it with historical capacity data to obtain the time-series physical residual; and substituting the nonlinear mapping features into the state space... The theoretical state evolution is calculated using the inter-equation method and compared with the observed state to obtain the state physical residual. The physical confidence of the three features is calculated based on the frequency domain physical residual, the time series physical residual, and the state physical residual. The smaller the physical residual, the higher the physical confidence of the corresponding feature. The three physical confidences are normalized to obtain the frequency domain feature weight coefficient, the time series feature weight coefficient, and the statistical feature weight coefficient. The frequency domain spatial feature is multiplied by the frequency domain feature weight coefficient, the degradation trend feature is multiplied by the time series feature weight coefficient, and the nonlinear mapping feature is multiplied by the statistical feature weight coefficient. The three weighted features are then concatenated to obtain the comprehensive feature vector under physical constraints.

[0041] In one example, frequency domain spatial features, degradation trend features, and nonlinear mapping features are input into the fully connected output layer of a physical information neural network model for nonlinear activation transformation, outputting health status predictions and remaining lifespan predictions, including: Frequency domain spatial features, degradation trend features, and nonlinear mapping features are concatenated into a comprehensive feature vector; The comprehensive feature vector is input into the fully connected output layer of the physical information neural network model for weighted summation and nonlinear activation transformation. The health status prediction value is calculated through the first output node, and the remaining lifespan prediction value is calculated through the second output node.

[0042] In this example, frequency domain spatial features, degradation trend features, and nonlinear mapping features are concatenated along the feature dimensions to generate a comprehensive feature vector of uniform length. Each dimension represents a different physical level, temporal stage, or operational state variable. The comprehensive feature vector is then input into the fully connected output layer of the physical information neural network model, where weighted summation and nonlinear activation operations are performed. Specifically, each element in the comprehensive feature vector is multiplied by its corresponding connection weight, a bias term is added, and the vector is then input into the activation function for nonlinear mapping processing. The output layer contains two parallel neuron nodes, corresponding to the battery health status prediction task and the remaining life prediction task, respectively. The first output node receives the result of the weighted aggregation of all comprehensive feature vectors and normalizes it using the Sigmoid activation function to output a health status prediction value between 0 and 100. The predicted value is equivalent to the ratio of the actual capacity to the rated capacity multiplied by a percentage, which is used to represent the current health level of the battery. The second output node, based on the same comprehensive features, uses a non-negative activation function such as ReLU or Softplus to ensure that the output result is zero or positive, and outputs the predicted remaining life value. This value reflects the number of cycles or time that the battery can continue to operate before reaching the set failure threshold (such as SOH dropping to 80%) under the current aging trend.

[0043] In one example, after outputting the predicted health status and remaining life, the battery remaining life prediction method based on the physical information neural network algorithm model also includes: The battery is controlled to perform constant current discharge by remote capacity control, and the discharge current and discharge time are collected in real time. The product of discharge current and discharge time is integrated to calculate the actual discharge capacity. Divide the actual discharge capacity by the rated capacity of the battery and multiply by 100% to obtain the actual capacity measurement value; Calculate the prediction error between the predicted health status value and the actual capacity measurement value; When the prediction error exceeds a set threshold, a fine-tuning dataset is constructed to perform transfer learning updates on the physical information neural network model, resulting in an updated physical information neural network model.

[0044] In this example, a capacity control command is issued to the battery management system via a remote monitoring platform, causing the system to enter a constant current discharge mode and initiating the discharge test process. During this process, the lower-level acquisition device continuously records the discharge current I(t) and the corresponding discharge time t at a fixed sampling period, forming a set of aligned time series data. The current value during the discharge process is multiplied by the adjacent time intervals, and then summed over all time intervals, i.e., the total discharge charge ∫I(t)dt is calculated in an approximate integral form, to obtain the actual discharge capacity under the current operating condition, reflecting the battery's ability to release electrical energy under a set load. The actual discharge capacity is compared with the rated capacity marked on the battery nameplate, and then multiplied by 100% to obtain the actual capacity measurement value for the current cycle. The difference between the health status prediction value obtained by the physical information neural network model for the current cycle and the actual capacity measurement value is calculated to obtain the prediction error value. If the prediction error exceeds the set error tolerance threshold (e.g., 3% or 5%), it indicates that the generalization ability of the existing model can no longer accurately cover the current battery degradation characteristics, and the system automatically triggers the model update mechanism. In the model update process, the platform combines the currently collected battery input features (such as impedance spectrum, charging segment data, and polarization resistance) with their corresponding real health status labels to form a new sample set. It also randomly selects samples from the historical training set for mixing and concatenation to construct a fine-tuning dataset. Based on the originally trained physical information neural network, the existing structure and most weights are retained. A transfer learning strategy is used to retrain the high-level parameters of the network. Rapid adaptation is achieved by limiting the learning rate of the lower-level parameters to a small value or freezing the lower-level modules, and only increasing the learning rate of the output layer and specific intermediate layers. During fine-tuning, the original loss function structure is used, still including data fitting terms, physical constraint terms, and boundary condition terms. Backpropagation is performed by recalculating the error, and the network is iterated through a limited number of rounds using a micro-batch approach. Once the expected accuracy is achieved on the validation set, the model is replaced, and the updated neural network is deployed to the running system, enabling the model to adaptively optimize over time.

[0045] Before merging the newly acquired electrochemical impedance spectroscopy data, polarization resistance data, voltage, current, and temperature data, along with the corresponding actual capacity measurements, with the historical training data, the battery remaining life prediction method based on the physical information neural network algorithm model further includes: decomposing the prediction error value into a data fitting error component and a physical constraint violation error component; calculating the physical constraint violation error component by substituting the prediction output into the state-space equation; subtracting the physical constraint violation error component from the prediction error value to obtain the data fitting error component; and determining that the error mainly originates from physical constraints when the proportion of the physical constraint violation error component to the prediction error value is greater than a first threshold. If the weighting coefficients are set improperly, the physical constraint weighting coefficients are adjusted and the model is retrained. When the proportion of the data fitting error component to the prediction error value is greater than the second threshold, it is determined that the error mainly comes from insufficient training data or changes in data distribution, and the incremental learning process is initiated. The similarity of the feature distributions of the newly collected data and the historical training data is calculated. Cosine similarity or Euclidean distance is used to measure the distance between the feature vectors of the new data and the feature vectors of the historical data. When the distance is greater than the distribution drift threshold, it is determined that the battery operating condition has changed significantly. The new data is marked as a high-weight sample, and the loss function of the high-weight sample is multiplied by the weight amplification factor during fine-tuning training.

[0046] After calculating the prediction error between the predicted health status and the actual capacity measurement, the battery remaining life prediction method based on the physical information neural network algorithm model further includes: extracting operating condition features from the current battery's operating parameters, including cumulative cycle count, average operating temperature, charge / discharge rate, and current capacity range; normalizing the cumulative cycle count to an aging degree index, the average operating temperature to a temperature stress index, the charge / discharge rate to a current stress index, and the current capacity range to a capacity decay index; and combining the aging degree index, temperature stress index, current stress index, and capacity decay index into an operating condition feature vector, using... K-means clustering or Gaussian mixture model is used to perform cluster analysis on the operating condition feature vectors of historical battery data, dividing the battery into multiple operating condition subclasses. Each operating condition subclass maintains a dedicated physical information neural network model. The distance between the current battery's operating condition feature vector and the center of each operating condition subclass is calculated, and the operating condition subclass with the smallest distance is selected as the current battery's category. The dedicated physical information neural network model corresponding to that operating condition subclass is called for prediction. When the prediction error value exceeds a set threshold and occurs repeatedly, it is determined that the current battery's operating condition has shifted. The operating condition feature vector is recalculated and its category is updated. The system then switches to the dedicated model corresponding to the new category.

[0047] In one example, when the prediction error exceeds a set threshold, a fine-tuning dataset is constructed to perform transfer learning updates on the physical information neural network model, resulting in an updated physical information neural network model, including: The newly acquired electrochemical impedance spectroscopy data, polarization internal resistance data, voltage, current and temperature data, and corresponding actual capacity measurements are merged with historical training data to obtain a fine-tuning dataset. Freeze the weight parameters of the physical constraint branch in the physical information neural network model, set the first bottom layer weight of the convolutional neural network and the second bottom layer weight of the long short-term memory network as the first learning rate, set the first weight of the multilayer perceptron and the second weight of the fully connected output layer as the second learning rate, and set the second learning rate to be greater than the first learning rate; A batch of samples is extracted from the fine-tuned dataset and input into the physical information neural network model to calculate the total loss function. Backpropagation is then used to update the weight parameters to obtain the updated physical information neural network model.

[0048] In this example, when the model update process is initiated, the latest acquired runtime data is uploaded in a structured format by the field edge device or the platform PHM system. The runtime data includes electrochemical impedance spectroscopy features, such as the real and imaginary parts of the complex impedance at 20 frequency points, and the charge transfer resistance and diffusion impedance coefficient obtained by fitting the Nyquist plot; it also includes dynamic features such as the polarization steady-state value, polarization time constant, and polarization voltage peak value extracted from the polarization internal resistance curve measured by the pulse method; and it contains voltage, current, and temperature time series features extracted from the constant current charging segment, as well as the actual capacity measurement values ​​obtained through remote capacitance measurement, i.e., the true SOH label. The newly acquired sample data is merged with a representative portion of the original training data according to a set ratio, for example, retaining 30% of the old samples to form a playback mechanism. The new and old samples are then combined to form a structurally unified fine-tuning dataset. The weight parameters of the intermediate branches related to physical constraints in the physical information neural network are frozen, especially the internal network structure used for electrochemical state-space modeling and differential constraint expression, and their weights remain unchanged. A hierarchical learning rate strategy is set for different network modules. The first layer weights of the convolutional neural network and the second layer weights of the long short-term memory network are set with a first learning rate, which is usually low, to preserve the original feature extraction capabilities and avoid destroying the existing general representation. The first weights of the multilayer perceptron and the second weights of the final output layer are set with a higher second learning rate because they have direct control over the prediction values ​​of the current task, so as to adapt to new data distributions and changes in SOH labels more quickly and achieve rapid correction of prediction bias. The entire network uses a training mechanism with a multi-objective loss structure. Its loss function consists of a data-driven term, a physical consistency term, and a boundary reasonableness term. The data-driven term uses the mean squared error between the new sample label and the prediction result. The physical consistency term retains the calculation results in the total loss under frozen parameters. The boundary term ensures that the prediction result is physically reasonable, for example, the output SOH is between 0 and 100%, RUL is non-negative, and all physical state parameters are positive. Training samples are progressively extracted from the fine-tuning dataset according to a set batch size. The three types of input features from these training samples are then fed into the convolutional, long short-term memory, and perceptron modules of the physical information neural network, respectively. This generates three deep features, which are concatenated into a comprehensive feature vector. After nonlinear mapping at the output layer, predicted values ​​for SOH and RUL are obtained. The predicted values ​​are then compared with the labels to calculate the loss. The weight parameters of different layers are independently updated using the backpropagation algorithm according to a set learning rate, thus achieving structured control of gradient adjustment. Throughout the training process, an optimizer such as Adam is used to perform targeted optimization on each parameter group. Simultaneously, a training epoch number and an early stopping mechanism are set. The fine-tuning process terminates when the validation loss tends to converge or there is no improvement for several consecutive epochs, resulting in an updated physical information neural network model with online adaptability, which replaces the original deployed model.

[0049] ReferenceFigure 2 This embodiment provides a battery remaining life prediction system based on a physical information neural network algorithm model, including: Acquisition module 1 is used to apply multi-frequency AC excitation and pulse current to the battery and acquire electrochemical impedance spectroscopy data, polarization internal resistance data, and voltage, current and temperature data. Create Module 2 to create feature vectors based on electrochemical impedance spectroscopy data, polarization internal resistance data, and voltage, current, and temperature data; Output module 3 is used to input feature vectors into the physical information neural network model and output health status prediction values ​​and remaining life prediction values.

[0050] In this embodiment, the specific implementation of each unit in the above system embodiment is described in the above method embodiment, and will not be repeated here.

[0051] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, system, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, system, article, or method. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, system, article, or method that includes that element.

[0052] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for predicting the remaining life of a battery based on a physical information neural network algorithm model, characterized in that, include: Multi-frequency AC excitation and pulsed current were applied to the battery, and electrochemical impedance spectroscopy data, polarization internal resistance data, and voltage, current and temperature data were collected. A feature vector is created based on the electrochemical impedance spectroscopy data, the polarization internal resistance data, and the voltage, current, and temperature data. The feature vector is input into the physical information neural network model, which outputs the predicted health status and the predicted remaining lifespan.

2. The method for predicting the remaining life of a battery based on a physical information neural network algorithm model according to claim 1, characterized in that, Multi-frequency AC excitation and pulsed current were applied to the battery, and electrochemical impedance spectroscopy data, polarization resistance data, and voltage, current, and temperature data were collected, including: Multi-frequency AC excitation and pulse current are applied to the battery, and impedance response, pulse voltage response and voltage, current and temperature data are collected; The impedance response is subjected to frequency domain transformation to obtain electrochemical impedance spectroscopy data, and the pulse voltage response is subjected to state estimation to obtain polarization internal resistance data.

3. The method for predicting the remaining life of a battery based on a physical information neural network algorithm model according to claim 1, characterized in that, A feature vector is created based on the electrochemical impedance spectroscopy data, the polarization resistance data, and the voltage, current, and temperature data, including: The complex impedance values ​​at each frequency point of the electrochemical impedance spectroscopy data are decomposed into real and imaginary parts to extract the ohmic internal resistance, charge transfer impedance and diffusion impedance parameters, and the ohmic internal resistance, charge transfer impedance and diffusion impedance parameters are used as the frequency domain impedance feature set. The steady-state value of polarization internal resistance, the polarization time constant, and the peak value of polarization voltage are extracted from the polarization internal resistance data, and the steady-state value of polarization internal resistance, the polarization time constant, and the peak value of polarization voltage are used as a time-varying feature set. The charging segment is extracted from the voltage, current and temperature data, and the mean voltage, standard deviation of voltage, rate of change of voltage gradient, peak current, current decay rate and temperature rise amplitude are calculated and used as the statistical feature set of the charging segment. A feature vector is constructed based on the frequency domain impedance feature set, the time-varying feature set, and the charging segment statistical feature set.

4. The method for predicting the remaining life of a battery based on a physical information neural network algorithm model according to claim 1, characterized in that, Before inputting the feature vector into the physical information neural network model, the battery remaining life prediction method based on the physical information neural network algorithm model further includes: The battery is characterized as a multi-level equivalent circuit with negative electrode impedance, electrolyte impedance and positive electrode impedance connected in series, and a state space equation is constructed. The state variables of the state space equation include solid electrolyte interface membrane impedance, charge transfer impedance and diffusion impedance. The predicted output features of the target neural network are used as the state variable estimates of the state space equation, and the physical residuals after substituting the state variable estimates into the state space equation are calculated to obtain the physical constraint loss term. The target neural network includes a convolutional neural network, a long short-term memory network, and a multilayer perceptron. A total loss function is constructed, which includes a data-driven loss term, a physical constraint loss term, and a boundary condition loss term. The target neural network is trained using the total loss function to obtain a physical information neural network model.

5. The method for predicting the remaining life of a battery based on a physical information neural network algorithm model according to claim 4, characterized in that, Construct a total loss function comprising a data-driven loss term, a physical constraint loss term, and a boundary condition loss term; train the target neural network using the total loss function to obtain a physical information neural network model, including: Construct a total loss function that includes the data-driven loss term, the physical constraint loss term, and the boundary condition loss term, based on the physical constraint weight coefficient and the boundary condition weight coefficient; Randomly extract a batch of samples and their true labels from the training set, input the sample vectors into the target neural network, calculate the predicted output features, and calculate the total loss function value based on the predicted output features and the true labels; The gradient of the weight parameters of each layer is calculated by backpropagation based on the total loss function value, the weight parameters are updated using an optimizer, and the learning rate is adjusted exponentially. The total loss function value is monitored on the validation set. When the total loss function value does not decrease for N consecutive rounds, the iteration is terminated, and the physical information neural network model is obtained.

6. The method for predicting the remaining life of a battery based on a physical information neural network algorithm model according to claim 1, characterized in that, The feature vector is input into the physical information neural network model, which outputs predicted health status and predicted remaining lifespan, including: The feature vectors are divided into impedance spectrum feature vectors, time series feature vectors, and statistical feature vectors. The impedance spectrum feature vector is input into the convolutional neural network in the physical information neural network model to extract frequency domain spatial features. The time series feature vector is input into the long short-term memory network in the physical information neural network model to extract degradation trend features. The statistical feature vector is input into the multilayer perceptron in the physical information neural network model to extract nonlinear mapping features. The frequency domain spatial features, the degradation trend features, and the nonlinear mapping features are input into the fully connected output layer of the physical information neural network model for nonlinear activation transformation, and the predicted health status and remaining lifespan are output.

7. The method for predicting the remaining life of a battery based on a physical information neural network algorithm model according to claim 6, characterized in that, The frequency domain spatial features, the degradation trend features, and the nonlinear mapping features are input into the fully connected output layer of the physical information neural network model for nonlinear activation transformation, outputting health status prediction values ​​and remaining lifespan prediction values, including: The frequency domain spatial features, the degradation trend features, and the nonlinear mapping features are concatenated into a comprehensive feature vector; The comprehensive feature vector is input into the fully connected output layer of the physical information neural network model for weighted summation and nonlinear activation transformation. The health status prediction value is calculated through the first output node, and the remaining lifespan prediction value is calculated through the second output node.

8. The method for predicting the remaining life of a battery based on a physical information neural network algorithm model according to claim 1, characterized in that, After outputting the predicted health status and remaining life, the battery remaining life prediction method based on the physical information neural network algorithm model further includes: The battery is controlled to perform constant current discharge by remote capacity control, and the discharge current and discharge time are collected in real time. The actual discharge capacity is obtained by integrating the product of the discharge current and the discharge time. Divide the actual discharge capacity by the rated capacity of the battery and multiply by 100% to obtain the actual capacity measurement value; Calculate the prediction error between the predicted health status value and the actual capacity measurement value; When the prediction error value exceeds a set threshold, a fine-tuning dataset is constructed to perform transfer learning updates on the physical information neural network model, resulting in an updated physical information neural network model.

9. The method for predicting the remaining life of a battery based on a physical information neural network algorithm model according to claim 8, characterized in that, When the prediction error exceeds a set threshold, a fine-tuning dataset is constructed to perform transfer learning updates on the physical information neural network model, resulting in an updated physical information neural network model, including: The newly acquired electrochemical impedance spectroscopy data, polarization internal resistance data, voltage, current and temperature data, and corresponding actual capacity measurements are merged with historical training data to obtain a fine-tuning dataset. The weight parameters of the physical constraint branch in the physical information neural network model are frozen. The first bottom layer weights of the convolutional neural network and the second bottom layer weights of the long short-term memory network are set as the first learning rate. The first weights of the multilayer perceptron and the second weights of the fully connected output layer are set as the second learning rate. The second learning rate is greater than the first learning rate. A batch of samples is extracted from the fine-tuned dataset and input into the physical information neural network model to calculate the total loss function. The weight parameters are updated by backpropagation to obtain the updated physical information neural network model.

10. A battery remaining life prediction system based on a physical information neural network algorithm model, characterized in that, The steps for implementing the battery remaining life prediction method based on the physical information neural network algorithm model according to any one of claims 1 to 9 include: The acquisition module is used to apply multi-frequency AC excitation and pulse current to the battery and acquire electrochemical impedance spectroscopy data, polarization internal resistance data, and voltage, current and temperature data. A module is created to generate feature vectors based on the electrochemical impedance spectroscopy data, the polarization internal resistance data, and the voltage, current, and temperature data. The output module is used to input the feature vector into the physical information neural network model and output the health status prediction value and the remaining lifespan prediction value.

Images