A health prediction method for aviation power battery digital twins integrated with deep learning

Abstract

The present invention discloses a health prediction method for aviation power batteries that integrates digital twins with deep learning, comprising the following steps: constructing a virtual battery model of the digital twin, collecting and processing battery data to obtain a new real-time data matrix; constructing a jointly trained multi-task deep learning model, inputting the new real-time data matrix into the jointly trained multi-task deep learning model for processing to obtain a state of health (SOH) value predicted by the jointly trained multi-task deep learning model; updating the battery data based on the SOH value predicted by the jointly trained multi-task deep learning model to obtain a new data scalar, inputting the new data scalar into the virtual battery model of the digital twin, and updating the virtual battery model of the digital twin. The present invention provides an efficient and reliable solution for health management and life prediction of aviation power batteries through the deep integration of the digital twin virtual battery model and the jointly trained multi-task deep learning model.

Description

Technical Field

[0001] The present invention relates to the technical field of battery health management, and in particular to a health prediction method for aviation power batteries integrating digital twins with deep learning. Background Art

[0002] Aviation power batteries are a key energy storage device in modern aviation, widely used in equipment such as drones and electric aircraft. Their performance directly impacts an aircraft's endurance, safety, and operating costs. However, operating under high power, high load, and complex environmental conditions, aviation power batteries face significant health management and reliability challenges, including capacity fade, shortened lifespan, and difficulty predicting faults. Battery health management methods primarily focus on battery condition monitoring (such as state of health (SOH)) and risk of unsafe operation (RUL) prediction. Therefore, accurately estimating the SOH and RUL of aviation power batteries can not only extend battery life but also optimize aircraft maintenance plans, improving the safety and cost-effectiveness of system operations.

[0003] Traditional battery modeling and health prediction methods can be divided into two categories: one is physical model methods based on electrochemical mechanisms (such as equivalent circuit models and pseudo-two-dimensional models), which achieve dynamic characterization by establishing mathematical equations for internal battery reactions. Their advantages are strong model interpretability and clear physical meaning; however, they rely on precise physical parameter identification and complex electrochemical equation solutions, and have technical drawbacks such as low computational efficiency and poor real-time performance. The other is machine learning methods based on big data analysis (such as neural networks and support vector machines), which achieve health status prediction by mining the statistical laws of battery operating data. Their advantages are that they do not require explicit mechanism modeling and can adapt to complex nonlinear relationships; however, they have the limitations of model generalization ability relying on the scale and quality of training data and insufficient interpretability of prediction results, making it difficult to guide in-depth optimization of battery aging mechanisms.

[0004] The existing Chinese patent CN119438958A introduces a lithium battery health status prediction method. This method first establishes a lithium battery health status prediction model consisting of a Transformer layer encoder, a multi-layer KAN network as a decoder, and an unscented Kalman filter. The lithium battery operation data is then input into the prediction model, and the prediction result is finally output. Another Chinese patent CN119310487A proposes a battery energy storage health status assessment method. It collects constant power charge and discharge process data of lithium batteries and obtains a health factor that is highly correlated with the battery energy storage health status through Pearson correlation analysis. Afterwards, the improved Kepler optimization algorithm is used to introduce a differential evolution strategy to optimize the bidirectional long short-term memory network model. Finally, the health factor is input into the bidirectional long short-term memory network model for processing, and an estimated value of the energy storage battery health status is output.

[0005] In the above patents, the battery health management method usually uses a single model for modeling. However, it is difficult for a single model to conduct in-depth mining of the collected data and real-time update of the model parameters. This has limited adaptability and efficiency under complex working conditions, and traditional data-driven methods often find it difficult to fully utilize the long-term dependency characteristics in time series data. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a health prediction method for aviation power battery digital twin fusion deep learning, which solves the problems mentioned in the background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a health prediction method for aviation power battery digital twin fusion deep learning, comprising the following steps:

[0008] Step S1: Build a virtual battery model of the digital twin;

[0009] Step S2: Collect battery data, process the battery data, and obtain a new real-time data matrix;

[0010] Step S3: Constructing a multi-task deep learning model for joint training; the multi-task deep learning model for joint training consists of a convolutional neural network layer, a multi-scale hierarchical bidirectional long short-term memory network layer, a multi-layer perceptron layer, and a task branching layer; inputting the new real-time data matrix into the convolutional neural network layer for processing to obtain the output of the pooling layer;

[0011] Step S4: Input the output of the pooling layer into the multi-scale hierarchical bidirectional long short-term memory network layer for processing to obtain the enhanced context vector;

[0012] Step S5: Input the enhanced context vector into the multi-layer perceptron layer for processing to obtain the output of the second fully connected layer;

[0013] Step S6: The output of the second fully connected layer is input into the task branch layer for processing to obtain the predicted value of the jointly trained multi-task deep learning model. value;

[0014] Step S7: and predict based on the multi-task deep learning model trained jointly The battery data is updated with the value to obtain a new data scalar, and the new data scalar is input into the virtual battery model of the digital twin to update the virtual battery model of the digital twin.

[0015] Furthermore, in step S1, a digital twin virtual battery model is constructed. The specific process is as follows:

[0016] Construct a dual RC equivalent circuit model as the virtual battery model of the digital twin. The virtual battery model of the digital twin consists of an ohmic resistor , two parallel circuits RC and open circuit voltage Composition: Two parallel circuits RC include a first parallel circuit RC 1 and the second parallel circuit RC 2 ; The first parallel circuit RC 1 Including the first resistor , the first capacitor ; Second parallel circuit RC 2 Including the second resistor and the second capacitor ;

[0017] The output of the digital twin virtual battery model is the battery terminal voltage , dynamic behavior, means:

[0018] ;

[0019] Where, Indicates the terminal voltage of the battery at time step t; Indicates the battery current at time step t; and They represent the first parallel circuit RC at time step t 1 and the second parallel circuit RC 2 Polarization voltage;

[0020] and The dynamic behavior of , means:

[0021] ;

[0022] ;

[0023] ;

[0024] ;

[0025] Where, Indicates the first parallel RC 1 The time constant of the loop, representing the fast polarization dynamics; Represents the second parallel RC 2 The time constant of the loop, representing the slow polarization dynamics; represents the base of natural logarithms; Represents a time variable.

[0026] Furthermore, in step S2, the new real-time data matrix is ​​generated in the following steps:

[0027] Collect battery data in real time through sensors and process them to build a real-time data matrix ,express:

[0028] ;

[0029] Where, The time step is Battery current at the moment; The time step is Battery current at the moment; Indicates the length of the time window; The time step is The temperature of the moment; The time step is The temperature of the moment; The time step is The temperature of the moment; represents the battery capacity at time step t; The time step is Battery capacity at the moment; The time step is Battery capacity at the moment;

[0030] The dynamic parameter S in the scalar of real-time battery data is recorded, which represents:

[0031] ;

[0032] Where, Indicates the depth of discharge; Indicates the initial battery capacity;

[0033] Real-time data matrix The dynamic parameter S in the scalar of the battery data is input into the virtual battery model of the digital twin for real-time status update to obtain a new real-time data matrix ,express:

[0034] ;

[0035] Where, Indicates the terminal voltage of the battery at time step t; The time step is The terminal voltage of the battery at the moment; The time step is The terminal voltage of the battery at the moment; Represents the new real-time data matrix Dimensions, A real number matrix; m represents the feature dimension.

[0036] Furthermore, the output of the pooling layer in step S3 is as follows:

[0037] New real-time data matrix Input into the input layer of the convolutional neural network layer to obtain the output of the input layer;

[0038] The output of the input layer is transferred to the convolution layer of the convolutional neural network layer, and the convolution kernel and dynamic gated segmentation activation function operations are performed to obtain the result of the convolution kernel operation. The result of the convolution kernel operation is processed to obtain the output result of the dynamic gated segmentation activation function. The result of the convolution kernel operation and the output result of the dynamic gated segmentation activation function are processed to obtain the output dimension of Output of the convolutional layer , Indicates the number of convolution kernels, which means:

[0039] ;

[0040] Where, Indicates the The output of the convolutional layer; Indicates the convolution kernel size;

[0041] The output of the convolutional layer Transfer it to the pooling layer of the convolutional neural network layer and perform pooling operation to obtain the output dimension of Output of the pooling layer ,express:

[0042] ;

[0043] ;

[0044] Where, represents the maximum pooling function; Indicates the number of time points after the pooling operation; Indicates the step length; Indicates the pooling layer window size.

[0045] Furthermore, the output result of the dynamic gated segmented activation function is as follows:

[0046] Calculate the result of the convolution kernel operation to obtain the mean and standard deviation ;

[0047] Based on the mean and standard deviation Generate dynamic thresholds, indicating:

[0048] ;

[0049] ;

[0050] Where, represents the first dynamic threshold; a learnable parameter representing a first dynamic threshold; represents the second dynamic threshold; a learnable parameter representing a second dynamic threshold;

[0051] Based on the mean and standard deviation Generate gating parameters, which means:

[0052] ;

[0053] Where, ; Indicates the Segment slopes; ; Indicates the intercepts; represents a lightweight fully connected network;

[0054] The result of the convolution kernel operation is divided into intervals according to the threshold, and a smooth transition is achieved through the activation function to obtain the output result of the dynamic gated segmented activation function, which is expressed as:

[0055] ;

[0056] ;

[0057] Where, Represents the output of the dynamic gated segmented activation function, according to the input The interval division is realized by interpolation to achieve smooth transition of different linear segments; Represents the Sigmoid activation function; Represents the result after the convolution operation; Represents the exponent part.

[0058] Furthermore, the context vector enhanced in step S4 is specifically processed as follows:

[0059] The output of the pooling layer The context vector is transmitted to the multi-scale hierarchical bidirectional long short-term memory network layer for processing to obtain the enhanced context vector; the multi-scale hierarchical bidirectional long short-term memory network layer includes a short-term layer and a long-term layer;

[0060] The output of the pooling layer Transmitted to the short-term layer, forward propagation and backward propagation are performed to obtain the short-term hidden state with a time step of t ;

[0061] The short hidden state of time step t Through the jump connection input to the long-term layer, it is calculated The long hidden state of the time step ;

[0062] Will The long hidden state of the time step Perform splicing and obtain the output dimension as The bidirectional hidden state matrix ; The number of hidden layer units representing a single direction;

[0063] For the bidirectional hidden state matrix Calculate the attention score at time step t and feature dimensions Feature attention score , combined with the spatiotemporal attention weights, to generate the enhanced context vector ,express:

[0064] ;

[0065] ;

[0066] ;

[0067] Where, express Activation function; represents the temporal attention weight matrix; represents the temporal attention bias term; Represents the feature dimension attention weight matrix; Represents the feature dimension attention bias term; Indicates that all time steps of the long-term layer are in the feature dimension The hidden state on ; Indicates that the long-term layer has a feature dimension of t at time step The hidden state value of Indicates the total number of time steps.

[0068] Furthermore, the output of the second fully connected layer in step S5 is as follows:

[0069] The enhanced context vector And the dynamic parameter S in the scalar of the battery data is spliced ​​to obtain a dimension of The concatenated column vector of ; Indicates the number of dynamic parameters, indicating:

[0070] ;

[0071] Concatenate column vectors Input into the multi-layer perceptron layer for processing to obtain the output of the second fully connected layer ;

[0072] The multi-layer perceptron layer includes a first fully connected layer and a second fully connected layer;

[0073] The first fully connected layer includes the weight matrix of the first fully connected layer and the bias term of the first fully connected layer ;

[0074] The second fully connected layer includes a weight matrix of the second fully connected layer and the bias term of the second fully connected layer ;

[0075] Concatenate column vectors and the weight matrix of the first fully connected layer Perform matrix multiplication to obtain the result of the first matrix multiplication operation, and add the result of the first matrix multiplication operation to the bias term of the first fully connected layer. Add them together and perform nonlinear transformation through the activation function ReLU to get the output of the first fully connected layer ,express:

[0076] ;

[0077] The output of the first fully connected layer and the weight matrix of the second fully connected layer Perform matrix multiplication to obtain the result of the second matrix multiplication, and convert the output of the first fully connected layer , the result of the second matrix multiplication operation and the bias term of the second fully connected layer Add them together and perform nonlinear transformation through the activation function ReLU to obtain the output of the second fully connected layer ,express:

[0078] .

[0079] Furthermore, the multi-task deep learning model jointly trained in step S6 predicts The specific process is:

[0080] The output of the second fully connected layer Input the task branch layer for processing and obtain the prediction of the jointly trained multi-task deep learning model value;

[0081] The task branch layer includes the SOH prediction branch; the SOH prediction branch includes the weight matrix of the first branch and the bias term of the first branch ;

[0082] The output of the second fully connected layer With the weight matrix of the first branch Perform matrix multiplication to obtain the result of the first matrix multiplication operation, and add the bias term of the first branch to the result of the first matrix multiplication operation. Finally, the activation function is used for nonlinear transformation to obtain the predicted value of the joint training multi-task deep learning model value ,express:

[0083] .

[0084] Furthermore, in step S7, the virtual battery model of the digital twin is updated. The specific process is as follows:

[0085] Predicted by a jointly trained multi-task deep learning model value Dynamically update the dynamic parameter S in the battery data scalar to obtain the new data scalar ,in, The first resistor after update The resistance value, The first capacitor after update The capacitance value; The second resistor after update The resistance value; The second capacitor after update The capacitance value;

[0086] Scalarize the new data Input into the virtual battery model of the digital twin and update the virtual battery model of the digital twin to indicate:

[0087] ;

[0088] ;

[0089] ;

[0090] ;

[0091] Where, represents the predictions of the jointly trained multi-task deep learning model Value for the first resistor Degraded sensitivity coefficient; represents the predictions of the jointly trained multi-task deep learning model Value for the first capacitor Degraded sensitivity coefficient; represents the predictions of the jointly trained multi-task deep learning model Value for the second resistor Degraded sensitivity coefficient; represents the predictions of the jointly trained multi-task deep learning model Value for the second capacitor Degraded sensitivity coefficient.

[0092] Compared with the existing technology, the present invention has the following beneficial effects:

[0093] (1) The present invention constructs a digital twin virtual battery model and uses sensors to collect real-time battery data; the digital twin can integrate the digital twin virtual battery model and real-time battery data to provide accurate support for the full life cycle management of battery health status.

[0094] (2) The present invention can effectively capture the long-term dependencies and dynamic change trends in the new real-time data matrix through the powerful temporal feature extraction capability of the multi-scale hierarchical bidirectional long short-term memory network layer; at the same time, combined with the advantages of the convolutional neural network layer in extracting local features, it can fully mine the detailed information and local features in the multi-dimensional battery data, and then realize the collaborative modeling of temporal global features and local features, thereby improving the expression ability and prediction accuracy of the jointly trained multi-task deep learning model for complex data.

[0095] (3) The present invention adds a multi-layer perceptron layer after the convolutional neural network layer and the multi-scale hierarchical bidirectional long short-term memory network layer. The multi-layer perceptron layer has a strong multi-modal feature fusion capability and can map the temporal features and static features to a higher-dimensional or lower-dimensional space, further optimizing the feature representation and improving the SOH branch prediction of the virtual battery model of the digital twin; the prediction of the multi-task deep learning model through joint training is Dynamically updating the dynamic parameters of the digital twin virtual battery model can effectively improve the modeling accuracy.

[0096] (4) The present invention deeply combines the virtual battery model of digital twins with the multi-task deep learning model of joint training, which can not only capture and reflect the health status of aviation power batteries in real time, but also optimize the input data quality through digital twins under complex working conditions, further improving the prediction accuracy, robustness and adaptability of the virtual battery model of digital twins to complex working conditions, and providing an efficient and reliable solution for the health management and life prediction of aviation power batteries. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0098] like Figure 1 As shown, the present invention provides a technical solution: a health prediction method for aviation power battery digital twin fusion deep learning, comprising the following steps:

[0099] Step S1: Build a virtual battery model of the digital twin;

[0100] Step S2: Collect battery data, process the battery data, and obtain a new real-time data matrix;

[0101] Step S3: Constructing a multi-task deep learning model for joint training; the multi-task deep learning model for joint training consists of a convolutional neural network layer, a multi-scale hierarchical bidirectional long short-term memory network layer, a multi-layer perceptron layer, and a task branching layer; inputting the new real-time data matrix into the convolutional neural network layer for processing to obtain the output of the pooling layer;

[0102] Step S4: Input the output of the pooling layer into the multi-scale hierarchical bidirectional long short-term memory network layer for processing to obtain the enhanced context vector;

[0103] Step S5: Input the enhanced context vector into the multi-layer perceptron layer for processing to obtain the output of the second fully connected layer;

[0104] Step S6: The output of the second fully connected layer is input into the task branch layer for processing to obtain the predicted value of the jointly trained multi-task deep learning model. value;

[0105] Step S7: and predict based on the multi-task deep learning model trained jointly The battery data is updated with the value to obtain a new data scalar, and the new data scalar is input into the virtual battery model of the digital twin to update the virtual battery model of the digital twin.

[0106] Among them, the virtual battery model of the digital twin is constructed in step S1, and the specific process is as follows:

[0107] Construct a dual RC equivalent circuit model as the virtual battery model of the digital twin. The virtual battery model of the digital twin consists of an ohmic resistor , two parallel circuits RC and open circuit voltage Composition: Two parallel circuits RC include a first parallel circuit RC 1 and the second parallel circuit RC 2 ; The first parallel circuit RC 1 Including the first resistor , the first capacitor ; Second parallel circuit RC2 Including the second resistor and the second capacitor ; Used to reflect the slow dynamic process of the battery;

[0108] The output of the digital twin virtual battery model is the battery terminal voltage , dynamic behavior, means:

[0109] ;

[0110] Where, Indicates the terminal voltage of the battery at time step t; Indicates the battery current at time step t; and They represent the first parallel circuit RC at time step t 1 and the second parallel circuit RC 2 Polarization voltage;

[0111] and The dynamic behavior of , means:

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] Where, Indicates the first parallel RC 1 The time constant of the loop, representing the fast polarization dynamics; Represents the second parallel RC 2 The time constant of the loop, representing the slow polarization dynamics; represents the base of natural logarithms; Represents a time variable.

[0117] The specific process of the new real-time data matrix in step S2 is as follows:

[0118] The sensors in the aviation battery management system collect battery data in real time for processing and build a real-time data matrix ,express:

[0119] ;

[0120] Where, The time step is Battery current at the moment; The time step is The battery current at the moment; T represents the length of the time window; Indicates the temperature at time step t; The time step is The temperature of the moment; The time step is The temperature of the moment; represents the battery capacity at time step t; The time step is Battery capacity at the moment; The time step is Battery capacity at the moment;

[0121] The dynamic parameter S in the scalar of real-time battery data is recorded, which represents:

[0122] ;

[0123] Where, Indicates the depth of discharge; Indicates the initial battery capacity;

[0124] Real-time data matrix The dynamic parameter S in the scalar of the battery data is input into the virtual battery model of the digital twin for real-time status update to obtain a new real-time data matrix ,express:

[0125] ;

[0126] Where, Indicates the terminal voltage of the battery at time step t; The time step is The terminal voltage of the battery at the moment; The time step is The terminal voltage of the battery at the moment; Represents the new real-time data matrix Dimensions, A real number matrix; m represents the feature dimension.

[0127] The output of the pooling layer in step S3 is as follows:

[0128] New real-time data matrix Input into the input layer of the convolutional neural network layer to obtain the output of the input layer;

[0129] The output of the input layer is transferred to the convolution layer of the convolutional neural network layer, and the convolution kernel and dynamic gated segmentation activation function (DGPA) operations are performed to obtain the result of the convolution kernel operation. The result of the convolution kernel operation is processed to obtain the output result of the dynamic gated segmentation activation function. The result of the convolution kernel operation and the output result of the dynamic gated segmentation activation function are processed to obtain an output dimension of Output of the convolutional layer , Indicates the number of convolution kernels, which means:

[0130] ;

[0131] Where, Indicates the The output of the convolutional layer; Indicates the convolution kernel size;

[0132] The output of the convolutional layer Transfer it to the pooling layer of the convolutional neural network layer and perform the pooling operation. The pooling operation selects the maximum pooling function, and the output dimension is Output of the pooling layer ,express:

[0133] ;

[0134] ;

[0135] Where MaxPooling1D represents the maximum pooling function; Indicates the number of time points after the pooling operation; Indicates the step length; Indicates the pooling layer window size.

[0136] Among them, the output result of the dynamic gated segmented activation function is as follows:

[0137] Calculate the result of the convolution kernel operation to obtain the mean and standard deviation ;

[0138] Based on the mean and standard deviation Generate dynamic thresholds, indicating:

[0139] ;

[0140] ;

[0141] Where, represents the first dynamic threshold; a learnable parameter representing a first dynamic threshold; represents the second dynamic threshold; a learnable parameter representing a second dynamic threshold;

[0142] Based on the mean and standard deviation Generate gating parameters, which means:

[0143] ;

[0144] Where, ; Indicates the Segment slopes; ; Indicates the intercepts; represents a lightweight fully connected network;

[0145] The result of the convolution kernel operation is divided into intervals according to the threshold, and a smooth transition is achieved through the activation function to obtain the output result of the dynamic gated segmented activation function, which is expressed as:

[0146] ;

[0147] ;

[0148] Where, Represents the output of the dynamic gated segmented activation function, according to the input The interval division is realized by interpolation to achieve smooth transition of different linear segments; Represents the Sigmoid activation function; Represents the result after the convolution operation; Represents the exponent part.

[0149] The context vector enhanced in step S4 is specifically processed as follows:

[0150] The multi-scale hierarchical bidirectional long short-term memory network layer BiLSTM includes a short-term layer and a long-term layer; the short-term layer time step is t=10; the long-term layer time step is =50;

[0151] The output of the pooling layer Transmitted to the short-term layer, forward propagation and backward propagation are performed to obtain the short-term hidden state with a time step of t ;

[0152] The short hidden state of time step t Through the jump connection input to the long-term layer, it is calculated The long hidden state of the time step ;

[0153] Will The long hidden state of the time step Perform splicing and obtain the output dimension as The bidirectional hidden state matrix ,in, The number of hidden layer units representing a single direction;

[0154] For the bidirectional hidden state matrix The calculation time step is Attention score and feature dimensions Feature attention score , combined with the spatiotemporal attention weights, to generate the enhanced context vector ,express:

[0155] ;

[0156] ;

[0157] ;

[0158] Where, express Activation function; represents the temporal attention weight matrix; represents the temporal attention bias term; Represents the feature dimension attention weight matrix; Represents the feature dimension attention bias term; Indicates that all time steps of the long-term layer are in the feature dimension The hidden state on ; Indicates that the long-term layer has a feature dimension of t at time step The hidden state value of Indicates the total number of time steps.

[0159] The output of the second fully connected layer in step S5 is as follows:

[0160] The enhanced context vector And the dynamic parameter S in the scalar of the battery data is spliced ​​to obtain a dimension of The concatenated column vector of ; Indicates the number of dynamic parameters, indicating:

[0161] ;

[0162] Concatenate column vectors Input into the multi-layer perceptron layer for processing to obtain the output of the second fully connected layer ;

[0163] The multi-layer perceptron layer includes a first fully connected layer and a second fully connected layer;

[0164] The first fully connected layer includes the weight matrix of the first fully connected layer and the bias term of the first fully connected layer ;

[0165] The second fully connected layer includes a weight matrix of the second fully connected layer and the bias term of the second fully connected layer ;

[0166] Concatenate column vectors and the weight matrix of the first fully connected layer Perform matrix multiplication to obtain the result of the first matrix multiplication operation, and add the result of the first matrix multiplication operation to the bias term of the first fully connected layer. Add them together and perform nonlinear transformation through the activation function ReLU to get the output of the first fully connected layer ,express:

[0167] ;

[0168] The output of the first fully connected layer and the weight matrix of the second fully connected layer Perform matrix multiplication to obtain the result of the second matrix multiplication, and convert the output of the first fully connected layer , the result of the second matrix multiplication operation and the bias term of the second fully connected layer Add them together and perform nonlinear transformation through the activation function ReLU to obtain the output of the second fully connected layer ,express:

[0169] .

[0170] Among them, the multi-task deep learning model jointly trained in step S6 predicts The specific process is:

[0171] The output of the second fully connected layer Input the task branch layer for processing and obtain the prediction of the jointly trained multi-task deep learning model value;

[0172] The task branch layer includes the SOH prediction branch (state of health) and the RUL prediction branch (remaining useful life); the SOH prediction branch includes the weight matrix of the first branch and the bias term of the first branch ;

[0173] The output of the second fully connected layer With the weight matrix of the first branch Perform matrix multiplication to obtain the result of the first matrix multiplication operation, and add the bias term of the first branch to the result of the first matrix multiplication operation. Finally, the activation function is used for nonlinear transformation to obtain the prediction of the joint training multi-task deep learning model value ,express:

[0174] .

[0175] Among them, the RUL prediction branch includes the weight matrix of the second branch and the bias term of the second branch ;

[0176] The output of the second fully connected layer and the weight matrix of the second branch Perform matrix multiplication to obtain the result of the second matrix multiplication operation, and add the bias term of the second branch to the result of the second matrix multiplication operation Finally, the activation function is used for nonlinear transformation to obtain the RUL value predicted by the joint training multi-task deep learning model ,express:

[0177] .

[0178] In step S7, the virtual battery model of the digital twin is updated. The specific process is as follows:

[0179] Predicted by a jointly trained multi-task deep learning model value Dynamically update the dynamic parameter S in the battery data scalar to obtain the new data scalar ,in, The first resistor after update The resistance value, The first capacitor after update The capacitance value; The second resistor after update The resistance value; The second capacitor after update The capacitance value;

[0180] Scalarize the new data Input into the virtual battery model of the digital twin, update the virtual battery model of the digital twin, and represent;

[0181] ;

[0182] ;

[0183] ;

[0184] ;

[0185] Where, represents the predictions of the jointly trained multi-task deep learning model Value for the first resistor Degraded sensitivity coefficient; represents the predictions of the jointly trained multi-task deep learning model Value for the first capacitor Degraded sensitivity coefficient; represents the predictions of the jointly trained multi-task deep learning model Value for the second resistor Degraded sensitivity coefficient; represents the predictions of the jointly trained multi-task deep learning model Value for the second capacitor Degraded sensitivity coefficient.

[0186] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A health prediction method for aviation power battery digital twin fusion deep learning, characterized by: The following steps are involved: Step S1: Build a virtual battery model of the digital twin; Step S2: Collect battery data, process the battery data, and obtain a new real-time data matrix; The new real-time data matrix in step S2 is as follows: Collect battery data in real time through sensors and process them to build a real-time data matrix ,express: ； Where, The time step is Battery current at the moment; The time step is Battery current at the moment; Indicates the length of the time window; The time step is The temperature of the moment; The time step is The temperature of the moment; The time step is The temperature of the moment; represents the battery capacity at time step t; The time step is Battery capacity at the moment; The time step is Battery capacity at the moment; The dynamic parameter S in the scalar of real-time battery data is recorded, which represents: ； Where, Indicates the depth of discharge; is the open circuit voltage; Indicates the initial battery capacity; is the ohmic resistance; is the first resistor; is the first capacitor; is the second resistor; is the second capacitor; Real-time data matrix The dynamic parameter S in the scalar of the battery data is input into the virtual battery model of the digital twin for real-time status update to obtain a new real-time data matrix ,express: ； Where, Indicates the terminal voltage of the battery at time step t; The time step is The terminal voltage of the battery at the moment; The time step is The terminal voltage of the battery at the moment; Represents the new real-time data matrix Dimensions, A real number matrix; m represents the feature latitude; Step S3: Constructing a multi-task deep learning model for joint training; the multi-task deep learning model for joint training consists of a convolutional neural network layer, a multi-scale hierarchical bidirectional long short-term memory network layer, a multi-layer perceptron layer, and a task branching layer; inputting the new real-time data matrix into the convolutional neural network layer for processing to obtain the output of the pooling layer; Step S4: Input the output of the pooling layer into the multi-scale hierarchical bidirectional long short-term memory network layer for processing to obtain the enhanced context vector; The context vector enhanced in step S4 is as follows: The output of the pooling layer is transferred to the multi-scale hierarchical bidirectional long short-term memory network layer for processing to obtain the enhanced context vector; the multi-scale hierarchical bidirectional long short-term memory network layer includes a short-term layer and a long-term layer; The output of the pooling layer is transferred to the short-term layer for forward and backward propagation to obtain the short-term hidden state with a time step of t. ; The short hidden state of time step t Through the jump connection input to the long-term layer, it is calculated The long hidden state of the time step ; Will The long hidden state of the time step Perform splicing and obtain the output dimension as The bidirectional hidden state matrix ; The number of hidden layer units representing a single direction; For the bidirectional hidden state matrix Calculate the attention score at time step t and feature dimensions Feature attention score , combined with the spatiotemporal attention weights, to generate the enhanced context vector ,express: ； ； ； Where, express Activation function; represents the temporal attention weight matrix; represents the temporal attention bias term; Represents the feature dimension attention weight matrix; Represents the feature dimension attention bias term; Indicates that all time steps of the long-term layer are in the feature dimension The hidden state on ; Indicates that the long-term layer has a feature dimension of t at time step The hidden state value of Indicates the total number of time steps; Step S5: Input the enhanced context vector into the multi-layer perceptron layer for processing to obtain the output of the second fully connected layer; The output of the second fully connected layer in step S5 is as follows: The enhanced context vector And the dynamic parameter S in the scalar of the battery data is spliced ​​to obtain a dimension of The concatenated column vector of ; Indicates the number of dynamic parameters, indicating: ； Concatenate column vectors Input into the multi-layer perceptron layer for processing to obtain the output of the second fully connected layer ; The multi-layer perceptron layer includes a first fully connected layer and a second fully connected layer; The first fully connected layer includes the weight matrix of the first fully connected layer and the bias term of the first fully connected layer ; The second fully connected layer includes a weight matrix of the second fully connected layer and the bias term of the second fully connected layer ; Concatenate column vectors and the weight matrix of the first fully connected layer Perform matrix multiplication to obtain the result of the first matrix multiplication operation, and add the result of the first matrix multiplication operation to the bias term of the first fully connected layer. Add them together and perform nonlinear transformation through the activation function ReLU to get the output of the first fully connected layer ,express: ； The output of the first fully connected layer and the weight matrix of the second fully connected layer Perform matrix multiplication to obtain the result of the second matrix multiplication, and convert the output of the first fully connected layer , the result of the second matrix multiplication operation and the bias term of the second fully connected layer Add them together and perform nonlinear transformation through the activation function ReLU to obtain the output of the second fully connected layer ,express: ； Step S6: The output of the second fully connected layer is input into the task branch layer for processing to obtain the predicted value of the jointly trained multi-task deep learning model. value; The multi-task deep learning model jointly trained in step S6 predicts The specific process is: The output of the second fully connected layer Input the task branch layer for processing and obtain the prediction of the jointly trained multi-task deep learning model value; The task branch layer includes the SOH prediction branch; The SOH prediction branch includes the weight matrix of the first branch and the bias term of the first branch ; The output of the second fully connected layer With the weight matrix of the first branch Perform matrix multiplication to obtain the result of the first matrix multiplication operation, and add the bias term of the first branch to the result of the first matrix multiplication operation. Finally, the activation function is used for nonlinear transformation to obtain the prediction of the joint training multi-task deep learning model value ,express: ； Step S7: and predict based on the multi-task deep learning model trained jointly The battery data is updated with the value to obtain a new data scalar, and the new data scalar is input into the virtual battery model of the digital twin to update the virtual battery model of the digital twin; In step S7, the virtual battery model of the digital twin is updated. The specific process is as follows: Predicted by a jointly trained multi-task deep learning model value Dynamically update the dynamic parameter S in the battery data scalar to obtain the new data scalar ,in, The first resistor after update The resistance value, The first capacitor after update The capacitance value; The second resistor after update The resistance value; The second capacitor after update The capacitance value; Scalarize the new data Input into the virtual battery model of the digital twin and update the virtual battery model of the digital twin to indicate: ； ； ； ； Where, represents the predictions of the jointly trained multi-task deep learning model Value for the first resistor Degraded sensitivity coefficient; represents the predictions of the jointly trained multi-task deep learning model Value for the first capacitor Degraded sensitivity coefficient; represents the predictions of the jointly trained multi-task deep learning model Value for the second resistor Degraded sensitivity coefficient; represents the predictions of the jointly trained multi-task deep learning model Value for the second capacitor Degraded sensitivity coefficient.

2. The health prediction method for aviation power battery digital twin fusion deep learning according to claim 1 is characterized by: In step S1, a digital twin virtual battery model is constructed. The specific process is as follows: Construct a dual RC equivalent circuit model as the virtual battery model of the digital twin. The virtual battery model of the digital twin consists of an ohmic resistor , two parallel circuits RC and open circuit voltage Composition: Two parallel circuits RC include a first parallel circuit RC 1 and the second parallel circuit RC 2 ; The first parallel circuit RC 1 Including the first resistor , the first capacitor ; Second parallel circuit RC 2 Including the second resistor and the second capacitor ; The output of the digital twin virtual battery model is the battery terminal voltage , dynamic behavior, means: ； Where, Indicates the terminal voltage of the battery at time step t; Indicates the battery current at time step t; and They represent the first parallel circuit RC at time step t 1 and the second parallel circuit RC 2 Polarization voltage; and The dynamic behavior of , means: ； ； ； ； Where, Indicates the first parallel RC 1 The time constant of the loop, representing the fast polarization dynamics; Represents the second parallel RC 2 The time constant of the loop, representing the slow polarization dynamics; represents the base of natural logarithms; Represents a time variable.

3. The health prediction method for aviation power battery digital twin fusion deep learning according to claim 2 is characterized by: The output of the pooling layer in step S3 is as follows: New real-time data matrix Input into the input layer of the convolutional neural network layer to obtain the output of the input layer; The output of the input layer is transferred to the convolution layer of the convolutional neural network layer, and the convolution kernel and dynamic gated segmentation activation function operations are performed to obtain the result of the convolution kernel operation. The result of the convolution kernel operation is processed to obtain the output result of the dynamic gated segmentation activation function. The result of the convolution kernel operation and the output result of the dynamic gated segmentation activation function are processed to obtain the output dimension of Output of the convolutional layer , Indicates the number of convolution kernels, which means: ； Where, Indicates the The output of the convolutional layer; Indicates the convolution kernel size; The output of the convolutional layer Transfer it to the pooling layer of the convolutional neural network layer and perform pooling operation to obtain the output dimension of Output of the pooling layer ,express: ； ； Where, represents the maximum pooling function; Indicates the number of time points after the pooling operation; Indicates the step length; Indicates the pooling layer window size.

4. The health prediction method for aviation power battery digital twin fusion deep learning according to claim 3 is characterized by: The output result of the dynamic gated segmented activation function is as follows: Calculate the result of the convolution kernel operation to obtain the mean and standard deviation ; Based on the mean and standard deviation Generate dynamic thresholds, indicating: ； ； Where, represents the first dynamic threshold; a learnable parameter representing a first dynamic threshold; represents the second dynamic threshold; a learnable parameter representing a second dynamic threshold; Based on the mean and standard deviation Generate gating parameters, which means: ； Where, ; Indicates the Segment slopes; ; Indicates the intercepts; represents a lightweight fully connected network; The result of the convolution kernel operation is divided into intervals according to the threshold, and a smooth transition is achieved through the activation function to obtain the output result of the dynamic gated segmented activation function, which is expressed as: ； ； Where, Represents the output of the dynamic gated segmented activation function, according to the input The interval division is realized by interpolation to achieve smooth transition of different linear segments; Represents the Sigmoid activation function; Represents the result after the convolution operation; Represents the exponent part.

Images