Lithium iron phosphate battery state of health prediction method, product, storage medium and equipment based on multi-modal double network

Abstract

The application discloses a lithium iron phosphate battery health state prediction method and product based on a multi-modal double network, a storage medium and equipment, relates to the technical field of battery detection, and solves the technical problem of poor prediction accuracy of the battery health state. The method comprises the following steps: obtaining and generating a recurrent graph and a gram angle and field graph based on a key feature sequence in a one-dimensional time sequence data tensor, and splicing to obtain a double-channel image feature; obtaining a battery health state prediction value by using a prediction model, wherein the prediction model comprises the following: a time sequence convolution network for extracting a deep feature sequence based on a one-dimensional time sequence data tensor; a residual neural network for extracting a global spatial feature vector based on the double-channel image feature; a fusion module for performing cross-attention processing on the deep feature sequence and the global spatial feature vector, explicitly injecting an absolute position encoding matrix obtained by learning into a key matrix, and obtaining a cross-modal fusion feature vector; and a first regression module for mapping the cross-modal fusion feature vector into the battery health state prediction value.

Description

Technical Field

[0001] This invention relates to the field of battery testing technology, and in particular to a method, product, storage medium, and device for predicting the health status of lithium iron phosphate batteries based on a multimodal dual-network architecture. Background Technology

[0002] Battery State of Health (SoH) is a core indicator for quantifying the degradation process of lithium iron phosphate (LiFePO4, LFP) batteries. During charging and discharging, the presence of a two-phase reaction between iron phosphate and lithium iron phosphate within the battery results in an extremely flat constant current and constant voltage plateau region. This means that during battery cycling and aging, accompanied by continuous capacity loss, the change in external terminal voltage is extremely weak and slow. However, once the hidden aging of the lithium iron phosphate battery exceeds the critical threshold, it can easily trigger a sudden drop in usable power, or even induce catastrophic accidents such as irreversible thermal runaway.

[0003] In existing battery health prediction methods, single-modal approaches utilize machine learning models to establish a mapping relationship between one-dimensional historical operating sequence data (such as voltage, current, and temperature sequences) and battery health status of lithium iron phosphate batteries. However, these approaches struggle to fully extract global topological changes, local similarities, and the characteristics of flat constant-current and constant-voltage plateau regions from the historical operating sequence data. While some approaches convert one-dimensional historical operating sequence data into two-dimensional images to enhance global spatial morphology perception, this can easily disrupt the temporal order and local evolution dynamics of the original sequence. Multimodal approaches fuse one-dimensional time-series data with two-dimensional image data, but the feature fusion method often employs end-connected layer splicing, failing to capture the deep nonlinear mapping between temporal dynamics and spatial structure. Treating all degradation processes equally can easily overlook the critical importance of key degradation nodes.

[0004] In the process of realizing this invention, the inventors discovered at least the following problems in the prior art: The inability to fully extract the degradation characteristics of lithium iron phosphate batteries leads to poor accuracy in predicting battery health status. Summary of the Invention

[0005] The purpose of this invention is to provide a method, product, storage medium, and device for predicting the state of health of lithium iron phosphate batteries based on a multimodal dual-network architecture, thereby solving the technical problem of poor accuracy in battery state of health prediction due to the inability to fully extract the degradation characteristics of lithium iron phosphate batteries in existing technologies. The various technical effects of the preferred solutions among the many technical solutions provided by this invention are detailed below.

[0006] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a method for predicting the health status of lithium iron phosphate (LFP) batteries based on a multimodal dual-network architecture. The method includes: acquiring a one-dimensional time-series data tensor of the LFP battery at a target time within a constant current-voltage plateau region, wherein the one-dimensional time-series data tensor includes multiple feature sequences; generating a recursive graph and a Gram angle and field graph based on pre-determined key feature sequences in the one-dimensional time-series data tensor; concatenating the recursive graph and the Gram angle and field graph to obtain dual-channel image features; inputting the one-dimensional time-series data tensor and the dual-channel image features into a trained prediction model to obtain a predicted value of the battery health status at the target time. The prediction model includes: a temporal convolutional network for extracting deep feature sequences based on the one-dimensional time-series data tensor; a residual neural network for extracting a global spatial feature vector based on the dual-channel image features; a fusion module for performing cross-attention processing on the deep feature sequences and the global spatial feature vector, explicitly injecting a learned absolute position encoding matrix into the key matrix to obtain a cross-modal fused feature vector; and a first regression module for mapping the cross-modal fused feature vector to a predicted value of the battery health status at the target time.

[0007] Preferably, the recursive graph and the Gram angle and field graph are stitched together in the channel dimension to obtain dual-channel image features.

[0008] Preferably, the temporal convolutional network includes: a causal dilated convolutional module, which is composed of multiple layers of causal dilated convolutional blocks stacked together, used to extract intermediate temporal feature sequences from the one-dimensional temporal data tensor; and a multilayer perceptron, used to map the intermediate temporal feature sequences to the deep feature sequences.

[0009] Preferably, the residual neural network includes: a primary feature extraction module, comprising a two-dimensional convolution, batch normalization, and activation unit connected in sequence, used to map the dual-channel image features into a primary spatial feature map; a cascaded multi-layer residual network, used to extract a high-dimensional spatial abstract feature map based on the primary spatial feature map; and a pooling compression module, which performs global average pooling on the high-dimensional spatial abstract feature map to obtain the global spatial feature vector.

[0010] Preferably, the fusion module includes: a first linear projection layer for mapping the global spatial feature vector to a query matrix; a second linear projection layer for mapping the deep feature sequence to a basic temporal feature matrix; an addition unit for adding the basic temporal feature matrix to the learned absolute position encoding matrix to obtain a key matrix; a calculation unit for using the basic temporal feature matrix as a value matrix and processing the query matrix, the key matrix, and the value matrix according to a cross-attention mechanism to obtain an attention output tensor; and a residual processing module for adding the attention output tensor and the query matrix element-wise to obtain an added feature vector, and performing layer normalization processing on the added feature vector to obtain a cross-modal fusion feature vector.

[0011] Preferably, the prediction model is trained based on an enhanced composite loss function; the enhanced composite loss function includes a final prediction loss calculated based on the battery health status prediction value output by the first regression module, and also includes at least one of monotonicity loss and difference loss; when the enhanced composite loss function includes the monotonicity loss: ;in, Indicates monotonicity loss. Indicates the training batch size. Indicates the sample index. , This indicates the output of the first regression module. Predicted battery health status for each sample. This indicates the output of the first regression module. Predicted battery health status for each sample. Indicates the relaxation factor. This represents the function that maximizes the result; when the enhanced composite loss function includes difference loss: ;in, Indicates the difference loss. Indicates the first The true value of the battery health status label for each sample. Indicates the first The true value of the battery health status label for each sample.

[0012] Preferably, the enhanced composite loss function further includes a one-dimensional processing loss and a two-dimensional processing loss; The deep feature sequence is mapped to a first auxiliary predicted value of battery health status using a second regression module, and the one-dimensional processing loss is the mean square error between the first auxiliary predicted value of battery health status and the true value of battery health status label of the sample. The global spatial feature vector is mapped to a second auxiliary predicted value of battery health status using a third regression module, and the two-dimensional processing loss is the mean square error between the second auxiliary predicted value of battery health status and the true value of battery health status label of the sample.

[0013] The present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the lithium iron phosphate battery health state prediction method based on a multimodal dual network provided by the present invention.

[0014] The present invention provides a computer-readable storage medium storing a computer program, which, when executed, implements the method for predicting the health status of lithium iron phosphate batteries based on a multimodal dual network provided by the present invention.

[0015] The present invention provides a processing device comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors are configured to execute the one or more computer programs stored in the memory, so that the one or more processors execute the lithium iron phosphate battery health status prediction method based on multimodal dual network provided by the present invention.

[0016] Implementing one of the above-described technical solutions of the present invention has the following advantages or beneficial effects: (1) This invention is based on the analysis of one-dimensional time-series data tensor of constant current voltage plateau region which is highly sensitive to the aging of LFP battery. One-dimensional time-series data tensor and dual-channel image features have time synchronization. The local dynamic time-series features at the time level are extracted by using a time-series convolutional network to obtain deep feature sequences. Simultaneously, the abstract features of macroscopic overall aging evolution and structural similarity at the high-dimensional space level are extracted by using a residual neural network to obtain global spatial feature vectors and form a dual-stream network architecture. (2) The dual-channel image features are obtained by increasing the dimensionality of the key feature sequences predetermined in the one-dimensional time series tensor. This can avoid non-key feature sequences from entering the residual neural network and causing interference. The small fluctuation patterns of the key feature sequences are captured based on the recursive graph, and the angle correlation changes of the key feature sequences are captured based on the Gram angle and field graph, thereby highlighting the changes of the key feature sequences and enhancing the recognizability of the global spatial feature vector output by the residual neural network. (3) The fusion module uses the cross-attention mechanism to fuse deep feature sequences and global spatial feature vectors. The learned absolute position encoding matrix is ​​explicitly injected into the key matrix, enabling the prediction model to accurately perceive the absolute order of each time step in the degradation process. This avoids the semantic fragmentation caused by traditional feature splicing, realizes cross-modal deep coupling, and obtains cross-modal fusion feature vectors. This invention, by constructing a one-dimensional time-series data tensor, generating dual-channel image features, a dual-stream network architecture, and a fusion module, can fully extract the degradation-related fluctuation characteristics of the battery in the constant current voltage plateau region, thereby improving the accuracy and stability of battery health status prediction. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 This is a flowchart illustrating the method for predicting the health status of lithium iron phosphate batteries based on multimodal dual networks in Embodiment 1 of the present invention. Figure 2 This is a schematic diagram of the recursion graph in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the Gram angle and field diagram in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the network structure of the prediction model in Embodiment 1 of the present invention; Figure 5 This is a scatter plot of the predicted battery health status and the actual battery health status label values ​​obtained by using a single-modal temporal convolutional network in the experimental verification. Figure 6 This is a scatter plot of the predicted battery health status obtained by using a single-modal residual neural network in the experimental verification, and the actual battery health status label value. Figure 7 This is a scatter plot of the predicted battery health status values ​​and the actual battery health status labels obtained using the prediction model of this invention in the experimental verification. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, various exemplary embodiments described below will be referenced to the accompanying drawings, which form part of the exemplary embodiments, illustrating various exemplary embodiments that may be used to implement the present invention. Unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. It should be understood that they are merely examples of processes, methods, and apparatuses consistent with some aspects of the present invention disclosed as detailed in the appended claims, and other embodiments may be used, or structural and functional modifications may be made to the embodiments listed herein without departing from the scope and spirit of the present invention.

[0019] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," etc., indicate the orientation or positional relationship based on the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the referred element must have a specific orientation, or be constructed and operated in a specific orientation. The terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. The term "multiple" means two or more. The terms "connected" and "linked" should be interpreted broadly, for example, they can be fixed connections, detachable connections, integral connections, mechanical connections, electrical connections, communication connections, direct connections, indirect connections through an intermediate medium, and can be the internal connection of two elements or the interaction relationship between two elements. The term "and / or" includes any and all combinations of one or more of the related listed items. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0020] To illustrate the technical solution described in this invention, specific embodiments are described below, showing only the parts related to the embodiments of this invention.

[0021] Example 1: like Figure 1 As shown, this invention provides a method for predicting the health status of lithium iron phosphate batteries based on a multimodal dual-network architecture, including: Step S1: Obtain a one-dimensional time-series data tensor of the LFP battery at the target time in the constant current voltage plateau region. The one-dimensional time-series data tensor includes multiple feature sequences.

[0022] LFP batteries, or lithium iron phosphate batteries, exhibit extremely significant long-term dynamic evolution characteristics during irregular charge-discharge cycles. A single sample's multidimensional features cannot describe the state transition of an LFP battery. Therefore, multidimensional feature data from the target time and multiple previous sampling times are arranged in chronological order to form a one-dimensional time-series tensor. The target time is not limited to the current sampling time or any historical sampling time. The multidimensional feature data for each sampling time includes... The characteristics include, but are not limited to, the local electrical parameters of LFP batteries in the constant current voltage plateau region, as well as global degradation indicators such as global average voltage, charging current / discharging current, and equivalent cycle count. It is a positive integer, which is 19 in one example. Therefore, a one-dimensional time series data tensor includes A feature sequence (a feature sequence represents a time series of features).

[0023] It should be noted that local electrical parameters are not limited to the voltage fluctuation amplitude of the LFP battery within a local time period, the peak height of the differential capacity at a certain sampling moment, and the local slope of the voltage curve. The global average voltage is for a fixed time length (as will be discussed later). The average voltage of the LFP battery within each time step. The equivalent cycle count refers to the number of partial cycles that the LFP battery has undergone under different operating conditions and different depths of charge and discharge at the target time, which are uniformly converted into the number of standard full cycles (the number of times the initial state of charge SoC is charged from 0% to 100% and then discharged to 0%), in order to quantify the cumulative aging equivalent of the LFP battery at the target time.

[0024] Because the initial state of charge (SoC) and charging rate vary for each charging cycle in actual operation, the number of discrete sampling points in a single cycle varies greatly, and the sampling length is inconsistent across different cycles. Therefore, this embodiment sets a fixed time length for the one-dimensional time-series data tensor. Specifically, a fixed length can be used. A sliding window, extracting lengths along the time axis. Multidimensional feature data sequences. It is a positive integer, which is 50 in one example.

[0025] Since the directly extracted multidimensional feature data sequence is heterogeneous tabular data, with differences in physical dimensions and absolute scale among its dimensions, directly inputting it into the network will inevitably lead to an extreme imbalance in the update of the weight matrix. Therefore, preferably, after obtaining a length of... After processing the multidimensional feature data sequence, the process also includes preprocessing it to obtain a one-dimensional time-series data tensor. Steps, The shape is The preprocessing method is not limited to min-max scaling. It compresses and maps the extracted multidimensional feature data sequences into a unified, strictly bounded normalized mathematical space, aligning heterogeneous table data, ensuring data quality, and eliminating differences in absolute magnitude. During the prediction model training phase, it also helps prevent the prediction model from over-relying on features with large absolute values, mitigating the drastic oscillations in the loss function during the initial stages of gradient backpropagation, and ensuring smooth convergence of the overall network optimization.

[0026] Step S2: Generate a recursive graph and a Gram angle and field graph based on the pre-determined key feature sequence in the one-dimensional time series data tensor, and stitch the recursive graph and the Gram angle and field graph to obtain dual-channel image features.

[0027] In this embodiment, it is possible to pre-select from experience One-dimensional feature is selected as the key feature from the dimensional features. Key features can also be determined based on the sample set used to train the prediction model; specifically, but not limited to, using existing principal component analysis methods to calculate... The importance of 2D features to the battery health status of samples is considered, and the features with the highest importance are selected as key features. In this embodiment, the global average voltage is chosen as the key feature. Although the terminal voltage (i.e., battery voltage) of LFP batteries in the constant current voltage plateau region is insensitive to internal capacity decay, the terminal voltage will also transiently change when the battery capacity suddenly drops, making it crucial for estimating battery health status. This embodiment selects the global average voltage as the key feature because it improves the signal-to-noise ratio, and also because temporal convolutional networks cannot extract weak degradation signals from the global average voltage sequence. Therefore, this invention upscales the global average voltage sequence (key feature sequence) to a 2D feature map to amplify the internal changes and subtle fluctuations of this feature. The key feature sequence is used... express.

[0028] One-dimensional key feature sequence The dimensionality is increased to a recursive graph and Gram angle and field graph with higher information density.

[0029] Specifically, the process of generating a recurrence plot (RP) is as follows: by calculating key feature sequences... The absolute distances at each sampling time point in the reconstructed phase space are used to generate a shape of recursive matrix This serves as a recursive graph. Its calculation formula is: ; in, Represents key feature sequences The Middle Key feature values ​​at each sampling time, Represents key feature sequences The Middle Key feature values ​​at each sampling time, This indicates taking the absolute value. Indicates the first Time and the The absolute difference of characteristics at time points. , All are time indices, and the value range is [missing information]. . Representing a recursive matrix (Recursive graph) at position The value of . Thus, the microscopic fluctuations and state recurrence patterns of the terminal voltage (a key characteristic value) in LFP batteries caused by increased polarization resistance or localized loss of active material are transformed into dense discrete points and diagonal segments in a recursive graph, specifically manifested as texture variations and diagonal line depths in the recursive graph. For example Figure 2 As shown, the recursion graph can characterize the local periodicity and recursive characteristics of the LFP cell state.

[0030] Specifically, the generation process of Gramian Angular Summation Field (GASF) includes: One-dimensional key feature sequence The key feature values ​​are normalized to the interval [-1, 1]. Using the inverse cosine function Key feature sequences The key eigenvalues ​​are strictly mapped to an angle sequence in polar coordinates, the first... The polar angle at time t is: ; Calculate the cosine of the sum of the polar angles at any two different times, thus generating a value that reflects the global temporal correlation of the sequence. The matrix, i.e., the Gram angular field matrix, is obtained. Also known as Gram Point and Field Map.

[0031] Position in the Gram angle field matrix value The calculation formula is: ; Represents the cosine function. Reflecting the The sampling time and the first sampling time Global correlation at each sampling time.

[0032] Gram angle and field plots preserve the absolute time dependence of the original signal, and each element in the Gram angle and field plots represents the global correlation of key feature sequences across different sampling times. As LFP cells age, such as... Figure 3 As shown, the small overall shift in the charge and discharge voltage curves is transformed into macroscopic changes in the distribution of specific color blocks and diagonal textures in the Gram angle and field plot, thus concretizing the hidden temporal correlation decay into an intuitive global structural evolution.

[0033] In this embodiment, preferably, the recursive graph and the Gram angle and field graph are stitched together along the channel dimension to obtain dual-channel image features. , shape and combine it with a one-dimensional time series data tensor Absolute alignment in time steps. Step S2 expands the information capacity of the originally single battery degradation time sequence process through dimensionality increase processing, realizes the intuitive manifestation of high-dimensional feature space, and provides a spatial structure carrier for the subsequent residual neural network. This allows the same physical aging process to be dually represented in one-dimensional temporal dynamics and two-dimensional spatial topology, providing prior features for the macroscopic guidance of the cross-attention mechanism of the fusion module.

[0034] Step S3: Input the one-dimensional time series data tensor and dual-channel image features into the trained prediction model to obtain the predicted value of the battery health status at the target time.

[0035] like Figure 4 As shown, the prediction model includes: Temporal convolutional networks extract deep feature sequences based on one-dimensional temporal data tensors; Residual neural networks extract global spatial feature vectors based on dual-channel image features; residual neural networks are not limited to existing deep residual networks such as ResNet18, ResNet34, or VGG networks (visual geometry group networks). The fusion module performs cross-attention processing on deep feature sequences and global spatial feature vectors, and explicitly injects the learned absolute position encoding matrix into the key matrix to obtain cross-modal fusion feature vectors. The first regression module maps the cross-modal fused feature vectors to the predicted battery health status at the target time.

[0036] In this embodiment, traditional unidirectional or mixed feature stacking often leads to a certain degree of semantic aliasing, and one-dimensional temporal signals easily mask global distribution shifts, while two-dimensional images lack sensitivity to absolute temporal order. Therefore, the prediction model utilizes a dual-stream parallel network structure, employing a parallel and complementary one-dimensional temporal data processing branch (i.e., temporal convolutional network) and a two-dimensional image transformation processing branch for simultaneous processing (i.e., residual neural network). The temporal convolutional network extracts the microscopic evolution patterns over long time scales, while the residual neural network synchronously captures the hidden macroscopic topological distortions in two-dimensional space.

[0037] In this embodiment, given the vanishing gradient problem in traditional recurrent neural networks when processing long sequences of battery degradation data, and the tendency of standard one-dimensional convolutional neural networks to destroy temporal causality due to simple stacking and expansion of receptive fields, this embodiment uses a deep temporal convolutional network (TCN) as its backbone architecture and makes logical improvements to its convolutional paradigm. For example... Figure 4 As shown, in a preferred embodiment of this example, the temporal convolutional network includes: The causal dilated convolution module, composed of multiple stacked causal dilated convolution blocks, is used to extract intermediate temporal feature sequences from one-dimensional temporal data tensors. For example... Figure 4 As shown, each causal dilated convolutional block includes a causal dilated convolution, a first batch of normalization units (batch normalization 1 shown in the diagram), and a smoothing nonlinear activation function unit connected sequentially. It also includes a first residual unit, which performs element-wise addition of the feature sequence from the input causal dilated convolution with the feature sequence from the output of the smoothing nonlinear activation function unit. The result is used as the output feature sequence of that causal dilated convolutional block. The feature sequence output by the last multi-layer causal dilated convolutional block is the intermediate temporal feature sequence. The smoothing nonlinear activation function is preferably GELU activation (Gaussian error linear unit).

[0038] A multilayer perceptron is used to map intermediate temporal feature sequences to deep feature sequences. The deep feature sequences are represented as follows: It retains the complete time dimension attributes, contains highly abstract physical features of each time step, and records the sequential evolution of LFP battery degradation features.

[0039] In this preferred embodiment, to prevent the leakage of future information to the current moment through mathematical mechanisms during the forward propagation of the network, this invention implements an asymmetric zero-padding strategy at the leading edge of the sequence's time dimension. When the data stream passes through the causal dilated convolution of each causal dilated convolution block, the one-sided padding length in the time dimension is dynamically determined by the kernel size and the dilation rate of the causal dilated convolution, calculated as follows: ; in, For the first The one-sided padding length of the causal dilated convolution of the layer causal dilated convolution block. For the first The hole expansion rate of the causal dilated convolution of the layer causal dilated convolution block is used to control the span of the convolution kernel sampling.

[0040] The above asymmetric zero-filling strategy ensures that the time step is correct. The Output feature sequence of layered causal dilated convolution blocks It only depends on the output feature sequence of the previous causal dilated convolution block. And its state at certain historical moments. The causal dilated convolutional block of the layer at time Output feature sequence for: ; in, This represents the internal index of the convolution kernel in a causal dilated convolution, with values ​​ranging from 0 to... , The kernel size for causal dilated convolution; Indicates the first The causal dilated convolution of the layer causal dilated convolution block at the current time Looking back at the position of a step in history, with The increase will be with This involves skipping samples across historical data to ensure that the current output depends only on the current and historical states of the previous causal dilated convolutional block. Mathematically, this prevents the leakage of future information to the current moment, thus guaranteeing causality. (If time index...) If zero is used, then zero-filling value is used, that is, in the input of the first... The feature sequence of the layer causal dilated convolution block is padded with actual values ​​of 0 at the front. , The first Learnable convolution weight matrix and bias vector of causal dilated convolution in layered causal dilated convolution blocks; For smoothing nonlinear activation function units; This is the batch normalization operation for the first batch of normalized units, used to eliminate internal covariate offsets to accelerate network convergence.

[0041] In order to ensure that the receptive field of the temporal convolutional network can cover all historical states within the entire sliding time window, this preferred implementation abandons the traditional pooling dimensionality reduction operation and instead sets an exponentially increasing hole expansion rate. (in The total number of layers in the causal dilated convolution block. This allows temporal convolutional networks to expand their temporal receptive field without losing the resolution of each dimension's feature sequence. Let the first layer to the second layer... The kernel size of the causal dilated convolution of the layer causal dilated convolution block is... The total receptive field of the temporal convolutional network is... The calculation formula is: .

[0042] As the number of cycles increases, the contour, local curvature, and similar structures between different aging stages of the voltage-capacity curve undergo certain spatial deformations. This embodiment utilizes a residual neural network to capture this global spatial variation that transcends local time point fluctuations. In the preferred embodiment of this first embodiment, as... Figure 4 As shown, the residual neural network includes: The primary feature extraction module, comprising sequentially connected 2D convolution, batch normalization, and activation units, is used to map dual-channel image features into a primary spatial feature map; the primary feature extraction module completes the process from raw physical pixels to a primary spatial feature map. The underlying mapping ensures that the discrete blob features representing local similarity in the recursive graph and the Gram angles and diagonal textures representing global correlations in the field graph are not lost. Primary spatial feature map The input is then processed by a subsequent cascaded multi-layer residual network.

[0043] A cascaded multi-layer residual network extracts high-dimensional spatial abstract feature maps based on primary spatial feature maps; The pooling compression module performs global average pooling on the high-dimensional spatial abstract feature map to obtain the global spatial feature vector.

[0044] In this preferred embodiment, such as Figure 4 As shown, each residual network layer includes a first convolution (convolution 1), a second batch normalization unit (batch normalization 2), a first ReLU activation function unit (ReLU1), a second convolution (convolution 2), and a third batch normalization unit (batch normalization 3), all connected in sequence. It also includes a second residual unit and a second ReLU activation function unit (ReLU2). The second residual unit performs element-wise addition of the features input to the first convolution and the output features of the third batch normalization unit, and inputs the result to the second ReLU activation function unit for further processing. The second ReLU activation function unit outputs a high-dimensional abstract feature map. ReLU activation represents a modified linear unit.

[0045] In this preferred embodiment, the forward propagation mathematical process of each residual network layer can be expressed as: ; in, , These represent the output feature map and input feature map of each residual network layer, respectively. This represents the ReLU activation function. This represents the set of learnable weight matrices for each layer of the residual network. The mapping function represents the residual master path composed of the first convolution (convolution 1 in the figure), the second batch normalization unit (batch normalization 2 in the figure), the first ReLU activation function unit (ReLU1 in the figure), the second convolution (convolution 2 in the figure), and the third batch normalization unit (batch normalization 3 in the figure), which is responsible for extracting deep topological deformations.

[0046] In the shallow residual stage of the cascaded multi-layer residual network, the main focus is on capturing edges, corners, and local fine-grained textures in the dual-channel image features. Physically, this corresponds to the near-diagonal evolution details in the Gram angle and field map, as well as the local similarity polarization spots in the recursive graph. In the deep residual stage, the network further performs deep nonlinear combinations of these local features, refining them into higher-order semantic morphologies. Its spatial receptive field focuses on the far-diagonal macrostructure representing overall temporal drift in the Gram angle and field map, and the large-area dynamic recursive patterns representing abrupt changes in battery aging in the recursive graph. After multi-level residual processing and convolutional stride downsampling, the cascaded multi-layer residual network outputs a high-dimensional spatial abstract feature map. Then, the pooling compression module performs a global average pooling operation to abstract the feature map from the two-dimensional high-dimensional space. Compression is performed in the spatial dimension, resulting in a one-dimensional global spatial feature vector after downsampling. A spatial topology map that macroscopically characterizes the overall aging of LFP batteries within the current time window.

[0047] In this embodiment, the two branches of the temporal convolutional network and the residual neural network are decoupled and operate in parallel, avoiding feature contamination that is prone to occur when processing single-modal hybrid processes, and realizing feature extraction more comprehensively.

[0048] In a preferred embodiment of this example, to realize the deep feature sequence output by the temporal convolutional network and the global spatial feature vector output by the residual neural network Efficient collaboration and information complementarity between the two modalities avoid semantic fragmentation and temporal information loss caused by undirected feature stacking. The fusion module introduces a cross-attention mechanism to utilize global spatial feature vectors. As a query benchmark, the deep feature sequence is dynamically defined. Differentiated importance weights are assigned to each sampling time point to accurately locate key moments strongly correlated with battery capacity degradation, while preserving the original sequence information of the time series. Therefore, the fusion module is configured to include: The first linear projection layer is used to transform the global spatial feature vector. Mapped to query matrix ; The second linear projection layer is used to transform the deep feature sequences Mapping to the basic time series feature matrix ; The summation unit adds the basic time series feature matrix. With the learned absolute position encoding matrix Add them together to obtain the key matrix. ; The computational unit is based on the fundamental time series feature matrix. Value matrix The query matrix is ​​processed according to the cross-attention mechanism. Key matrix Sum matrix Obtain the attention output tensor ; The residual processing module outputs the attention tensor. and query matrix Element-wise addition yields a summed feature vector, and layer normalization is applied to this summed feature vector to obtain a cross-modal fusion feature vector. .

[0049] In this preferred embodiment, the first linear projection layer is learnable during the training of the prediction model, and its specific processing is as follows: ; in, This represents the projection matrix of the first linear projection layer. The bias matrices representing the first linear projection layer are all learnable. The generated query matrix... It physically presents macroscopic information on the overall aging of LFP batteries within the current sliding time window. ,in, Let be the fusion hidden layer dimension of the cross-attention mechanism, and be a hyperparameter. To adapt to the batch processing format of multi-head attention mechanisms, Expand in sequence dimension to a sequence of length 1. .

[0050] In this preferred embodiment, the second linear projection layer is also learnable, and its specific processing procedure is as follows: ; in, , represents the projection matrix and bias matrix of the second linear projection layer, respectively, both of which are learnable. deep feature sequences The inverted matrix is ​​used to conform to the sequence dimension of multi-head attention pairs. This represents the basic time series feature matrix.

[0051] In this preferred embodiment, since the standard attention mechanism lacks the ability to perceive the absolute order of the sequence (i.e., permutation invariance) when calculating the dot product, if the basic temporal feature matrix is ​​directly used... Inputting computation will result in the loss of deep feature sequences. The temporal causal relationships within the matrix. This is to enable the macroscopic query matrix to accurately identify one-dimensional deep feature sequences. At the moment of occurrence of each micro-evolutionary feature, a learnable absolute position encoding matrix is ​​explicitly injected during the construction of the bond matrix. , Absolute position encoding matrix It is learnable. The addition unit executes: , .

[0052] Based on the basic time series feature matrix Value matrix , , .

[0053] After completing the isomorphic construction of the query matrix, key matrix, and value matrix, the computation unit performs adaptive cross-attention computation, using the query matrix... and key matrix with location information Perform matrix dot product calculations and apply a scaling factor. To prevent the gradient from entering the saturation region due to an excessively large dot product value, the calculation formula is as follows: ; in, For attention output tensors that incorporate key spatiotemporal information, Represents the normalized exponential function, Key matrix The transpose of .

[0054] Dot product operation in cross attention mechanism Used to measure the query matrix Representing macroscopic aging distortion and bond matrix This represents the nonlinear correlation between each micro-evolutionary node within a given time window. After... The attention weights, normalized by the activation function, represent the dynamic weight allocation adaptively performed by the prediction model within a complete one-dimensional sequence under the explicit guidance of two-dimensional macroscopic spatial features. A sequence node with a higher attention weight indicates a greater impact of physical fluctuations occurring at that moment (such as transient polarization abrupt changes at the end of constant current charging) on ​​the current capacity decay.

[0055] In this preferred embodiment, the attention output tensor that incorporates spatiotemporal key information is obtained. Subsequently, to prevent the loss of the original two-dimensional spatial global prior information in deep mapping, the residual processing module performs the following: converting the attention output tensor... and query matrix Element-wise addition yields a summed feature vector, and layer normalization is applied to this summed feature vector to obtain a cross-modal fusion feature vector. The internal feature distribution is stabilized through layer normalization.

[0056] In this embodiment, the first regression module is not limited to using the existing multilayer perceptron structure; it fuses feature vectors across modalities. This is mapped to the predicted battery health status value SoH at the target time.

[0057] Example 2: This embodiment discloses a method for predicting the health status of lithium iron phosphate batteries based on a multimodal dual network. The difference from Embodiment 1 is that the prediction model is trained based on an enhanced composite loss function. To better understand the principle and technical effects of the enhanced composite loss function, this embodiment will detail the training process of the prediction model of this invention.

[0058] The training process for the prediction model includes: Step 1, construct the sample set. This specifically includes: Step 1-1: Extract the feature time-series matrices of multiple LFP batteries as they evolve with the cycle period (containing core physical quantities such as operating voltage and charge / discharge current, with a time length greater than...). The data includes the actual maximum releasable capacity for each cycle. During the extraction process, reasonable boundary conditions based on the physical limits of LFP batteries are used to eliminate measurement anomalies and high-frequency spike noise caused by sensor failures or communication packet loss. Specifically, the aforementioned feature time-series matrix and its corresponding actual maximum releasable capacity can be extracted from an existing publicly available LFP battery degradation dataset. This LFP battery degradation dataset contains 77 commercial LFP battery cells, totaling 146,122 charge-discharge cycles (referred to as cycles). The fast charging protocol is the same for all cycles, but the multi-stage discharge protocols differ.

[0059] Steps 1-2 convert the actual maximum releasable capacity extracted in each cycle into a standardized battery state of health label (SoH label) for that cycle.

[0060] No. The formula for calculating the SoH tag for each cycle is: ; The index is a circular index, and it is a positive integer. For the first Each cycle corresponds to the actual maximum release capacity of the LFP battery. Indicates the first Each cycle corresponds to the nominal capacity of the LFP battery.

[0061] Steps 1-3: In the feature time series matrix of each cycle, using a fixed time length... Sliding window along the time axis Extracting the LFP cell in the constant current voltage plateau region Multidimensional feature data from each sampling time point are used to form a one-dimensional time-series data tensor. Based on the pre-determined key feature sequences in the one-dimensional time-series data tensor of each cycle, a recursive graph and Gram angle and field graph for that cycle are generated. The recursive graph and Gram angle and field graph are then stitched together in the channel dimension to obtain the dual-channel image features of that cycle.

[0062] Steps 1-4: Combine the one-dimensional time-series data tensor and dual-channel image features of each cycle into a sample, and associate the battery health status label (SoH label) of each cycle with the sample (i.e. label the sample) to complete the sample set construction.

[0063] Step 2: Construct the network architecture of the prediction model. The prediction model includes a temporal convolutional network, a residual neural network, a fusion module, and a first regression module. The specific connection relationships are detailed in Example 1. Figure 4 Initialize the network parameters of the prediction model.

[0064] Step 3: Divide the sample set into training set, test set and validation set according to the prediction ratio (e.g., 8:1:1).

[0065] Step 4: Iteratively train the prediction model built in Step 2 using the training set in batches. After each batch is trained, calculate the loss according to the augmented composite loss function, and use the Adam optimizer (adaptive moment estimator) to optimize the network parameters of the prediction model by gradient descent. Stop training when the value of the augmented composite loss function is less than the preset loss threshold, or when the number of training rounds reaches the preset maximum number of training rounds.

[0066] Step 5: Test and validate the prediction model trained in Step 4 using the test set and validation set respectively. If the test and validation pass, the prediction model trained in Step 4 is used as the final prediction model (i.e., the trained prediction model in Step S3 of Example 1). Otherwise, adjust the training parameters and return to Step 4.

[0067] In this embodiment, the composite loss function is enhanced. This includes the final predicted loss calculated based on the battery health state prediction value output by the first regression module. It also includes monotonicity loss. Sum of difference loss At least one of them. Preferably, monotonicity loss. Sum of difference loss All are weighted and compared with the final predicted loss. Integration.

[0068] Final predicted loss The calculation formula is: ; in, Indicates the training batch size. Indicates the sample index. , This indicates the output of the first regression module. Predicted battery health status for each sample. Indicates the first The true value of the battery health status label for each sample.

[0069] When enhancing the composite loss function Including monotonicity loss hour: ; in, Indicates monotonicity loss. This indicates the output of the first regression module. Predicted battery health status for each sample. Indicates the relaxation factor. This represents a function that takes the maximum value; preferably, This means permission There is a reasonable localized rebound of 0 to 2%. When enhancing the composite loss function Including differential loss hour: ; in, Indicates the difference loss. Indicates the first The true value of the battery health status label for each sample. Indicates the first The true value of the battery health status label for each sample.

[0070] In one example of this embodiment, the composite loss function is enhanced. for: ; Indicates the monotonicity loss weight. These represent the differential loss weights, which are all hyperparameters and can be set based on experience. , The values ​​range from 0 to 1.

[0071] In this embodiment, the final predicted loss The first regression module output is used to measure the... The predicted battery health status of the sample is compared with the predicted value of the first sample. The difference between the true values ​​of the battery health status labels for each sample.

[0072] In this embodiment, monotonicity loss The predicted battery health status value output by the first regression module should follow the irreversible degradation law of LFP batteries. When the predicted battery health status value of the prediction model deviates from this law and increases significantly, the monotonicity loss... This will generate a penalty gradient, causing the decision boundary to revert. Furthermore, the battery health state value of a normal LFP battery may experience a slight rebound during degradation due to practical reasons such as sensor hysteresis. To accommodate this situation, monotonicity loss... A small constant relaxation factor was introduced. This design allows the prediction model to make local fine-tuning of the battery health state prediction within a reasonable physical range. While maintaining the macroscopic decreasing trend of the battery health state SoH throughout its entire life cycle, this design gives the prediction model the ability to capture local fine physical evolution.

[0073] In this embodiment, to ensure that the prediction model can capture the dynamic rate consistency of LFP battery degradation, the present invention introduces differential loss. This loss term is used to constrain the degradation rate. It calculates the deviation between the first derivative of the predicted trajectory and the first derivative of the actual aging trajectory, allowing the prediction model to learn the transient dynamics of degradation. Compared to simple numerical fitting, difference loss... The introduction of this loss term makes the prediction model more sensitive to the slope of battery capacity decay, especially in the accelerated aging stage of lithium iron phosphate batteries at the end of the cycle. This loss term ensures that the prediction model accurately tracks the abrupt changes in the degradation rate, enabling the prediction model to learn the correct local degradation rate, thereby more closely approximating the actual aging trajectory in local intervals and improving the high fidelity of the prediction results in the time-series evolution.

[0074] In a preferred embodiment of this example, the enhanced composite loss function further includes a one-dimensional processing loss. and 2D processing loss ; The deep feature sequence is processed using the second regression module. The mapping is to the first auxiliary prediction value of battery health status, and the one-dimensional processing loss is the mean squared error between the first auxiliary prediction value of battery health status and the true value of battery health status label for the sample. ,in, This indicates that the second regression module will... Deep feature sequence of each sample The first mapped The first auxiliary prediction value for the battery health status of each sample. The second regression module is not limited to existing multilayer perceptrons, but is pre-calculated based on the deep feature sequences of multiple samples. The battery health status labels were obtained through training.

[0075] Using the third regression module to transform the global spatial feature vector Mapped to a second auxiliary prediction value of battery health status, with two-dimensional processing loss. The mean squared error between the second auxiliary predicted value of the battery health status and the true value of the battery health status label for the sample is: ,in, This indicates that the third regression module will... Global spatial feature vector of each sample The first mapped The second auxiliary prediction value for the battery health status of each sample. The third regression module is not limited to the existing multilayer perceptron, but is pre-calculated based on the global spatial feature vectors of multiple samples. The battery health status labels were obtained through training.

[0076] Preferably, the composite loss function is enhanced at this time. for: ; This indicates the one-dimensional processing loss weights. The weights represent the two-dimensional processing loss, all of which are hyperparameters and can be set empirically. The first three loss terms together form a multi-head collaborative supervision mechanism. By introducing a second and third regression module to assist training, the one-dimensional temporal convolutional network and the two-dimensional residual neural network possess preliminary independent prediction capabilities before fusion, effectively solving the gradient vanishing problem common in deep multimodal networks and accelerating the convergence efficiency of the model in the early stages. Preferably, , The values ​​range from 0 to 1.

[0077] The above-mentioned enhanced composite loss function This approach not only ensures the high fitting ability of the prediction model to battery health status label data, but also enhances the temporal consistency, monotonicity, and degradation rate accuracy of the prediction sequence through auxiliary supervision and physical constraints. Ultimately, this allows the prediction model to demonstrate superior robustness and generalization ability during the testing phase.

[0078] Experiments were conducted to validate the performance of the trained prediction model: Based on the publicly available LFP battery degradation dataset, the prediction model of this invention was compared with existing single-modal methods to demonstrate the battery health state prediction performance of this invention. The single-modal methods refer to both one-dimensional temporal convolutional network (TCN) and two-dimensional residual neural network (improved deep residual network ResNet18) methods. All models were trained and tested under the same conditions: training, validation, test set partitioning, hyperparameter configuration (including learning rate, batch size, training epochs, etc.), and hardware environment to ensure fairness in the comparison.

[0079] The quantitative experiment used mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE) as evaluation indicators. The experimental results are shown in Table 1 below: Table 1 Comparative Experimental Results The quantitative experimental results show that, compared with the single-modal temporal convolutional network (TCN) model, the prediction model of this invention reduces the root mean square error (RMSE) by approximately 53.6% and the mean absolute percentage error (MAPE) by approximately 58.5%; compared with the single-modal ResNet18 model, the RMSE is reduced by approximately 75.1% and the MAPE by approximately 74.7%. Considering all four error metrics, this invention achieves an average performance improvement of approximately 62.6% compared to the optimal single-modal baseline model. This significant quantitative improvement demonstrates that the dual-stream parallel network (i.e., the prediction model) constructed in this invention can overcome the information bottleneck of single-modality networks, and through the deep coupling of one-dimensional temporal dynamics and two-dimensional spatial topology, greatly enhances the model's comprehensive perception of the complex degradation trajectory of LFP batteries.

[0080] To further examine the predictive stability of the prediction model of this invention throughout its entire lifecycle, Figures 5 to 7 Scatter plots of the actual battery health state (True SoH) and predicted battery health state (Predicted SoH) for the single-modal ResNet18 model, the single-modal TCN model, and the prediction model of this invention are presented as results of the qualitative experiment. In the scatter distribution evaluation, the closer the data points are to the ideal diagonal, the higher the absolute prediction accuracy. Observation Figure 5 and Figure 6It can be seen that both the single-modal ResNet18 model and the single-modal TCN model exhibit significant outlier divergence in their prediction scatter plots. The single-modal ResNet18 model shows a large prediction bias in local degradation regions (especially from the middle to the end of the cycle), with the scatter plots deviating significantly from the diagonal. Although the single-modal TCN model outperforms the single-modal ResNet18, it still shows obvious discrete bands in the later stages when the predicted battery health state (SoH) is below 80%. In contrast, Figure 7 The prediction scatter points of the prediction model of this invention converge closely to the vicinity of the diagonal, the number of marginal outliers is greatly reduced, and the overall error distribution is highly concentrated.

[0081] The qualitative experimental results not only validated the accuracy advantage of the quantitative indicators, but also intuitively verified the core advantage of the cross-attention fusion mechanism proposed in this invention. Specifically, the global spatial feature vector extracted by the two-dimensional residual neural network branch serves as a macroscopic query benchmark, effectively guiding the deep feature sequence of the one-dimensional temporal convolutional network to dynamically locate key degradation nodes, thereby compensating for the inherent defects of pure temporal features in global morphological perception. Simultaneously, the enhanced monotonicity loss and differential loss constraints in the composite loss function effectively suppress non-physical prediction jitter caused by sensor noise or operating condition fluctuations, enabling the prediction model to maintain excellent robustness and accuracy even when facing complex, non-stationary degradation data.

[0082] Example 3: The present invention also provides a computer program product, which, when executed by a processor, implements the steps of a method for predicting the state of health of lithium iron phosphate batteries based on a multimodal dual-network architecture. The computer program can be stored on a physical medium (e.g., optical disc, flash drive, portable hard drive, read-only memory) or a non-transient computer-readable medium, or can be provided in a downloadable form via a network.

[0083] Example 4: Those skilled in the art will understand that all or part of the features / steps of the above-described method embodiments can be implemented by methods, data processing systems, or computer programs. These features may be implemented without hardware, entirely in software, or in a combination of hardware and software. The aforementioned computer program may be stored in one or more computer-readable storage media. When the computer program is executed (e.g., by a processor), it performs the steps of the above-described embodiments of the lithium iron phosphate battery health status prediction method based on a multimodal dual-network.

[0084] The aforementioned storage media capable of storing program code include: static disks, solid-state drives, random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), optical storage devices, magnetic storage devices, flash memory, magnetic disks or optical disks, and / or combinations of the above devices, that is, they can be implemented by any type of volatile or non-volatile storage devices or combinations thereof.

[0085] Example 5: The present invention also provides a processing device embodiment, including one or more processors and a memory; wherein the memory is used to store one or more computer programs, and the one or more processors are used to execute the one or more computer programs stored in the memory, so that the processors execute the features / steps of the above-described embodiment of the lithium iron phosphate battery health status prediction method based on multimodal dual networks.

[0086] The embodiment is merely a specific example and does not indicate that this is the only way to implement the present invention.

[0087] The above description is merely a preferred embodiment of the present invention. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the present invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.

Claims

1. A method for predicting the health status of lithium iron phosphate batteries based on a multimodal dual-network, characterized in that, The method includes: Obtain a one-dimensional time-series data tensor of the LFP battery at the target time in the constant current voltage plateau region, wherein the one-dimensional time-series data tensor includes multiple feature sequences; Based on the predetermined key feature sequence in the one-dimensional time series data tensor, a recursive graph and a Gram angle and field graph are generated, and the recursive graph and the Gram angle and field graph are stitched together to obtain dual-channel image features; Input the one-dimensional time-series data tensor and the dual-channel image features into the trained prediction model to obtain the predicted battery health status at the target time. The prediction model includes: Temporal convolutional networks extract deep feature sequences based on the one-dimensional temporal data tensor; The residual neural network extracts a global spatial feature vector based on the dual-channel image features. The fusion module performs cross-attention processing on the deep feature sequence and the global spatial feature vector, and explicitly injects the learned absolute position encoding matrix into the key matrix to obtain the cross-modal fusion feature vector. The first regression module maps the cross-modal fusion feature vector to the predicted battery health status value at the target time.

2. The method for predicting the health status of lithium iron phosphate batteries based on multimodal dual networks according to claim 1, characterized in that, By stitching the recursive graph and the Gram angle and field graph along the channel dimension, dual-channel image features are obtained.

3. The method for predicting the health status of lithium iron phosphate batteries based on multimodal dual networks according to claim 1, characterized in that, The temporal convolutional network includes: The causal dilated convolution module, composed of multiple stacked causal dilated convolution blocks, is used to extract intermediate temporal feature sequences from the one-dimensional temporal data tensor. A multilayer perceptron is used to map the intermediate temporal feature sequence to the deep feature sequence.

4. The method for predicting the health status of lithium iron phosphate batteries based on multimodal dual networks according to claim 1, characterized in that, The residual neural network includes: The primary feature extraction module includes a two-dimensional convolution, batch normalization and activation unit connected in sequence, used to map the dual-channel image features into a primary spatial feature map; A cascaded multi-layer residual network is used to extract a high-dimensional spatial abstract feature map based on the primary spatial feature map; The pooling compression module performs global average pooling on the high-dimensional spatial abstract feature map to obtain the global spatial feature vector.

5. The method for predicting the health status of lithium iron phosphate batteries based on multimodal dual networks according to claim 1, characterized in that, The fusion module includes: The first linear projection layer is used to map the global spatial feature vector into a query matrix; The second linear projection layer is used to map the deep feature sequence to the basic temporal feature matrix; The addition unit adds the basic temporal feature matrix to the learned absolute position encoding matrix to obtain the key matrix; The computing unit uses the basic temporal feature matrix as the value matrix and processes the query matrix, the key matrix, and the value matrix according to the cross-attention mechanism to obtain the attention output tensor. The residual processing module adds the attention output tensor and the query matrix element-wise to obtain a summed feature vector, and performs layer normalization on the summed feature vector to obtain a cross-modal fusion feature vector.

6. The method for predicting the health status of lithium iron phosphate batteries based on multimodal dual networks according to claim 1, characterized in that, The prediction model is trained based on an enhanced composite loss function; the enhanced composite loss function includes the final prediction loss calculated based on the battery health status prediction value output by the first regression module, and also includes at least one of monotonic loss and difference loss. When the enhanced composite loss function includes the monotonic loss: ； in, Indicates monotonicity loss. Indicates the training batch size. Indicates the sample index. , This indicates the output of the first regression module. Predicted battery health status for each sample. This indicates the output of the first regression module. Predicted battery health status for each sample. Indicates the relaxation factor. This represents the function that takes the maximum value. When the enhanced composite loss function includes difference loss: ； in, Indicates the difference loss. Indicates the first The true value of the battery health status label for each sample. Indicates the first The true value of the battery health status label for each sample.

7. The method for predicting the health status of lithium iron phosphate batteries based on a multimodal dual network according to claim 6, characterized in that, The enhanced composite loss function also includes one-dimensional processing loss and two-dimensional processing loss; The deep feature sequence is mapped to the first auxiliary prediction value of battery health status using the second regression module. The one-dimensional processing loss is the mean square error between the first auxiliary prediction value of battery health status and the true value of battery health status label of the sample. The global spatial feature vector is mapped to a second auxiliary prediction value of battery health status using the third regression module. The two-dimensional processing loss is the mean square error between the second auxiliary prediction value of battery health status and the true value of battery health status label.

8. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method for predicting the health status of lithium iron phosphate batteries based on a multimodal dual network as described in any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed, implements the method for predicting the health status of lithium iron phosphate batteries based on a multimodal dual network as described in any one of claims 1-7.

10. A processing apparatus, characterized in that, include: One or more processors; A memory for storing one or more computer programs, and one or more processors for executing the one or more computer programs stored in the memory to cause the one or more processors to perform the method for predicting the state of health of lithium iron phosphate batteries based on a multimodal dual network as described in any one of claims 1-7.

Images