A fuel cell life prediction method, device, equipment and medium

Abstract

This application discloses a method, apparatus, device, and medium for predicting fuel cell lifetime, relating to the field of battery lifetime prediction. The method includes: determining a multimodal bidirectional sequence of the target vehicle; inputting the multimodal bidirectional sequence into a feature extraction network for blind decoupling to obtain transient fluctuation features and degradation trend features; the feature extraction network is obtained by adjusting a pre-trained time-series large model base using a joint loss function that incorporates a contrastive loss function; the pre-trained time-series large model base is constructed based on a cross-modal cross-attention mechanism; obtaining trend prediction data based on degradation trend features and a degradation trend prediction model, and obtaining fluctuation prediction data based on transient fluctuation features and a transient fluctuation prediction model; reconstructing a voltage decay prediction sequence based on the trend prediction data and fluctuation prediction data. This application can improve the long-term prediction accuracy and system robustness of fuel cells under complex dynamic operating environments.

Description

Technical Field

[0001] This application relates to the field of battery life prediction, and in particular to a method, apparatus, device and medium for predicting fuel cell life. Background Technology

[0002] Proton exchange membrane fuel cells (PEMFCs) have been widely used in the field of new energy vehicles due to their zero carbon emissions, eco-friendliness, and high efficiency. However, their lifespan is susceptible to both material degradation and complex operating conditions. Accurately predicting their voltage degradation trajectory and remaining lifespan is a core challenge in the health management of fuel cell vehicles.

[0003] In practical on-board hydrogen-electric dual-source power systems, the voltage output of the fuel cell is deeply coupled with the vehicle's driving power demand, i.e., the current load. When vehicles are driving on real roads, they face complex driving cycles such as frequent starts, rapid acceleration, cruising, and deceleration, as well as conditions such as the Worldwide Harmonized Light Vehicles Test Cycle (WLTC) and the China Light-duty Vehicle Test Cycle (CLTC). This high-frequency, unsteady-state load variation operation triggers a violent hydrothermal dynamic response within the fuel cell stack, resulting in significant transient fluctuations in the output voltage. Therefore, the actual voltage sequence of an on-board fuel cell is essentially a superposition of the "slow degradation trend" representing the absolute lifespan of the battery and the "high-frequency transient fluctuations" representing the dynamic load response.

[0004] In recent years, with the development of artificial intelligence technology, large-scale time series models based on the Transformer architecture have demonstrated strong generalization ability in long-series prediction and have gradually become the mainstream of time series analysis. However, most existing large-scale time series models are general-purpose, purely data-driven architectures, and directly applying them to the prediction of vehicle fuel cell lifespan has significant engineering and theoretical shortcomings, as detailed below.

[0005] First, general-purpose large models lack awareness of the specific physical degradation mechanisms of fuel cells. The models are prone to misinterpreting transient voltage drops caused by high-current load commands as permanent physical aging within the fuel cell stack, leading to severe accumulation of multi-step prediction errors and divergence in lifetime prediction curves.

[0006] Secondly, multimodal feature fusion lacks physical semantic alignment. When faced with multi-sensor data such as voltage and current, traditional methods often use simple channel matrix stitching, which fails to achieve high-dimensional semantic alignment between "vehicle driving commands" and "pile physical state response" in the deep feature space, resulting in low feature decoupling accuracy.

[0007] In addition, a few algorithms that attempt to introduce frequency domain decomposition, such as empirical mode decomposition, are susceptible to mode aliasing and endpoint effects. Furthermore, some studies rely excessively on specific action tags provided by external battery management systems, such as purging commands, for manual segmentation, making it difficult to achieve online real-time blind decoupling of features under unknown and complex driving conditions.

[0008] Therefore, improving the long-range prediction accuracy and system robustness of fuel cells under complex dynamic operating environments has become an urgent problem to be solved. Summary of the Invention

[0009] The purpose of this application is to provide a method, apparatus, device, and medium for predicting the lifespan of a fuel cell, which can improve the long-term prediction accuracy and system robustness of fuel cells under complex dynamic operating environments.

[0010] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for predicting the lifespan of a fuel cell, including: A multimodal bidirectional sequence of the target vehicle is determined; the vehicle-mounted multimodal sequence includes a multimodal forward sequence matrix and a multimodal backward sequence matrix; the multimodal forward sequence matrix characterizes the actual continuous driving physical process of the target vehicle; The multimodal bidirectional sequence input feature extraction network is blindly decoupled to obtain transient fluctuation features and degradation trend features; the feature extraction network is obtained by adjusting the pre-trained temporal large model base with a joint loss function that incorporates a contrastive loss function; the pre-trained temporal large model base is constructed based on a cross-modal cross-attention mechanism; Based on the degradation trend characteristics and degradation trend prediction model, trend prediction data characterizing the long-term aging baseline of the target vehicle battery is obtained, and fluctuation prediction data characterizing the dynamic operating condition response is obtained based on the transient fluctuation characteristics and transient fluctuation prediction model; the degradation trend prediction model is constructed based on a sparse attention large model; the transient fluctuation prediction model is constructed based on a time-series fusion large model. The voltage degradation prediction sequence is reconstructed based on the trend prediction data and the fluctuation prediction data; the voltage degradation prediction sequence is used to characterize the actual driving load information of the target vehicle and the absolute life baseline of the target vehicle battery.

[0011] Secondly, this application provides a fuel cell life prediction device, comprising: A multimodal bidirectional sequence determination module is used to determine the multimodal bidirectional sequence of the target vehicle; the vehicle-mounted multimodal sequence includes: a multimodal forward sequence matrix and a multimodal backward sequence matrix; the multimodal forward sequence matrix represents the actual continuous driving physical process of the target vehicle; A blind decoupling module is used to blindly decouple the multimodal bidirectional sequence input feature extraction network to obtain transient fluctuation features and degradation trend features. The feature extraction network is obtained by adjusting the pre-trained temporal large model base with a joint loss function that incorporates a contrastive loss function. The pre-trained temporal large model base is constructed based on a cross-modal cross-attention mechanism. The trend and fluctuation prediction module is used to obtain trend prediction data characterizing the long-term aging baseline of the target vehicle battery based on the degradation trend characteristics and degradation trend prediction model, and to obtain fluctuation prediction data characterizing the dynamic operating condition response based on the transient fluctuation characteristics and transient fluctuation prediction model; the degradation trend prediction model is constructed based on a sparse attention large model; the transient fluctuation prediction model is constructed based on a time-series fusion large model. The voltage degradation prediction module is used to reconstruct a voltage degradation prediction sequence based on the trend prediction data and the fluctuation prediction data; the voltage degradation prediction sequence is used to characterize the actual driving load information of the target vehicle and the absolute life baseline of the target vehicle battery.

[0012] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the computer program to implement the fuel cell life prediction method described in any one of the above.

[0013] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the fuel cell lifetime prediction method described in any one of the above descriptions.

[0014] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a method, device, equipment, and medium for predicting fuel cell lifetime. It introduces a pre-trained time-series large model base and achieves cross-modal semantic alignment through a cross-modal cross-attention mechanism. At the same time, it introduces a contrastive loss function as a physical prior target to efficiently adjust the parameters of the large model. Without the intervention of external labels, this application guides the large model to automatically complete the physical blind decoupling of degradation trend and transient fluctuations, thereby significantly improving the long-term prediction accuracy and system robustness of on-board fuel cells in complex dynamic operating environments. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A schematic flowchart illustrating a fuel cell lifetime prediction method provided in this application embodiment; Figure 2 A schematic diagram of the time-reversal twin network decoupling architecture and prediction model structure provided in the embodiments of this application; Figure 3 This is a schematic diagram of the internal structure of the dual-branch prediction model provided in the embodiments of this application; Figure 4 A schematic diagram of the functional modules of a fuel cell life prediction device provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0019] To address the engineering challenges of interference under varying load conditions in vehicles and the lack of physical constraints in general time-series models, this application proposes a method, device, equipment, and medium for fuel cell lifetime prediction under vehicle operating conditions, specifically addressing blind decoupling of fuel cell degradation and large-scale model lifetime prediction. It introduces a pre-trained time-series large-scale model foundation and achieves efficient semantic alignment between "current condition commands" and "voltage state response" through a cross-modal cross-attention mechanism. Simultaneously, it innovatively constructs a contrastive loss function based on the difference in time-reversal symmetry and uses it as a physical prior objective for efficient parameter fine-tuning of the large-scale model. This application leverages the physical and thermodynamic differences between actual degradation and operating condition fluctuations to guide the large-scale model to automatically complete the physical blind decoupling of "degradation trend" and "transient fluctuations" without external label intervention, thereby significantly improving the long-term prediction accuracy and system robustness of on-board fuel cells under complex dynamic operating environments.

[0020] In one exemplary embodiment, such as Figure 1 As shown, a method for predicting the lifespan of a fuel cell is provided, including: Step 101: Determine the multimodal bidirectional sequence of the target vehicle.

[0021] The vehicle-mounted multimodal sequence includes a multimodal forward sequence matrix and a multimodal backward sequence matrix; the multimodal forward sequence matrix represents the actual continuous driving physical process of the target vehicle.

[0022] Step 102: Blindly decouple the multimodal bidirectional sequence input feature extraction network to obtain transient fluctuation features and degradation trend features; the feature extraction network is obtained by adjusting the pre-trained temporal large model base with a joint loss function that introduces a contrastive loss function; the pre-trained temporal large model base is constructed based on a cross-modal cross-attention mechanism.

[0023] Step 103: Based on the degradation trend characteristics and degradation trend prediction model, obtain trend prediction data characterizing the long-term aging baseline of the target vehicle battery; based on the transient fluctuation characteristics and transient fluctuation prediction model, obtain fluctuation prediction data characterizing the dynamic operating condition response; the degradation trend prediction model is constructed based on a sparse attention large model; the transient fluctuation prediction model is constructed based on a time-series fusion large model.

[0024] Step 104: Reconstruct the voltage degradation prediction sequence based on the trend prediction data and the fluctuation prediction data; the voltage degradation prediction sequence is used to characterize the actual driving load information of the target vehicle and the absolute lifespan baseline of the target vehicle battery.

[0025] In another exemplary embodiment of this application, step 101 specifically includes: acquiring voltage data and drive current operating condition sequence output by the target vehicle; constructing a two-dimensional multimodal vector based on the voltage data and drive current operating condition sequence; sampling the two-dimensional multimodal vector using a sliding window to obtain a multimodal forward sequence matrix; and reversing the multimodal forward sequence matrix in the time dimension using a time inversion operator to obtain a multimodal inverse sequence matrix.

[0026] In another exemplary embodiment of this application, the pre-trained temporal large model base includes: a multimodal embedding layer and a cross-modal cross-attention layer.

[0027] The multimodal embedding layer is used for: The input multimodal forward sequence matrix is ​​mapped to forward voltage and forward current features of a set dimension using a one-dimensional convolution or linear projection matrix.

[0028] The input multimodal inverse sequence matrix is ​​mapped to inverse voltage and inverse current features of a set dimension using a one-dimensional convolution or linear projection matrix.

[0029] The cross-modal attention layer is used for: A positive query matrix is ​​obtained by linearly transforming the positive current features using a first projection matrix. A positive key matrix and a positive value matrix are obtained by linearly transforming the positive voltage features using a second projection matrix and a third projection matrix, respectively. A positive multimodal alignment feature is obtained by cross-attention calculation based on the positive query matrix, the positive key matrix, and the positive value matrix. The positive multimodal alignment feature is mapped to a positive feature vector of a set length using a feedforward network. The positive feature vector is then segmented into positive transient fluctuation features and positive degradation trend features using a set segmentation dimension.

[0030] A reverse query matrix is ​​obtained by linearly transforming the reverse current features using a first projection matrix. A reverse key matrix and a reverse value matrix are obtained by linearly transforming the reverse voltage features using a second projection matrix and a third projection matrix, respectively. A reverse multimodal alignment feature is obtained by cross-attention calculation based on the reverse query matrix, the reverse key matrix, and the reverse value matrix. The reverse multimodal alignment feature is mapped to a reverse feature vector of a set length using a feedforward network. The reverse feature vector is then segmented into reverse transient fluctuation features and reverse degradation trend features using a set segmentation dimension.

[0031] The transient fluctuation characteristics include: positive transient fluctuation characteristics and negative transient fluctuation characteristics; the degradation trend characteristics include: positive degradation trend characteristics and negative degradation trend characteristics.

[0032] In another exemplary embodiment of this application, the contrast loss function is a contrast loss function based on the difference in time-reversal characteristics, and its expression is: .

[0033] in, To compare loss functions; and This is the penalty weighting coefficient; It exhibits positive transient fluctuation characteristics; It exhibits characteristics of reverse transient fluctuations; It exhibits a positive degradation trend. It is characterized by a reverse degradation trend; The feature distance boundary threshold is used to avoid feature collapse caused by over-optimization of the network. It is dynamically updated based on the variance distribution of feature vectors within the training batch. The squared Euclidean distance between the eigenvectors. Physical prior fine-tuning update: The transient fluctuations in hydrothermal energy caused by frequent vehicle load changes are local stationary oscillations, possessing statistical time symmetry. This can be addressed by minimizing... This forces large models to categorize such features as transient fluctuations; physical degradation caused by material damage is a thermodynamic dissipative process with strict temporal unidirectionality, which can be addressed by maximizing... It forces the extraction of monotonous aging components with huge differences between the forward and reverse directions.

[0034] By combining the aforementioned contrastive loss function with the original reconstruction loss of the temporal large model base as a joint fine-tuning objective, the weights of the large model base are updated through backpropagation using an efficient parameter fine-tuning technique, achieving accurate physical blind decoupling of hybrid multimodal features without external labels.

[0035] Specifically, the aforementioned contrastive loss function is used as the physical prior objective, combined with the original autoregressive prediction loss of the large model. Construct a joint fine-tuning objective and a joint loss function. In the formula Fine-tuning the weighting coefficients to balance the contrast loss.

[0036] The mechanism for efficient parameter fine-tuning involves using low-rank adaptive techniques to update the large model through backpropagation. During training, the massive original pre-trained dense weight matrix of the time-series model is frozen. Only inject and update the low-rank reduced-dimensional matrix in the side path. and the increasing dimension matrix (rank At this point, the forward propagation computation becomes In the formula The input is the hidden layer feature vector. This is the output vector for forward propagation.

[0037] This fine-tuning mechanism retains the general large model's ability to perceive time series macroscopically, while forcing the low-rank weights of the network to evolve in accordance with the physical law of "thermodynamic time unidirectionality," automatically completing the precise physical blind decoupling of degradation trends and transient fluctuations.

[0038] In another exemplary embodiment of this application, in step 103, the sparse attention model includes a sparse self-attention layer and a fully connected layer. The sparse self-attention layer is used to calculate the degradation trend features of the input using a probabilistic sparse self-attention mechanism to obtain global temporal features; the fully connected layer is used to perform linear dimensionality reduction on the global temporal features to obtain trend prediction data.

[0039] In another exemplary embodiment of this application, in step 103, the temporal fusion large model includes: a gated residual network and a temporal self-attention layer. The gated residual network is used to extract features from the input transient fluctuation features; the temporal self-attention layer is used to perform self-attention calculation on the features extracted by the gated residual network to obtain fluctuation prediction data.

[0040] In another exemplary embodiment of this application, step 104 specifically includes: adding the trend prediction data and the fluctuation prediction data element by element to reconstruct the voltage decay prediction sequence.

[0041] An implementation process of the fuel cell lifetime prediction method in the above embodiments can be summarized as follows: (1) Obtain historical output voltage data of the vehicle fuel cell power system and driving current operating condition data that synchronously characterizes the dynamic demand of vehicle load; preprocess the acquired data and construct voltage-current multimodal forward and reverse sequences based on time axis flipping; (2) Input the multimodal forward and reverse sequences into the pre-established time series large model base respectively, perform feature semantic alignment through cross-modal cross attention mechanism, and extract forward and reverse feature vectors; and divide the feature vectors into two independent feature components along the channel dimension: degradation trend features and transient fluctuation features; (3) Construct a system based on The contrast loss function of time reversal symmetry difference is used as the physical prior target to perform unsupervised fine-tuning of the time series large model base; using the thermodynamic time unidirectional law of the physical degradation process of the internal material of the fuel cell stack, and the local time symmetry law of the system's response to the vehicle load change command, the pure degradation trend feature and transient fluctuation feature are blindly decoupled from the semantically aligned multimodal features; (4) The decoupled degradation trend feature sequence and transient fluctuation feature sequence are input into the pre-established long-range trend prediction large model (sparse attention large model) and transient fluctuation prediction large model (time series fusion large model) respectively, and the corresponding baseline degradation trajectory data and dynamic operating condition response data are obtained through multi-step deduction. (5) The above two sets of prediction data are summed and reconstructed point by point, and the high-fidelity voltage degradation prediction result of the final vehicle fuel cell is output.

[0042] The following is combined with Figure 2 and Figure 3 Taking a proton exchange membrane fuel cell as an example, a more specific implementation process of the fuel cell lifetime prediction method in the above embodiments will be described in further detail.

[0043] To address the existing problems of low prediction accuracy under complex variable load conditions in vehicles, semantic misalignment of multimodal time series data, and the ease with which general time series models fall into black box fitting, the detailed steps of the fuel cell lifetime prediction method in this embodiment are as follows.

[0044] Step 1: Sliding window sampling and time inversion operation of vehicle-mounted multimodal data.

[0045] In actual operation of fuel cell vehicles, voltage response and drive current load exhibit strong dynamic coupling characteristics. This application simultaneously collects historical output voltage data of the on-board system and the drive current operating condition sequence output by the motor controller.

[0046] Data preprocessing and standardization: At preset time intervals, voltage and drive current data points of the on-board fuel cell are acquired synchronously, and noise reduction and outlier removal are performed to construct a voltage-current multimodal vector. ,in For the first Voltage response state value at time t. For the first The current command value at each moment. Specifically, firstly, a one-dimensional median filtering algorithm is used to perform high-frequency hardware noise reduction on the acquired raw data. Specifically, a fixed-length filtering window slides along the time series, and the median of the data sequence within the window is used to replace the value at the center point, thereby eliminating isolated abnormal spikes caused by high-frequency sampling from the sensor, while perfectly preserving the edges of real step abrupt signal changes caused by rapid vehicle acceleration. Subsequently, the Z-score normalization method is introduced to map voltage response data and current command data in different magnitude ranges to the same standard normal distribution range. This operation completely eliminates the dimensional differences between the two, ensuring the numerical stability and gradient convergence efficiency of subsequent large-scale model attention weight calculations. A two-dimensional multimodal vector containing normalized voltage and current information is constructed. The formula for calculating the Z-score using the normalization method is: ,in The mean of the sequence. Standard deviation, To standardize the data, This is the standardized data.

[0047] Construction of time-reversal sequences: Using the sliding window technique to truncate sequences of length... By using a time window, the multimodal vectors are arranged in chronological order to construct a multimodal positive sequence matrix that characterizes the actual continuous driving physical process of the vehicle. , Define the time inversion operator. By strictly reversing the order of the multimodal forward sequence matrix in the time dimension, a multimodal inverse sequence matrix with time inversion properties is constructed. .

[0048] Step 2: Cross-modal semantic alignment and high-dimensional space segmentation based on a large temporal model.

[0049] A pre-trained temporal large model is introduced as the feature extraction network, which internally contains a multimodal embedding layer and a cross-modal attention layer. The multimodal positive sequence matrix is ​​then used as the feature extraction network. and multimodal inverse sequence matrix The time-series large model base is input independently. After mapping the voltage and current sequences to a high-dimensional latent space in the multimodal embedding layer, a cross-modal cross-attention layer is used. The current condition features representing the vehicle's driving intention are used as the query vector, and the voltage state response features are used as the key and value for dynamic cross-attention calculation. By calculating the attention weight distribution between sequences, the high-dimensional alignment and feature fusion of "vehicle load change command semantics" and "pile state evolution semantics" in the latent variable space are achieved, and a fixed-length output is generated. 3D positive eigenvector and inverse feature vectors. A feature segmentation dimension is predefined. ( The feature vector is segmented along the channel dimension: the index is... to The dimension is defined as the positive / negative transient fluctuation characteristics. ; index as to The dimension is defined as positive / negative degradation trend characteristics. The specific implementation process is as follows.

[0050] ① High-dimensional mapping: transforming multimodal forward sequence matrices and multimodal inverse sequence matrix Each input is independently fed into a multimodal embedding layer. The two-dimensional scalar sequence is independently mapped to a latent variable dimension using independent shallow one-dimensional convolution or linear projection matrices. The high-dimensional hidden state sequence, i.e., high-dimensional voltage features With high-dimensional current characteristics .

[0051] ② Cross-modal semantic alignment: To overcome the dimensional barrier caused by simple channel splicing in traditional algorithms, this application explicitly establishes a physical causal relationship between "instruction and response" in the cross-attention layer. This is achieved using high-dimensional current features to characterize "vehicle driving load change intentions". via projection matrix Linear transformation as a query matrix High-dimensional voltage characteristics to characterize the "internal state response of the fuel cell stack" via projection matrix and After transformation, it becomes the key matrix. Sum matrix ,Will and , Dynamic cross-attention is calculated using the following formula.

[0052] .

[0053] Multimodal alignment features are calculated using the scaled dot product attention formula. .

[0054] .

[0055] The specific semantic alignment mechanism is: dot product operation. Calculate the dynamic similarity score between "specific current jump command" and "corresponding voltage transient drop" within a local time window frame by frame; The function normalizes the score to an attention weight distribution of physical causality; finally, it multiplies it by a value matrix. The voltage response features are dynamically aggregated according to this weight. This mechanism accurately captures the instantaneous physical correlation of "command-induced response" in the latent variable space, realizing deep semantic alignment of multi-source time series data in high-dimensional space.

[0056] The aligned features are mapped by the feedforward network, and the output length is... positive eigenvectors and inverse eigenvectors Set the segmentation dimension. ( ), index Truncation is defined as transient fluctuation characteristics. , index Truncation is defined as a characteristic of degradation trend. .

[0057] Step 3: Fine-tuning of the physical prior large model based on time-reversal contrastive learning.

[0058] Material damage-induced fuel cell stack degradation is an irreversible thermodynamic dissipation process. Its time-reversal sequence exhibits a non-physical phenomenon of spontaneous performance recovery, with significant differences between forward and reverse characteristics. Hydrothermal fluctuations caused by high-current load variations possess statistical time symmetry. A contrastive loss function is constructed and used as a physical prior objective. Combined with the original autoregressive prediction loss of the large model, a joint fine-tuning objective is constructed. An efficient parameter fine-tuning mechanism is used to backpropagate and update the large model. This ensures that the resulting feature extraction network retains the macroscopic perception of time series of the general large model while forcing the low-rank weights of the network to evolve in accordance with the physical law of "thermodynamic time unidirectionality." This automatically completes the precise physical blind decoupling of degradation trends and transient fluctuations. The specific expression of the loss function is not elaborated here.

[0059] Step 4: Multi-step prediction and high-fidelity reconstruction based on dual Transformer architecture.

[0060] During the online inference phase, as the sampling time window of the vehicle's Battery Management System (BMS) continuously slides, the real-time collected positive multimodal sequences are... The continuously fine-tuned time-series model is continuously input. The positive degradation trend features and transient fluctuation features extracted from each time step are concatenated frame by frame in chronological order to construct degradation trend feature sequences and transient fluctuation feature sequences with time step dimensions, respectively. These sequences are then fed into the dual-branch large model prediction module.

[0061] (1) Long-term extrapolation of baseline lifetime.

[0062] The degradation trend feature sequence is input into a sparse attention large model, utilizing a computationally complex model of O(log n). The sparse self-attention mechanism reduces the memory usage of ultra-long sequence computation, calculates the global time dependency under the operating span of fuel cells, extracts the monotonic evolution trajectory of irreversible physical aging, and deduces trend prediction data characterizing the long-term absolute aging baseline of the battery.

[0063] Specifically, global temporal features are extracted by introducing a probabilistic sparse self-attention mechanism. The calculation formula is as follows.

[0064] .

[0065] in, This is a subset of active queries filtered by Kullback-Leibler divergence evaluation. , 1 and 1 represents the query matrix, key matrix, and value matrix generated by linear mapping from the degradation trend feature sequence, respectively. This mechanism accurately eliminates redundant calculations, reducing the computational complexity of ultra-long sequences from... sudden drop This breakthrough overcomes the memory bottleneck of large models, enabling the system to directly input long-span historical characteristics of fuel cells spanning thousands of hours, accurately capturing the monotonic evolution baseline of irreversible physical aging. This represents the time step length of the input degradation trend feature sequence. Subsequently, the extracted global temporal features are linearly reduced through a fully connected mapping layer to generate trend prediction data characterizing the long-term aging baseline of the battery.

[0066] (2) Dynamic residual deduction under working conditions.

[0067] The transient fluctuation feature sequence is input into the time-series fusion large model (TFT). The gated residual network and variable selection network inside the TFT are used to dynamically focus and amplify the nonlinear voltage change response caused by sudden large current load changes by the driver, such as rapid acceleration and sudden braking. Then, the local dynamic time-series dependence is captured by the time-series self-attention layer, and the fluctuation prediction data characterizing the dynamic operating condition response is deduced.

[0068] Specifically, it utilizes a core component within the TFT—the gated residual network (GRN, defined as...) Where GLU is a gated linear unit, , The weight matrix is ​​a learnable matrix. This is the bias term. Using this structure, the model can adaptively skip or suppress irrelevant steady-state historical background noise, and adjust based on the real-time variable load condition vector. The nonlinear voltage surge response triggered by the driver's sudden high-current command is amplified exponentially. Then, the extracted high-frequency features are fed into a time-series self-attention layer to further capture the dynamic time-series dependence of local surge signals, and finally, fluctuation prediction data characterizing the dynamic operating condition response is generated.

[0069] (3) Predict physical reconstruction.

[0070] The baseline degradation prediction data output from the dual-branch system is summed element-wise with the dynamic operating condition prediction data, and the statistical distribution parameters (mean) of the corresponding historical window are applied. with standard deviation ) Perform reverse Z-score restoration operation ( The reconstructed output includes both the dynamic details of actual driving load changes and accurately reflects the final voltage decay prediction sequence of the stack absolute lifetime baseline.

[0071] The decoupled architecture and prediction model structure of the time-reversal twin network are as follows: Figure 2 As shown, the internal structure of the dual-branch prediction model is as follows: Figure 3 As shown.

[0072] Based on the same inventive concept, this application also provides a fuel cell life prediction device for implementing the fuel cell life prediction method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more fuel cell life prediction device embodiments provided below can be found in the limitations of the fuel cell life prediction method described above, and will not be repeated here.

[0073] In one exemplary embodiment, such as Figure 4 As shown, a fuel cell life prediction device is provided, comprising: The multimodal bidirectional sequence determination module 401 is used to determine the multimodal bidirectional sequence of the target vehicle; the vehicle-mounted multimodal sequence includes a multimodal forward sequence matrix and a multimodal backward sequence matrix; the multimodal forward sequence matrix represents the actual continuous driving physical process of the target vehicle.

[0074] The blind decoupling module 402 is used to blindly decouple the multimodal bidirectional sequence input feature extraction network to obtain transient fluctuation features and degradation trend features. The feature extraction network is obtained by adjusting the pre-trained temporal large model base with a joint loss function that introduces a contrastive loss function. The pre-trained temporal large model base is constructed based on a cross-modal cross-attention mechanism.

[0075] The trend and fluctuation prediction module 403 is used to obtain trend prediction data characterizing the long-term aging baseline of the target vehicle battery based on the degradation trend characteristics and degradation trend prediction model, and to obtain fluctuation prediction data characterizing the dynamic operating condition response based on the transient fluctuation characteristics and transient fluctuation prediction model; the degradation trend prediction model is constructed based on a sparse attention large model; the transient fluctuation prediction model is constructed based on a time-series fusion large model.

[0076] The voltage degradation prediction module 404 is used to reconstruct a voltage degradation prediction sequence based on the trend prediction data and the fluctuation prediction data; the voltage degradation prediction sequence is used to characterize the actual driving load information of the target vehicle and the absolute life baseline of the target vehicle battery.

[0077] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 5 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores voltage degradation prediction sequences. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a fuel cell lifetime prediction method.

[0078] Those skilled in the art will understand that Figure 5The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment to which the present application is applied. Specific computer equipment may include, for example, [the following is a list of possible additional structures]. Figure 5 The embodiments show more or fewer components, combinations of certain components, or different component arrangements. In one exemplary embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, which the processor executes to implement the steps in the above-described method embodiments.

[0079] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0080] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0081] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0082] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0083] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, etc., and are not limited to these.

[0084] This application acquires the output voltage and dynamic current operating condition data of on-board fuel cells and constructs multimodal forward and reverse sequences with time-reversal characteristics. The sequences are input into a pre-trained time-series large-scale model base. A cross-modal cross-attention mechanism is used to semantically align voltage and current, and the feature vectors are divided along the channel dimension into degradation trend features and transient fluctuation features. A contrastive loss function based on the time-reversal symmetry difference is constructed and used as a physical prior objective to efficiently fine-tune the parameters of the large-scale model. Without external labels, the degradation trend features representing unidirectional aging and the transient fluctuation features representing variable load response are blindly decoupled. The decoupled features are input into a sparse attention large-scale model to extrapolate the long-term aging baseline, and into a time-series fusion large-scale model to capture high-frequency dynamic variable load responses. Finally, the degradation results are reconstructed by summing the features point by point. This application deeply integrates physical degradation laws with the large-scale model base, eliminating the interference of complex variable load conditions in vehicles on life prediction and significantly improving online prediction accuracy and robustness.

[0085] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0086] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method of predicting the lifetime of a fuel cell, characterized by, include: Determine the multimodal bidirectional sequence of the target vehicle; The vehicle-mounted multimodal sequence includes: a multimodal forward sequence matrix and a multimodal backward sequence matrix; the multimodal forward sequence matrix characterizes the actual continuous driving physical process of the target vehicle; The multimodal bidirectional sequence input feature extraction network is blindly decoupled to obtain transient fluctuation features and degradation trend features; the feature extraction network is obtained by adjusting the pre-trained temporal large model base with a joint loss function that incorporates a contrastive loss function; the pre-trained temporal large model base is constructed based on a cross-modal cross-attention mechanism; Based on the degradation trend characteristics and degradation trend prediction model, trend prediction data characterizing the long-term aging baseline of the target vehicle battery is obtained, and fluctuation prediction data characterizing the dynamic operating condition response is obtained based on the transient fluctuation characteristics and transient fluctuation prediction model; the degradation trend prediction model is constructed based on a sparse attention large model; the transient fluctuation prediction model is constructed based on a time-series fusion large model. The voltage degradation prediction sequence is reconstructed based on the trend prediction data and the fluctuation prediction data; the voltage degradation prediction sequence is used to characterize the actual driving load information of the target vehicle and the absolute life baseline of the target vehicle battery.

2. The fuel cell life prediction method according to claim 1, characterized by, Determine the multimodal bidirectional sequence of the target vehicle, specifically including: Acquire the voltage data and drive current operating condition sequence output by the target vehicle; A two-dimensional multimodal vector is constructed based on the voltage data and the driving current operating condition sequence; A sliding window is used to sample the two-dimensional multimodal vector to obtain a multimodal positive sequence matrix; The multimodal forward sequence matrix is ​​reversed in the time dimension using a time inversion operator to obtain a multimodal reverse sequence matrix.

3. The fuel cell life prediction method according to claim 1, characterized by, The pre-trained temporal large model base includes: a multimodal embedding layer and a cross-modal cross-attention layer; The multimodal embedding layer is used for: The input multimodal forward sequence matrix is ​​mapped into forward voltage and forward current features of a set dimension using a one-dimensional convolution or linear projection matrix. The input multimodal inverse sequence matrix is ​​mapped into inverse voltage and inverse current features of a set dimension using a one-dimensional convolution or linear projection matrix. The cross-modal attention layer is used for: A positive query matrix is ​​obtained by linearly transforming the positive current features using a first projection matrix. A positive key matrix and a positive value matrix are obtained by linearly transforming the positive voltage features using a second projection matrix and a third projection matrix, respectively. A positive multimodal alignment feature is obtained by cross-attention calculation based on the positive query matrix, the positive key matrix, and the positive value matrix. A feedforward network is used to map the positive multimodal alignment feature into a positive feature vector of a set length. The positive feature vector is then segmented into positive transient fluctuation features and positive degradation trend features using a set segmentation dimension. A reverse query matrix is ​​obtained by linearly transforming the reverse current features using a first projection matrix. A reverse key matrix and a reverse value matrix are obtained by linearly transforming the reverse voltage features using a second projection matrix and a third projection matrix, respectively. A reverse multimodal alignment feature is obtained by cross-attention calculation based on the reverse query matrix, the reverse key matrix, and the reverse value matrix. The reverse multimodal alignment feature is mapped to a reverse feature vector of a set length using a feedforward network. The reverse feature vector is then segmented into reverse transient fluctuation features and reverse degradation trend features using a set segmentation dimension. The transient fluctuation characteristics include: positive transient fluctuation characteristics and negative transient fluctuation characteristics; the degradation trend characteristics include: positive degradation trend characteristics and negative degradation trend characteristics.

4. The fuel cell life prediction method according to claim 1, characterized by, The expression for the contrastive loss function is: ； in, To compare loss functions; and This is the penalty weighting coefficient; It exhibits positive transient fluctuation characteristics; It exhibits characteristics of reverse transient fluctuations; It exhibits a positive degradation trend. It is characterized by a reverse degradation trend; The feature distance boundary threshold; The squared Euclidean distance between the eigenvectors.

5. The fuel cell lifetime prediction method according to claim 1, characterized in that, The sparse attention model includes: a sparse self-attention layer and a fully connected layer; The sparse self-attention layer is used to calculate the degradation trend features of the input by introducing a probabilistic sparse self-attention mechanism to obtain global temporal features; the fully connected layer is used to perform linear dimensionality reduction on the global temporal features to obtain trend prediction data.

6. The fuel cell lifetime prediction method according to claim 1, characterized in that, The temporal fusion model includes: a gated residual network and a temporal self-attention layer; The gated residual network is used to extract features from the input transient fluctuation features; the temporal self-attention layer is used to perform self-attention calculation on the features extracted by the gated residual network to obtain fluctuation prediction data.

7. The fuel cell lifetime prediction method according to claim 1, characterized in that, Reconstructing the voltage decay prediction sequence based on the trend prediction data and the fluctuation prediction data specifically includes: The trend prediction data and the fluctuation prediction data are added element by element to reconstruct the voltage decay prediction sequence.

8. A fuel cell life prediction device, characterized in that, include: A multimodal bidirectional sequence determination module is used to determine the multimodal bidirectional sequence of the target vehicle; The vehicle-mounted multimodal sequence includes: a multimodal forward sequence matrix and a multimodal backward sequence matrix; the multimodal forward sequence matrix characterizes the actual continuous driving physical process of the target vehicle; A blind decoupling module is used to blindly decouple the multimodal bidirectional sequence input feature extraction network to obtain transient fluctuation features and degradation trend features. The feature extraction network is obtained by adjusting the pre-trained temporal large model base with a joint loss function that incorporates a contrastive loss function. The pre-trained temporal large model base is constructed based on a cross-modal cross-attention mechanism. The trend and fluctuation prediction module is used to obtain trend prediction data characterizing the long-term aging baseline of the target vehicle battery based on the degradation trend characteristics and degradation trend prediction model, and to obtain fluctuation prediction data characterizing the dynamic operating condition response based on the transient fluctuation characteristics and transient fluctuation prediction model; the degradation trend prediction model is constructed based on a sparse attention large model; the transient fluctuation prediction model is constructed based on a time-series fusion large model. The voltage degradation prediction module is used to reconstruct a voltage degradation prediction sequence based on the trend prediction data and the fluctuation prediction data; the voltage degradation prediction sequence is used to characterize the actual driving load information of the target vehicle and the absolute life baseline of the target vehicle battery.

9. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that the processor executes the computer program to implement the fuel cell lifetime prediction method according to any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the fuel cell lifetime prediction method according to any one of claims 1-7.

Images