Method for predicting the remaining useful life of a solid oxide fuel cell system on board a vehicle

Abstract

The application discloses a kind of remaining useful life prediction methods of vehicle-mounted solid oxide fuel cell system, it is applied to vehicle-mounted battery technical field, the method includes: according to the real-time data of multiple sensors of vehicle-mounted solid oxide fuel cell system, the sliding window timing characteristics and cumulative degradation characteristics of battery system are obtained;Model input tensor is obtained according to sliding window timing characteristics and cumulative degradation characteristics;The model input tensor is input into long short-term memory network, gated recurrent unit and support vector regression respectively three target models carry out remaining useful life prediction, obtain three predicted values;According to the three predicted values and the fusion weight obtained by swarm intelligence optimization algorithm, the prediction result of remaining useful life prediction is obtained.The technical scheme, by increasing cumulative degradation characteristics, multi-model intelligent optimization fusion solves the technical problem that the remaining useful life prediction accuracy of vehicle-mounted SOFC based on data driving in prior art is poor.

Description

Technical Field

[0001] This invention relates to the field of vehicle battery technology, and in particular to a method for predicting the remaining service life of vehicle-mounted solid oxide fuel cell systems. Background Technology

[0002] Solid oxide fuel cells (SOFCs) are high-temperature electrochemical energy conversion devices that directly convert the chemical energy of fuel into electrical energy. They have advantages such as high energy conversion efficiency, strong fuel adaptability, and maintenance-free all-solid-state structure. They have application value in fields such as distributed power generation, combined heat and power, auxiliary power systems, large stationary power stations, and bus power systems.

[0003] SOFCs are affected by a combination of degradation mechanisms during operation: coarsening of nickel particles at the anode leads to a reduction in the three-phase interface; strontium segregation at the cathode causes a decrease in surface catalytic activity; oxidation of the connector increases contact resistance; and microcracks in the electrolyte propagate. Furthermore, processes such as carbon deposition and sulfur poisoning also cause reversible or irreversible performance degradation. These degradations cause the battery system performance to decline monotonically over time until the voltage drops below the set failure threshold. Therefore, predicting the Remaining Useful Life (RUL) of automotive SOFC battery systems is of significant engineering value for ensuring the safe operation of automotive battery systems, optimizing maintenance strategies, and reducing total lifecycle costs.

[0004] The prediction methods for the remaining useful life (RUL) of automotive SOFCs are currently transitioning from traditional empirical models to intelligent data-driven methods. Current data-driven methods predict RUL based on sensor readings at the current moment. A battery system that has been operating for thousands of hours and a newly started battery system might have the same instantaneous sensor reading at a certain moment, resulting in identical RUL predictions. This contradicts the physical laws of SOFC degradation and leads to significant prediction errors. Therefore, existing data-driven methods for predicting the remaining useful life of automotive SOFCs suffer from poor accuracy. Summary of the Invention

[0005] This invention provides a method for predicting the remaining service life of a vehicle-mounted solid oxide fuel cell system, which addresses the technical problems in the prior art.

[0006] In a first aspect, embodiments of this application provide a method for predicting the remaining service life of an on-board solid oxide fuel cell system, including: Based on real-time data from multiple sensors of the on-board solid oxide fuel cell system, the sliding window timing characteristics and cumulative degradation characteristics of the battery system are obtained. The cumulative degradation characteristics include the cumulative electrical performance characteristics, cumulative operating event characteristics, cumulative carbon deposition index, and the cumulative temperature gradient of the burner within the sliding window from the moment the system starts up and runs. The model input tensor is obtained based on the sliding window temporal features and the cumulative degradation features; The input tensor of the model is input into three target models respectively, and the remaining service life of the battery system is predicted by the three target models respectively to obtain three predicted values. The three target models include a long short-term memory network, a gated recurrent unit, and a support vector regression. Based on the three predicted values ​​and the fusion weights, the prediction result of the remaining useful life is obtained. The fusion weights are obtained through a swarm intelligence optimization algorithm during the model training phase.

[0007] Optionally, the cumulative characteristics of the running events include: The cumulative number of start-stop cycles and the duration of high-temperature exposure, wherein the cumulative number of start-stop cycles is the cumulative number of times the battery system has started since the system started running, and the duration of high-temperature exposure is the cumulative time during which the battery system temperature is greater than a preset threshold.

[0008] Optionally, the cumulative electrical performance characteristics include: The cumulative operating time, total voltage decay, voltage degradation rate, and voltage degradation acceleration of the battery system.

[0009] Optionally, obtaining the model input tensor based on the sliding window temporal features and the cumulative degradation features includes: The cumulative degradation features are broadcast to the time dimension of the sliding window to obtain the cumulative degradation time-series features of the cumulative degradation features; The model input tensor is obtained by fusing the sliding window temporal features and the cumulative degradation temporal features.

[0010] Optionally, the prediction method further includes: During model training, the fusion weights of the three target models are obtained by searching on the validation dataset using the Grey Wolf Optimization Algorithm.

[0011] Optionally, obtaining the fusion weights of the three target models by searching the validation dataset using the Grey Wolf optimization algorithm includes: The static fusion weights of the three target models are obtained by searching the entire validation dataset using the Grey Wolf optimization algorithm; or, The entire validation dataset is divided into multiple validation dataset subsets according to current density and / or temperature range. The Gray Wolf optimization algorithm is then used to search each validation dataset subset to obtain multiple dynamic fusion weights.

[0012] Optionally, obtaining the prediction result of the remaining useful life prediction based on the three predicted values ​​and the fusion weight includes: The current current density and / or current temperature range of the battery system are determined based on the sliding window timing characteristics, and the target dynamic fusion weights corresponding to the current current density and / or current temperature range are obtained. Based on the three predicted values ​​and the target dynamic fusion weight, the prediction result of the remaining useful life is obtained.

[0013] Optionally, obtaining the prediction result of the remaining useful life prediction based on the three predicted values ​​and the fusion weight includes: The three predicted values ​​are weighted and summed according to the fusion weights to obtain the target prediction duration for the remaining useful life; The confidence interval of the target prediction duration is calculated based on the target prediction duration and the dispersion of the three prediction values. The target prediction duration and the confidence interval are obtained as the prediction result.

[0014] Optionally, after obtaining the target prediction duration and the confidence interval as the prediction result, the prediction method further includes: The health status level of the battery system is determined based on the upper and lower bound thresholds corresponding to the target prediction duration and the battery system design life. Output the prediction result and the health status level.

[0015] Secondly, embodiments of this application provide an electronic device, including a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors to implement any of the methods described in the first aspect.

[0016] The above-described one or more technical solutions in the embodiments of this application have at least the following technical effects: This application provides a method for predicting the remaining service life of a vehicle-mounted solid oxide fuel cell system. It acquires the sliding window time-series characteristics and cumulative degradation characteristics of the battery system to distinguish battery system states with similar instantaneous sensor readings but different cumulative operating times using the cumulative degradation characteristics. A model input tensor is obtained based on the sliding window time-series characteristics and cumulative degradation characteristics. This model input tensor is then input into three target models—a Long Short-Term Memory network, a gated recurrent unit, and a support vector regression network—to predict the remaining service life, achieving heterogeneous prediction across multiple models. The three predicted values ​​from the models are then fused using swarm intelligence optimization algorithms with fusion weights to obtain the prediction result for the remaining service life. Firstly, this solution adds cumulative degradation characteristics to the input data, enabling the model input to carry full historical degradation information, thus accurately distinguishing battery systems with the same instantaneous readings but different operating stages, thereby improving the accuracy of the remaining service life prediction. Secondly, in terms of model selection, this solution employs a multi-model heterogeneous approach, utilizing Long Short-Term Memory (LSTM) networks, gated recurrent units (GRUs), and Support Vector Regression (SVR) to improve the robustness and accuracy of battery system remaining life prediction from three aspects: long-term degradation trends, short-term degradation rate fluctuations, and enhanced generalization ability with small samples and sparse data. Thirdly, regarding model fusion, this solution integrates multiple model predictions through swarm intelligence optimization, avoiding manually preset thresholds and achieving a globally optimal search, further improving the robustness and accuracy of model predictions. Therefore, this solution addresses the technical problem of poor accuracy in data-driven remaining life prediction of automotive SOFCs in existing technologies through these three aspects, improving both prediction accuracy and robustness. Attached Figure Description

[0017] Figure 1 A flowchart illustrating a method for predicting the remaining useful life of an on-board solid oxide fuel cell system provided in this application embodiment; Figure 2 A functional block diagram for predicting remaining useful life provided in the embodiments of this application; Figure 3a A schematic diagram of the model structure of the LSTM model provided in the embodiments of this application; Figure 3b A schematic diagram of the model structure of the GRU model provided in the embodiments of this application; Figure 3c A schematic diagram of the model structure of the SVR model provided in the embodiments of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0018] Before introducing the embodiments of this disclosure, it should be noted that: Some embodiments of this disclosure are described as processing flows. Although the various operational steps of the flow may be numbered sequentially, the operational steps may be performed in parallel, concurrently, or simultaneously.

[0019] The term “and / or” may be used in embodiments of this disclosure, and “and / or” includes any and all combinations of one or more of the associated features listed.

[0020] It should be understood that when describing the connection or communication relationship between two components, unless it is explicitly stated that the two components are directly connected or communicate directly, the connection or communication between the two components can be understood as a direct connection or communication, or it can be understood as an indirect connection or communication through an intermediate component.

[0021] To make the technical solutions and advantages of the embodiments of this disclosure clearer, the exemplary embodiments of this disclosure will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this disclosure, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this disclosure can be combined with each other.

[0022] To address the technical problem of poor accuracy in predicting the remaining useful life of vehicle-mounted SOFCs based on data-driven methods in the prior art, this application embodiment optimizes the technology from one or more of the following aspects: Firstly, this solution adds cumulative degradation features to the input data, enabling the model input to carry full historical degradation information. This allows for accurate differentiation of battery systems with the same instantaneous readings but different operating stages, thereby improving the accuracy of remaining service life prediction.

[0023] Secondly, in terms of model selection, this scheme uses a multi-model heterogeneity, including long short-term memory networks, gated recurrent units, and support vector regression, to improve the robustness and accuracy of predictions by addressing long-term degradation trends, short-term degradation rate fluctuations, and generalization capabilities for small samples and sparse data.

[0024] Thirdly, in terms of model fusion, this solution uses swarm intelligence optimization to fuse predictions from multiple models, avoiding manually preset thresholds and achieving global optimal search, thereby further improving the robustness and accuracy of model predictions.

[0025] The specific scheme of this application will be described in detail below.

[0026] Example:

[0027] Please refer to Figure 1 This embodiment provides a method for predicting the remaining service life of an on-board solid oxide fuel cell system. The method includes: S11. Based on the real-time data from multiple sensors of the on-board solid oxide fuel cell system, obtain the sliding window timing characteristics and cumulative degradation characteristics of the battery system. The cumulative degradation characteristics include the cumulative electrical performance characteristics, cumulative operating event characteristics, cumulative carbon deposition index, and the cumulative temperature gradient of the burner within the sliding window of the battery system since the system starts running. S12. Obtain the model input tensor based on the sliding window temporal features and the cumulative degradation features; S13. Input the model input tensor into three target models respectively, and use the three target models to predict the remaining service life of the battery system to obtain three predicted values. The three target models include a long short-term memory network, a gated recurrent unit, and a support vector regression. S14. Based on the three predicted values ​​and the fusion weight, obtain the prediction result of the remaining useful life. The fusion weight is obtained through a swarm intelligence optimization algorithm during the model training phase.

[0028] For ease of explanation, the embodiments of this application divide the above method into four functional layers: a data input layer, a multi-model prediction layer, a swarm intelligence adaptive fusion layer, and an output display layer, as follows: Figure 2 As shown.

[0029] (1) Data Input Layer: Executes S11 and S12 to process the received real-time data from multiple sensors of the SOFC battery system. After preprocessing, the sliding window time-series features and cumulative degradation features are extracted respectively. The two are then concatenated according to the feature dimensions to form a unified model input tensor. This embodiment of the invention only requires conventional sensor data (voltage, current, temperature, flow rate, pressure, etc.), reducing the data acquisition requirements.

[0030] (2) Multi-model prediction layer: It contains three structurally heterogeneous parallel prediction models: Long Short-Term Memory Network (LSTM), Gated Recurrent Unit (GRU) and Support Vector Regression (SVR). S13 is executed to input the same model input tensor into each model, and each model outputs the RUL prediction value independently. Parallel prediction is performed from different modeling perspectives under the same input, and the integration diversity is achieved through structural heterogeneity.

[0031] (3) Swarm Intelligence Adaptive Fusion Layer: The fusion weights of the models are obtained using a swarm intelligence optimization algorithm. S14 is executed to obtain the prediction results of the remaining useful life based on the predicted values ​​output by the three models and the fusion weights. Unlike the heuristic weight switching strategy based on the operational stage in the prior art, the fusion weights in this embodiment are obtained by data-driven optimization and do not rely on human experience or prior rules.

[0032] (4) Output display layer (optional): Outputs prediction results, confidence intervals calculated based on the three-model dispersion, health status level and / or alarms. If no output display layer is set, the prediction results can be output by the swarm intelligence adaptive fusion layer.

[0033] S11 Feature Extraction By using two parallel data processing channels, sliding window time-series features and cumulative degradation features are obtained from two dimensions: recent dynamics and full historical degradation, respectively, to characterize the health status of SOFC battery systems.

[0034] (1) Channel 1: Timing characteristics of sliding window For real-time data from multiple sensors in the battery system, a sliding window of preset width W is used to extract the sensor reading sequence of the most recent W time steps, which serves as the sliding window timing feature. Assume the SOFC battery system is equipped with F sensors, and channel one outputs a two-dimensional timing matrix of shape (W, Fsensor). The window width W is selected based on the following principle: W = 100 when the sampling interval is 2 seconds, and W = 60 when the sampling interval is 1 minute. The original timing readings of the sensors are retained within the window, and the subsequent deep learning model autonomously extracts the timing dynamic features. This channel differs from existing Kalman filtering smoothing methods; this method does not preprocess or filter the original signal to preserve degradation details in the original signal. The sliding window timing feature is denoted as: .

[0035] Channel 2: Cumulative Degradation Characteristics The cumulative degradation characteristics include the cumulative electrical performance characteristics, cumulative operational event characteristics, cumulative carbon buildup indicators, and the cumulative temperature gradient of the burner within the sliding window, all starting from the system's initial operation time (t=0). The cumulative electrical performance characteristics include: the cumulative operating time of the battery system, total voltage decay, voltage degradation rate, and / or voltage degradation acceleration. The cumulative operational event characteristics include: the cumulative number of start-stop cycles and / or the duration of high-temperature exposure. The cumulative degradation characteristics monotonically increase over time, carrying the complete historical degradation information of the battery system from its initial state to the present moment. The parameter representations and meanings of each cumulative degradation characteristic are as follows: F1: Cumulative runtime: t current -t0, where t0 represents the start-up and operation time of the battery system, t current Indicates the current time; F2: Total voltage attenuation: V initial -V current V initial V represents the initial voltage of the battery system. current Indicates the current voltage of the battery system; F3: Voltage degradation rate: , Indicates the data sampling interval of the on-board SOFC battery system. This represents the voltage at the current moment. This represents the voltage at the previous sampling time. F4: Voltage degradation acceleration: The mean of voltage degradation acceleration within the sliding window; F5: Cumulative Start-Stop Count: The cumulative number of times the battery system has started since the system was first started; F6: High temperature exposure duration: The cumulative time during which the battery system temperature exceeds a preset threshold; F7: Cumulative carbon deposit index: The cumulative integral of the carbon deposit sensor readings from the moment the battery system starts up to the present moment; F7=Σ[max(T_VD208(t i -400°·△t / 100] T_VD208(t i ) represents the temperature reading (in °C) at the top of the burner (position VD208) at the i-th sampling time, and Δt represents the sensor sampling interval of the battery system; F8: Temperature gradient accumulation: The average value of the sliding window of the difference between the upper burner temperature (VD208 sensor reading) and the exhaust temperature (VD224 sensor reading); F8=mean{T_VD208(t i )-T_VD224(t i )},t i ∈[t-W+1,t] T_VD208(t i T_VD208(t_VD224) represents the temperature at the top of the burner (°C) at the i-th sampling time. This location is above the SOFC stack burner and reflects the thermal field of the combustion zone. T_VD224(t_VD224) represents the exhaust port temperature (°C) at the same time. This location is located in the exhaust gas passage and reflects the exhaust gas temperature after heat exchange. T_VD208-T_VD224 represents the temperature difference between the top of the burner and the exhaust port, reflecting the uniformity of heat distribution inside the battery system. The larger the temperature difference, the lower the thermal management efficiency and the greater the internal thermal stress. W represents the width of the sliding window. mean{·} represents the arithmetic mean of all temperature differences within this window.

[0036] The aforementioned cumulative degradation features differ from existing "integral domain features" in the following ways: (i) Different purposes: The cumulative degradation features provided in this embodiment are specifically designed for modeling the degradation process, changing monotonically over time and directly reflecting the degree of degradation; the integral features of existing technologies (cumulative power generation, discharge capacity, etc.) are statistical results of conventional physical quantities. (ii) Different compositions: In addition to basic quantities such as operating time, this embodiment of the invention also includes features reflecting degradation dynamics such as voltage degradation rate (F3) and voltage degradation acceleration (F4), as well as cumulative amounts of operating events such as start-stop times (F5) and high-temperature exposure (F6), and cumulative carbon deposition index (F7) reflecting the cumulative degree of carbon deposition degradation since the battery system has been in operation, and degree gradient accumulation (F8) reflecting the current stage of thermal field degradation level.

[0037] In summary: F1-F4 characterize degradation from an electrical performance perspective, F5-F6 from an operational event perspective, and F7-F8 from the chemical and thermomechanical degradation mechanisms, respectively. These multiple features together constitute a multi-dimensional degradation characterization system, enabling the model to distinguish battery system states under different degradation modes, thus achieving more accurate RUL predictions. The absence of F7 and F8 will prevent the model from recognizing the two key degradation drivers: carbon buildup and thermal stress, resulting in a significant decrease in prediction accuracy when generalizing across multiple degradation modes.

[0038] S12 Feature Fusion Sliding window temporal features extracted from channel 1 The cumulative degradation features are extracted from channel two as a two-dimensional vector. It is a one-dimensional vector. During feature fusion, accumulated degenerate features can be... Broadcasting to the time dimension W of the sliding window yields the cumulative degradation time-series features; fusing the sliding window time-series features and the cumulative degradation time-series features by concatenating them along the feature dimension yields the model input tensor. The dimension of the model input tensor is... , For the number of sensor channels, This represents the cumulative number of degradation features.

[0039] For example, battery system A (which has been running for 3000 hours) and battery system B (which has been running for 5 minutes) have the same sensor readings within a sliding window at a certain moment. = However, due to cumulative degradation characteristics The input tensors of the different models are different, resulting in differentiated RUL predictions that conform to the physical laws of degradation.

[0040] S13 Multi-model Prediction This embodiment employs three complementary prediction models working in parallel. Each model receives the same model input tensor and independently outputs RUL prediction values. Under the same full information input, predictions are made in parallel from different modeling perspectives, achieving complementarity through the heterogeneity of the model architecture.

[0041] (1) Model 1: Long Short-Term Memory Network (LSTM) LSTM captures long-range temporal dependencies through three gating mechanisms: forget gate, input gate, and output gate, making it suitable for modeling the long-term degradation trend of SOFC. Please refer to [reference needed]. Figure 3a The network structure of this model is as follows: the first LSTM layer contains 128 hidden units, the second LSTM layer contains 64 hidden units (compressing the time sequence into a fixed-length vector), the Dropout layer (dropout probability 0.3), the fully connected layer Dense regression head 32→1, and the output prediction value RUL1.

[0042] LSTM forward propagation process: .

[0043] in, The input vector at the current time step t contains the temporal features and cumulative specialization features of time step t within the sliding window; This represents the hidden state vector of the previous time step, which encodes the degradation dynamics information of the most recent W-1 time steps; It represents the cell state vector of the previous time step, characterizing the accumulated information of the degradation trend of the SOFC battery system across multiple time steps; The parameters of the LSTM / fully connected layer are learned from the training set data and encode the mapping relationship between SOFC sensor readings and degradation trends. The hidden state vector at the current time step represents the degradation state features extracted by the LSTM from the SOFC sensor input at the current moment. It integrates recent dynamic and long-term memory information and is used for subsequent RUL regression prediction. The cell state vector at the current time step represents the updated long-term memory, which integrates new information from the SOFC sensor at the current moment with historical degradation memory, characterizing the cumulative degradation trajectory features of the battery system up to the current moment.

[0044] The loss function of the LSTM model can be the mean squared error plus L2 regularization. The optimizer is Adam (learning rate 0.001), the batch size is 64, the maximum number of training epochs is 200, and the early stopping patience value is 20.

[0045] It should be noted that the above-described LSTM structure is only an example, and the embodiments of this application do not limit the specific structure of the model.

[0046] (2) Model 2: Gated Cyclic Unit (GRU) GRU combines the forget gate and input gate of LSTM into an update gate, reducing the number of parameters by about 25% compared to LSTM, achieving faster convergence, and exhibiting higher sensitivity to short-term degradation rate fluctuations. Figure 3b As shown, the network structure can be: a first GRU layer with 64 hidden units, a second GRU layer with 32 hidden units, a Dropout layer (dropout probability 0.2), a fully connected layer with a 16→1 regression head, and an output predicted value RUL2. The differences between GRU and LSTM in gating mechanisms and parameter count ensure model diversity. Its gating update formula is: .

[0047] in, This represents the input vector at the current time step t. This is an update gate used to control the update ratio of the SOFC degenerate state at adjacent sampling times; This is a reset gate used to control the degree to which SOFC degenerate state information from the previous time step is retained when calculating candidate states; The hidden state at time t represents the short-term degradation rate feature of SOFC extracted by GRU at sampling time t. This represents the hidden state at time t-1; These represent candidate hidden states; σ is the Sigmoid function, and ⊙ represents element-wise multiplication; , , The weight matrix is ​​the trainable weight matrix of the GRU network. , , This is the bias vector that the GRU network can train.

[0048] (3) Model 3: Support Vector Regression (SVR) SVR demonstrates strong generalization ability under conditions of small sample size and sparse data. The model input is a tensor resulting from the fusion of two-channel features. After global average pooling, RBF kernel function, hyperparameter search, and SVR optimization, the output predicted value RUL3 is shown below. Figure 3c .

[0049] The prediction results of the three models, RUL1, RUL2, and RUL3, provide complementary information from different modeling perspectives: LSTM is good at capturing long-term degradation trends, GRU is sensitive to short-term degradation rate fluctuations, and SVR provides stable estimates under small sample conditions. The integration of diversity is achieved through the heterogeneity of the models themselves rather than the splitting of the input data.

[0050] S14 Multi-model Fusion The three predicted values ​​obtained from the three models are fused using fusion weights calculated by a swarm intelligence optimization algorithm to obtain the final predicted value RUL. final , will RUL final The prediction result is output. Among them, the swarm intelligence optimization algorithm includes: Artificial Bee Colony Algorithm (ABC), Grey Wolf Optimization (GWO), Whale Optimization (WOA), and Sparrow Search (SSA), etc. Any of the optimization algorithms can be used in the embodiments of this application.

[0051] In the process of implementing this application, the inventors of this application selected the Grey Wolf Optimization (GWO) algorithm, which simulates the gray wolf hierarchy and hunting behavior, based on the progressive characteristics of the three models in the degradation relationship acquisition. The fusion weights of the three target models were obtained by searching on the validation dataset, without the need for manually preset switching thresholds or stage division rules.

[0052] The following section elaborates on the specific application mechanism of the GWO algorithm in determining the fusion weights.

[0053] The Gray Wolf Optimization Algorithm simulates the social hierarchy of a gray wolf pack, dividing candidate solutions in the solution space into four levels: α, β, δ, and ω. α, β, and δ represent the current optimal, second-best, and third-best solutions, respectively, corresponding to the search directions of the three model weights. During iteration, other wolf individuals update their positions based on the positions of these three leaders, thus achieving efficient optimization of the optimal weight combination. This ensures that the final fusion result fully combines the advantages of each model, improving the accuracy and robustness of RUL prediction.

[0054] Specifically, during GWO encoding, each candidate solution (i.e., the gray wolf position) is represented as a three-dimensional continuous vector x=(x1,x2,x3). The original vector x is mapped to a weight combination w satisfying Σw_i=1 through the Softmax function. i For i=1,2,3, to ensure that all three models contribute substantially to the ensemble result (avoiding a single model's weight approaching 1 while other models' weights approach 0), a minimum contribution constraint w is applied. i ≥0.1. If a weight is below this threshold, it is truncated to 0.1 and renormalized.

[0055] The goal of GWO search is to maximize the following composite fitness function: .

[0056] in, The mean absolute prediction error on the validation set. The standard deviation of the predicted values ​​on the validation set (used to penalize unstable weight combinations of predictions) is used, with a balance coefficient λ = 0.1. This represents the predicted remaining lifetime of the j-th base model (such as GRU, LSTM, or SVR) for the i-th validation set sample. It represents the true remaining lifetime (label truth value) of the i-th validation set sample.

[0057] GWO iterative optimization includes the following six steps: Step 1 (Initialization): Set the gray wolf population size N wolves =30, maximum number of iterations T max =100. Randomly initialize the position vector x of each wolf. k ∈[-2,2], k=1,2,...,30.

[0058] Step 2 (Fitness Evaluation): After mapping the position of each wolf to weights {w1, w2, w3} using Softmax, calculate the ensemble prediction and fitness on the validation set.

[0059] Step 3 (Social Hierarchy Determination): Sort the wolves from highest to lowest fitness, select the three best wolves and label them α wolf, β wolf, and δ wolf, respectively, and denote their positions as x. α x β x δ .

[0060] Step 4 (Position Update): The remaining ω wolves update their own positions based on the positions of α, β, and δ, simulating the process of the wolf pack surrounding the prey: .

[0061] Where x=(x1,x2,x3) represents the gray wolf position vector, which corresponds to the contribution ratio of the three models LSTM, GRU and SVR in the RUL fusion prediction respectively; , , , , , This is the coefficient vector.

[0062] Step 5 (Constrained Projection): Map the updated position back to the weight space and apply w. i ≥0.1 constraint.

[0063] Step Six (Convergence Criterion): Reaching the maximum number of iterations T max Or, the fitness change is less than the threshold for several consecutive rounds. =10 -6 If the iteration stops, output the optimal weights (w1*, w2*, w3*) corresponding to wolf α.

[0064] In specific implementation, this embodiment uses two modes to obtain the fusion weights through the GWO optimization algorithm: Mode 1: During the offline training phase, the static fusion weights (w1*, w2*, w3*) of the three target models are obtained by searching the entire validation dataset using the Grey Wolf Optimization Algorithm. These weights remain unchanged during the online inference phase. This mode is suitable for applications with relatively stable operating conditions.

[0065] Mode 2: The entire validation dataset is divided into multiple validation data subsets according to the operating conditions of the SOFC battery system, such as current density and / or temperature range. The Grey Wolf optimization algorithm is used to search each validation data subset to obtain multiple dynamic fusion weights (w1*, w2*, w3*), thus obtaining the optimal weights for each range. During the online inference phase, the weight combination is switched according to the current operating condition category, suitable for application scenarios with frequently changing operating conditions.

[0066] During the prediction phase, the fusion weights are obtained, and RUL1, RUL2, and RUL3 are calculated independently for each of the three models: LSTM, GRU, and SVR. S14 then performs a weighted fusion. .

[0067] The target prediction duration is set for the remaining lifespan of the battery system. In the application scenario corresponding to Mode 1, the fusion weights are static weights (w1*, w2*, w3*). In the application scenario corresponding to Mode 2, the current current density and / or current temperature range of the battery system are determined based on the sliding window timing characteristics, and the target dynamic fusion weights corresponding to the current current density and / or current temperature range are obtained. The target prediction duration for the remaining lifespan is obtained by weighted summation of the three predicted values ​​and the target dynamic fusion weights.

[0068] In the specific implementation process, in addition to the target prediction duration, the prediction results can be further analyzed by calculating the confidence interval of the target prediction duration based on the target prediction duration and the dispersion of the three predicted values; the target prediction duration and the confidence interval are then used as the prediction results. Furthermore, based on the target prediction duration and the upper and lower bound thresholds corresponding to the battery system's design life, the health status level of the battery system can be determined; the prediction results and the health status level are then output.

[0069] Among them, the confidence interval calculation: the 95% confidence interval CI can be directly calculated based on the dispersion of the predicted values ​​of the three models, without the need for additional bootstrap sampling or Bayesian inference. .

[0070] The target prediction duration for the remaining lifespan of the battery system. This is the predicted value output by the j-th model out of the three models.

[0071] Health Status Grading: The battery system health status is divided into three levels based on the RUL predicted value: normal: > ; Notice: ; Warning: .

[0072] in, This is the upper limit threshold corresponding to the design life of the battery system. This represents the lower bound threshold corresponding to the design life of the battery system. For a battery system with a design life of 5000 hours, we can take... =4150h、 =2500h.

[0073] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific examples. The following embodiments are based on semi-realistic full-life degradation data of a bus-mounted SOFC battery system. The initial operating parameters (voltage, current, power, and temperature of each channel) are derived from 470 hours of measured data, and the degradation trajectory is generated by the Hill physical degradation model, forming a complete lifetime dataset covering five modes from fast degradation to slow degradation.

[0074] Example 1: Health diagnosis and life prediction of on-board SOFC battery system using static weighting model (I) Experimental Platform and Data Sources This embodiment takes the SOFC battery system on a bus as the research object. The sensor configuration includes 15 operating sensors, including: 1 DC output voltage VD (range 0~60V, accuracy ±0.5%), 1 DC output current ID (range 0~20A, accuracy ±0.5%), 13 temperature sensors (VD208 upper burner, VD212 lower burner, VD216 reformer, VD224 exhaust port, VD232 heat exchanger inlet, VD236 heat exchanger outlet, VD240 fuel cell stack inlet, VD244 fuel cell stack outlet, VD248 air preheater, VD256 fuel cell stack center, VD272 fuel inlet, VD276 fuel preheater, VD280 air inlet, VD284 air preheater, VD288 exhaust gas, VD316 ambient, VD320 cooling water, 13 of which are effective from a total of 17 temperature channels), and 1 DC bus power sensor (calculated by the product of voltage and current). The sampling interval is 5 minutes, and the data acquisition continuously covers the entire lifespan of the battery system from initial operation to failure.

[0075] The semi-real data used in this embodiment was generated based on the following method: initial operating parameters for the stable operation period were extracted from 470 hours of measured data, namely V0=46.32V (initial DC voltage), I0=5.37A (initial DC current), P0=255W (initial DC power), as well as the initial temperature T0 and degradation amplitude ΔT of 17 temperature channels; the degradation trajectory adopted the Hill degradation model. The failure criteria for the battery system are a degradation factor d ≥ 0.95 (corresponding to a voltage drop to approximately 85% V0 and a power drop to approximately 70% P0).

[0076] To verify the robustness of the method of the present invention under different degradation rates, semi-realistic full-lifetime data for five degradation modes were generated, and the parameters of each mode are as follows: Mode A (rapid degradation): τ=912.7h, β=5.35, failure life≈1582.5h, corresponding to the degradation characteristics of buses under extreme conditions (scorching sun in midsummer, frequent start-stop, heavy-load uphill route); Mode B (faster degradation): τ=1456.3h, β=4.72, failure lifetime≈2717.5h, corresponding to high-intensity urban bus routes; Mode C (moderate degradation): τ=2038.5h, β=4.15, failure lifetime≈4144.2h, corresponding to ordinary mixed bus routes; Mode D (normal degradation, baseline): τ=2607.8h, β=3.88, failure life≈5569.9h, corresponding to standard operation and maintenance of conventional urban public transport; Mode E (slow degradation): τ=3142.1h, β=3.41, failure lifetime≈7451.1h, corresponding to flat suburban lines and mild driving conditions.

[0077] This embodiment uses mode D (normal degradation, failure lifetime 5569.9h) as the benchmark for detailed verification, and then uses the other four modes for robustness verification.

[0078] (II) Training Data Construction and Dual-Channel Feature Extraction Based on the 5569.9h lifetime data of pattern D, the dataset is divided according to the running time: the training set is the first 60% of the running time (t=0~3341.9h), the validation set is 60%~70% (t=3341.9~3898.9h), and the test set is 70%~100% (t=3898.9~5569.9h).

[0079] Set the sliding window width W = 60 sampling points (corresponding to 5 hours of recent dynamic information). Construct training samples from the training set with a sliding stride of 5 sampling points (25 minutes). Each sample's input consists of two parts: Channel 1, sliding window timing characteristics: a matrix of 15 sensor readings over the most recent 60 time steps, with dimensions 60×15. The 15 sensor channels include: DC voltage VD, DC current ID, DC power PDC, and 12 temperature channels (VD208 upper burner, VD212 lower burner, VD224 exhaust port, VD236 heat exchanger outlet, VD240 fuel cell stack inlet, VD244 fuel cell stack outlet, VD256 fuel cell stack center, VD272 fuel inlet, VD276 fuel preheating, VD280 air inlet, VD288 exhaust gas, VD316 environment).

[0080] Channel 2, Cumulative Degradation Features: Eight types of degradation feature vectors F1~F8 accumulated from t=0 to the current time. The definitions of each cumulative degradation feature are as described above, and the actual values ​​at key time points in Mode D are shown in Table 1 below: Table 1: Actual values ​​of each cumulative degradation feature in Mode D (normal degradation) at critical degradation stages

[0081] The 60×15 matrix from channel one and the 1×8 vector from channel two are concatenated along their feature dimensions, resulting in an input tensor of 60×23 for each sample. The label for each sample is the true RUL value at the current time step: RUL(t) = 5569.9 -t (when t≤5569.9h).

[0082] Taking the dual-channel input features at four different time points in the test set as an example, Tables 2 and 3 are shown below: Table 2: Sliding window temporal features of the test set at four time points (channel 1) Feature values

[0083] Table 3: Cumulative degradation features of the test set at four time points (Channel 2) Feature values

[0084] As shown in the table above, the cumulative degradation characteristics at the four time points are significantly different: taking F2 (total voltage decay) as an example, it increases from 1265 V·h at t=4000h to 10846 V·h at t=5300h, an increase of 8.6 times; F5 (cumulative start-stop count) increases from 2257 times to 5659 times, an increase of 2.5 times. This indicates that even if the sliding window sensor readings at two different times are similar, the difference in their cumulative degradation characteristics is sufficient for the model to distinguish different degradation stages. This is precisely the significant difference between the embodiments of the present invention and existing technologies that rely solely on sensor readings at the current time.

[0085] (III) Model Training Three prediction models are independently trained on the training set, and early stopping is monitored by the MSE of the validation set.

[0086] LSTM network: two-layer structure, the first layer has 128 hidden units, the second layer has 64 hidden units, which compresses the time sequence into a fixed-length vector; a Dropout layer (with a dropout probability of 0.3); a fully connected regression head of 32→1, which outputs a scalar RUL estimation value. The loss function is the mean square error MSE plus L2 regularization (regularization coefficient λ=0.001). The optimizer is Adam, the learning rate is 0.001, the batch size is 64, the maximum number of training epochs is 200, and the early stopping patience value is 20.

[0087] GRU network: two-layer structure, the first layer has 64 hidden units, the second layer has 32 hidden units; a Dropout layer (with a dropout probability of 0.2); a fully connected regression head of 16→1. The remaining hyperparameters are the same as those of the LSTM.

[0088] SVR model: using a radial basis function RBF kernel, the input is a fixed-dimensional vector obtained by performing global average pooling on the dual-channel features. The C parameter is searched from {0.1, 1, 10, 100}, the γ parameter is searched from {0.001, 0.01, 0.1, 1, scale}, and the optimal combination (C=100, γ=0.01) is determined through 5-fold cross-validation.

[0089] After the three models are independently trained, the MAE (Mean Absolute Error) of RUL prediction on the validation set is shown in Table 4 below: Table 4: MAE Mean Error

[0090] The error ranking of the three models on the validation set (LSTM<GRU<SVR) is consistent with the order of their model capacity and time sequence modeling capability. The MAE of GRU is 21.2% higher than that of LSTM, and the MAE of SVR is relatively large, which meets expectations, because SVR has no time sequence memory capability and is weaker than LSTM and GRU in fitting the nonlinear S-shaped feature of the Hill degradation curve. However, the physical significance of SVR lies in that: its prediction is robust, it is not prone to extreme deviation under unknown working conditions, and it plays an "anchoring" role in the GWO integration.

[0091] (4) GWO Weight Search Run GWO on the validation set to search for the optimal fusion weights of the three models. GWO parameter settings: population size N=30, maximum number of iterations is 100, search space w i ∈[0.1, 0.6] and Σw i=1 (guaranteed by normalization). The fitness function is the MAE on the validation set. After 100 iterations, the GWO convergence process is as follows: the optimal fitness of the initial population (i.e., the MAE of the best individual) is 376h, which converges to 258h after about 40 iterations. In the subsequent 60 iterations, the fitness stabilizes in the range of 258±5h, indicating that the algorithm has converged to near the global optimum.

[0092] The optimal weight combination obtained from the search is (w1*, w2*, w3*) = (0.45, 0.32, 0.23), meaning that LSTM contributes the most (0.45, consistent with its lowest single-model MAE), followed by GRU (0.32), and SVR provides a stable baseline (0.23, preventing extreme prediction biases from LSTM and GRU under unfamiliar conditions). The validation set MAE corresponding to this weight combination is 258h, which is 12.8% lower than that of the equal-weighted ensemble (w1=w2=w3=1 / 3, MAE=296h) and 17.3% lower than that of the optimal single-model LSTM (MAE=312h).

[0093] (v) Test set evaluation and robustness verification of the five modes Weighted fusion prediction was performed on the test set (t=3898.9~5569.9h) with optimal weights (w1*,w2*,w3*)=(0.45,0.32,0.23). The comparison with the single model and equal-weighted fusion is shown in Table 5 below: Table 5: Performance Comparison of Mode D (Normal Degradation) Test Set

[0094] As can be seen from the table above: GWO adaptive integration is achieved in MAE, RMSE, and R... 2 It outperforms all benchmark methods across all three metrics. The MAE is 21.3% lower than the best single-model LSTM and 16.8% lower than equal-weighted ensemble. R 2 The result of 0.954 indicates that the model explains 95.4% of the RUL variance.

[0095] (vi) Online operation process The three trained model files, along with the optimal weights (w1*, w2*, w3*) = (0.45, 0.32, 0.23) obtained through GWO search, are deployed to the in-vehicle online inference system. A new set of sensor readings is acquired every 5 minutes. (1) Data update: Add new readings to the sliding window (remove the oldest data point), and update all 8 cumulative degradation features according to the F1~F8 definitions; (2) Parallel inference of three models: The updated 60×23 dual-channel input is fed into LSTM, GRU and SVR respectively, and each model independently calculates the RUL estimate; (3) GWO weighted fusion: The final RUL is obtained by weighted summation of (0.45, 0.32, 0.23). final Predicted value; (4) Confidence interval and health status: 95% confidence interval is calculated based on the dispersion of the predicted values ​​of the three models, and the health status is output according to the three-level classification strategy.

[0096] Example 2: Health diagnosis and lifespan prediction of on-board SOFC battery system using dynamic weighting mode (a) Differences from Example 1 Example 1 applies to application scenarios with relatively stable and uniform operating conditions. However, buses experience frequent switching of operating conditions in actual operation—idling (vehicle speed ≈ 0 km / h, low current), starting acceleration (vehicle speed 5~15 km / h, medium current), cruising (vehicle speed 15~25 km / h, high current), and deceleration braking. The voltage-current characteristics and degradation performance of the battery system differ under each operating condition. Example 2 adopts a dynamic weighting model: automatically switching the optimal fusion weight combination according to the current operating condition category to achieve adaptive model integration based on operating conditions.

[0097] (II) Working condition segmentation and GWO weight search for each segment Using current density and voltage as operating condition characteristics, the training and validation data of Mode D are divided into three categories of operating conditions according to the following rules: Low load range: Current I < 4.0A (corresponding to idle / stop conditions), accounting for approximately 18% of the total data; Medium load range: 4.0A≤I<6.0A (corresponding to starting / low speed driving), accounting for approximately 47% of the total data; High load segment: I≥6.0A (corresponding to acceleration / cruising / hill climbing), accounting for approximately 35% of the total data.

[0098] The validation set was divided into three subsets according to the three types of working conditions mentioned above. GWO (with the same parameters as in Example 1) was run independently on each subset, and the dynamic fusion weights were obtained as shown in Table 6 below: Table 6: Optimal GWO Weights and Verification MAE for Each Operating Condition

[0099] The most significant feature in the table above is that under high load conditions, the GRU weight (w2=0.42) surpasses LSTM (w1=0.38) to become the leading contributing model. The physical reason for this is that under high load conditions, current fluctuations are large (±15%) and voltage fluctuations are severe, making short-term degradation rate information relatively more important for RUL prediction; GRU's high sensitivity to short-term fluctuations is particularly advantageous in this scenario. This contrasts with the result in Example 1 where the global weight LSTM consistently ranks first—the global weight reflects the average optimality across all operating conditions, while under high load conditions, the optimal fusion method of the model undergoes a structural change.

[0100] (III) Online dynamic switching During online inference, the system first determines the operating condition category k∈{low load, medium load, high load} based on the current current value, and then loads the optimal weight combination {w1,w2,w3} corresponding to that operating condition for model fusion. Operating condition determination and weight switching are performed each time the sensor reading is updated (every 5 minutes). The switching calculation involves only one current threshold comparison and one table lookup, and the computational overhead is negligible, having no impact on the real-time performance of online inference. The rest of the process is exactly the same as in Example 1.

[0101] Testing showed that the MAE accuracy of the dynamic weighting mode decreased from 312h to 265h under high load conditions, a reduction of 15.1%, validating the effectiveness of the adaptive weighting switching strategy. During the high-load acceleration / climbing phase, when accurate RUL prediction is most crucial for buses, dynamic weighting significantly improved the adaptability to drastic current fluctuations by increasing the contribution ratio of the GRU (Gross Flow Unit). Notably, the reduction in low load conditions was only 1.6% (187h → 184h), because the operating conditions in low load conditions change gradually, and the additional gain from dynamic switching is limited.

[0102] In summary, Example 1 (static weights) provides a stable, universal baseline scheme applicable to most operating conditions; Example 2 (dynamic weights), based on the baseline scheme, achieves a 15.1% improvement in prediction accuracy for high-load segments by dynamically switching the optimal fusion weights among the three operating conditions unique to buses. Both modes can be flexibly selected according to the actual deployment scenario.

[0103] The technical solutions provided in the above embodiments have at least one or more of the following beneficial effects: (1) Accurately distinguish battery systems in different operating stages by using cumulative degradation characteristic channels. This invention constructs eight types of degradation features that accumulate continuously from t=0, enabling the model input to include full-history degradation information. When the sensor readings of two battery systems are the same within the sliding window, they obtain different model inputs due to the differences in their accumulated features, resulting in differentiated RUL predictions that conform to the physical laws of degradation.

[0104] (2) Simultaneously sensing the degree of degradation, degradation rate and degradation acceleration The sliding window channel (channel one) provides recent degradation dynamics information, while the cumulative degradation feature channel (channel two) provides overall degradation process information. This embodiment of the invention uses a dual-channel structure to comprehensively characterize degradation from three dimensions (degree, rate, and acceleration), providing more complete information.

[0105] (3) Swarm intelligence algorithm realizes data-driven global optimal fusion The Global Optimization Word (GWO) method automatically searches for the optimal fusion weights of the three models on validation data, replacing the fixed weight rules in existing technologies. GWO's global optimization capability ensures convergence to a near-global optimal weight combination on given data, and the dynamic weight mode further enables adaptive model integration based on operating conditions.

[0106] (4) Multi-model heterogeneous integration improves prediction accuracy and robustness The LSTM, GRU, and SVR models differ structurally in their gating mechanisms, parameter scales, and modeling principles, complementing each other in modeling the SOFC degradation process from different perspectives. Compared to a single model, the prediction accuracy and robustness of the integrated three models are superior to any single model.

[0107] (5) Utilize model dispersion to provide convenient prediction confidence intervals The confidence interval is directly calculated based on the dispersion of the predicted values ​​from the three models, without the need for additional Monte Carlo sampling or Bayesian inference. The computational cost is significantly lower than that of the quantile regression plus Bootstrap sampling scheme.

[0108] (6) It has inherent robustness to sensor noise. The long-term cumulative effect of the cumulative characteristics smooths out single measurement fluctuations (e.g., the effect of voltage fluctuations of a few millivolts on F2 (total voltage decay) is numerically negligible). Combined with the three-model integrated averaging effect, the final RUL prediction value is insensitive to sensor noise and exhibits a physically reasonable trend of slow and monotonically decreasing. Compared to the prior art (e), where RUL jumps for hundreds of hours due to single reading fluctuations, the prediction stability of this invention is significantly improved.

[0109] Figure 4 This is a block diagram of an electronic device 400 for a method of predicting the remaining useful life of an on-board solid oxide fuel cell system according to an exemplary embodiment. For example, the electronic device 400 may be an industrial control computer, a computer, an edge server, an edge computing device, etc.

[0110] Reference Figure 4The electronic device 400 may include one or more of the following components: a processing component 402, a memory 404, a power supply component 406, an input / output (I / O) interface 408, and a communication component 410.

[0111] Processing component 402 typically controls the overall operation of electronic device 400, such as operations associated with data calculation, control, command issuance, and camera triggering. Processing component 402 may include one or more processors 420 to execute instructions to complete all or part of the steps described above. Furthermore, processing component 402 may include one or more modules to facilitate interaction between processing component 402 and other components.

[0112] Memory 404 is configured to store various types of data to support the operation of device 400. Examples of such data include instructions for any application or method operating on electronic device 400, image data, associated data, configuration data, etc. Memory 404 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static 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), magnetic storage, flash memory, magnetic disk, or optical disk.

[0113] Power supply component 406 provides power to various components of electronic device 400. Power supply component 406 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to electronic device 400.

[0114] Communication component 410 is configured to facilitate wired or wireless communication between electronic device 400 and other devices. Electronic device 400 can access wireless networks based on communication standards, such as Wi-Fi, 2G, or 3G, or combinations thereof. In one exemplary embodiment, communication component 410 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, communication component 410 also includes a near-field communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on radio frequency identification (RFID) technology, Infrared Data Association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0115] In an exemplary embodiment, the electronic device 400 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform one or more methods of the above embodiments.

[0116] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 404 including instructions, which can be executed by a processor 420 of an electronic device 400 to perform the above-described methods. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device, etc. When the instructions in this non-transitory computer-readable storage medium are executed by the processor 420 of the electronic device 400, one or more methods in the above embodiments can be implemented.

[0117] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed in these embodiments. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the following claims.

[0118] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the present invention is limited only by the appended claims. The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining service life of an on-board solid oxide fuel cell system, characterized in that, include: Based on real-time data from multiple sensors of the on-board solid oxide fuel cell system, the sliding window timing characteristics and cumulative degradation characteristics of the battery system are obtained. The cumulative degradation characteristics include the cumulative electrical performance characteristics, cumulative operating event characteristics, cumulative carbon deposition index, and the cumulative temperature gradient of the burner within the sliding window from the moment the system starts up and runs. The model input tensor is obtained based on the sliding window temporal features and the cumulative degradation features; The input tensor of the model is input into three target models respectively, and the remaining service life of the battery system is predicted by the three target models respectively to obtain three predicted values. The three target models include a long short-term memory network, a gated recurrent unit, and a support vector regression. Based on the three predicted values ​​and the fusion weights, the prediction result of the remaining useful life is obtained. The fusion weights are obtained through a swarm intelligence optimization algorithm during the model training phase.

2. The prediction method as described in claim 1, characterized in that, The cumulative characteristics of the runtime events include: The cumulative number of start-stop cycles and the duration of high-temperature exposure, wherein the cumulative number of start-stop cycles is the cumulative number of times the battery system has started since the system started running, and the duration of high-temperature exposure is the cumulative time during which the battery system temperature is greater than a preset threshold.

3. The prediction method as described in claim 2, characterized in that, The cumulative electrical performance characteristics include: The cumulative operating time, total voltage decay, voltage degradation rate, and voltage degradation acceleration of the battery system.

4. The prediction method as described in claim 1, characterized in that, The step of obtaining the model input tensor based on the sliding window temporal features and the cumulative degradation features includes: The cumulative degradation features are broadcast to the time dimension of the sliding window to obtain the cumulative degradation time-series features of the cumulative degradation features; The model input tensor is obtained by fusing the sliding window temporal features and the cumulative degradation temporal features.

5. The prediction method as described in claim 1, characterized in that, The prediction method further includes: During model training, the fusion weights of the three target models are obtained by searching on the validation dataset using the Grey Wolf Optimization Algorithm.

6. The prediction method as described in claim 5, characterized in that, The process of obtaining the fusion weights of the three target models by searching on the validation dataset using the Grey Wolf optimization algorithm includes: The static fusion weights of the three target models are obtained by searching the entire validation dataset using the Grey Wolf optimization algorithm; or, The entire validation dataset is divided into multiple validation dataset subsets according to current density and / or temperature range. The Gray Wolf optimization algorithm is then used to search each validation dataset subset to obtain multiple dynamic fusion weights.

7. The prediction method as described in claim 6, characterized in that, The step of obtaining the prediction result of the remaining useful life prediction based on the three predicted values ​​and the fusion weight includes: The current current density and / or current temperature range of the battery system are determined based on the sliding window timing characteristics, and the target dynamic fusion weights corresponding to the current current density and / or current temperature range are obtained. Based on the three predicted values ​​and the target dynamic fusion weight, the prediction result of the remaining useful life is obtained.

8. The prediction method according to any one of claims 1-7, characterized in that, The step of obtaining the prediction result of the remaining useful life prediction based on the three predicted values ​​and the fusion weight includes: The three predicted values ​​are weighted and summed according to the fusion weights to obtain the target prediction duration for the remaining useful life; The confidence interval of the target prediction duration is calculated based on the target prediction duration and the dispersion of the three prediction values. The target prediction duration and the confidence interval are obtained as the prediction result.

9. The prediction method as described in claim 8, characterized in that, After obtaining the target prediction duration and the confidence interval as the prediction result, the prediction method further includes: The health status level of the battery system is determined based on the upper and lower bound thresholds corresponding to the target prediction duration and the battery system design life. Output the prediction result and the health status level.

10. An electronic device, characterized in that, It includes a memory and one or more programs, wherein one or more programs are stored in the memory and are configured to be executed by one or more processors to implement the prediction method as described in any one of claims 1-9.

Images