A fuel cell system time series data anomaly detection method based on residual analysis

Abstract

This invention relates to the field of anomaly detection technology for fuel cell systems, specifically a method for anomaly detection of time-series data in fuel cell systems based on residual analysis. The specific implementation steps include: acquiring multi-source design simulation data of the system and constructing it as a multi-source design simulation data tensor; constructing a computer-aided design model, performing behavioral simulation on the simulation data tensor, and outputting simulated design behavior values; comparing the simulation data as simulation observations with the simulated design behavior values ​​to calculate a multi-dimensional residual sequence, and performing online adaptive updates to the design model; establishing a physical reachability matrix, and adjusting dynamic thresholds based on the matrix and real-time simulation conditions; weighted fusing of the multi-dimensional residual sequences to generate collaborative design constraint indices, and comparing them with the dynamic thresholds to generate computer-aided design verification results; inputting the design verification results into a preset residual-control mapping rule, and outputting design parameter optimization instructions to optimize the design parameters.

Description

Technical Field

[0001] This invention relates to the field of anomaly detection technology for fuel cell systems, specifically a method for anomaly detection of time-series data in fuel cell systems based on residual analysis. Background Technology

[0002] In complex systems like fuel cells, building highly sensitive anomaly detection capabilities is crucial for ensuring their performance and safety throughout their entire lifecycle. Existing technologies primarily provide a solid foundation by integrating analytical model-based residual analysis with data-driven methods. Analytical model-based residual analysis establishes mathematical or physical models of the system, simulating and verifying typical failure modes such as flooding and drying, providing a clear physical basis for the design of diagnostic logic. Simultaneously, data-driven methods, such as machine learning, allow designers to utilize massive amounts of historical or simulation data to model the nonlinear and time-varying characteristics of the system under various complex simulation conditions, enhancing the comprehensiveness of computer-aided design verification. By comprehensively applying these technologies, especially using residual analysis as a bridge connecting model simulation and real-world observation, a powerful tool is provided for evaluating whether the system design meets expected performance indicators.

[0003] In current system design and verification practices, the effectiveness of using residual analysis to construct a highly sensitive anomaly detection system is limited by the fidelity and adaptability of the simulation model. On the one hand, to improve computational efficiency, the physical model used in the system is usually simplified. This simplification makes it difficult for the model to accurately reproduce the complex behavior of fuel cells under dynamic simulation conditions, multivariate coupling, and aging degradation, thus generating deviations between the model and the physical entity. This results in a low residual signal-to-noise ratio generated in the simulation environment, making it difficult to accurately reflect early fault signals and affecting the reliability of the detection logic design. On the other hand, existing methods, including data-driven methods, generally suffer from high computational complexity, difficulty in meeting the real-time requirements of the system, and the separation between the detection module and the control execution module, failing to form a closed-loop design of "detection-diagnosis-optimization," which limits the overall adaptability and robustness of the system.

[0004] In summary, although the existing design and verification framework provides design ideas for diagnostic functions of fuel cell systems, there are still significant shortcomings in building high-fidelity simulation models, generating high-quality residuals, and verifying and optimizing the system's ability to detect early weak anomalies. Therefore, this invention proposes a method for detecting time-series data anomalies in fuel cell systems based on residual analysis. Summary of the Invention

[0005] The purpose of this invention is to provide a method for detecting anomalies in time-series data of fuel cell systems based on residual analysis, so as to achieve accurate detection of early and subtle anomalies in fuel cell systems and proactive and safe online optimization and adjustment.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for detecting anomalies in time-series data of a fuel cell system based on residual analysis, comprising:

[0008] Acquire multi-source design simulation data and construct it as a multi-source design simulation data tensor;

[0009] A machine learning-based computer-aided design model is constructed to simulate the behavior of multi-source design simulation data tensors and output simulated design behavior values ​​that characterize the behavior of design specifications. The multi-source design simulation data is compared with the simulated design behavior values ​​as simulation observations to calculate a multidimensional residual sequence. The multidimensional residual sequence is used to perform online adaptive updates of the computer-aided design model.

[0010] A physical reachability matrix is ​​established to characterize the design dependency paths between various design parameters. The dynamic threshold is adjusted based on the matrix and the real-time simulation conditions. The multidimensional residual sequence is weighted and fused using a physical characteristic adaptive weighting method to generate a collaborative design constraint index. This index is then compared with the dynamic threshold to generate computer-aided design verification results.

[0011] The computer-aided design verification results are input into a preset residual-control mapping rule, and design parameter optimization instructions are output to optimize the design parameters.

[0012] Preferably, the step of acquiring multi-source design simulation data and constructing it into a multi-source design simulation data tensor includes:

[0013] The multi-source design simulation data includes: multi-point temperature data, operating voltage, load current, energy storage operating parameters, and design control parameters. The acquired multi-source design simulation data is converted into structured time-series data by parsing timestamps, identifiers, and data fields. Continuous data is constructed into tensors according to time step, variable dimension, and spatial dimension, while discrete event data is stored separately as event sequences. The multi-source design simulation data tensor is obtained by associating timestamps with the tensors.

[0014] Preferably, the construction of the computer-aided design model based on machine learning includes:

[0015] Historical design simulation data is acquired, processed using tensor decomposition, and an initial spatiotemporal tensor is constructed. Based on this initial spatiotemporal tensor, a machine learning-based computer-aided design model is trained using a neural network algorithm. During design verification, an incremental tensor decomposition method satisfying sparsity constraints is used to perform low-rank updates on the initial tensor model. Combined with feedback information from multidimensional residual sequences, the computer-aided design model is subjected to online learning and adaptive updates to improve simulation accuracy under dynamically changing simulation conditions. The model update frequency is adaptively adjusted according to the magnitude of residual changes, increasing the update frequency under dynamic simulation conditions and decreasing it under steady-state simulation conditions.

[0016] Preferably, the step of performing behavioral simulation on the multi-source design simulation data tensor and outputting simulated design behavior values ​​characterizing the design specification behavior includes:

[0017] The multi-source design simulation data tensor is input into the computer-aided design model, and forward reasoning is performed on each time step of the multi-source design simulation data tensor. Based on the historical sequence of the multi-source design simulation data, the simulation of the state of the next time step is output. The simulation output constitutes a simulation tensor that is consistent with the input tensor structure in terms of variables and spatial dimensions, which represents the simulated design behavior value of the ideal behavior state that conforms to the specifications under the current simulation conditions.

[0018] Preferably, the multi-source design simulation data is compared with the simulated design behavior values ​​as simulation observations to calculate a multidimensional residual sequence, including:

[0019] The multi-source design simulation data is used as the simulation observation value and the simulated design behavior value. The difference is calculated element by element to obtain the initial residual tensor with the same structure as the previous two. The initial residual tensor is grouped according to the physical characteristics and dimensional type of the design parameters to obtain a multidimensional residual sequence. The multidimensional residual sequence is a quantitative index of the degree of deviation between the actual behavior and the design specification behavior.

[0020] Preferably, the step of establishing a physical reachability matrix characterizing the design dependency paths between various design parameters, adjusting the dynamic threshold based on the matrix and real-time simulation conditions, and generating a collaborative design constraint index by weighting and fusing the multidimensional residual sequence through a physical characteristic adaptive weighting method, and comparing it with the dynamic threshold to generate computer-aided design verification results, includes:

[0021] A system-designed accessibility matrix is ​​established to clearly mark the design dependency paths between various design parameters. The statistically generated causal graph is used as a preliminary reference causal relationship, and the causal edge relationships with design dependency paths are retained through screening and weighting by the physical accessibility matrix. A dynamic graph update mechanism is designed to dynamically adjust the graph update frequency according to the simulation process: the update frequency is automatically increased when the rate of change of relevant data increases, and the update frequency is decreased under steady-state simulation conditions. Through a physical characteristic adaptive weighting method, relevant residuals are assigned weights according to system characteristics, and weighted fusion is used to generate collaborative design constraint indicators. An adaptive dynamic threshold is set based on the real-time simulation conditions, and computer-aided design verification results are generated based on the comparison results between the indicators and the thresholds.

[0022] Preferably, the computer-aided design verification results are input into a preset residual-control mapping rule, and design parameter optimization instructions are output to optimize the design parameters, including:

[0023] The calculation of the numerical value and rate of change of the computer-aided design verification result exceeding the dynamic threshold is performed. The numerical value and rate of change exceeding the threshold are input into the preset residual-control mapping rule, and the design parameter optimization instruction is output. When the residual corresponding to the data in the multi-source design simulation data tensor of the computer-aided design verification result exceeds the limit, the design parameter optimization instruction generates a progressive adjustment sequence. The controller executes the design parameter update in segments and triggers topology adjustment after verifying the stability condition synchronously. When the energy storage-related residual and the system state residual exceed the preset threshold, while other residuals remain within the normal range, the preset physical characteristic model is called to update the compensation threshold, and the state parameter safety window is constrained in the controller. After all design parameter optimization instructions are executed, a graded anomaly warning is triggered. The warning level is positively correlated with the magnitude and rate of change of the residual exceeding the limit.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] 1. This invention constructs a multi-source design simulation data tensor and combines it with an incremental tensor decomposition method that satisfies sparsity constraints to perform online low-rank updates on the design model. This solves the problem that simplified physical models in the prior art are difficult to accurately reproduce dynamic, coupling, and aging characteristics, enabling computer-aided design models to more accurately reflect the real behavior of the system under complex and variable simulation conditions, and improving the signal-to-noise ratio and reliability of subsequent residual analysis.

[0026] 2. This invention establishes a Granger causal graph based on physical reachability matrix verification and performs adaptive weighting based on physical characteristics. This effectively distinguishes between real physical causal chains and statistical pseudo-correlation, solving the problem that existing methods are unable to identify early and weak fault signals from a strong noise background, and improving the system's predictability and sensitivity to potential risks.

[0027] 3. This invention directly transforms design verification results into specific, hierarchical design parameter optimization instructions through a preset residual-control mapping rule. This overcomes the defect of the detection module and control execution module being isolated from each other in the prior art, and constructs a complete closed loop from deviation detection to adaptive adjustment of design parameters, thereby improving the robustness, operating efficiency and safety of the system throughout its entire life cycle.

[0028] 4. The control logic of this invention is based on the comprehensive mode judgment of multiple residual channels and combined with the synchronous verification of safety boundary conditions such as pressure and flow. Before adjusting the compensation threshold of the DC-DC converter, an independent simulation model is called and a safety window is constrained, which avoids misjudgment and unsafe control actions that may be caused by a single abnormal index, and ensures that the decision-making process for optimizing design parameters is more accurate, reliable and safe. Attached Figure Description

[0029] Figure 1 This is an overall flowchart of the time-series data anomaly detection method for fuel cell systems based on residual analysis according to the present invention;

[0030] Figure 2 This is a schematic diagram illustrating the construction of multi-source design simulation data tensors and the generation of residuals in an example of the present invention.

[0031] Figure 3 This is a schematic diagram of the residual-based closed-loop control decision-making process in an example of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention. Other embodiments that can be obtained by those skilled in the art based on the content disclosed in this specification without creative effort are all within the scope of protection claimed by the present invention.

[0033] Example 1:

[0034] This embodiment provides a residual analysis-based method for detecting anomalies in time-series fuel cell system data for unmanned aerial vehicles (UAVs) equipped with fuel cell systems. This method aims to generate a time-series data for an upcoming flight mission cycle, enabling early identification of potential system anomalies and proactive parameter optimization while ensuring flight safety and meeting power requirements. (Refer to...) Figure 1The specific implementation steps of the method of the present invention in this scenario are as follows: S1. Obtain multi-source design simulation data of the system and construct it as a multi-source design simulation data tensor; S2. Construct a computer-aided design model based on machine learning, simulate the behavior of the multi-source design simulation data tensor, and output simulated design behavior values ​​that characterize the system design specification behavior; S3. Compare the multi-source design simulation data as simulation observation values ​​with the simulated design behavior values ​​to calculate a multi-dimensional residual sequence; S4. Use the multi-dimensional residual sequence to perform online adaptive updates on the computer-aided design model; S5. Establish a physical reachability matrix that characterizes the design dependency paths between various design parameters of the system, and adjust the dynamic threshold based on the matrix and real-time simulation conditions; use a physical characteristic adaptive weighting method to weight and fuse the multi-dimensional residual sequence to generate a collaborative design constraint index, and compare it with the dynamic threshold to generate a computer-aided design verification result; S6. Input the computer-aided design verification result into a preset residual-control mapping rule, and output a design parameter optimization instruction to optimize the design parameters.

[0035] Furthermore, the multi-source design simulation data of the acquired system is constructed into a multi-source design simulation data tensor, referring to... Figure 2 Corresponding to step S1 above, the specific implementation process includes:

[0036] The fuel cell system includes a fuel cell stack, a lithium battery energy storage module, a three-way valve cooling subsystem, a bidirectional DC-DC converter, a data acquisition unit, and a controller. The system acquires multi-source design simulation data of the fuel cell system. The multi-source design simulation data includes: the inlet temperature (range 50-80℃), outlet temperature (range 60-90℃), and critical area temperature (10 key points, range 55-85℃) of the fuel cell stack acquired by the temperature monitoring unit; the bus voltage (range 24-28V), load current (range 10-30A), energy storage unit status (SOC range 20%-100%), and temperature (range 25-45℃) monitored by the data acquisition unit; and the three-way valve opening (range 0%-100%) and cooling water circuit topology status (open / closed state) recorded by the controller.

[0037] Multi-source design simulation data is acquired at a frequency of 1Hz. The acquired multi-source design simulation data is converted into structured time-series data by parsing timestamps accurate to milliseconds, sensor IDs, and fields. For a continuous inspection task lasting 2 hours, continuous data such as temperature and voltage are constructed into a 7200×7×10 three-dimensional tensor according to time step (7200 seconds), variable dimensions (7 variables: inlet temperature, outlet temperature, zone temperature, operating voltage, load current, SOC, battery temperature), and spatial dimensions (10 stack locations). At the same time, discrete event data such as cooling water circuit topology switching are stored separately as event sequences. By associating the tensor with timestamps, a multi-source design simulation data tensor that can completely describe the spatiotemporal state of the system is obtained.

[0038] This embodiment establishes a unified processing flow from multi-source heterogeneous data to structured multi-source design simulation data tensors, generating a high-dimensional data representation that can comprehensively and synchronously characterize the dynamic characteristics of complex systems. This solves the problem of incomplete and inaccurate state descriptions caused by the single data perspective or asynchronous time in traditional methods, thereby enhancing the input signal-to-noise ratio for subsequent anomaly detection.

[0039] Furthermore, the construction of a machine learning-based computer-aided design model, which simulates the behavior of multi-source design simulation data tensors and outputs simulated design behavior values ​​characterizing the system design specification behavior, corresponds to step S2 above. The specific implementation process includes:

[0040] Six months of high-altitude inspection historical design simulation data were acquired. The data were processed using the CP tensor decomposition method to construct an initial spatiotemporal tensor of 7200×7×10. Based on the initial spatiotemporal tensor, a neural network algorithm containing 128 hidden units was used for training to learn the dynamic behavior of the system under normal operation mode. The training process iterated for 100 cycles with a learning rate of 0.001, ultimately resulting in a machine learning computer-aided design model that can predict the future state of the system.

[0041] The data-driven computer-aided design model is specifically a neural network containing two Long Short-Term Memory (LSTM) layers and a fully connected output layer. The model takes as input a tensor of multi-source design simulation data from the past 60 time steps and outputs a tensor predicting the system state for the next time step (1 second later). The first LSTM layer contains 128 hidden units with a tanh activation function and returns a sequence output. The second LSTM layer also contains 128 hidden units with a tanh activation function, returning only the output of the last time step. A fully connected layer then maps the LSTM output to the final prediction dimension (7 × 10 = 70 nodes). The training process uses the Adam optimizer with mean squared error as the loss function, iterates for 100 epochs, and sets the initial learning rate to 0.001.

[0042] During system operation, an incremental tensor decomposition method satisfying sparsity constraints is employed, such as online robust principal component analysis. When the 5-second moving average of any channel in the multidimensional residual sequence exceeds 1.5 times its historical standard deviation, a model update is triggered. During the update, the latest residual tensor is treated as a sparse noise term, and the OR-PCA algorithm is used to perform low-rank correction on the weight matrix of the last layer (fully connected layer) of the neural network. This allows for rapid adaptation to minor changes in system characteristics without retraining the entire model. Combined with real-time feedback information from the multidimensional residual sequence, the design model undergoes online low-rank updates. The update frequency is adaptively adjusted based on the moving average of the residuals: when this value increases by more than 3% within 5 seconds, the update frequency is increased to once every 60 seconds; under steady-state simulation conditions, it is reduced to once every 600 seconds.

[0043] This embodiment establishes a data-driven design model trained on historical data, generating a dynamic simulation benchmark that can accurately reproduce the complex nonlinear behavior of the system. This solves the problem that traditional analytical models suffer from low simulation fidelity and are unable to adapt to changing simulation conditions due to oversimplification, thus ensuring the accuracy of subsequent residual calculations.

[0044] Real-time multi-source design simulation data tensors are input into a trained computer-aided design model. The model runs on the controller and performs forward inference for each time step (1 second). Based on the historical sequence of the past 7200 seconds, it outputs a simulation tensor of the system state for the next time step. The output simulation tensor is structurally consistent with the input tensor (7200×7×10) and includes predictions of all temperatures, operating voltages, load currents, and SOC values ​​for the next second. It represents the ideal behavior state of the system under high-altitude inspection simulation conditions, which conforms to the specifications. That is, the simulated design behavior value. The time taken for a single complete inference process is less than 0.1 seconds.

[0045] This embodiment utilizes a design model for real-time online simulation, generating a multi-dimensional dynamic simulation benchmark that adapts to the simulation conditions. This solves the problem of frequent false alarms and missed alarms caused by traditional detection methods based on static or segmented thresholds, which cannot track the highly dynamic characteristics of the system. This improves the applicability and robustness of anomaly detection.

[0046] Furthermore, the multi-source design simulation data is compared with the simulated design behavior values ​​as simulation observations to calculate a multidimensional residual sequence, which is then referred to... Figure 2 Corresponding to step S3 above, the specific implementation process includes:

[0047] The real-time multi-source design simulation data tensor is used as the simulation observation value. The simulated design behavior value output by the model is subtracted element-wise to generate an initial residual tensor of 7200×7×10. This tensor is then processed by grouping:

[0048] For the temperature-related residual set at 10 locations with spatial distribution characteristics, principal component analysis (PCA) was used to extract the top 3 principal components as core change pattern features (requiring a cumulative contribution rate greater than 95%).

[0049] For the voltage and current related residual set that characterizes the overall state of the system, calculate its standardized statistical characteristics (mean and standard deviation).

[0050] For the state-related residual set of the energy storage unit, the residual deviation is calculated by combining it with the offline design parameter simulation model. In this embodiment, the offline design parameter simulation model is specifically a state estimator based on extended Kalman filter (EKF). The core of this estimator is a state-space model established based on a second-order RC equivalent circuit model. Its state prediction step is to calculate the prior state estimate at the current moment based on the state vector at the previous moment (including the voltage of the SOC and the two-stage RC network) and the input current at the current moment through a preset state transition matrix and input matrix. The prior state estimate is substituted into the observation equation to obtain the theoretical terminal voltage. The theoretical terminal voltage is compared with the actual measured terminal voltage. The prior state estimate is corrected by combining the Kalman gain calculated by the covariance matrix to obtain the optimal posterior state estimate at the current moment. Among them, the design parameters of each matrix in the state-space model are pre-identified and determined by the battery factory data and experimental calibration. The process noise covariance and observation noise covariance are also pre-set according to the experimental data. For example, the diagonal elements take values ​​in the range of 1e-6 to 1e-4 and 1e-4 to 1e-2 respectively. The model finally outputs the theoretical SOC after optimal estimation, and the difference is calculated with the simulation observation value to obtain the residual deviation.

[0051] Finally, the results of these three types of feature processing are output as independent channels in parallel to form the final multidimensional residual sequence. The entire calculation process is executed on the controller, and the time taken for a single operation is less than 0.2 seconds.

[0052] This embodiment calculates the difference between the actual value and the high-fidelity simulation value, and groups and extracts features from the obtained residuals to generate a fault feature sequence with more physical meaning and a higher signal-to-noise ratio. This solves the problem that weak anomalies in the original measurement signal are easily submerged by strong noise and are difficult to use directly for diagnosis, thereby enhancing the detection sensitivity of early and weak faults.

[0053] Furthermore, the computer-aided design model is adaptively updated online using the multidimensional residual sequence. Corresponding to step S4 above, the specific implementation process includes:

[0054] The calculated multidimensional residual sequence is used as a real-time, high-value feedback input signal and passed to the update module of the computer-aided design model. This residual sequence quantifies the deviation between the model's current predicted behavior and the system's actual operating behavior. The update module uses an incremental tensor decomposition method that satisfies sparsity constraints to perform model updates, processing the received multidimensional residual sequence into a sparse noise term. Then, an online, low-rank correction is performed on the weight tensor of the computer-aided design model. The correction process is completed without retraining the entire neural network model, only fine-tuning the key weight parameters of the model, rather than a global gradient backpropagation that consumes a lot of computational resources.

[0055] To further improve the system's intelligence level and resource utilization efficiency, the execution frequency of the model update is adaptively adjusted based on the change amplitude of the multidimensional residual sequence within a preset time window. Specifically, when the moving average or volatility of the residuals increases significantly in the short term, the system judges that the current operating conditions are changing drastically and will automatically increase the model update frequency; conversely, when the system is in steady-state operation and the residuals change slowly, the update frequency will be automatically reduced.

[0056] Furthermore, the establishment of a physical reachability matrix for design dependency paths among various design parameters of the characterization system, and the adjustment of dynamic thresholds based on the matrix and real-time simulation conditions; the weighted fusion of the multidimensional residual sequence through a physical characteristic adaptive weighting method to generate collaborative design constraint indices, which are then compared with the dynamic thresholds to generate computer-aided design verification results. Corresponding to step S5 above, the specific implementation process includes:

[0057] Establish an N×N physical reachability matrix, where N is the total number of system design parameters. A matrix element (i, j) of 1 indicates that a change in design parameter i can directly physically affect design parameter j, otherwise it is 0. Filter and weight the statistically generated Granger causality graph (screening condition: p-value less than 0.05) with this physical reachability matrix: if the Granger causality graph shows that i has a statistical causal relationship with j, and the reachability matrix (i, j) is 1, then retain the causal relationship; if the reachability matrix (i, j) is 0, then set the weight of the statistical causal edge to a low value, such as 0.05.

[0058] Weights are assigned to each residual channel: the weight of temperature-related residuals ranges from 0.4 to 0.6, the weight of voltage and current-related residuals ranges from 0.3 to 0.5, and the weight of energy storage unit state-related residuals is adaptively adjusted based on the cumulative statistic of the root mean square value of SOC residuals over the past 10 flight hours. A collaborative design constraint index is generated by weighted fusion of all information and compared with a dynamic threshold set based on real-time simulation conditions (e.g., set to deviate from the normal value by 0.5 standard deviations) to generate computer-aided design verification results.

[0059] Furthermore, the computer-aided design verification results are input into a preset residual-control mapping rule, and design parameter optimization instructions are output to optimize the design parameters. Corresponding to step S6 above, the specific implementation process includes:

[0060] The computer-aided design verification results are calculated to exceed the dynamic threshold and their rate of change. This information is then input into a preset residual-control mapping rule. In this embodiment, the mapping rule is specifically implemented as a decision tree model trained through offline simulation. Its input features are the normalized amplitude and first-order difference value (representing the rate of change) of each residual channel. Its leaf nodes directly output a pre-set sequence of design parameter optimization instructions for the specific residual mode, such as a structure containing key-value pairs, where the key is the name of the design control parameter (e.g., "target opening") and the value is the specific control value (e.g., "45%)".

[0061] When a specific value corresponding to the fuel cell temperature gradient data in the multi-source design simulation data tensor indicates that the residual exceeds the limit (e.g., the temperature gradient is greater than 5°C) in the computer-aided design verification results, the design parameter optimization command generates a "progressive three-way valve opening adjustment sequence," specifically: within 1.5 seconds, the three-way valve opening is linearly adjusted from the current 30% to 45%. During the adjustment, the controller synchronously verifies that the pressure fluctuation of the cooling water circuit is less than 0.02 MPa / s and the flow stability fluctuation is less than 5%. Only after these conditions are met is the final switching of the cooling water circuit topology triggered.

[0062] When the residuals related to the state of the energy storage unit and the residuals of the load current fluctuate drastically at the same time, while the residuals of the temperature remain within the normal range, the system calls an independent energy storage unit state simulation model to update the compensation threshold of the DC-DC converter. When the controller executes this update, it can also add an additional state parameter safety window constraint (e.g., maintaining the SOC between 20% and 90%) to enhance safety. All design parameter optimization commands trigger graded anomaly warnings after execution. The warning level is positively correlated with the residual over-limit magnitude and rate of change, and is visualized through the UAV control interface.

[0063] This implementation constructs a complete technical solution, from data acquisition and high-fidelity online adaptive modeling to multi-dimensional residual feature analysis, and finally to closed-loop adaptive control based on physical causal fusion. It solves the problem that the system is unable to effectively detect early and subtle deviations in system behavior due to the discrepancy between the simulation model and the physical entity and the low signal-to-noise ratio. It also solves the functional defect that prevents the system from autonomously optimizing itself based on real-time status because the detection and analysis module and the control execution module are isolated from each other. Therefore, it can detect deviations from the system's specifications earlier and more accurately, and enables the system to autonomously adjust its operating design parameters according to real-time status. This allows the system to maintain a high-efficiency, safe, and stable operating state under changing simulation conditions, thereby improving the reliability of the system throughout its entire life cycle.

[0064] Example 2:

[0065] This embodiment focuses on the detection of abnormal voltage in the fuel cell system of a drone equipped with a fuel cell system. The key point is the specific process of performing deep causal analysis on the residual sequence to achieve diagnosis and generating closed-loop control commands.

[0066] The system acquires multi-source sensor data and constructs it into a multi-source design simulation data tensor. Then, it loads a data-driven design model that has been pre-trained based on massive historical data. Subsequently, it uses the model to simulate the behavior of real-time data to obtain ideal simulated design behavior values. Finally, it compares the simulated observation values ​​with the simulated values ​​and calculates and extracts information-rich multidimensional residual sequences.

[0067] Once the multidimensional residual sequence is calculated and an initial anomaly is detected, the system immediately initiates the following in-depth diagnostic and closed-loop optimization control process, referring to... Figure 3 :

[0068] Furthermore, the process involves establishing a physical reachability matrix for design dependency paths among various design parameters of the characterization system, adjusting dynamic thresholds based on the matrix and real-time simulation conditions, and generating collaborative design constraint indices by weighted fusion of the multidimensional residual sequences using an adaptive weighting method based on physical characteristics. These indices are then compared with the dynamic thresholds to generate computer-aided design verification results. The specific implementation process includes:

[0069] During the system design phase, based on the physical topology, energy flow, and control logic of the fuel cell system, an N×N Boolean physical reachability matrix is ​​constructed, where N is the total number of all monitored key design parameters in the system. If a change in design parameter i can directly physically affect design parameter j, then the matrix element M(i,j) is set to 1 (e.g., "three-way valve opening" affects "stack outlet temperature"); if there is no direct physical influence path, it is set to 0. For example, if the measurement point of design parameter i (e.g., "stack inlet temperature") is at... If a fluid pipeline or heat conduction path is upstream of a measurement point for design parameter j (e.g., "pile outlet temperature"), then M(i,j) = 1; if design parameter i (e.g., "load current") and design parameter j (e.g., "bus voltage") are in the same electrical loop and have a direct influence relationship according to Kirchhoff's laws, then M(i,j) = 1; if design parameter i is an actuator state (e.g., "three-way valve opening"), and design parameter j is a physical quantity directly controlled by that actuator (e.g., "cooling water temperature"), then M(i,j) = 1. This matrix provides the first-principles-based physical constraints required for causal judgment.

[0070] During real-time operation, the system selects a predetermined sliding time window (e.g., containing data from the most recent 600 time steps) and performs the Granger causality test algorithm on the multidimensional residual sequences within the window. This algorithm outputs a preliminary statistical causal graph, which contains the precedence / lag relationships between all statistically significant (e.g., p-value less than 0.05) residual pairs. Then, the system performs cross-validation: it iterates through each causal edge in the statistical causal graph and queries the physical reachability matrix. Only when the corresponding element in the physical reachability matrix is ​​1 is the causal relationship considered "high confidence" and retained, forming the final causal graph. Statistically significant but physically unreachable edges are considered spurious correlations and removed.

[0071] The method for generating computer-aided design verification results includes a causal mediation analysis enhancement mechanism based on the front-door criterion:

[0072] After filtering the statistically generated causal graph through the physical reachability matrix, the causal relationship graph that has been physically verified is automatically searched for and identified to satisfy the front door criterion. A path that satisfies the front door criterion means that there is a mediating design parameter that blocks the direct path between the causal design parameter and the result design parameter, and there is no backdoor path from the causal design parameter to the mediating design parameter, nor is there a backdoor path from the mediating design parameter to the result design parameter.

[0073] The system verifies the authenticity of the causal chain by quantitatively calculating the causal effects transmitted through the mediating design parameters and comparing them with the total effect reflected by the performance deviation residual sequence, thereby filtering out spurious causal associations caused by unobserved confounding factors.

[0074] The causal mediation enhancement mechanism based on the front-door criterion can effectively eliminate spurious causal associations caused by unobserved variables by actively seeking and verifying "clean" causal paths free from confounding factors, thereby improving the accuracy of fault diagnosis and the ability to identify the root cause of the system.

[0075] After confirming a high-confidence causal relationship, the system adaptively adjusts the weights of different residual channels. Specifically, the system dynamically updates the weights of each residual channel based on statistical measures such as the root mean square (RMS) or variance within the recent time window; the channel with greater fluctuations has a correspondingly higher weight. Subsequently, all weighted residual information is used to calculate a "co-design constraint index" that can comprehensively reflect the overall deviation of the system through a weighted fusion algorithm.

[0076] Finally, the system compares this Co-design Constraint Index (CCI) with an adaptively changing dynamic threshold, which is adjusted in real time based on the current macroscopic operating simulation conditions of the system (such as load level and ambient temperature). When the CCI exceeds the dynamic threshold for a period of time, the system generates a structured computer-aided design verification result. The result data structure includes anomaly level, anomaly parameters, residual amplitude, and high-confidence causal chain information determined through cross-validation operations, such as: "{Anomaly level: Warning, Root cause path: [Abnormal three-way valve opening → Overheating of fuel cell outlet temperature → Voltage drop]}.

[0077] The adjustment of the dynamic threshold can also employ an adaptive dynamic threshold partitioning generation mechanism based on clustering of simulation working condition features:

[0078] From historical design simulation data, a multi-dimensional simulation condition feature vector that comprehensively describes the system's operating state is extracted. This vector includes load, load change rate, and temperature gradient. Using the K-Means clustering algorithm, unsupervised learning is performed on the historical simulation condition feature vector, and it is objectively divided into several simulation condition clusters with similar intrinsic characteristics. For each simulation condition cluster, the system calculates and stores a set of dedicated and optimal residual alarm thresholds offline. During real-time design verification, the system determines the simulation condition cluster to which the feature vector calculated based on the current real-time operating conditions belongs, and loads the complete set of thresholds dedicated to that partition as the current dynamic compliance threshold.

[0079] By matching a "tailor-made" threshold to each complex operating mode identified by data-driven analysis, the adaptive accuracy of dynamic thresholds under all simulation conditions is improved, and the false alarm and false negative rates are reduced.

[0080] This application embodiment establishes a diagnostic mechanism that combines data-driven statistical causality with physical prior knowledge for dual verification. This generates a high-confidence diagnostic conclusion that can distinguish between true causality and false correlation. It solves the problem that traditional diagnostic methods rely solely on statistics or models, resulting in inaccurate diagnosis and poor interpretability. This improves the accuracy of identifying the root cause of the fault.

[0081] Furthermore, the computer-aided design verification results are input into a preset residual-control mapping rule, and design parameter optimization instructions are output to optimize the design parameters. The specific implementation process includes:

[0082] The system has a built-in preset mapping rule base, which is a decision tree model trained with simulation data or a rule set built based on expert experience. It can be implemented in the following ways:

[0083] Method 1: Based on decision tree model

[0084] Decision tree models are trained on a large amount of simulation data with fault labels. A typical decision tree logic is as follows:

[0085] Root node: Determine if the "Collaborative Design Constraint Index" exceeds the dynamic threshold?

[0086] If (an anomaly occurs), proceed to branch 1: determine which residual channel has the largest contribution to the normalized amplitude.

[0087] If it is "temperature-related residual", then proceed to branch 1.1: determine whether "pile temperature gradient residual" or "outlet temperature residual" exceeds the limit.

[0088] If the "pile temperature gradient residual" is > 5℃ (leaf node): Output design parameter optimization instruction A: {"Instruction type": "progressive adjustment", "target design parameter": "three-way valve opening", "target value": "45%", "adjustment time": "1.5s", "safety check": ["pressure fluctuation < 0.02MPa / s", "flow stability"]}.

[0089] ... (other leaf nodes)

[0090] If it is "energy storage unit related residual", then proceed to branch 1.2: determine if "SOC residual" is negative and "load current residual" is positive?

[0091] If it is a leaf node: Output design parameter optimization instruction B: {"Instruction type": "Threshold update", "Target design parameter": "DC-DC compensation threshold", "Update method": "Call independent simulation model", "Constraint": "SOC > 20%"}.

[0092] ... (other leaf nodes)

[0093] If not (normal), proceed to the leaf node: no operation.

[0094] Method 2: Rule sets based on expert knowledge

[0095] This rule set is stored in IF-THEN format, as shown in the table below:

[0096] Table 1: Rule set table based on expert knowledge

[0097]

[0098] The rule base takes the structured "design verification results" (especially the root cause path information) generated above as input and outputs a set of design parameter optimization instructions for the specific failure mode.

[0099] Following the output design parameter optimization instructions, a closed-loop control instruction progressive execution and verification mechanism based on safety margin is also included:

[0100] Before executing the design parameter optimization instruction, a safety margin pre-verification is performed. The pre-verification includes: calculating the expected values ​​of the key safety design parameters after executing the design parameter optimization instruction.

[0101] Based on the expected value, calculate the relative position within the preset safe working range to obtain a safety margin value; compare the safety margin value with a preset safety margin threshold, and execute the design parameter optimization instruction only when the safety margin value is greater than the safety margin threshold.

[0102] For design parameter optimization instructions with an adjustment range greater than the preset value, the mechanism automatically decomposes this design parameter optimization instruction into multiple sub-instructions with smaller adjustment ranges, forming an adjustment sequence, and performs it step by step in a progressive manner;

[0103] After executing each sub-instruction in the adjustment sequence, the mechanism pauses the execution of subsequent sub-instructions and performs a computer-aided design verification again. After confirming that the system state is stable, it continues to execute the next sub-instruction in the adjustment sequence.

[0104] The progressive execution and verification mechanism of closed-loop control commands based on safety margin adds an independent "safety gatekeeper" and "circuit breaker mechanism" to the execution link of automatic control, thereby improving the reliability and process safety of the entire automation system.

[0105] Within the aforementioned safety framework, the system executes corresponding control strategies based on different diagnostic conclusions. For example, when the diagnostic conclusion points to a clear, single actuator misadjustment (such as "abnormal three-way valve opening"), the mapping rule generates a "progressive adjustment sequence." Its execution process fully follows the step-by-step execution and inter-step verification logic of this safety mechanism, such as fine-tuning the opening every 200 milliseconds and simultaneously verifying safety constraints such as "whether the cooling pipeline pressure change rate is less than 0.01 MPa / s" in each interval. When the diagnostic conclusion points to a complex fault caused by the coupling of multiple factors (such as "the energy storage unit's SOC is too low and the load current residual fluctuates violently at the same time"), the mapping rule generates a collaborative control strategy. Each specific instruction in this strategy also needs to pass a safety margin pre-verification before execution. Furthermore, the "state design parameter safety window" included in this strategy (such as constraining the SOC to be maintained within the range of 20% to 90%) can also be regarded as a specific implementation of this safety margin verification mechanism.

[0106] When a specific value corresponding to the fuel cell temperature gradient data in the multi-source design simulation data tensor indicates that the residual exceeds the limit (e.g., the temperature gradient is greater than 5°C) in the computer-aided design verification results, the design parameter optimization command generates a "progressive three-way valve opening adjustment sequence," specifically: within 1.5 seconds, the three-way valve opening is linearly adjusted from the current 30% to 45%. During the adjustment, the controller synchronously verifies that the pressure fluctuation of the cooling water circuit is less than 0.02 MPa / s and the flow stability fluctuation is less than 5%. Only after these conditions are met is the final switching of the cooling water circuit topology triggered.

[0107] When the residuals related to the state of the energy storage unit and the residuals of the load current fluctuate drastically at the same time, while the residuals of the temperature remain within the normal range, the system calls an independent energy storage unit state simulation model to update the compensation threshold of the DC-DC converter. When the controller executes this update, it can also add an additional state design parameter safety window constraint (e.g., maintaining SOC between 20% and 90%) to enhance safety. Finally, after all design parameter optimization instructions are executed, the system will automatically trigger a graded abnormal warning (e.g., divided into "information", "warning", and "serious" levels) based on the over-limit magnitude and rate of change of the initial residuals and the severity level of the diagnosed fault mode. This warning will be displayed to the operator in a visual manner (e.g., highlighting the fault path) through the UAV control interface or remote monitoring platform.

[0108] This application embodiment constructs a closed-loop control paradigm with safety constraints, from high-confidence diagnosis to hierarchical control, resulting in an intelligent adjustment system capable of autonomously optimizing operating design parameters. This solves the problem of traditional detection and control modules being isolated from each other, leading to the inability to adaptively adjust when the system malfunctions, thus achieving a stable improvement in the system's operational robustness and overall lifecycle efficiency.

[0109] While the invention has been described in detail herein with reference to specific embodiments, those skilled in the art should recognize that these descriptions are illustrative only and not restrictive. Various adjustments, combinations, or equivalent substitutions can be made to the disclosed embodiments without departing from the core principles disclosed herein, and all such modifications should be considered to be included within the scope of the invention. The final scope of protection of the invention should be determined by the appended claims and their legal equivalents.

Claims

1. A method for detecting anomalies in time-series data of a fuel cell system based on residual analysis, characterized in that, include: Acquire multi-source design simulation data and construct it as a multi-source design simulation data tensor; The multi-source design simulation data includes: multi-point temperature data, operating voltage, load current, energy storage operating parameters, and design control parameters. The acquired multi-source design simulation data is converted into structured time-series data by parsing timestamps, identifiers, and data fields. Continuous data is constructed into tensors according to time step, variable dimension, and spatial dimension, while discrete event data is stored separately as event sequences. The multi-source design simulation data tensor is obtained by associating timestamps with the tensors. A machine learning-based computer-aided design model is constructed. Historical design simulation data is acquired, and tensor decomposition is used to process the historical design simulation data and construct an initial spatiotemporal tensor. Based on the initial spatiotemporal tensor, a neural network algorithm is used to train the machine learning-based computer-aided design model. During the design verification process, an incremental tensor decomposition method satisfying sparsity constraints is used to perform low-rank updates on the initial tensor model. Combined with multidimensional residual sequence feedback information, the computer-aided design model is subjected to online learning and adaptive updates to improve the simulation accuracy under dynamically changing simulation conditions. The model update frequency is adaptively adjusted according to the residual change amplitude, increasing the update frequency of the computer-aided design model under dynamic simulation conditions and decreasing the update frequency under steady-state simulation conditions. Behavioral simulation is performed on the multi-source design simulation data tensor, and simulated design behavior values ​​representing the design specification behavior are output. The multi-source design simulation data is compared with the simulated design behavior values ​​as simulation observations to calculate a multidimensional residual sequence. The computer-aided design model is then subjected to online adaptive updates using the multidimensional residual sequence. A physical reachability matrix is ​​established to characterize the design dependency paths between various design parameters. The dynamic threshold is adjusted based on the matrix and the real-time simulation conditions. The multidimensional residual sequence is weighted and fused using a physical characteristic adaptive weighting method to generate a collaborative design constraint index. This index is then compared with the dynamic threshold to generate computer-aided design verification results. The computer-aided design verification results are input into a preset residual-control mapping rule, and design parameter optimization instructions are output to optimize the design parameters.

2. The method for detecting time-series data anomalies in a fuel cell system based on residual analysis according to claim 1, characterized in that, The process of simulating the behavior of multi-source design simulation data tensors and outputting simulated design behavior values ​​that characterize the behavior of design specifications includes: The multi-source design simulation data tensor is input into the computer-aided design model, and forward reasoning is performed on each time step of the multi-source design simulation data tensor. Based on the historical sequence of the multi-source design simulation data, the simulation of the state of the next time step is output. The simulation output constitutes a simulation tensor that is consistent with the input tensor structure in terms of variables and spatial dimensions, which represents the simulated design behavior value of the ideal behavior state that conforms to the specifications under the current simulation conditions.

3. The method for detecting time-series data anomalies in a fuel cell system based on residual analysis according to claim 1, characterized in that, The multi-source design simulation data is compared with the simulated design behavior values ​​as simulation observations to calculate a multidimensional residual sequence, including: The multi-source design simulation data is used as the simulation observation value and the simulated design behavior value. The difference is calculated element by element to obtain the initial residual tensor with the same structure as the previous two. The initial residual tensor is grouped according to the physical characteristics and dimensional type of the design parameters to obtain a multidimensional residual sequence. The multidimensional residual sequence is a quantitative index of the degree of deviation between the actual behavior and the design specification behavior.

4. The method for detecting time-series data anomalies in a fuel cell system based on residual analysis according to claim 1, characterized in that, The process involves establishing a physical reachability matrix characterizing the design dependency paths between various design parameters, adjusting dynamic thresholds based on the matrix and real-time simulation conditions, and generating collaborative design constraint indices by weighting and fusing the multidimensional residual sequences using an adaptive weighting method based on physical characteristics. These indices are then compared with the dynamic thresholds to generate computer-aided design verification results, including: A system-designed accessibility matrix is ​​established to clearly mark the design dependency paths between various design parameters. The statistically generated causal graph is used as a preliminary reference causal relationship, and the causal edge relationships with design dependency paths are retained through screening and weighting by the physical accessibility matrix. A dynamic graph update mechanism is designed to dynamically adjust the graph update frequency according to the simulation process: the update frequency is automatically increased when the rate of change of relevant data increases, and the update frequency is decreased under steady-state simulation conditions. Through a physical characteristic adaptive weighting method, relevant residuals are assigned weights according to system characteristics, and weighted fusion is used to generate collaborative design constraint indicators. An adaptive dynamic threshold is set based on the real-time simulation conditions, and computer-aided design verification results are generated based on the comparison results between the indicators and the thresholds.

5. The method for detecting time-series data anomalies in a fuel cell system based on residual analysis according to claim 1, characterized in that, The computer-aided design verification results are input into a preset residual-control mapping rule, and design parameter optimization instructions are output to optimize the design parameters, including: The calculation of the numerical value and rate of change of the computer-aided design verification result exceeding the dynamic threshold is performed. The numerical value and rate of change exceeding the threshold are input into the preset residual-control mapping rule, and the design parameter optimization instruction is output. When the residual corresponding to the data in the multi-source design simulation data tensor of the computer-aided design verification result exceeds the limit, the design parameter optimization instruction generates a progressive adjustment sequence. The controller executes the design parameter update in segments and triggers topology adjustment after verifying the stability condition synchronously. When the energy storage-related residual and the system state residual exceed the preset threshold, while other residuals remain within the normal range, the preset physical characteristic model is called to update the compensation threshold, and the state parameter safety window is constrained in the controller. After all design parameter optimization instructions are executed, a graded anomaly warning is triggered. The warning level is positively correlated with the magnitude and rate of change of the residual exceeding the limit.

Images