Fuel cell system time sequence data anomaly detection method based on residual analysis

Abstract

The invention relates to the technical field of fuel cell system anomaly detection, in particular to a fuel cell system time series data anomaly detection method based on residual analysis, and the method comprises the specific implementation steps: obtaining multi-source design simulation data of a system, and constructing a multi-source design simulation data tensor; constructing a computer aided design model, performing behavior simulation on the simulation data tensor, and outputting a simulation design behavior value; the simulation data is used as a simulation observation value to be compared with a simulation design behavior value, a multi-dimensional residual sequence is obtained through calculation, and online self-adaptive updating is conducted on the design model; establishing a physical reachability matrix, and adjusting a dynamic threshold value based on the matrix and a real-time simulation working condition; performing weighted fusion on the multi-dimensional residual sequence to generate a collaborative design constraint index, and comparing the collaborative design constraint index with the dynamic threshold to generate a computer aided design verification result; and inputting the design verification result into a preset residual-control mapping rule, and outputting a design parameter optimization instruction to optimize the design parameters.

Description

Technical Field

[0001] The present invention relates to the technical field of fuel cell system anomaly detection, and in particular to a fuel cell system time series data anomaly detection method based on residual analysis. Background Art

[0002] Building highly sensitive anomaly detection capabilities in complex systems like fuel cells is a key component in ensuring their performance and safety throughout their lifecycle. Existing technologies primarily provide a solid foundation for this by integrating analytical model-based residual analysis with data-driven approaches. Analytical model-based residual analysis, for example, establishes mathematical or physical models of the system to simulate and verify typical failure modes such as flooding and drying, providing a clear physical basis for the design of diagnostic logic. Meanwhile, data-driven approaches, such as machine learning, allow designers to leverage 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. The combined application of these techniques, particularly residual analysis as a bridge between model simulation and real-world observations, provides a powerful tool for evaluating whether a system design meets expected performance indicators.

[0003] In current system design and verification practices, the effectiveness of using residual analysis to build 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 the fuel cell under dynamic simulation conditions, multivariable coupling, and aging degradation, which in turn causes deviations between the model and the physical entity. This results in a low signal-to-noise ratio of the residuals generated in the simulation environment, making it difficult to accurately reflect the signals that indicate early faults, affecting the reliability of the detection logic design. On the other hand, existing methods, including data-driven methods, generally have high computational complexity and are difficult to meet the real-time requirements of the system. In addition, the detection module and the control execution module are separated from each other, 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 the diagnostic function of the fuel cell system, there are still obvious deficiencies in building high-fidelity simulation models, generating high-quality residuals, and verifying and optimizing the system's ability to detect early weak anomalies. To this end, the present invention proposes a fuel cell system time series data anomaly detection method based on residual analysis. Summary of the Invention

[0005] The purpose of the present invention is to provide a fuel cell system time series data anomaly detection method based on residual analysis, so as to achieve accurate detection of early and weak anomalies of the fuel cell system and active and safe online optimization and adjustment.

[0006] To achieve the above object, the present invention provides the following technical solutions: A method for detecting abnormality in time series data of a fuel cell system based on residual analysis, comprising: Acquire multi-source design simulation data and construct it into a multi-source design simulation data tensor; A computer-aided design model based on machine learning is constructed to simulate the behavior of multi-source design simulation data tensors and output simulated design behavior values ​​representing the design specification behavior; the multi-source design simulation data is used as simulation observation values ​​to compare with the simulated design behavior values ​​to calculate a multi-dimensional residual sequence; and the computer-aided design model is adaptively updated online using the multi-dimensional residual sequence; Establishing a physical reachability matrix that characterizes the design dependency paths between various design parameters, and adjusting the dynamic threshold based on the matrix and real-time simulation conditions; weighting and fusing the multidimensional residual sequence using a physical property adaptive weighting method to generate collaborative design constraint indicators, and comparing them with the dynamic threshold to generate computer-aided design verification results; The computer-aided design verification result is input into a preset residual-control mapping rule, and a design parameter optimization instruction is output to optimize the design parameters.

[0007] Preferably, the acquiring of multi-source design simulation data and constructing the multi-source design simulation data tensor includes: 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. Among them, continuous data is constructed as tensors according to time steps, variable dimensions, and spatial dimensions, and discrete event data is stored separately as event sequences. The multi-source design simulation data tensor is obtained by associating the tensors with the timestamps.

[0008] Preferably, the constructing of a computer-aided design model based on machine learning includes: Historical design simulation data is obtained, and the tensor decomposition method is used to process the historical design simulation data and construct an initial space-time tensor. Based on the initial space-time tensor, a neural network algorithm is used to train a computer-aided design model based on machine learning. During the design verification process, an incremental tensor decomposition method that satisfies sparsity constraints is used to perform low-rank updates on the initial tensor model. Combined with multi-dimensional 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, which increases the computer-aided design model update frequency under dynamic simulation conditions and reduces the computer-aided design model update frequency under steady-state simulation conditions.

[0009] Preferably, the performing behavioral simulation on the multi-source design simulation data tensor and outputting a simulated design behavior value representing the design specification behavior includes: The multi-source design simulation data tensor is input into a 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, a simulation of the state of the next time step is output. The output result of the simulation constitutes a simulation tensor that is consistent with the input tensor structure in both variable and spatial dimensions, that is, it represents the simulated design behavior value of the ideal behavior state that meets the specifications under the current simulation working conditions.

[0010] Preferably, the multi-source design simulation data is used as simulation observation values ​​and compared with the simulated design behavior values ​​to calculate a multi-dimensional residual sequence, including: The multi-source design simulation data are used as simulation observation values ​​and simulated design behavior values ​​to perform element-by-element difference calculation to obtain an 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, which is a quantitative indicator of the degree of deviation between the actual behavior and the design specification behavior.

[0011] Preferably, the step of establishing a physical reachability matrix representing the design dependency paths between the design parameters, adjusting the dynamic threshold based on the matrix and the real-time simulation working condition, and weightedly fusing the multidimensional residual sequence using a physical property adaptive weighting method to generate a collaborative design constraint index, comparing the index with the dynamic threshold, and generating a computer-aided design verification result includes: Establish a reachability matrix based on the system design architecture, and clearly mark the design dependency paths between each design parameter; use the statistically generated causal graph as a preliminary reference for causal relationships, and screen and weight them through the physical reachability matrix to retain causal edge relationships with design dependency paths; design a dynamic graph update mechanism to dynamically adjust the graph update frequency according to the simulation process: automatically increase the update frequency when the change rate of related data increases, and reduce the update frequency under steady-state simulation conditions; through the physical property adaptive weighting method, the relevant residuals are weighted according to the system characteristics, and the weighted fusion generates collaborative design constraint indicators. The dynamic threshold of adaptive change is set based on the real-time simulation conditions, and the computer-aided design verification results are generated based on the comparison results of the indicators and thresholds.

[0012] Preferably, the computer-aided design verification result is input into a preset residual-control mapping rule, and a design parameter optimization instruction is output to optimize the design parameters, including: The numerical value and the rate of change of the computer-aided design verification result exceeding the dynamic threshold are calculated, the numerical value and the rate of change exceeding the threshold are input into a preset residual-control mapping rule, and a design parameter optimization instruction is output; when the residual corresponding to the computer-aided design verification result and the data in the multi-source design simulation data tensor exceeds the limit, the design parameter optimization instruction generates a progressive adjustment sequence, the controller executes the design parameter update in segments, and triggers the topology adjustment after synchronously verifying the stability condition; 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 property 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 abnormality warning is triggered, and the warning level is positively correlated with the residual excess amplitude and change rate.

[0013] Compared with the prior art, the present invention has the following beneficial effects: 1. The present 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 in the existing technology that simplified physical models are difficult to accurately reproduce dynamic, coupling, and aging characteristics. This enables the computer-aided design model to more accurately reflect the true behavior of the system under complex and changeable simulation conditions, thereby improving the signal-to-noise ratio and reliability of subsequent residual analysis.

[0014] 2. By establishing a Granger causality graph based on physical reachability matrix verification and adaptively weighting it in combination with physical characteristics, the present invention can effectively distinguish between real physical causal chains and statistical pseudo-correlations, solving the problem that existing methods are difficult to identify early and weak fault signals from a strong noise background, and improving the system's predictability and sensitivity to potential risks.

[0015] 3. The present invention directly converts the design verification results into specific, hierarchical design parameter optimization instructions through preset residual-control mapping rules, overcoming the defect of the existing technology that the detection module and the control execution module are separated from each other, and constructs a complete closed loop from deviation detection to adaptive adjustment of design parameters, thereby improving the system's robustness, operating efficiency and safety throughout its life cycle.

[0016] 4. The control logic of the present invention is based on comprehensive mode judgment of multiple residual channels and combines synchronous verification of safety boundary conditions such as pressure and flow. Before adjusting the DC-DC converter compensation threshold, an independent simulation model will be called and the safety window will be constrained, avoiding misjudgment and unsafe control actions that may be caused by abnormalities in a single indicator, ensuring that the decision-making process for design parameter optimization is more accurate, reliable and safe. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1This is an overall flow chart of the fuel cell system time series data anomaly detection method based on residual analysis of the present invention; Figure 2 A schematic diagram of multi-source design simulation data tensor construction and residual generation for an example of the present invention; Figure 3 Schematic diagram of the closed-loop control decision process based on residual in an example of the present invention. DETAILED DESCRIPTION

[0018] In order to make the objectives, technical solutions and advantages of the present invention more clear, 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 only used to explain the present invention and do not constitute a limitation to the present invention. Based on the contents disclosed in this specification, other embodiments that can be obtained by those skilled in the art without paying any creative work are all within the scope of protection required by the present invention.

[0019] Example 1: This embodiment provides a method for detecting abnormalities in fuel cell system time series data based on residual analysis, which is applied to a UAV equipped with a fuel cell system. This method aims to generate an early identification of potential abnormalities in the system and proactively optimize design parameters while ensuring flight safety and meeting power requirements for the upcoming flight mission cycle. Figure 1 The specific implementation steps of the method of the present invention in this scenario are as follows: S1. Acquire multi-source design simulation data of the system and construct it into a multi-source design simulation data tensor; S2. Construct a computer-aided design model based on machine learning, perform behavioral simulation on the multi-source design simulation data tensor, and output a simulated design behavior value that characterizes the system design specification behavior; S3. Compare the multi-source design simulation data with the simulated design behavior value as a simulation observation value to calculate a multi-dimensional residual sequence; S4. Use the multi-dimensional residual sequence to perform online adaptive update on the computer-aided design model; S5. Establish a physical reachability matrix that characterizes the design dependency path between each design parameter of the system, and adjust the dynamic threshold based on the matrix and the real-time simulation working condition; Through the physical property adaptive weighting method, weightedly 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.

[0020] Furthermore, the multi-source design simulation data of the system is obtained and constructed as a multi-source design simulation data tensor, referring to Figure 2 , corresponding to the above step S1, the specific implementation process includes: 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 obtains multi-source design simulation data of the fuel cell system, which includes: the fuel cell stack inlet temperature (range 50-80°C), outlet temperature (range 60-90°C) and key area temperature (10 key points, range 55-85°C) obtained 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°C) monitored by the data acquisition unit; and the three-way valve opening (range 0%-100%) and cooling water path topology structure status (on / off state) recorded by the controller.

[0021] Multi-source design simulation data is acquired at a frequency of 1 Hz. The acquired multi-source design simulation data is converted into structured time series data by parsing the timestamps accurate to milliseconds, sensor IDs, and fields. For a 2-hour inspection task, continuous data such as temperature and voltage are constructed into a 7200×7×10 three-dimensional tensor based on the time step (7200 seconds), variable dimension (7 variables: inlet temperature, outlet temperature, regional temperature, operating voltage, load current, SOC, battery temperature), and spatial dimension (10 stack positions). At the same time, discrete event data such as cooling water path topology switching are stored separately as event sequences and associated with the tensor through timestamps to obtain a multi-source design simulation data tensor that can fully describe the system's spatiotemporal state.

[0022] 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 a single data perspective or time asynchrony in traditional methods, thereby enhancing the input signal-to-noise ratio of subsequent anomaly detection.

[0023] Furthermore, the construction of a computer-aided design model based on machine learning performs behavioral simulation on multi-source design simulation data tensors and outputs simulated design behavior values ​​representing the system design specification behavior, corresponding to the above-mentioned step S2. The specific implementation process includes: Six months of historical design simulation data of high-altitude inspections were obtained. The CP tensor decomposition method was used to process this data and construct an initial space-time tensor of 7200×7×10. Based on this initial space-time tensor, a neural network algorithm with 128 hidden units was used for training to learn the dynamic behavior of the system under normal operating mode. The training process was iterated for 100 cycles with a learning rate set to 0.001. Finally, a machine learning computer-aided design model capable of predicting the future state of the system was obtained.

[0024] The data-driven computer-aided design model is a neural network consisting of two long short-term memory (LSTM) layers and a fully connected output layer. The model inputs a tensor of multi-source design simulation data from the past 60 time steps and outputs a tensor predicting the system state at the next time step (one second later). The first LSTM layer contains 128 hidden units, uses the tanh activation function, and returns a sequence of outputs. The second LSTM layer also contains 128 hidden units, uses the tanh activation function, and returns only the output of the last time step. A fully connected layer is then connected to the model, mapping the LSTM output to the final prediction dimension (7 × 10 = 70 nodes). Training is performed using the Adam optimizer with a mean squared error loss function. The training process is repeated for 100 cycles, with an initial learning rate of 0.001.

[0025] During system operation, incremental tensor decomposition methods that satisfy sparsity constraints, such as online robust principal component analysis, are employed. 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 a low-rank correction on the weight matrix of the final layer (fully connected layer) of the neural network. This allows for rapid adaptation to minor changes in system characteristics without retraining the entire model. In combination with real-time multidimensional residual sequence feedback, the design model is updated online with low rank information. 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 every 60 seconds; under steady-state simulation conditions, it is reduced to every 600 seconds.

[0026] This embodiment establishes a data-driven design model based on historical data training to produce a dynamic simulation benchmark that can accurately reproduce the complex nonlinear behavior of the system. This solves the problem of traditional analytical models with low simulation fidelity and inability to adapt to changing simulation conditions due to oversimplification, thereby ensuring the accuracy of subsequent residual calculations.

[0027] Real-time multi-source design simulation data tensors are input into a trained computer-aided design model. The model runs on the controller, performs forward reasoning for each time step (1 second), and outputs a simulation tensor for the system state at the next time step based on the historical sequence of the past 7200 seconds. 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 and standardized behavior state of the system under high-altitude inspection simulation conditions, that is, the simulated design behavior value. A single complete reasoning process takes less than 0.1 second.

[0028] This embodiment uses the design model for real-time online simulation to generate 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 the inability of traditional detection methods based on static or segmented thresholds to track the highly dynamic characteristics of the system, thereby improving the applicability and robustness of anomaly detection.

[0029] Furthermore, the multi-source design simulation data is used as simulation observation value and compared with the simulated design behavior value to calculate the multi-dimensional residual sequence. Figure 2 , corresponding to the above S3 step, the specific implementation process includes: The real-time multi-source design simulation data tensor is used as the simulation observation value, and the simulated design behavior value output by the model is subtracted element by element to generate an initial residual tensor of 7200×7×10. The tensor is then grouped: For the temperature-related residual groups of 10 locations with spatial distribution characteristics, the principal component analysis (PCA) method was used to extract the first three principal components as the core change pattern characteristics (the cumulative contribution rate was required to be greater than 95%); For the voltage and current related residual group that represents the overall state of the system, calculate its standardized statistical characteristics (mean and standard deviation); For the state-related residual group of the energy storage unit, the residual deviation is calculated by combining it with the offline design parameter simulation model. For the state-related residual group 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 the extended Kalman filter (EKF). The core of the 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 SOC and the voltage of the two-stage RC network) and the input current at the current moment through the preset state transfer matrix and input matrix; the prior state estimate is substituted into the observation equation to obtain the theoretical terminal voltage, and the theoretical terminal voltage is compared with the actual measured terminal voltage. The prior state estimate is corrected in combination with 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 of the state-space model are pre-identified and determined by battery factory data and experimental calibration. Its process noise covariance and observation noise covariance are also pre-set according to 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 subtracts it from the simulated observation value to obtain the residual deviation.

[0030] Finally, the three types of feature processing results are output in parallel as independent channels to form the final multi-dimensional residual sequence. The entire calculation process is executed on the controller, and the single calculation takes less than 0.2 seconds.

[0031] This embodiment calculates the difference between the actual value and the high-fidelity simulation value, and groups and extracts features from the resulting residuals, thereby generating a fault feature sequence with greater physical significance and a higher signal-to-noise ratio. This solves the problem that weak anomalies in the original measurement signal are easily drowned out by strong noise and difficult to use directly for diagnosis, thereby enhancing the detection sensitivity of early-stage, weak faults.

[0032] Furthermore, the multidimensional residual sequence is used to perform online adaptive updating on the computer-aided design model, corresponding to the above step S4. The specific implementation process includes: The calculated multidimensional residual sequence is passed to the update module of the computer-aided design model as a real-time, high-value feedback input signal. 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 the key weight parameters of the model are fine-tuned, rather than global gradient backpropagation that consumes a lot of computing resources.

[0033] In order to further improve the intelligence level and resource utilization efficiency of the system, the execution frequency of the model update is adaptively adjusted according to the change amplitude of the multidimensional residual sequence within a preset time window. Specifically, when the moving average or volatility of the residual increases significantly in a short period of time, the system determines that the current operating conditions have changed dramatically and will automatically increase the update frequency of the model; conversely, when the system is in steady-state operation and the residual changes are smooth, the update frequency will be automatically reduced.

[0034] Furthermore, the physical reachability matrix characterizing the design dependency paths between the design parameters of the system is established, and the dynamic threshold is adjusted based on the matrix and the real-time simulation working condition; the multi-dimensional residual sequence is weighted and fused by a physical property adaptive weighting method to generate a collaborative design constraint index, which is compared with the dynamic threshold to generate a computer-aided design verification result. Corresponding to the above step S5, the specific implementation process includes: Establish an N×N physical reachability matrix, where N is the total number of system design parameters. The matrix element (i, j) is 1, indicating that the change of design parameter i can directly physically affect design parameter j, otherwise it is 0. The statistically generated Granger causality graph (screening condition: p-value less than 0.05) is screened and weighted with the physical reachability matrix: if the Granger causality graph shows that i has a statistical causal effect on j, and the reachability matrix (i, j) is 1, then the causal relationship is retained; if the reachability matrix (i, j) is 0, then the weight of the statistical causal edge is set to a lower value, such as 0.05.

[0035] A weight is assigned to each residual channel: the weight range of temperature-related residuals is 0.4-0.6, the weight range of voltage and current-related residuals is 0.3-0.5, and the weight of the energy storage unit state-related residuals is adaptively adjusted based on the cumulative statistic of the root mean square value of the SOC residuals in the past 10 flight hours. By weighted fusion of all information, a collaborative design constraint indicator is generated and compared with a dynamic threshold set based on real-time simulation conditions (for example, set to 0.5 standard deviations from the normal value) to generate a computer-aided design verification result.

[0036] Furthermore, the computer-aided design verification result is input into a preset residual-control mapping rule, and a design parameter optimization instruction is output to optimize the design parameters. Corresponding to the above step S6, the specific implementation process includes: The numerical value and the rate of change of the computer-aided design verification result exceeding the dynamic threshold are calculated, and this information is input into a preset residual-control mapping rule. In this embodiment, the mapping rule is specifically implemented as a decision tree model obtained through offline simulation training, whose input features are the normalized amplitude and first-order difference value (representing the rate of change) of each residual channel, and its leaf node directly outputs a preset design parameter optimization instruction sequence for the specific residual pattern, such as a structure containing a key-value pair, where the key is the design control parameter name (such as "target opening") and the value is the specific control value (such as "45%").

[0037] When the computer-aided design verification results indicate that a specific value corresponding to the stack temperature gradient data in the multi-source design simulation data tensor indicates an excessive residual (for example, a temperature gradient greater than 5°C), the design parameter optimization instruction 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 this adjustment, the controller simultaneously 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 when these conditions are met will the cooling water circuit topology switch be finally triggered.

[0038] When the energy storage unit state-related residuals and the load current residuals fluctuate violently at the same time, while the temperature residuals remain within the normal range, the system calls an independent energy storage unit state simulation model to update the DC-DC converter compensation threshold. When executing this update, the controller can also additionally superimpose a state parameter safety window constraint (for example, the SOC is maintained between 20% and 90%) to enhance safety. After the execution of all design parameter optimization instructions, a graded abnormality warning is triggered. The warning level is positively correlated with the residual excess amplitude and change rate, and is visualized through the drone control interface.

[0039] This implementation solves the problem of the system being unable to effectively detect early and subtle deviations from system behavior due to the deviation between the simulation model and the physical entity and the low signal-to-noise ratio by constructing a complete technical solution from data acquisition, high-fidelity online adaptive modeling, multi-dimensional residual feature analysis, to the final closed-loop adaptive control based on physical causal fusion. It also solves the functional defect of the system being unable to autonomously optimize according to the real-time status due to the separation between the detection and analysis module and the control execution module. Therefore, it can detect the system's deviation from its normative behavior earlier and more accurately, and enable the system to have the ability to autonomously adjust the operating design parameters according to the real-time status, so as to always maintain an efficient, safe and stable operating state under changing simulation conditions, thereby improving the reliability of the system throughout its life cycle.

[0040] Example 2: This embodiment targets a UAV equipped with a fuel cell system and detects abnormal voltage in the fuel cell system. The focus is on performing deep causal analysis on the residual sequence to achieve diagnosis and generate the specific process of closed-loop control instructions.

[0041] 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; then it uses the model to simulate the behavior of real-time data to obtain the ideal simulated design behavior value; finally, it compares the simulated observation value with the simulation value to calculate and extract an information-rich multi-dimensional residual sequence.

[0042] When the multidimensional residual sequence is calculated and the initial anomaly is detected, the system immediately starts the following in-depth diagnosis and closed-loop optimization control process, refer to Figure 3 : Furthermore, the physical reachability matrix representing the design dependency paths between the design parameters of the system is established, and the dynamic threshold is adjusted based on the matrix and the real-time simulation working condition; the multi-dimensional residual sequence is weighted and fused using a physical property adaptive weighting method to generate a collaborative design constraint index, which is compared with the dynamic threshold to generate a computer-aided design verification result. The specific implementation process includes: During the system design phase, based on the physical topology, energy flow, and control logic of the fuel cell system, an N×N-dimensional Boolean physical reachability matrix is ​​constructed, where N is the total number of all monitored key design parameters in the system. If the change of design parameter i can directly physically affect design parameter j, the matrix element M(i, j) is set to 1 (for example, "three-way valve opening" affects "stack outlet temperature"); if there is no direct physical impact path, it is set to 0, for example: if the measurement point of design parameter i (such as "stack inlet temperature") is If the fluid pipeline or heat conduction path is upstream of the measurement point of design parameter j (such as "stack outlet temperature"), then M(i, j) = 1. If design parameter i (such as "load current") and design parameter j (such as "bus voltage") are in the same electrical circuit and have a direct influence according to Kirchhoff's law, then M(i, j) = 1. If design parameter i is an actuator state (such as "three-way valve opening") and design parameter j is a physical quantity directly controlled by the actuator (such as "cooling water line temperature"), then M(i, j) = 1. This matrix provides the physical constraints based on first principles required for causal judgment.

[0043] During real-time operation, the system selects a predetermined sliding time window (for example, containing data from the most recent 600 time steps) and applies the Granger causality test algorithm to the multidimensional residual sequence within the window. This algorithm outputs a preliminary statistical causal graph containing all leading / lagging relationships between pairs of residuals that are statistically significant (for example, with a p-value less than 0.05). The system then performs a cross-validation operation: it iterates over each causal edge in the statistical causal graph and queries the physical accessibility matrix. Only when the corresponding element in the physical accessibility matrix for that edge is 1 is the causal relationship determined to be "high confidence" and retained, forming the final causal relationship graph. Statistically significant but physically unreachable edges are considered spurious correlations and are removed.

[0044] The method for generating computer-aided design verification results includes a causal mediation analysis enhancement mechanism based on a front-door criterion: After filtering the statistically generated causal graph using the physical reachability matrix, automatically searching and identifying a causal path that satisfies a front-door criterion in the physically verified causal relationship graph; a path that satisfies the front-door criterion means that there is an intermediary design parameter that blocks a direct path between a cause design parameter and a result design parameter, and there is no back-door path from the cause design parameter to the intermediary design parameter, nor is there a back-door path from the intermediary design parameter to the result design parameter; The system verifies the authenticity of the causal chain by quantitatively calculating the causal effect transmitted through the mediating design parameters and comparing it with the total effect reflected by the performance deviation residual sequence, thereby filtering out false causal associations caused by unobserved confounding factors.

[0045] The causal mediation analysis enhancement mechanism based on the front-door criterion can effectively eliminate false causal relationships caused by unobserved variables by actively searching for and verifying "clean" causal paths without confounding factors, thereby improving the accuracy of fault diagnosis and the ability to identify the root causes of the system.

[0046] After confirming a causal relationship with high confidence, the system adaptively adjusts the weights of different residual channels. Specifically, the system dynamically updates the weight of each residual channel based on statistics such as the root mean square value (RMS) or variance in a recent time window. The greater the fluctuation, the higher the weight of the channel. Subsequently, all weighted residual information is calculated through a weighted fusion algorithm to produce a "collaborative design constraint index" that can comprehensively reflect the overall deviation degree of the system.

[0047] Finally, the system compares this collaborative design constraint indicator CCI with an adaptively changing dynamic threshold, which is adjusted in real time based on the current system's macro-operation simulation conditions (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 contains the anomaly level, anomaly parameters, residual amplitude, and high-confidence causal chain information determined through cross-validation operations, for example: "{Anomaly level: Warning, root cause path: [abnormal three-way valve opening → excessively high stack outlet temperature → voltage drop]}".

[0048] The dynamic threshold can also be adjusted by using an adaptive dynamic threshold partition generation mechanism based on clustering of simulation working condition characteristics: A multi-dimensional simulation condition feature vector is extracted from historical design simulation data, which can comprehensively describe the system operating status. This vector includes load, load change rate and temperature gradient. The K-Means clustering algorithm is used to perform unsupervised learning on the historical simulation condition feature vectors and objectively divide them into several simulation condition clusters with similar intrinsic characteristics. For each simulation condition cluster, the system calculates and stores a set of exclusive and optimal residual alarm thresholds offline. During the real-time design verification process, the system determines the simulation condition cluster to which it belongs based on the feature vector calculated from the current real-time condition, and loads the full set of thresholds exclusive to the partition as the current dynamic compliance threshold.

[0049] By matching a "tailor-made" threshold for each complex operating mode identified by data-driven analysis, the adaptive accuracy of the dynamic threshold under all simulation conditions is improved, and the false alarm and missed alarm rates are reduced.

[0050] The embodiment of the present application establishes a diagnostic mechanism that dually confirms data-driven statistical causality and physical prior knowledge, producing 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, and improves the accuracy of identifying the root cause of the fault.

[0051] Furthermore, the computer-aided design verification result is input into a preset residual-control mapping rule, and a design parameter optimization instruction is output to optimize the design parameters. The specific implementation process includes: The system has a built-in preset mapping rule base, which is a decision tree model trained through simulation data or a rule set built based on expert experience. It can be implemented in the following ways: Method 1: Based on decision tree model The decision tree model is trained on a large amount of simulation data with fault labels. A typical decision tree logic is as follows: Root node: Determine whether the "collaborative design constraint index" exceeds the dynamic threshold? If (an exception occurs), enter branch 1: Determine which residual channel has the largest normalized amplitude contribution? If it is "temperature-related residual", then go to branch 1.1: Determine whether the "stack temperature gradient residual" exceeds the limit or the "outlet temperature residual" exceeds the limit? If the "stack temperature gradient residual" is greater than 5°C (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"]}.

[0052] ...(other leaf nodes) If it is the "energy storage unit related residual", go to branch 1.2: Determine whether the "SOC residual" is negative and the "load current residual" is positive? If yes (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%”}.

[0053] ...(other leaf nodes) If not (normal), go to the leaf node: no operation.

[0054] Method 2: Rule Set Based on Expert Knowledge The rule set is stored in the form of IF-THEN, as shown in the following table: Table 1: Rule set based on expert knowledge

[0055] The rule base takes the structured "design verification results" generated above (especially the root cause path information) as input and outputs a set of design parameter optimization instructions for this specific failure mode.

[0056] After the output design parameter optimization instruction, a closed-loop control instruction progressive execution and verification mechanism based on safety margin is also included: Before executing the design parameter optimization instruction, performing a safety margin pre-check, the pre-check comprising: calculating a value that a key safety design parameter is expected to reach after executing the design parameter optimization instruction; Calculating a relative position of the expected value within a preset safe operating range to obtain a safety margin value based on the expected value; comparing the safety margin value with a preset safety margin threshold, and executing the design parameter optimization instruction only when the safety margin value is greater than the safety margin threshold; For a design parameter optimization instruction with an adjustment range greater than a preset value, the mechanism automatically decomposes the design parameter optimization instruction into multiple sub-instructions with smaller adjustment ranges, forming an adjustment sequence, and performs progressive step-by-step execution; After executing each sub-instruction in the adjustment sequence, the mechanism suspends the execution of subsequent sub-instructions and re-performs computer-aided design verification. After confirming that the system state is stable, the mechanism continues to execute the next sub-instruction in the adjustment sequence.

[0057] The closed-loop control instruction progressive execution and verification mechanism based on safety margin adds an independent "safety gatekeeper" and "fuse mechanism" to the execution link of automatic control, improving the reliability and process safety of the entire automation system.

[0058] 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 rules generate a "progressive adjustment sequence" whose execution fully adheres to the safety mechanism's step-by-step execution and inter-step verification logic. For example, the valve opening is fine-tuned every 200 milliseconds, and safety constraints such as "whether the rate of change of cooling line pressure is less than 0.01 MPa / s" are simultaneously verified at each interval. When the diagnostic conclusion points to a complex fault caused by the coupling of multiple factors (such as "low energy storage unit SOC and severe fluctuations in load current residual"), the mapping rules generate a coordinated control strategy. Each specific instruction in this strategy also requires a safety margin pre-verification before execution. Furthermore, the "state design parameter safety window" included in this strategy (such as constraining the SOC to remain between 20% and 90%) can also be considered a specific implementation of this safety margin verification mechanism.

[0059] When the computer-aided design verification results indicate that a specific value corresponding to the stack temperature gradient data in the multi-source design simulation data tensor indicates an excessive residual (for example, a temperature gradient greater than 5°C), the design parameter optimization instruction 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 this adjustment, the controller simultaneously 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 when these conditions are met will the cooling water circuit topology switch be finally triggered.

[0060] When the energy storage unit state-related residuals and the load current residuals fluctuate violently at the same time, while the temperature residuals remain within the normal range, the system calls an independent energy storage unit state simulation model to update the DC-DC converter compensation threshold. When the controller performs this update, it can also additionally superimpose a state design parameter safety window constraint (for example, SOC is maintained between 20% and 90%) to enhance safety. Finally, after all design parameter optimization instructions are executed, the system will automatically trigger a graded abnormality warning (for example, divided into "information", "warning" and "serious") according to the over-limit amplitude and change rate of the initial residuals and the severity of the diagnosed fault mode, and display it to the operator in a visual manner (such as highlighting the fault path) through the drone control interface or remote monitoring platform.

[0061] The embodiment of the present application constructs a closed-loop control paradigm from high-confidence diagnosis to classification with safety constraints, creating an intelligent adjustment system that can autonomously optimize operating design parameters. It solves the problem of the traditional detection module and control module being separated from each other, resulting in the inability to adaptively adjust when the system is abnormal, and thus steadily improves the system's operational robustness and full life cycle efficiency.

[0062] Although the present invention has been described in detail herein through specific embodiments, those skilled in the art will recognize that these descriptions are illustrative only and not restrictive. The disclosed embodiments may be modified, combined, or replaced with equivalents in various ways without departing from the core principles disclosed herein, and all such modifications are considered within the scope of this invention. The ultimate scope of protection of this invention shall be determined by the appended claims and their legal equivalents.

Claims

1. A method for detecting abnormality in fuel cell system time series data based on residual analysis, characterized in that: include: Acquire multi-source design simulation data and construct it into a multi-source design simulation data tensor; Build a machine learning-based computer-aided design model to simulate the behavior of multi-source design simulation data tensors and output simulated design behavior values ​​that represent the design specification behavior; Comparing the multi-source design simulation data as simulation observation values ​​with the simulated design behavior values ​​to calculate a multi-dimensional residual sequence; Performing online adaptive updating on the computer-aided design model using the multidimensional residual sequence; Establishing a physical reachability matrix that characterizes the design dependency paths between design parameters, and adjusting dynamic thresholds based on the matrix and real-time simulation conditions; The multidimensional residual sequence is weighted and fused by a physical property adaptive weighting method to generate a collaborative design constraint index, and the index is compared with the dynamic threshold to generate a computer-aided design verification result; The computer-aided design verification result is input into a preset residual-control mapping rule, and a design parameter optimization instruction is output to optimize the design parameters.

2. The method for detecting anomaly in fuel cell system time series data based on residual analysis according to claim 1, characterized in that: The step of acquiring multi-source design simulation data and constructing the multi-source design simulation data tensor includes: 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. Among them, continuous data is constructed as tensors according to time steps, variable dimensions, and spatial dimensions, and discrete event data is stored separately as event sequences. The multi-source design simulation data tensor is obtained by associating the tensors with the timestamps.

3. The method for detecting anomaly in fuel cell system time series data based on residual analysis according to claim 1, characterized in that: The construction of a computer-aided design model based on machine learning includes: Historical design simulation data is obtained, and the tensor decomposition method is used to process the historical design simulation data and construct an initial space-time tensor. Based on the initial space-time tensor, a neural network algorithm is used to train a computer-aided design model based on machine learning. During the design verification process, an incremental tensor decomposition method that satisfies sparsity constraints is used to perform low-rank updates on the initial tensor model. Combined with multi-dimensional 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, which increases the computer-aided design model update frequency under dynamic simulation conditions and reduces the computer-aided design model update frequency under steady-state simulation conditions.

4. The method for detecting anomaly in fuel cell system time series data based on residual analysis according to claim 1, characterized in that: The performing behavioral simulation on the multi-source design simulation data tensor and outputting a simulated design behavior value representing the design specification behavior includes: The multi-source design simulation data tensor is input into a 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, a simulation of the state of the next time step is output. The output result of the simulation constitutes a simulation tensor that is consistent with the input tensor structure in both variable and spatial dimensions, that is, it represents the simulated design behavior value of the ideal behavior state that meets the specifications under the current simulation working conditions.

5. The method for detecting anomaly in fuel cell system time series data based on residual analysis according to claim 1, characterized in that: The multi-source design simulation data is used as simulation observation values ​​and compared with the simulated design behavior values ​​to calculate a multi-dimensional residual sequence, including: The multi-source design simulation data are used as simulation observation values ​​and simulated design behavior values ​​to perform element-by-element difference calculation to obtain an 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, which is a quantitative indicator of the degree of deviation between the actual behavior and the design specification behavior.

6. The method for detecting anomaly in fuel cell system time series data based on residual analysis according to claim 1, characterized in that: The method includes establishing a physical reachability matrix that characterizes the design dependency paths between the design parameters, adjusting the dynamic threshold based on the matrix and the real-time simulation working condition, weightedly fusing the multidimensional residual sequence using a physical property adaptive weighting method, generating a collaborative design constraint index, and comparing the index with the dynamic threshold to generate a computer-aided design verification result, including: Establish a reachability matrix based on the system design architecture, and clearly mark the design dependency paths between each design parameter; use the statistically generated causal graph as a preliminary reference for causal relationships, and screen and weight them through the physical reachability matrix to retain causal edge relationships with design dependency paths; design a dynamic graph update mechanism to dynamically adjust the graph update frequency according to the simulation process: automatically increase the update frequency when the change rate of related data increases, and reduce the update frequency under steady-state simulation conditions; through the physical property adaptive weighting method, the relevant residuals are weighted according to the system characteristics, and the weighted fusion generates collaborative design constraint indicators. The dynamic threshold of adaptive change is set based on the real-time simulation conditions, and the computer-aided design verification results are generated based on the comparison results of the indicators and thresholds.

7. The method for detecting anomaly in fuel cell system time series data based on residual analysis according to claim 1, characterized in that: Inputting the computer-aided design verification result into a preset residual-control mapping rule and outputting a design parameter optimization instruction to optimize the design parameters includes: The numerical value and the rate of change of the computer-aided design verification result exceeding the dynamic threshold are calculated, the numerical value and the rate of change exceeding the threshold are input into a preset residual-control mapping rule, and a design parameter optimization instruction is output; when the residual corresponding to the computer-aided design verification result and the data in the multi-source design simulation data tensor exceeds the limit, the design parameter optimization instruction generates a progressive adjustment sequence, the controller executes the design parameter update in segments, and triggers the topology adjustment after synchronously verifying the stability condition; 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 property 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 abnormality warning is triggered, and the warning level is positively correlated with the residual excess amplitude and change rate.

Images