Actuator system hierarchical fault diagnosis method, storage medium and device

Abstract

The application discloses an actuator system hierarchical fault diagnosis method, a storage medium and equipment, belongs to the technical field of actuator fault diagnosis, obtains multi-channel time sequence data and carries out data preprocessing; based on a dynamic differential threshold model, steady-state detection and diagnosis of multi-channel data are carried out, when detecting that a steady-state threshold exceeds a value and an index is not empty, parameter information of the steady-state threshold exceeding the value and corresponding fault modes are output; when detecting that the steady-state threshold exceeds the value and the index is empty, a fault diagnosis process is carried out based on an LSTM prediction model library. The application effectively integrates the advantages of threshold detection and prediction models, can simultaneously locate fault parameters and fault sensor channels, and reduces the fault omission rate under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the technical field of actuator fault diagnosis, specifically relating to a hierarchical fault diagnosis method, storage medium, and device for actuator systems. Background Technology

[0002] Actuator systems are widely used in automation, robotics, aerospace, automotive, and precision instrumentation to drive robot joints and limbs, control the position of control surfaces (such as ailerons, elevators, and rudders), and provide precise motion control. The normal operation of these robots, aerospace equipment, automobiles, and industrial automation equipment highly depends on the proper functioning of actuators, which must provide specified functions and meet performance parameter requirements. Under the redundancy requirements of the control system, the multi-channel redundancy design of actuators ensures high reliability. High reliability requirements often result in insufficient and scattered fault samples and difficulties in fault mode acquisition during actual testing. Fault data acquisition for actuator systems is difficult, annotation is costly, the number of samples is limited, and the feature attributes have high dimensionality, making it a small sample set problem. Introducing expert systems and data-driven methods for fault diagnosis under small sample conditions easily encounters a series of problems such as knowledge acquisition bottlenecks and narrow knowledge thresholds. When accurate modeling is not possible, data-driven diagnosis shows its advantages.

[0003] In the current process of identifying and classifying root causes of failures, it is difficult to accurately distinguish between actuator body failures and failures in related components. Existing systems mainly rely on expert experience to set static threshold parameters for testing and judgment, and these threshold ranges are generally set too broadly due to considerations of system integration error compensation. The performance parameter fluctuations of components from the same batch and source are smaller than the set threshold range, and the existing threshold system has significant sensitivity redundancy. If the threshold can be accurately extracted from normal data, the test sensitivity can be improved, and anomalies in the system can be more easily detected from the test results.

[0004] For example, Chinese patent 201910114471.8 discloses a fault diagnosis method for aileron actuators based on a Simulink model. It constructs a simulation model for a direct-drive valve-type redundant aileron actuator, collects fault data of key fault modes by embedding different types of faults into the simulation model, and performs fault diagnosis by classifying and applying data difference thresholds and threshold analysis based on model residuals. However, this method requires building an accurate model for the complex nonlinear system of the actuator, which is difficult to obtain. The residual library established by numerical changes or the introduction of biases from the normal model differs from the actual faults of this complex nonlinear system. During fault diagnosis, the Pearson correlation coefficient between the residual signal to be diagnosed and the residual signals in the residual library is used as the classification basis. Since zero-bias deviation and position accuracy deviation cannot be distinguished from the listed residual signals of the collected data, and combined fault modes cannot be identified, the fault diagnosis suffers from dependence on the accuracy of the model and difficulty in overcoming the influence of the correlation between fault modes.

[0005] For example, Chinese patent 202110195845.0 discloses a method and apparatus for fault diagnosis of aileron actuators based on the AMESim model. This method involves constructing an AMESim simulation model of the actuator, using fault data obtained from the model simulation for training, and then obtaining a deep neural network model for fault diagnosis. However, this method suffers from the problem that fault data, fault characteristics, and diagnostic model are all highly dependent on the accuracy of the constructed model, making it difficult to guarantee accuracy when applied to diagnosing faults under real-world operating conditions of the actuator.

[0006] For example, Chinese patent 201810260434.3 discloses an actuator fault detection and diagnosis method based on a deep random forest algorithm. This method uses the actuator's input-output data under normal operating conditions as training data, trains a Radial Basis Function (RBF) network algorithm to obtain the analytical redundancy of the monitored actuator, extracts residual data features from the actuator's actual output and analytical signals as a feature dataset, and uses this historical residual feature dataset to train a deep random forest for fault mode recognition. However, the RBF network training suffers from complexity, instability, and generalization issues, requiring thorough verification and adjustment. It relies on a large amount of data for continuous learning, making it difficult to quickly build a reliable and accurate actuator analytical model. The high reliability design of actuators and the diverse fault modes result in a small number of real fault samples, leading to overfitting and generalization problems in the trained classifier model. Furthermore, this classifier can only classify faults into three main categories, making it difficult to guarantee the accuracy of fault diagnosis, and the fault diagnosis results are insufficient for fault localization. Summary of the Invention

[0007] The purpose of this invention is to provide a hierarchical fault diagnosis method, storage medium, and device for actuator systems, in order to solve the above-mentioned problems.

[0008] This invention is mainly achieved through the following technical solutions: A hierarchical fault diagnosis method for actuator systems includes the following steps: Step S1: Acquire multi-channel timing data and perform data preprocessing; Step S2: Perform steady-state detection and diagnosis of multi-channel data based on the dynamic differential threshold model. When a steady-state threshold error is detected and the index is not empty, output the parameter information of the steady-state threshold error and the corresponding fault mode; when a steady-state threshold error is detected and the index is empty, proceed to step S3. Step S3: Fault diagnosis process based on LSTM prediction model library.

[0009] To better realize the present invention, step S2 further includes the following steps: Step S21: Heterogeneous data fusion processing: Standardize the multi-channel time series data, and then construct a two-dimensional channel-parameter mapping matrix; Step S22: Construct a dynamic threshold model; construct a dynamic threshold model based on the initial set of steady-state differential thresholds of the excitation parameters in the feature parameters, and implement dynamic threshold adjustment in combination with the system working mode; Step S23: Identification of continuous steady-state intervals and threshold detection; Step S231: First, noise suppression processing is performed on the standardized multi-channel time-series data to be diagnosed; Step S232: Preliminary detection, using differential signal threshold to trigger steady-state marking, and initially screening candidate intervals; Step S233: Secondary detection, perform extreme value analysis on the interval that passed the primary detection, and verify whether the detection threshold is exceeded based on the signal peak value; Step S24: Output the steady-state threshold deviation value and index to obtain the parameter information of the steady-state threshold deviation value and the corresponding fault mode.

[0010] To better realize the present invention, in step S21, during the standardization process, the data length of each channel is detected and time synchronization alignment is performed, and a truncation / compensation strategy is adopted to eliminate sampling deviation.

[0011] To better realize the present invention, further, in step S231, the differential signal is subjected to dual smoothing filtering processing, and signal noise reduction is achieved by a convolution algorithm based on a sliding window; in step S232, steady-state evaluation is performed based on a dynamic sliding window, and a hard constraint of minimum steady-state duration τ is applied; in step S233, the spatial resolution accuracy is controlled by the window width parameter ΔT, and a dual constraint condition is constructed in conjunction with the minimum duration τ to form a composite criterion that takes into account both temporal continuity and fluctuation tolerance.

[0012] To better realize the present invention, in step S22, by fusing qualitative model and normal / abnormal operating condition data value, machine learning algorithm is used to perform deep feature extraction on historical normal / abnormal operating condition data to construct a dynamic threshold model based on multi-channel data fusion.

[0013] To better realize the present invention, further, in step S3, the construction of the LSTM model includes the following steps: Step T11: Obtain training data; obtain multi-channel time-series data of actuator historical tests, including normal data and fault data; Step T12: Train a normal actuator model based on normal data to represent the normal mode of the actuator, the input-output relationship of each channel, and its dynamic behavior characteristics; Step T13: Train an actuator anomaly model based on fault data to describe the actuator's historical fault modes, the input-output relationship of each channel, and its dynamic behavior characteristics; Step T14: Construct an LSTM prediction model library: Integrate the trained normal actuator model and abnormal actuator model into the database and manage them through a configuration table to form an LSTM prediction model library; the configuration table records information such as model version, path, status, and parameters.

[0014] To better realize the present invention, step T12 further includes the following steps: Step A1: During training, the data is divided into channels, and each channel's data is trained independently; Step A2: Split the raw normal data by sensor channel and establish channel index identifiers; after normalizing the normal data by channel, divide it into training set, validation set and test set according to the proportion; Step A3: Design a parallel data loader, create an independent data window for each channel, and use a sliding window synchronization strategy to align time steps; Step A4: Adopt a multi-head LSTM network architecture, build a dedicated LSTM sub-network for each channel, and realize cross-channel information interaction through a gated attention fusion layer; adopt a distributed optimization strategy, configure an independent Adam optimizer for each LSTM sub-network, and design a MAE loss function based on channel variance weighting. Step A5: Introduce a dynamic early stop mechanism and monitor and verify the loss curves for each channel; Step A6: Finally, perform stratified cross-validation to maintain the consistency of channel data distribution during K-fold partitioning and ensure that the validation set contains corresponding samples for all channels.

[0015] To better implement this invention, further, in step S3, fault diagnosis based on the LSTM prediction model library includes the following steps: Step T21: Multi-channel data fusion processing: Input multi-channel time-series data, detect the data length of each channel, and truncate and align the data based on the shortest length to avoid data misalignment; Step T22: Data Prediction: Input the working conditions corresponding to the multi-channel time series data, and based on the configuration table, call the corresponding actuator normal model and actuator abnormal model in the LSTM prediction model library to predict the corresponding response parameters in the multi-channel time series data, and obtain the multi-dimensional vector of prediction results of each model. Step T23: Weighted distance calculation and fault matching: Calculate the weighted Euclidean distance between the predicted multidimensional vector and the actual data multidimensional vector, and match the most likely fault mode and corresponding sensor channel with the minimum distance; determine the optimal matching fault mode and its corresponding sensor channel through global optimization. Step T24: Output a multi-dimensional interpretation of the diagnostic results.

[0016] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described hierarchical fault diagnosis method for an actuator system.

[0017] An electronic device includes a memory and a processor; the memory stores a computer program; the processor is configured to execute the computer program in the memory to implement the above-described hierarchical fault diagnosis method for an actuator system.

[0018] The beneficial effects of this invention are as follows: (1) This invention integrates qualitative models and normal / abnormal operating condition data, and uses machine learning algorithms to perform deep feature extraction on historical normal / abnormal operating condition data to construct a dynamic threshold model based on multi-channel data fusion. By judging steady-state thresholds, model residuals, and model distances, it can not only effectively determine the working state of complex systems and whether the working state of complex systems is faulty, but also deeply analyze the causes of faults and predict potential impacts, helping to solve the problem of low fault diagnosis accuracy caused by small fault samples and diverse fault modes of redundant actuators. It breaks through the diagnostic bottleneck of traditional methods in small sample and multi-mode fault scenarios, and significantly improves the accuracy and predictability of fault diagnosis. On the other hand, compared with traditional anomaly detection methods, this invention improves data fusion capabilities, establishes an adaptive alignment mechanism for multi-channel data, eliminates sampling bias in time-series data, develops a channel-parameter two-dimensional mapping matrix, realizes multi-state and multi-dimensional correlation analysis of composite parameters such as current and displacement in actuator system scenarios, and effectively analyzes coupled faults in nonlinear control scenarios.

[0019] (2) Compared with traditional fault diagnosis methods, this invention shortens the steady-state threshold detection response time, constructs a dynamic threshold model based on sliding window convolution, and designs a multi-level processing architecture of "noise suppression → steady-state judgment → threshold verification" to improve dynamic diagnosis performance. This invention uses a gated attention-enhanced LSTM network, which accurately fits through multi-layer nonlinear activation stacking, effectively capturing the nonlinear characteristics in the actuator command-response sequence and improving the fault identification accuracy; the model has good anti-interference and generalization performance, especially for faults in the square wave response data generated by step commands. This invention proposes a fault parameter-sensor channel dual localization method to achieve accurate localization of the fault source.

[0020] (3) This invention constructs a hierarchical fault diagnosis method that integrates feature matching and model prediction, effectively combining the advantages of threshold detection and prediction models. It can simultaneously locate fault parameters and fault sensor channels, reducing the false alarm rate under complex operating conditions. An extensible threshold configuration dictionary is developed to support incremental updates of fault modes and reduce model reconstruction time when adding new fault modes. A multi-dimensional visual diagnostic report is constructed, and the interpretability of diagnostic results is improved through time-domain comparison curves and confidence radar charts.

[0021] (4) This invention fully utilizes qualitative models and normal data value to propose a dynamic threshold modeling method based on multi-channel steady-state differential thresholds. By constructing a channel-parameter two-dimensional mapping matrix, it achieves structured representation of heterogeneous data and realizes multi-parameter threshold collaborative optimization in hydraulic and channel modes. Secondly, this invention proposes a multi-level processing architecture of "noise suppression → steady-state determination → threshold verification". It adopts differential signal dual smoothing filtering, sliding window convolution algorithm and dynamic window width adjustment strategy to establish a composite criterion that takes into account both temporal continuity and fluctuation tolerance. On the other hand, this invention uses a multi-channel LSTM network architecture to train the prediction model, breaking through the dimensional limitations of traditional single-channel modeling.

[0022] This invention hierarchically integrates dynamic threshold detection with an LSTM prediction model. A steady-state threshold triggering mechanism enables dynamic switching of diagnostic strategies. Combining dynamic differential thresholding with LSTM prediction residual analysis, it achieves collaborative detection of transient / steady-state faults, reducing the error rate of fault detection. This enables hierarchical fault diagnosis through the fusion of feature matching and model prediction, addressing the imbalance in sensitivity between abrupt and gradually changing faults using a single method. Furthermore, this invention establishes a weighted Euclidean distance matching algorithm, introducing parameter weight coefficients to dynamically optimize fault identification and improve the confusion problem in multi-mode fault diagnosis. It also proposes a hierarchical labeling strategy for fault mode identifiers, using dual identifiers of operating condition and mode to effectively aggregate multiple batches of data, improving the utilization rate of fault samples. This invention can be rapidly applied based on qualitative models, achieving fault detection in redundant actuator systems, and reducing computational resource consumption and single-diagnosis time through a hierarchical diagnostic strategy. Attached Figure Description

[0023] Figure 1 This is a flowchart of steady-state detection and diagnosis based on a dynamic differential threshold model; Figure 2 This is a flowchart for fault diagnosis based on an LSTM prediction model library. Figure 3 This is a flowchart of the hierarchical fault diagnosis method for actuator systems according to the present invention. Detailed Implementation Example 1:

[0024] A hierarchical fault diagnosis method for actuator systems is proposed, which calls a multi-channel data steady-state detection and diagnosis method based on a dynamic differential threshold model for diagnosis. When a steady-state threshold deviation is detected and the index is empty, a fault diagnosis method based on an LSTM prediction model library is initiated; when a steady-state threshold deviation is detected and the index is not empty, the parameter information of the steady-state threshold deviation and the corresponding fault mode are output.

[0025] Preferably, the multi-channel data steady-state detection and diagnosis method based on a dynamic differential threshold model includes the following steps: Step S21: Heterogeneous data fusion processing: The multi-channel time series data to be diagnosed is standardized, the length of each channel data is automatically detected, and time synchronization alignment is performed. A truncation / compensation strategy is used to eliminate sampling bias. Then, a channel-parameter two-dimensional mapping matrix is ​​constructed to realize the structured representation of heterogeneous data.

[0026] Step S22: Dynamic Threshold Modeling: A dynamic threshold model is constructed based on the initial set of steady-state differential thresholds for the excitation parameters (P1-PN) in the feature parameters. A dynamic threshold adjustment mechanism is implemented in conjunction with the system operating modes (M1-N1, M2-N2, M3-N3). Multi-parameter threshold rules are specifically constructed for the hydraulic mode and channel mode to achieve coordinated optimization of thresholds for current parameters, displacement composite parameters, etc.

[0027] Step S23: Continuous Steady-State Interval Identification and Threshold Detection: A multi-level processing architecture of "noise suppression → steady-state determination → threshold verification" is adopted. First, the standardized multi-channel time-series data to be diagnosed undergoes differential signal double smoothing filtering, and signal noise reduction is achieved through a convolution algorithm based on a sliding window (default window width ΔT). Threshold detection is achieved through a two-level verification mechanism. The primary detection uses the differential signal threshold to trigger steady-state marking and preliminarily screens candidate intervals, i.e., a dynamic sliding window with an adjustable window width ΔT is used for continuous evaluation, while a hard constraint of the minimum steady-state duration τ is applied to ensure that the identified interval meets the time continuity requirement. The secondary detection performs extreme value analysis to verify whether the signal peak value exceeds the detection threshold. The spatial resolution accuracy is controlled by the window width parameter ΔT, and a dual constraint condition is constructed in conjunction with the minimum duration τ to form a composite criterion that takes into account both time continuity and fluctuation tolerance.

[0028] Step S24: Diagnostic result output: Output the steady-state threshold deviation value and index obtained from the multi-channel data steady-state detection and diagnosis, and give the parameter information of the steady-state threshold deviation value and the corresponding fault mode.

[0029] Preferably, the fault diagnosis method based on the LSTM prediction model library includes the following steps: Step T1: Construct a multi-mode LSTM prediction model library based on multi-channel time series data of actuators.

[0030] Step T11: Training Data Preparation: Prepare historical actuator test data, which is multi-channel time-series data containing normal and fault data, used to train normal and fault models respectively. When aggregating multi-source data, such as merging actuator data from different batches under the same condition (both normal or the same fault condition), identifiers are added to merge the multi-source data. For multiple normal datasets under the same operating condition, add a normal mode identifier and an operating condition identifier, then merge all test data to form an actuator normal response feature dataset. For datasets under the same operating condition and with the same anomaly, add the corresponding fault mode identifier and an operating condition identifier, then merge all test data to form an actuator fault response feature dataset for the corresponding fault mode.

[0031] Step T12: Actuator Normal Model Training: Train the actuator normal response feature dataset using the following method to form a normal model based on multi-channel time-series data of the actuator, which is used to represent the normal mode of the actuator, the input-output relationship of each channel, and its dynamic behavior characteristics.

[0032] Specifically, the excitation parameters in the actuator test normal mode basic parameter model are used as training parameters, and the response parameters are used as prediction parameters. Training is performed on the actuator normal response feature dataset for each operating condition, with training divided by channel and each channel's data trained independently. The original dataset is split by sensor channel, and channel index identifiers (e.g., ChM0-ChMn) are established. After channel-specific normalization, the training and test sets are proportionally divided. A parallel data loader is designed to create an independent data window for each channel, and a sliding window synchronization strategy is used to align time steps. A multi-head LSTM network architecture is adopted, with a dedicated LSTM sub-network built for each channel, and cross-channel information interaction is achieved through a gated attention fusion layer. A distributed optimization strategy is employed, configuring an independent Adam optimizer for each sub-network and designing a MAE loss function based on channel variance weighting. A dynamic early stop mechanism is introduced, and the loss curve is monitored and verified for each channel, with a dedicated patience counter set for each channel. Finally, hierarchical cross-validation is implemented to maintain the consistency of channel data distribution during K-fold partitioning, ensuring that the validation set contains corresponding samples for all channels.

[0033] Step T13: Actuator anomaly model training: Train the actuator fault response feature dataset using the same method as training the normal response feature dataset to form an anomaly model that corresponds one-to-one with "operating condition + fault mode" to describe the actuator's historical fault modes, the input-output relationship of each channel, and its dynamic behavior characteristics.

[0034] Specifically, a corresponding actuator anomaly model is constructed for each fault mode in the initial set of actuator test fault modes. The excitation parameters from the corresponding actuator test fault feature parameters model are used as training parameters, and the response parameters are used as prediction parameters. The actuator fault response feature dataset for each operating condition is trained sequentially, with training divided by channel and each channel's data trained independently. The original dataset is split by sensor channel, and channel index identifiers (e.g., ChM0-ChMn) are established. A parallel data loader is designed to create an independent data window for each channel, and a sliding window synchronization strategy is used to align time steps. A multi-head LSTM network architecture is adopted, with a dedicated LSTM sub-network built for each channel, and cross-channel information interaction is achieved through a gated attention fusion layer. A distributed optimization strategy is adopted, configuring an independent Adam optimizer for each sub-network and designing a MAE loss function based on channel variance weighting. A dynamic early stop mechanism is introduced, and the loss curve is monitored and verified for each channel, with a dedicated patience counter set for each channel. Finally, hierarchical cross-validation is implemented to maintain the consistency of channel data distribution during K-fold partitioning, ensuring that the validation set contains corresponding samples from all channels.

[0035] Step T14: Constructing the LSTM prediction model library: By integrating the trained actuator normal and abnormal models into the database and managing them through a configuration table, an LSTM prediction model library is formed. The configuration table records information such as model version, path, status, and parameters.

[0036] Step T2: Fault diagnosis method based on LSTM prediction model library: Step T21: Multi-channel data fusion processing: Input multi-channel diagnostic data, automatically detect the length of each channel's data, and truncate and align it based on the shortest length to avoid data misalignment.

[0037] Step T22: Data Prediction: Input the operating conditions, and based on the configuration table, automatically match the model subset in the LSTM prediction model library according to the operating conditions, call the corresponding normal and abnormal models, use the input and output parameters of the corresponding model in the LSTM prediction model library, predict the corresponding response parameters in the multi-channel time series data to be diagnosed, and obtain the multi-dimensional vector of prediction results of each model, which includes the prediction result vector of multiple sensor channels and multiple response parameters.

[0038] Step T23: Weighted Distance Calculation and Fault Matching: A weighting coefficient k is introduced to configure the importance of parameters in the multidimensional vector of the prediction result when calculating the weighted Euclidean distance. Adjusting the value of k dynamically optimizes the fault identification accuracy under different operating conditions. The weighted Euclidean distance between the multidimensional vector of the prediction result and the multidimensional vector of the actual data is calculated. The minimum distance criterion is used to match the most probable fault mode and its corresponding sensor channel with the minimum distance. Global optimization determines the optimal matching fault mode and its corresponding sensor channel.

[0039] Step T24: Multi-dimensional interpretation and output of diagnostic results: Automatically generate multi-dimensional Excel reports containing predicted values, actual values, and distances for each model, summarize the fault modes and channels corresponding to the global minimum distance, and output structured reports to enhance the interpretability of the results. In addition to outputting the best-matching fault mode and its corresponding sensor channel, the structured reports may also include time-domain comparison curves of raw data and predicted values, a sorted list of distance values ​​for each model, a weight distribution histogram, a confidence radar chart of the fault mode, and a similar fault comparison matrix.

[0040] Step T25: Anomaly Detection Mechanism: The fault diagnosis process based on the LSTM prediction model library includes anomaly detection mechanisms such as file occupancy detection, abnormal model skipping, and dynamic parameter configuration to ensure the integrity of the fault diagnosis execution process based on the LSTM prediction model library.

[0041] Preferably, the data source for the above method is as follows: (1) Constructing the normal mode parameter model of actuator testing: Based on the function, performance and test range of actuator system testing, integrate multi-channel timing test data, extract the excitation parameters, response parameters and their corresponding thresholds of each tested function, and construct the basic parameter model of the normal mode of actuator testing.

[0042] (2) Constructing the excitation parameter model in the actuator test fault characteristic parameters: Based on the failure modes, impacts, and hazards analysis of the actuator system test, combined with the test requirements and historical failure samples in the test data, an initial set of actuator test failure modes is defined. The excitation and response parameters corresponding to the functions of the failure modes in the initial set of actuator test failure modes, and their corresponding set of out-of-tolerance values, are used as the initial set of excitation parameters in the actuator test fault characteristic parameters for that failure mode. Based on the basic model of the actuator test normal mode, the initial set of actuator test failure modes, and the initial set of excitation parameters in the actuator test fault characteristic parameters, a corresponding excitation parameter model in the actuator test fault characteristic parameters is constructed for each failure mode in the initial set of actuator test failure modes.

[0043] (3) Set an initial set of abnormal parameter thresholds according to the parameter thresholds required by the test, including the steady-state threshold, transient threshold, steady-state differential threshold and other abnormal parameter thresholds. Example 2:

[0044] A hierarchical fault diagnosis method for actuator systems, taking an aircraft actuator system with N hydraulic system modes and M channels as an example, achieves single-channel fault diagnosis and multi-state, multi-channel fault localization of a single actuator using qualitative models, normal data, and historical fault samples. Figure 3 As shown, the specific steps include: (1) Acquire data and perform data preprocessing.

[0045] 1.1 Constructing a normal mode parameter model for actuator testing: according to the functions, performance and test scope of the actuator system test, integrate multi-channel time series test data, extract the excitation parameters, response parameter sets and their corresponding thresholds of each tested function, construct a basic parameter model of the normal mode for actuator testing, and save it in XML format.

[0046] Select normal actuator test data from multiple sorties, including position sensor data, current, etc. There is a nonlinear relationship between the actuator current and displacement. Under multi-system conditions, the position error of A / C channels shall be within ±X1 degrees, the position error of B / D channels shall be within ±X2 degrees, and the current fluctuation amplitude of the driving motors of M channels shall be less than Y A. An example of the XML file of the basic parameter model of the normal mode for actuator testing is as follows: <Actuator Test Normal Mode Parameter Model> <Channel Information> <Channel Name>Channel A< / Channel Name> <Excitation Parameter> <Parameter Name>Control Signal< / Parameter Name> <Parameter Type>Voltage< / Parameter Type> <Unit>V< / Unit> <Description>Control signal for actuator Channel A< / Description> <Threshold>< / Threshold> < / Excitation Parameter> <Excitation Parameter> <Parameter Name>Supply Pressure< / Parameter Name> <Parameter Type>Pressure< / Parameter Type> <Unit>bar< / Unit> <Description>Supply pressure for actuator Channel A< / Description> <Threshold>< / Threshold> < / Excitation Parameter> <Response Parameter> <Parameter Name>Displacement< / Parameter Name> <Parameter Type>Position< / Parameter Type> <Unit>mm< / Unit> <Description>Displacement of actuator Channel A< / Description> <Threshold>< / Threshold> < / Response Parameter> …… < / Response Parameter> < / Channel Information> < / Actuator test normal mode parameter model>.

[0047] 1.2 Construction of excitation parameter model in actuator test fault characteristic parameters: according to the failure mode, effects and criticality analysis of actuator system testing, combined with test requirements and historical fault samples in test data, an initial set of actuator test failure modes is defined. The excitation and response parameters corresponding to the failure mode functions in the initial set of actuator test failure modes and their corresponding out-of-tolerance value sets are taken as the initial set of excitation parameters in the actuator test fault characteristic parameters of the failure mode. Based on the basic model of the actuator test normal mode, the initial set of actuator test failure modes and the initial set of excitation parameters in the actuator test fault characteristic parameters, a corresponding excitation parameter model in the actuator test fault characteristic parameters is constructed for each failure mode in the initial set of actuator test failure modes, and saved in XML format.

[0048] As shown in Table 1, typical failure modes may include: jamming, response delay, derated operation, leakage, etc.

[0049] Table 1 Typical failure modes

[0050] Actuator data in a fault state is extracted from test data marked as fault in historical actuator test data.

[0051] According to the failure mode set and the qualitative model, the extracted features and labels are determined, including key input and output data of various failure modes; based on the failure mode set, the features and labels that need to be monitored are determined. Key input and output data of various failure modes are extracted to form a label data set.

[0052] An example of the XML file of the excitation parameter model in actuator test fault characteristic parameters is as follows: <faultfeatureparameters> <faultmode name="机械卡滞"> <parameter name="响应时间" threshold="超过设定阈值" / > <parameter name="位移传感器反馈" threshold="异常" / > < / faultmode> <faultmode name="电气故障"> <parameter name="电压" threshold="超出正常范围" / > <parameter name="电流" threshold="超出正常范围" / > <parameter name="信号噪声" threshold="增加" / > < / faultmode> <faultmode name="液压故障"> <parameter name="液压压力" threshold="低于下限或超过上限" / > < / faultmode> <faultmode name="过载故障"> <parameter name="电机电流" threshold="超过额定值" / > < / faultmode> <faultmode name="控制系统故障"> <parameter name="控制信号频率" threshold="偏离正常值" / > <parameter name="控制信号幅度" threshold="偏离正常值" / > < / faultmode> < / faultfeatureparameters> .

[0053] 1.3 According to the parameter threshold required by the test, an initial set of abnormal parameter thresholds is set. As shown in Table 2, it includes abnormal parameter thresholds such as steady-state threshold, transient threshold and steady-state difference threshold, and is saved in XML format. The state corresponds to the hydraulic system mode and effective channel conditions.

[0054] Table 2 Example of initial set of abnormal parameter thresholds

[0055] An example of the XML file of the initial set of abnormal parameter thresholds is as follows: <abnormal_parameter_thresholds> <fault_category> <name> Mechanical jamming< / name> <monitoring_parameters> <parameter> Response time< / parameter> <parameter> Displacement deviation< / parameter> < / monitoring_parameters> <threshold_criteria> <threshold> <type> steady-state threshold< / type> <condition> Response time > d1ms< / condition> <status> M0CN0S< / status> <channel> A< / channel> < / threshold> <threshold> <type> steady-state threshold< / type> <condition> Displacement deviation > d2mm< / condition> <status> M0CN0S< / status> <channel> C< / channel> < / threshold> <threshold> <type> Steady-state differential threshold< / type> <condition> Response time difference > d4ms< / condition> <status> M0CN0S< / status> <channel> A< / channel> < / threshold> < / threshold_criteria> < / fault_category> <fault_category> <name> Electrical fault< / name> <monitoring_parameters> <parameter> Current noise figure< / parameter> < / monitoring_parameters> <threshold_criteria> <threshold> <type> steady-state threshold< / type> <condition> SNR <d3dB < / condition> <status>M0CN0S< / status> <channel> A< / channel> < / threshold> <threshold> <type> steady-state threshold< / type> <condition> Current noise > d5A< / condition> <status> M0CN0S< / status> <channel> C< / channel> < / threshold> <threshold> <type> Steady-state differential threshold< / type> <condition> Poor current noise <d6dB < / condition> <status>M0CN0S< / status> <channel> A< / channel> < / threshold> < / threshold_criteria> < / fault_category> ...< / abnormal_parameter_thresholds> .

[0056] (ii) such as Figure 3 As shown, firstly, steady-state detection is performed using a dynamic differential threshold model. When the steady-state threshold deviation and index are empty, fault diagnosis is performed based on the LSTM prediction model library. When the steady-state threshold deviation and index are not empty, the parameter information of the steady-state threshold deviation and the corresponding fault mode are output.

[0057] Preferably, 1.4 steady-state detection and diagnosis are based on a dynamic differential threshold model, such as... Figure 1 As shown, it includes the following steps: 1.4.1 Heterogeneous data fusion processing: The multi-channel time series data to be diagnosed is standardized, the length of each channel data is automatically detected, and time synchronization alignment is performed. A truncation / compensation strategy is used to eliminate sampling bias. As shown in Table 3, a channel-parameter two-dimensional mapping matrix is ​​then constructed to realize the structured representation of heterogeneous data.

[0058] Table 3 Example of a two-dimensional channel-parameter mapping matrix

[0059] 1.4.2 Dynamic Threshold Modeling: A dynamic threshold model is constructed based on the initial set of steady-state differential thresholds for the excitation parameters (P1, P2, ..., PN) among the characteristic parameters. A dynamic threshold adjustment mechanism is implemented in conjunction with the system operating modes (MOCN0S / 2C2S / 2C1S...). Multi-parameter threshold rules are specifically developed for dual-hydraulic and channel modes to achieve coordinated optimization of thresholds for current parameters, displacement composite parameters, etc.

[0060] 1.4.3 Continuous steady-state interval identification and threshold detection: A multi-level processing architecture of "noise suppression → steady-state determination → threshold verification" is adopted.

[0061] ① Noise Suppression: The standardized multi-channel time-series data to be diagnosed undergoes differential signal double smoothing filtering, and signal denoising is achieved through a convolution algorithm based on a sliding window (default window width ΔT). That is, the first-order difference is calculated and a moving average filter is applied to achieve noise suppression.

[0062] ② Steady-state determination: Steady-state marking is triggered using a differential signal threshold to initially screen candidate intervals. Continuity is evaluated using a dynamic sliding window, with a minimum steady-state duration constraint applied. Specifically, a dynamic sliding window with an adjustable window width ΔT is used for continuous evaluation. The spatial resolution is controlled by the window width parameter ΔT, while a hard constraint of the minimum steady-state duration τ is applied, constructing a dual constraint condition to form a composite criterion that considers both temporal continuity and fluctuation tolerance.

[0063] ③ Threshold verification: Perform extreme value analysis on the interval that passes the primary detection, and verify whether the signal peak value exceeds the secondary threshold.

[0064] The resulting multidimensional anomalies will be recorded using a dictionary array, and the output format is shown in the example below: { "Parameter name": [ { "system": "Hydraulic system status", "channel": "channel_1", "exceedance_indices": [List of abnormal indexes], "exceedance_values": [List of abnormal values] }, ... ] } 1.4.4 Diagnostic Result Output: Outputs the steady-state threshold deviation value and index obtained from multi-channel data steady-state detection and diagnosis, and provides the parameter information of the steady-state threshold deviation value and the corresponding fault mode.

[0065] Preferably, 1.5 a multi-mode LSTM prediction model is constructed based on multi-channel time-series data of the actuator.

[0066] 1.5.1 Training Data Preparation: Prepare historical actuator test data, which is multi-channel time-series data containing normal and fault data, used to train normal and fault models respectively. When aggregating multi-source data, such as merging actuator data from different test batches in the same state (all normal or the same fault state), identifiers are added to merge the multi-source data. For multiple normal datasets under the same operating condition, add a normal mode identifier and an operating condition identifier, then merge all test data to form an actuator normal response feature dataset. For datasets with the same operating condition and the same anomaly, add the corresponding fault mode identifier and operating condition identifier, then merge all test data to form an actuator fault response feature dataset for the corresponding fault mode.

[0067] 1.5.3 Actuator Normal Model Training: The actuator normal response feature dataset is trained using the following method to form a normal model based on multi-channel time-series data of the actuator, which is used to represent the normal mode of the actuator, the input-output relationship of each channel, and its dynamic behavior characteristics.

[0068] 1.5.3.1 The excitation parameters in the basic parameter model of the actuator test normal mode are used as training parameters, and the response parameters are used as prediction parameters. The actuator normal response feature dataset for each operating condition is trained sequentially, with training divided by channel and each channel's data trained independently. The original dataset is split by sensor channel, and channel index identifiers (e.g., ChM0-ChMn) are established. After normalization processing for each channel, the training and test sets are proportionally divided. A parallel data loader is designed to create an independent data window for each channel, and a sliding window synchronization strategy is used to align time steps. A multi-head LSTM network architecture is adopted, with a dedicated LSTM sub-network built for each channel, and cross-channel information interaction is achieved through a gated attention fusion layer. A distributed optimization strategy is adopted, configuring an independent Adam optimizer for each sub-network and designing a MAE loss function based on channel variance weighting. A dynamic early stop mechanism is introduced, and the loss curve is monitored and verified for each channel, with a dedicated patience counter set for each channel. Finally, hierarchical cross-validation is implemented to maintain the consistency of channel data distribution during K-fold partitioning, ensuring that the validation set contains corresponding samples for all channels.

[0069] 1.5.4 Actuator Anomaly Model Training: The actuator fault response feature dataset is trained using the same method as the normal response feature dataset, forming an anomaly model that corresponds one-to-one with "operating condition + fault mode" to describe the actuator's historical fault modes, the input-output relationship of each channel, and its dynamic behavior characteristics.

[0070] 1.5.4.1 For each fault mode in the initial set of actuator test fault modes, a corresponding actuator anomaly model is constructed. The excitation parameters in the corresponding actuator test fault feature parameters model are used as training parameters, and the response parameters are used as prediction parameters. The actuator fault response feature dataset for each operating condition is trained sequentially, with training divided by channel and each channel's data trained independently. The original dataset is split by sensor channel, and channel index identifiers (e.g., ChM0-ChMn) are established. A parallel data loader is designed to create an independent data window for each channel, and a sliding window synchronization strategy is used to align time steps. A multi-head LSTM network architecture is adopted, constructing a dedicated LSTM sub-network for each channel, and achieving cross-channel information interaction through a gated attention fusion layer. A distributed optimization strategy is adopted, configuring an independent Adam optimizer for each sub-network and designing a MAE loss function based on channel variance weighting. A dynamic early stop mechanism is introduced, and the loss curve is monitored and verified for each channel, with a dedicated patience counter set for each channel. Finally, hierarchical cross-validation is implemented, maintaining the consistency of channel data distribution during K-fold partitioning to ensure that the validation set contains corresponding samples for all channels.

[0071] 1.5.5 Building the LSTM Prediction Model Library: By integrating the trained actuator normal and abnormal models into the database and managing them through a configuration table, an LSTM prediction model library is formed. As shown in Table 4, the configuration table records information such as model version, path, status, and parameters.

[0072] Table 4 Configuration Examples

[0073] Preferably, 1.6 fault diagnosis is performed based on an LSTM prediction model library, such as... Figure 2 As shown, it includes the following steps: 1.6.1 Multi-channel data fusion processing: Input multi-channel diagnostic data, automatically detect the length of each channel's data, and truncate and align it based on the shortest length to avoid data misalignment.

[0074] 1.6.2 Data Prediction: Based on the input operating conditions, the system automatically matches a subset of models from the LSTM prediction model library according to the configuration table, calls the corresponding normal and abnormal models, and uses the input and output parameters of the corresponding models in the LSTM prediction model library to predict the corresponding response parameters in the multi-channel time series data to be diagnosed, thus obtaining a multi-dimensional vector of prediction results for each model, which includes the prediction result vectors of multiple sensor channels and multiple response parameters.

[0075] Let the model index be i (i=0,1,...,m), the parameter index be j (j=1,...,p), and the time index be t (t=0,1,...,n). Let Y be the prediction vector output by each model. i It contains predicted values ​​for n time steps t and p response parameters. The vector structure of the prediction results for each model is as follows: Y i[t][j] = p i_{j,t} ∈ R^{n×p} in: p represents the number of response parameters; m represents the total number of models; n represents the total number of time steps.

[0076] The specific prediction result vectors of each model can be represented as follows: Y0 = [[p 0_{1,t0} , p 0_{2,t0} , ..., p 0_{p,t0} ], [p 0_{1,t1} , p 0_{2,t1} , ..., p 0_{p,t1} ], ... [p 0_{1,tn} , p 0_{2,tn}, ..., p 0_{p,tn} ]]^T ∈ R^{n×p} ... Y m = [[p m_{1,t0} , p m_{2,t0} , ..., p m_{p,t0} ], [p m_{1,t1} , p m_{2,t1} , ..., p m_{p,t1} ], ... [p m_{1,tn} , p m_{2,tn} , ..., p m_{p,tn} ]]^T ∈ R^{n×p} 1.6.3 Weighted Distance Calculation and Fault Matching: A weighting coefficient k is introduced to configure the importance of parameters in the multidimensional vector of the prediction result when calculating the weighted Euclidean distance. Adjusting the value of k dynamically optimizes the fault identification accuracy under different operating conditions. The weighted Euclidean distance between the multidimensional vector of the prediction result and the multidimensional vector of the actual data is calculated. The minimum distance criterion is used to match the most probable fault mode and its corresponding sensor channel with the minimum distance. Global optimization determines the optimal matching fault mode and its corresponding sensor channel.

[0077] Real data multidimensional vector Y real and the predicted data multidimensional vector Y pre The structures are as follows: Y real = [[y real_{1,t0} , y real_{2,t0} , ..., y real_{p,t0} ], [y real_{1,t1} , y real_{2,t1} , ..., y real_{p,t1} ], ... [y real_{1,tn} , y real_{2,tn} , ..., y real_{p,tn} ]]^T ∈ R^{n×p} Y pre = [[y pre_{1,t0} , y pre_{2,t0} , ..., y pre_{p,t0} ], [y pre_{1,t1} , y pre_{2,t1} , ..., y pre_{p,t1} ], ... [ypre_{1,tn} , y pre_{2,tn} , ..., y pre_{p,tn} ]]^T ∈ R^{n×p} (t) represents the time step (t = 0, 1, ..., n); (j) represents the response parameter, (j = 1, 2, ..., p); (n) represents the total number of time steps; (p) represents the number of response parameters.

[0078] The weighted Euclidean distance of model i is defined as: .

[0079] in: y pre j Let j be the multidimensional vector of the predicted data; y real j Let j be the j-th actual data multidimensional vector; k j Let j be the weight.

[0080] 1.6.4 Multi-dimensional Interpretation Output of Diagnostic Results: Automatically generates multi-dimensional Excel reports containing predicted values, actual values, and distances for each model. It summarizes the fault modes and channels corresponding to the global minimum distance, outputting structured reports to enhance the interpretability of the results. In addition to outputting the best-matching fault mode and its corresponding sensor channel, the structured reports may also include time-domain comparison curves of raw data and predicted values, a sorted list of distance values ​​for each model, a weight distribution histogram, a confidence radar chart of the fault mode, and a similar fault comparison matrix.

[0081] 1.6.5 Anomaly Detection Mechanism: The fault diagnosis process based on the LSTM prediction model library includes anomaly detection mechanisms such as file occupancy detection, abnormal model skipping, and dynamic parameter configuration to ensure the integrity of the fault diagnosis execution process based on the LSTM prediction model library.

[0082] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. An actuator system hierarchical fault diagnosis method characterized by, Includes the following steps: Step S1: Acquire multi-channel timing data and perform data preprocessing; the timing data includes response time, displacement sensor data, voltage, current, signal noise, hydraulic pressure, control signal frequency or amplitude, and status; Step S2: Perform steady-state detection and diagnosis of multi-channel data based on the dynamic differential threshold model. When a steady-state threshold error is detected and the index is not empty, output the parameter information of the steady-state threshold error and the corresponding fault mode; when a steady-state threshold error is detected and the index is empty, proceed to step S3. Step S21: Heterogeneous data fusion processing: Standardize the multi-channel time series data, and then construct a two-dimensional channel-parameter mapping matrix; Step S22: Construct a dynamic threshold model; construct a dynamic threshold model based on the initial set of steady-state differential thresholds of the excitation parameters in the feature parameters, and implement dynamic threshold adjustment in combination with the system working mode; Step S23: Identification of continuous steady-state intervals and threshold detection; Step S231: First, noise suppression processing is performed on the standardized multi-channel time-series data to be diagnosed; Step S232: Preliminary detection, using differential signal threshold to trigger steady-state marking, and initially screening candidate intervals; Step S233: Secondary detection, perform extreme value analysis on the interval that passed the primary detection, and verify whether the detection threshold is exceeded based on the signal peak value; Step S24: Output the steady-state threshold deviation value and index to obtain the parameter information of the steady-state threshold deviation value and the corresponding fault mode; Step S3: Fault diagnosis process based on LSTM prediction model library.

2. The actuator system hierarchical fault diagnosis method according to claim 1, characterized in that, In step S21, during the standardization process, the data length of each channel is detected and time synchronization alignment is performed. A truncation / compensation strategy is used to eliminate sampling bias.

3. The actuator system hierarchical fault diagnosis method of claim 1, wherein In step S231, the differential signal undergoes dual smoothing filtering, and signal noise reduction is achieved through a convolution algorithm based on a sliding window. In step S232, steady-state evaluation is performed based on a dynamic sliding window, while a hard constraint of the minimum steady-state duration τ is applied. In step S233, the spatial resolution accuracy is controlled by the window width parameter ΔT, and a dual constraint condition is constructed in conjunction with the minimum duration τ to form a composite criterion that takes into account both temporal continuity and fluctuation tolerance.

4. The hierarchical fault diagnosis method for an actuator system according to claim 1, characterized in that, In step S22, by fusing the qualitative model and the value of normal / abnormal operating condition data, a machine learning algorithm is used to perform deep feature extraction on historical normal / abnormal operating condition data, and a dynamic threshold model based on multi-channel data fusion is constructed.

5. The hierarchical fault diagnosis method for an actuator system according to claim 1, characterized in that, In step S3, the construction of the LSTM model includes the following steps: Step T11: Obtain training data; obtain multi-channel time-series data of actuator historical tests, including normal data and fault data; Step T12: Train a normal actuator model based on normal data to represent the normal mode of the actuator, the input-output relationship of each channel, and its dynamic behavior characteristics; Step T13: Train an actuator anomaly model based on fault data to describe the actuator's historical fault modes, the input-output relationship of each channel, and its dynamic behavior characteristics; Step T14: Construct an LSTM prediction model library: Integrate the trained normal actuator model and abnormal actuator model into the database and manage them through a configuration table to form an LSTM prediction model library; the configuration table records information such as model version, path, status, and parameters.

6. The hierarchical fault diagnosis method for an actuator system according to claim 5, characterized in that, Step T12 includes the following steps: Step A1: During training, the data is divided into channels, and each channel's data is trained independently; Step A2: Split the raw normal data by sensor channel and establish channel index identifiers; after normalizing the normal data by channel, divide it into training set, validation set and test set according to the proportion; Step A3: Design a parallel data loader, create an independent data window for each channel, and use a sliding window synchronization strategy to align time steps; Step A4: Adopt a multi-head LSTM network architecture, build a dedicated LSTM sub-network for each channel, and realize cross-channel information interaction through a gated attention fusion layer; adopt a distributed optimization strategy, configure an independent Adam optimizer for each LSTM sub-network, and design a MAE loss function based on channel variance weighting. Step A5: Introduce a dynamic early stop mechanism and monitor and verify the loss curves for each channel; Step A6: Finally, perform stratified cross-validation to maintain the consistency of channel data distribution during K-fold partitioning and ensure that the validation set contains corresponding samples for all channels.

7. A hierarchical fault diagnosis method for an actuator system according to claim 5 or 6, characterized in that, In step S3, fault diagnosis is performed based on the LSTM prediction model library, including the following steps: Step T21: Multi-channel data fusion processing: Input multi-channel time-series data, detect the data length of each channel, and truncate and align the data based on the shortest length to avoid data misalignment; Step T22: Data Prediction: Input the working conditions corresponding to the multi-channel time series data, and based on the configuration table, call the corresponding actuator normal model and actuator abnormal model in the LSTM prediction model library to predict the corresponding response parameters in the multi-channel time series data, and obtain the multi-dimensional vector of prediction results of each model. Step T23: Weighted distance calculation and fault matching: Calculate the weighted Euclidean distance between the predicted multidimensional vector and the actual data multidimensional vector, and match the most likely fault mode and corresponding sensor channel with the minimum distance; determine the optimal matching fault mode and its corresponding sensor channel through global optimization. A weighting coefficient k is introduced to configure the importance of parameters in the multi-dimensional vector of the prediction result when calculating the weighted Euclidean distance. Adjusting the value of k dynamically optimizes the fault identification accuracy under different operating conditions; the weighted Euclidean distance of model i... D i Defined as: ； in: y pre j Let j be the multidimensional vector of the predicted data; y real j Let j be the j-th actual data multidimensional vector; k j Let j be the weight; p represents the total number of response parameters; Step T24: Output a multi-dimensional interpretation of the diagnostic results.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements a hierarchical fault diagnosis method for an actuator system as described in any one of claims 1-7.

9. An electronic device, characterized in that, It includes a memory and a processor; the memory stores a computer program; the processor is used to execute the computer program in the memory to implement the hierarchical fault diagnosis method for an actuator system according to any one of claims 1-7.

Images