A thermal control system actuator fault diagnosis method

Abstract

The application relates to the technical field of big data processing, and discloses a thermal control system actuator fault diagnosis method, time sequence signals of a thermal actuator under multiple running states are collected and labeled, and a training sample set is constructed; a physical reference model is constructed based on normal category samples, an enhanced time sequence matrix and a global statistical feature vector are obtained; time sequence graph convolution coding based on a physical variable correlation graph; modulation parameters and attention query vectors are generated by using the global statistical features, coding process adaptive adjustment and time step weighted pooling are realized respectively, and a fusion fault feature vector is obtained; an actuator state category is output through a classification layer. The normal physical reference model is identified offline, a multivariate deviation channel such as a following residual, a pressure reference deviation and a current reference deviation is constructed, fault representation is converted from original signal amplitude detection into physical law deviation degree measurement, and weak fault signals are enhanced.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically a method for diagnosing actuator faults in a thermal control system. Background Technology

[0002] Actuators in thermal control systems, such as pneumatic control valves and hydraulic actuators, are crucial links connecting control decisions and physical processes; their health directly impacts the safe and economical operation of the unit. With the increasing demands for automation and intelligent maintenance in industrial systems, traditional strategies relying solely on periodic maintenance or simple threshold alarms are insufficient to meet the urgent need for early warning and accurate identification of sudden faults. Data-driven intelligent diagnostic methods have begun to receive extensive research; however, relying solely on deep learning models to manually extract fault characteristics from raw time-series signals such as commands, valve positions, pressures, and currents is not a viable approach.

[0003] In practical engineering, we face many fundamental challenges, such as:

[0004] 1. Traditional methods rely directly on the original time-series signals without taking into account physical biases, forcing the model to implicitly identify nonlinear physical laws in high-noise data, resulting in low sensitivity to early or minor faults.

[0005] 2. Conventional multivariate fusion treats different physical quantities as unconstrained independent channels, ignoring the causal topological relationships between variables. It is easy to learn false correlation features on a limited number of fault samples.

[0006] 3. Fixed-structure networks do not distinguish between steady-state and transitional phases and use uniform time pooling, which dilutes the characteristics of short-term abrupt faults and makes it difficult for the same model to effectively capture long-term trend anomalies and local details. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for diagnosing actuator faults in thermal control systems.

[0008] Methods are proposed to address the technical problems mentioned in the background section.

[0009] To achieve the above objectives, the present invention employs the following technical solution:

[0010] A method for diagnosing actuator faults in a thermal control system includes the following steps:

[0011] S1. Collect timing signals of the thermal actuator under different operating conditions, label them, and construct a training sample set for the thermal actuator; the timing signals include control commands, valve position feedback, valve downstream pressure, and winding current; the labels are used to characterize the state category of the thermal actuator.

[0012] S2. Construct a physical baseline model based on time-series signal samples of the normal category. Construct new time-series channels according to the deviation of the measured values ​​of the time-series signals from the physical baseline model. Extract global physical statistical features according to the change process of control commands to obtain the enhanced thermal actuator time-series matrix and global physical statistical feature vector.

[0013] S3. Define multiple channels of the enhanced thermal actuator time series matrix as multiple nodes in the physical variable association graph. Predefine the physical prior adjacency matrix according to the physical action relationship of the actuator. Under the constraint of the physical prior adjacency matrix, introduce trainable edge coefficients so that the multidimensional enhanced features of each time step are first fused with node information on the physical variable association graph, and then enter the subsequent multi-scale time series network along the time dimension.

[0014] S4. Generate modulation parameters and attention query vectors based on the global physical statistical feature vector. The multiple modulation parameters are used to adaptively adjust the encoding process of the multi-scale temporal network, and the attention query vector is used to perform weighted pooling on the time step features output by the multi-scale temporal network to obtain the fused fault feature vector.

[0015] S5. Based on the fused fault feature vector, the classification layer outputs the state category of the thermal actuator.

[0016] Furthermore, the labels include normal, stuck, internal leakage, and abnormal pneumatic or hydraulic power source.

[0017] Furthermore, the physical reference model includes an actuator nominal follower model, a pressure-valve position reference function, and a current-command reference function;

[0018] The deviations include valve position feedback deviation, valve downstream pressure deviation, and winding current deviation.

[0019] The multiple channels of the enhanced thermal actuator timing matrix include: control command, valve position feedback, valve downstream pressure, winding current, valve position feedback deviation, valve downstream pressure deviation, winding current deviation, valve position change rate, and pressure change rate.

[0020] Furthermore, S2 specifically refers to:

[0021] The nominal follower model of the actuator is identified using control commands and valve position feedback from normal category samples, and the follower residual is calculated step by step to ensure that the model input includes the actuator action deviation.

[0022] Construct pressure-valve position reference function and current-command reference function, and calculate the deviation of measured values ​​from normal reference functions;

[0023] A steady-state-transition segmented mask is generated based on the rate of change of the control command, and global physical statistical features are extracted under the constraints of the segmented mask.

[0024] Furthermore, S3 specifically refers to:

[0025] A physical variable association graph containing 9 nodes is established. The physical variable association graph is a graph structure with the 9 timing channels in the timing matrix of the enhanced thermal actuator as nodes and the physical interaction relationship of the thermal actuator as edges.

[0026] Based on the physical prior adjacency matrix, a trainable edge coefficient matrix is ​​introduced, and self-loops are added to obtain an adaptive physical adjacency matrix;

[0027] For each time step Take the first one from the timing matrix of the enhanced thermal actuator Line up and transpose to construct the node feature matrix;

[0028] Perform two layers of graph convolution on the node feature matrix to obtain the node-level graph convolution feature matrix;

[0029] The node-level graph convolutional feature matrices are converged along the node dimension to obtain the first... Graph convolutional feature vectors at each time step;

[0030] The node feature matrix construction, graph convolution calculation, and node dimension aggregation are repeated for all time steps to obtain the graph-enhanced temporal feature map.

[0031] Furthermore, the modulation parameters include multi-scale convolutional branch weights and channel modulation parameters, which are generated based on the global physical statistical feature vector, specifically:

[0032] Using graph-enhanced temporal feature maps as input feature maps, multiple dynamically modulated dilated convolutional blocks are stacked to extract temporal features.

[0033] The global physical statistical feature vector is input into a branch weight generation network consisting of a linear layer, a GELU activation function, a linear layer, and a Softmax function to obtain the branch fusion weight vector of the dynamically modulated dilated convolutional block.

[0034] The outputs of multiple dilated convolution branches are weighted and fused according to the branch fusion weight vector to obtain the multi-scale candidate feature map of the dynamically modulated dilated convolution block. The multi-scale candidate feature map is a candidate temporal feature map after fusing receptive field features from different times.

[0035] The channel modulation scaling vector and channel modulation translation vector of the dynamically modulated dilated convolutional block are generated using global physical statistical feature vectors.

[0036] Multiply the feature value of each channel of the multi-scale candidate feature map by the scaling parameter of the corresponding channel, and add the translation parameter of the corresponding channel to obtain the modulated feature value;

[0037] Residual connections, layer normalization, and GELU activation are performed on the modulated feature map to obtain the output feature map of the dynamically modulated dilated convolutional block.

[0038] After sequentially completing multiple dynamically modulated dilated convolutional blocks, a deep temporal feature map is obtained;

[0039] Furthermore, an attention query vector is generated based on the global physical statistical feature vector. This attention query vector is used to perform weighted pooling of the time-step features output by the multi-scale temporal network to obtain a fused fault feature vector, specifically:

[0040] Input the global physical statistical feature vector into the trainable linear projection layer to generate a global query vector;

[0041] Calculate the similarity between the global query vector and the feature vector at each time step in the deep temporal feature map;

[0042] The time step features of the deep temporal feature map are weighted and summed according to the time attention weight to obtain the deep fault feature vector.

[0043] The deep fault feature vector and the global physical statistics feature vector are concatenated along the feature dimension to obtain the fused fault feature vector.

[0044] Furthermore, based on the deep fault feature vector, the classification layer outputs the state category of the thermal actuator, specifically:

[0045] The fused fault feature vector is input into the classification fully connected layer to obtain the state classification logical value vector;

[0046] Perform Softmax normalization on the state classification logical value vector to obtain the posterior probability vector of the state category;

[0047] The diagnostic result is output based on the posterior probability vector of the state category. If the maximum posterior probability is less than the confidence threshold, a low confidence warning is output, and the subsequent diagnostic window is observed. If the maximum posterior probability is not less than the confidence threshold, the state category corresponding to the maximum posterior probability is used as the candidate diagnostic result for the current diagnostic window, and the confidence threshold is set. This represents the minimum probability threshold for a diagnosis to be accepted.

[0048] Furthermore, it also includes the consistency confirmation of the execution time of candidate diagnostic results in multiple consecutive diagnostic windows. When the same fault category is the maximum a posteriori probability category in multiple consecutive diagnostic windows, and the corresponding maximum a posteriori probability is not less than the confidence threshold, a formal fault alarm is output. If the candidate diagnostic results of adjacent diagnostic windows change frequently, no alarm is set off, and subsequent diagnostic windows are collected.

[0049] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0050] 1. This invention identifies normal physical reference models offline and constructs multi-dimensional deviation channels such as following residual, pressure reference deviation and current reference deviation, transforming fault characterization from the detection of original signal amplitude to the measurement of deviation of physical law, thereby enhancing weak fault signals.

[0051] 2. This invention predefines the enhanced physical variables as graph nodes and constructs a physical prior adjacency matrix based on the causal mechanism of thermal actuators. Under constraints, graph convolution information is fused to avoid introducing false correlations in unconstrained fully connected fusion.

[0052] 3. This invention extracts global physical statistical features of the transition segment and steady-state segment, uses macroscopic statistics to adaptively modulate the branch weights and channel responses of multi-scale convolution, and guides attention pooling to focus on key time steps, so that the model structure can be dynamically adjusted with the running state. Attached Figure Description

[0053] Figure 1 This is a flowchart of the overall process of this invention;

[0054] Figure 2 This is a schematic diagram of the thermal generator training sample process of the present invention;

[0055] Figure 3 This is a schematic diagram of the pressure reference deviation of the present invention;

[0056] Figure 4 This is a schematic diagram of the fault diagnosis process of the present invention. Detailed Implementation

[0057] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0058] like Figure 1 As shown, this invention proposes a method for diagnosing actuator faults in a thermal control system, the main contents of which are as follows:

[0059] S1, Multi-source timing sample acquisition of thermal actuator

[0060] like Figure 2 As shown, four types of timing signals are collected from the thermal actuator under different operating conditions: control commands, valve position feedback, valve downstream pressure, and winding current. Among them, the control commands are used to characterize the opening or displacement target that the control system expects the actuator to achieve, the valve position feedback is used to characterize the actual valve opening achieved by the actuator, the valve downstream pressure is used to characterize the change in the working fluid pressure downstream of the valve, and the winding current is used to characterize the electromagnetic load and driving force state of the actuator drive mechanism.

[0061] During data acquisition, a single thermal actuator sample is recorded as the original thermal actuator timing matrix. , dimension Where 128 represents the number of consecutive time steps in a single sample, and 4 represents the number of signal acquisition channels; the original thermal actuator timing matrix. The four channels are, in order, control command, valve position feedback, valve downstream pressure, and winding current.

[0062] No. The control instructions for each time step are denoted as , This refers to the original thermal actuator timing matrix. In the The control command sample value at the first time step; The valve position feedback at each time step is recorded as follows: , This refers to the original thermal actuator timing matrix. In the The actual valve position sampled value at each time step; The post-valve pressure at each time step is denoted as... , This refers to the original thermal actuator timing matrix. In the The valve post-pressure sampling value at the first time step; The winding current at each time step is denoted as... , This refers to the original thermal actuator timing matrix. In the The current sampling value of the drive winding at each time step; This represents the time step number, with a value range from 1 to 128.

[0063] In practical implementation, for each signal channel (control command, valve position feedback, downstream pressure, winding current) in the original thermal actuator timing matrix, the minimum and maximum values ​​of all samples in the training set for that channel are used to linearly map the sampled values ​​of that channel to the [0,1] interval using a minimum-maximum normalization method, resulting in a normalized thermal actuator timing matrix. This normalized matrix is ​​used to avoid the impact of differences in the dimensions of different engineering quantities on subsequent physical deviation calculations and neural network training. The maximum, minimum, mean, or standard deviation used for normalization are all obtained statistically from the training set and remain fixed during the verification, testing, and online inference phases, without being recalculated based on a single sample to be diagnosed.

[0064] Furthermore, each thermal actuator sample is labeled with a state category. The number of state categories is denoted as... In this invention, four state categories can be selected, namely normal, jammed, internal leakage, and pneumatic or hydraulic source abnormality. The normal category corresponds to samples where the control command, valve position feedback, downstream pressure, and winding current all conform to the normal actuator operation pattern. The jammed category corresponds to samples where the valve position feedback exhibits partial stagnation, jumps, increased hysteresis, or an abnormally high winding current. The internal leakage category corresponds to samples where the downstream pressure is abnormally high in the valve's closed section or small-opening steady-state section. The pneumatic or hydraulic source abnormality category corresponds to samples where the actuator's overall following speed slows down, the driving force is insufficient, or the overall winding current deviates.

[0065] This ultimately forms a training sample set for thermal actuators. , recorded as ,in, Indicates the sample index. This represents the total number of training samples. Indicates the first The original thermal actuator timing matrix, Indicates the first The actual state category label corresponding to the timing matrix of each original thermal actuator.

[0066] S2. Enhanced characterization of multivariate deviations based on normal physical benchmarks

[0067] When a thermal actuator is operating normally, there is a dynamic response relationship between the control command and the valve position feedback determined by the mechanical following characteristics. There is a nonlinear correspondence between the valve position feedback and the downstream pressure determined by the valve flow characteristics. There is a correspondence between the control command and the winding current determined by the electromagnetic drive characteristics. When jamming, internal leakage, or abnormality of the pneumatic or hydraulic source occurs, the fault information in the original thermal actuator timing matrix often manifests as a deviation from normal physical laws, rather than a simple abnormality in the absolute amplitude of a single channel.

[0068] This invention uses normal category samples to offline construct the actuator nominal following model, pressure-valve position reference function and current-command reference function, and then constructs the deviation of the measured value from the normal physical reference as a new timing channel, and extracts global physical statistical features according to the control command change process to obtain the enhanced thermal actuator timing matrix and global physical statistical feature vector. The specific steps are as follows.

[0069] S201, Construction of Following Residuals and Residual Differences

[0070] The response from control commands to valve position feedback is not instantaneous; it involves mechanical inertia and a pure time lag. In a normal thermal actuator, after receiving a control command, the valve position feedback changes according to a relatively stable dynamic pattern. However, when the thermal actuator jams or the pneumatic or hydraulic power source malfunctions, the valve position feedback will deviate significantly from the normal dynamic pattern.

[0071] This invention identifies the nominal following model of the actuator offline using normal category samples and calculates the following residuals step by step, so that the model input includes the actuator action deviation. The specific steps are as follows:

[0072] 1> From the training sample set of thermal actuators Normal category samples are selected, and the control commands and valve position feedback from the normal category samples are used to identify the actuator nominal following model. The actuator nominal following model refers to the dynamic response model from control command to valve position feedback obtained by fitting normal category samples. It is used to characterize the valve position response law that a fault-free thermal actuator should have under the same control command.

[0073] In the specific implementation, a set of candidate pure time delay steps is pre-set. , This refers to the candidate set used to search for the pure time delay steps of the executor, for example, the set of possible values. For each candidate pure time delay step... The first normal category sample Valve position feedback at each time step and the Control commands for each time step As the independent variable, the first Valve position feedback at each time step As the dependent variable, the first-order dynamic relationship is fitted using the least squares method. Furthermore, the number of pure time delays for this candidate variable is calculated. The average prediction error of all normal category samples is used, and the candidate pure lag step with the smallest average prediction error is selected as the identification pure lag step. .

[0074] In practical implementation, the actuator nominal follower model can be represented as follows: ,in, Indicates the first The nominal predicted valve position feedback for each time step refers to the valve position feedback value predicted according to the normal dynamic law of the actuator. This represents the inertia retention coefficient, used to characterize the inertial influence of the valve position feedback from the previous time step on the valve position feedback from the current time step. Example values ​​can be 0.80 to 0.95. This represents the control drive coefficient, used to characterize the driving effect of control commands on valve position feedback changes. Example values ​​can be from 0.05 to 0.20. This represents the bias coefficient of the following model, used to compensate for the zero-point offset of the valve position feedback; This indicates the number of pure time delay steps for identification, which refers to the number of time steps after the control command is issued before the valve position feedback begins to produce an effective response. Example values ​​can be 2 to 4. Furthermore, , , and All samples were determined offline using normal category samples and were used consistently during subsequent model training and online inference.

[0075] It should be noted that, , and After identifying pure time delay steps Once determined, normal category sample data is used to... and For independent variable, The optimal coefficients are obtained by fitting the linear equation using the least squares method, with as the dependent variable.

[0076] In one embodiment, as an example, to illustrate the identification process of the actuator nominal follower model, assume that all "normal" category samples are selected from the training set, and a candidate pure lag step set is preset. For candidate values Take a time step from a normal sample ,but:

[0077] Independent variable 1: the first Valve position feedback at each time step ;

[0078] Independent variable 2: the first Control commands for each time step ;

[0079] Dependent variable: the first Valve position feedback at each time step ;

[0080] Using all normal samples in Batch data constructed below Solve for a set of coefficients using the least squares method, such that The sum of squared prediction errors is minimized; for each All values ​​were fitted in this way, and their average prediction error was calculated; by comparison, it was found that... If the average prediction error is minimized at a given time, then the number of pure time delay steps can be identified. The value is determined to be 3.

[0081] 2> For each thermal actuator sample to be processed, calculate the nominal predicted valve position feedback step-by-step according to the actuator's nominal following model. Specifically, for the first... The time step, using the first Actual valve position feedback at each time step and the Control commands for each time step Calculate nominal predictive valve position feedback (Calculated using a pre-defined actuator nominal following model). For The starting time step can be determined using the control command of the first time step. Perform boundary completion; for Time lack The situation can make This ensures that the nominal predicted valve position feedback can be obtained at each of the 128 time steps.

[0082] Furthermore, the difference between the measured valve position feedback and the nominal predicted valve position feedback is taken as the following residual, and expressed as:

[0083] ;

[0084] in, Indicates the first The following residual at each time step refers to the deviation of the measured valve position feedback from the actuator's nominal following model prediction. It can directly characterize the difference between the actual action of a thermal actuator and its normal following behavior; when the thermal actuator experiences jamming, hysteresis, or insufficient power source, the following residual is... It will continue to increase or show a directional shift during the corresponding time period.

[0085] In one embodiment, for example, if , , , Valve position feedback from the previous time step at the 20th time step The control command after lag compensation (i.e., the first) Control commands for each time step ) The nominal predicted valve position feedback is If the measured valve position feedback Then follow the residual This indicates that the measured valve position feedback is lower than the predicted value of the normal dynamic model, which may correspond to slow start-up or insufficient driving force.

[0086] 3> Calculate the residual difference based on the change of the following residual in adjacent time steps. The residual difference is used to characterize the rate of change of the following residual between adjacent time steps, which can highlight dynamic boundaries such as sudden jump after valve position stagnation, local jamming release or transient impact.

[0087] In specific implementation, the first The residual difference for each time step is calculated as follows: .in, Indicates the first The residual difference at the nth time step refers to the residual difference at the nth time step. The time step follows the residual relative to the first time step Each time step follows the change in the residual; Indicates the first The following residuals at each time step. For the first time step, we can let .

[0088] It should be noted that, following the residual It mainly reflects the deviation range and residual difference. This primarily reflects the rate of deviation from the change. A jamming fault often manifests as slow or stagnant valve position feedback changes over a period of time, followed by a sudden jump; only the following residual is used. It can reflect the degree of deviation, but it is not sensitive enough to mutation boundaries, so residual differencing is introduced. This can enhance the ability of subsequent neural networks to recognize step shifts and intermittent stagnation.

[0089] 4> Will follow the residual and residual difference These are respectively treated as two new timing channels and concatenated in time step order to the normalized thermal actuator timing matrix to obtain the first enhanced thermal actuator timing matrix. , dimension First enhanced thermal actuator timing matrix It is a 6-channel timing matrix formed by adding a follower residual channel and a residual differential channel on the basis of the four channels of control command, valve position feedback, valve downstream pressure and winding current.

[0090] It should be noted that both the residual channel and the residual difference channel are calculated by the actuator's nominal following model and do not rely on additional sensors. Through this processing, subsequent models do not need to rely entirely on training data to learn the normal dynamic response relationship on their own, but can directly receive the "deviation of the measured response from the normal response", thereby improving the sensitivity to early jamming, slight hysteresis, and power source anomalies.

[0091] S202, Combined structure of pressure reference deviation and current reference deviation

[0092] Under normal operating conditions, there is a relatively stable nonlinear relationship between the downstream pressure and the valve position feedback, and a relatively stable driving relationship between the winding current and the control command. Internal leakage faults usually cause the downstream pressure in the valve closing section or the steady-state section with a small opening to deviate from the normal pressure-valve position relationship; jamming faults may cause winding current spikes under the same control command; abnormalities in the pneumatic or hydraulic power source may lead to an overall low winding current or insufficient driving force.

[0093] This invention constructs a pressure-valve position reference function and a current-command reference function, and then calculates the deviation of the measured value relative to the normal reference function, thereby transforming the physical deviation caused by valve internal leakage and actuation abnormalities into deviation features that are easy for neural networks to learn. The specific steps are as follows:

[0094] 1> Construct a pressure-valve position reference function using normal category samples The pressure-valve position reference function is a function that takes valve position feedback as input and normal expected downstream pressure as output. It is used to calculate the downstream pressure reference level of a normal thermal actuator under different valve opening degrees.

[0095] In one implementation, the normalized valve position feedback value range is... Divided into Each valve position range The number of valve position intervals representing the pressure-valve position reference function can be between 20 and 50. Then, for each valve position interval, all time steps in the normal category samples where the valve position feedback falls within that interval are collected, and the median or mean of the downstream pressure corresponding to these time steps is calculated as the pressure reference point for that valve position interval. For valve position intervals lacking sample points, linear interpolation can be performed using the pressure reference points of adjacent valve position intervals. All pressure reference points are connected in valve position order to form a piecewise linear curve representation of the pressure-valve position reference function.

[0096] In one embodiment, as an example, when constructing the pressure-valve position reference function, it is assumed that the number of valve position intervals is... Set to 20. For a certain valve position range. By iterating through all time steps of all normal category samples, 200 time steps were found where the valve position feedback fell within this interval, and the median downstream pressure corresponding to these time steps was 0.22. Therefore, the pressure reference point for this valve position interval is... (Midpoint of the interval) corresponds to the pressure All Connecting the reference points in each interval forms the curve of the pressure-valve position reference function.

[0097] In this implementation, to reduce the impact of random pressure noise on the pressure-valve position reference function, a moving average smoothing can be performed on the pressure reference points along the valve position sequence. The sliding window length can be 3 or 5 valve position intervals. It should be noted that the pressure-valve position reference function is only established using normal class samples during the offline phase and is consistently used during subsequent training and online inference.

[0098] 2> Constructing a current-command reference function using normal category samples The current-command reference function is a function that takes the control command as input and the normal expected winding current as output. It is used to characterize the winding current reference level of a normal thermal actuator under different control commands.

[0099] In one implementation, the normalized control command value range is... Divided into One control command range, The number of control command intervals representing the current-command reference function can be between 20 and 50. Then, for each control command interval, all time steps in the normal category samples where the control command falls within that interval are collected, and the median or mean of the winding current corresponding to these time steps is calculated as the current reference point for that control command interval. For control command intervals lacking sample points, linear interpolation can be performed using the current reference points of adjacent control command intervals. All current reference points are connected in the order of the control commands to form a piecewise linear curve representation of the current-command reference function.

[0100] In one embodiment, as an example, when constructing the current-command reference function, it is assumed that the number of control command intervals will be... Set to 20. For a certain instruction range By iterating through all time steps of all normal category samples, 300 time steps were found where the control command fell within this interval, and the average winding current corresponding to these time steps was 0.60. Therefore, the current reference point for this command interval is... Current corresponding to (midpoint of interval) All Connecting the reference points in each interval forms the curve of the current-command reference function.

[0101] In another implementation, to distinguish between the short-term current rise caused by normal operation transition and the current spike caused by abnormal jamming, the control command can be... The absolute value of the difference between the control command and the control command Using these as inputs, a two-dimensional current reference table is established. The first dimension of the two-dimensional current reference table is the control command range, and the second dimension is the range of absolute values ​​of the control command differences. The current reference value within each two-dimensional grid cell is determined by the mean or median of the winding currents falling into that grid cell in the normal category samples. This method enables the current reference to better adapt to the normal current differences during steady-state maintenance and operational transitions.

[0102] In one embodiment, as an example, assuming a two-dimensional current reference table is used, the first dimension of the table represents the control command range, and the second dimension represents the range of absolute values ​​of the control command differences. For example, the normalized control command value range... Divide the control command into 20 intervals and calculate the absolute value of the difference after normalization. The range of values ​​is also divided into 10 intervals, forming a A two-dimensional grid. For each grid cell, control commands are collected from all normal category samples. The absolute value of the difference between the control command and the control command Simultaneously, the time steps falling within the cell are used to calculate the median of the winding current corresponding to these time steps, which is then used as the current reference value for that cell. If a sample point for a cell is missing, two-dimensional linear interpolation is performed using the current reference values ​​of adjacent cells.

[0103] In this embodiment, during online computation, if at a certain time step... , If the current falls within the 11th control command interval and the 4th differential absolute value interval respectively, then the current reference value within that grid cell is obtained by querying the two-dimensional current reference table as the normal expected winding current for the current time step. In this method, for the same control command... ,when (Corresponding to the transition phase of the action) and (During the steady-state phase), different current reference values ​​can be retrieved, thus avoiding misinterpreting reasonable current increases during normal operation transitions as abnormal current spikes. Among these, Indicates the first The absolute value of the control command difference for each time step.

[0104] 3> For each thermal actuator sample to be processed, calculate the pressure reference deviation step by step. Specifically, the first time step... Valve position feedback at each time step By inputting the pressure-valve position reference function, the expected downstream pressure corresponding to the valve position feedback under normal operating conditions can be obtained. Then the first Measured post-valve pressure at each time step Subtract the normal expected valve pressure , obtained the The pressure reference deviation at each time step, i.e. .

[0105] in, Indicates the first The pressure reference deviation at each time step refers to the deviation of the measured downstream pressure from the pressure-valve position reference function. It refers to the first The normal expected downstream pressure corresponding to each time step. Pressure reference deviation. A positive value that continues to increase at small openings or during steady-state closure indicates that the downstream pressure is higher than the normal reference level, suggesting internal leakage or abnormal flow; pressure reference deviation Severe fluctuations in the transition section can indicate instability in the actuator flow regulation process.

[0106] In one embodiment, as an example, valve position feedback at a certain time step The normal expected downstream pressure output by the pressure-valve position reference function is: Actual measured pressure after valve Then the pressure reference deviation Since the valve position feedback is close to closed or slightly open, and the downstream pressure is significantly higher than the normal reference level, this time step is more likely to contain information related to internal leakage.

[0107] 4> For each thermal actuator sample to be processed, calculate the current reference deviation step by step. Specifically, the first... Control commands for each time step The input current-command reference function yields the expected normal winding current corresponding to the control command under normal operating conditions. Then the first Measured winding current at each time step Subtract the normal expected winding current , obtained the The current reference deviation at each time step, i.e. .

[0108] in, Indicates the first The current reference deviation at each time step refers to the deviation of the measured winding current from the current-command reference function. Indicates the first The normal expected winding current corresponding to each time step. Current reference deviation. The local positive spikes can reflect the increased frictional load caused by jamming; current reference deviation A sustained negative offset can reflect insufficient driving force caused by abnormal pneumatic or hydraulic sources.

[0109] In the implementation using a two-dimensional current reference table, the first... Control commands for each time step The absolute value of the difference between the control command and the control command By inputting a two-dimensional current reference table, the expected winding current is obtained, and then the current reference deviation is calculated. This method can avoid misjudging reasonable current increases during normal operation transitions as abnormal current spikes.

[0110] In one embodiment, as an example, in an implementation using a two-dimensional current reference table, let the first... Control commands for each time step Control command differential absolute value By inputting both into a two-dimensional current reference table, the normal expected winding current can be obtained. If the measured winding current at that time step Then the current reference deviation This positive deviation indicates that, under the current command and action intensity, the measured current is higher than the normal reference level, possibly corresponding to increased frictional load caused by jamming. If a one-dimensional current-command reference function that depends solely on the control command is used, the normal expected winding current under this command might be... The calculated deviation is This may include a reasonable current increase due to normal operational transitions. A two-dimensional current reference table introduces... Distinguishing between steady-state and transient states allows for more accurate deviation calculations. Among these, Indicates the first Current reference deviation at each time step This represents the expected normal winding current obtained by jointly querying the two-dimensional current reference table based on the control command and its differential absolute value. Indicates the first The measured winding current at each time step.

[0111] 5> Adjust the pressure reference deviation and current reference deviation These are respectively used as two new timing channels, and are concatenated into the timing matrix of the first enhanced thermal actuator in time step order. Then, the timing matrix of the second enhanced thermal actuator is obtained. , dimension Second enhanced thermal actuator timing matrix It is in the timing matrix of the first enhanced thermal actuator Based on this, an 8-channel timing matrix is ​​formed by adding a pressure reference deviation channel and a current reference deviation channel.

[0112] It should be noted that the pressure reference deviation channel and the current reference deviation channel are not simple repetitions of the valve downstream pressure and winding current, but rather abnormal response quantities obtained by comparing the measured physical quantities with normal reference characteristics. They can reflect the physical deviations corresponding to internal leakage, jamming, and abnormalities of pneumatic or hydraulic sources, reducing the difficulty for the neural network to learn complex nonlinear reference relationships from the original thermal actuator timing matrix.

[0113] In one embodiment, such as Figure 3 As shown, the pressure reference deviation is analyzed. For the internal leakage sample, the upper subplot shows the time step on the horizontal axis and the normalized value on the vertical axis, displaying the comparison between the measured valve position feedback, the measured downstream pressure, and the normal expected downstream pressure. The lower subplot shows the pressure reference deviation in the form of an area plot, with the time step on the horizontal axis and the normalized deviation on the vertical axis. The deviation channel obtained by comparing the measured pressure with the normal pressure-valve position reference function can capture the pressure anomalies caused by internal leakage.

[0114] S203, Steady-state-transition segmentation and global physical statistical feature extraction

[0115] Thermal actuator failures manifest differently in the steady-state and transition phases. Jamming usually causes valve position start-up delay in the early stages of control command changes. Abnormalities in pneumatic or hydraulic sources usually prolong the transition process. Internal leakage usually manifests as pressure abnormalities in the small opening or closed steady state. If a fixed window statistical method is used, steady-state noise and transition fault signals will be mixed, resulting in insufficient stability of sample-level characteristics.

[0116] This invention generates a steady-state-transition segmented mask based on the rate of change of control commands, and extracts global physical statistical features under the constraints of the segmented mask. This allows the extracted statistical features to more specifically capture fault characteristics at different operating stages, providing more stable sample-level information for subsequent diagnosis. The specific steps are as follows:

[0117] 1> Calculate the absolute value of the control command difference step by step. The absolute value of the control command difference is used to characterize the magnitude of the change in the control command between adjacent time steps, i.e. .

[0118] in, Indicates the first The absolute value of the control command difference at each time step Indicates the first The control commands for each time step are differential; for the first time step, it can be set as follows: The larger the absolute value of the control command difference, the more likely the thermal actuator is in the command-driven transition phase; the smaller the absolute value of the control command difference, the more likely the thermal actuator is in the steady-state holding phase.

[0119] Furthermore, to reduce frequent state switching caused by minor jitter in control commands, a dual-threshold hysteresis segmentation method is used to generate a binary segmented mask. Binary segmentation mask It refers to the first The status markers for each time step Indicates the first This time step belongs to the transition period. Indicates the first This time step belongs to the steady state period.

[0120] In practical implementation, let the low change threshold be... High change threshold is ,in, This indicates the lower limit threshold of the control command change when returning from the transition phase to the steady-state phase. This indicates the upper limit threshold of the control command change when transitioning from the steady-state phase to the transition phase, and .when At that time, Marked as 1; when At that time, Marked as 0; when lie in and In between, make Maintain the state flag from the previous time step. If the control commands have been normalized to... , 0.02 is acceptable. 0.10 is acceptable.

[0121] In one embodiment, for example, if the absolute value of the control command difference at a certain time step is 0.15, then because 0.15 is greater than the high change threshold of 0.10, this time step is marked as a transition segment; if the absolute value of the control command difference at a certain time step is 0.01, then because 0.01 is less than the low change threshold of 0.02, this time step is marked as a steady-state segment; if the absolute value of the control command difference at a certain time step is 0.05, then because 0.05 is between 0.02 and 0.10, this time step retains the state label of the previous time step. This hysteresis mechanism avoids the use of binary segmentation masks. It frequently jumps when the instruction fluctuates slightly.

[0122] 2> Based on the binary segmentation mask Determine the set of time steps for the transition period and the set of time steps for the steady state period. The set of time steps for the transition period is denoted as... That is, all that satisfy The set of time steps in steady state; the set of time steps in steady state is denoted as That is, all that satisfy A set consisting of time steps.

[0123] Furthermore, the number of steps during the transition period was statistically analyzed. and the number of steady-state time steps ,in, Refers to the set of time steps during the transition period The number of time steps in the middle. Refers to the set of steady-state time steps The number of time steps in the time step.

[0124] 3> Set of time steps during the transition period Calculate the absolute value of the average residual of the transition section. absolute value of average residual in the transition section This refers to the following residual at all transition time steps. The statistic obtained by averaging the absolute values ​​is used to reflect the overall following deviation intensity of the thermal actuator relative to the nominal following model during command changes. The absolute value of the average residual in the transition section... The larger the value, the more difficult it is for the thermal actuator to respond to control commands according to normal dynamic laws during the transition process, and the more likely there will be jamming or abnormality of the pneumatic or hydraulic source.

[0125] In practical implementation, the absolute value of the average residual of the transition section It is the set of transition time steps Follow-up residuals at all time steps Take the absolute value and then calculate the arithmetic mean, that is, the arithmetic mean of the values. within Sum and divide by the number of transition time steps .

[0126] Furthermore, during the transition period, the time step set The average pressure reference deviation of the transition section is calculated above. Average pressure reference deviation in the transition section This refers to the deviation from the pressure reference at all transition time steps. The statistic obtained by averaging is used to measure the degree to which the downstream pressure of a thermal actuator deviates from the pressure-valve position reference function during dynamic regulation. (Average pressure reference deviation in the transition section) It can provide supplementary descriptions of pressure anomalies during the flow regulation process.

[0127] In practical implementation, the average pressure reference deviation of the transition section It is the set of transition time steps Pressure reference deviation at all time steps Find the arithmetic mean, that is, the average of the two sides. within Sum and divide by the number of transition time steps .

[0128] 4> Based on the number of time steps in the transition period The transition duration is calculated as a percentage of the total time steps (128). The proportion of the duration of the transition phase This refers to the proportion of time steps in the transition phase out of the total number of time steps in the sample, used to describe the dynamic adjustment time of the thermal actuator in that sample. Under normal operating conditions, the duration of the transition section accounts for a certain percentage. Typically at a lower level; when a pneumatic or hydraulic source malfunction causes the actuator response to slow down, the transition period duration accounts for a significant portion. It will rise.

[0129] In practical implementation, the duration of the transition segment accounts for a certain percentage. It is the number of time steps in the transition period. The ratio obtained by dividing by the total number of time steps, 128, is... .

[0130] Furthermore, during the steady-state time step set The average current reference deviation in the steady-state section is calculated above. Steady-state average current reference deviation This refers to the deviation of the current reference over all steady-state time steps. The averaged statistic reflects the degree to which the drive current of the thermal actuator deviates from the current-command reference function during the holding phase. The average current reference deviation in the steady-state phase is... A persistently high level can indicate increased frictional load or jamming; the average current reference deviation in the steady-state range... A persistently low level can indicate insufficient driving force.

[0131] In practical implementation, the average current reference deviation in the steady-state range It is the set of steady-state time steps Current reference deviation at all time steps Find the arithmetic mean, that is, the average of the two sides. within Summing and dividing by the number of steady-state time steps .

[0132] It should be noted that if a sample does not have a transition time step, then the absolute value of the average residual of the transition time step will be used. Deviation of average pressure reference in the transition section Set to 0, and through This indicates that the sample has no obvious transition period; if a sample does not have a steady-state time step, then the steady-state average current reference deviation will be... Set it to 0 to avoid calculation uncertainty caused by averaging an empty set.

[0133] 5> The absolute value of the average residual of the transition section Average pressure reference deviation in the transition section The proportion of the duration of the transition phase Deviation of average current reference in steady-state region The global physical statistical feature vector is formed in a fixed order. ,Right now ,in, Used to characterize the intensity of the follow-through deviation during the transition segment; Used to characterize the degree of pressure deviation in the transition section; Used to characterize the duration of the transition segment; Used to characterize the degree of deviation of the current in the steady-state segment.

[0134] 6> The binary segmented mask consisting of 128 time steps The resulting mask sequence serves as the 9th timing channel and is concatenated into the timing matrix of the second enhanced thermal actuator according to the time step order. Then, the timing matrix of the enhanced thermal actuator is obtained. , dimension Enhance the timing matrix of thermal actuators It is a timing matrix containing nine channels: control command, valve position feedback, valve downstream pressure, winding current, follower residual, residual differential, pressure reference deviation, current reference deviation, and binary segmented mask.

[0135] It should be noted that the timing matrix of the enhanced thermal actuator... Provides time-step level physical deviation information and global physical statistical feature vectors. It provides macroscopic operating condition information at the sample level, and both are fed into the subsequent diagnostic model, enabling the model to focus on both local fault waveforms and the overall operating distribution of steady-state and transition sections.

[0136] S3. Temporal Graph Convolutional Coding Based on Physical Variable Relationship Graph

[0137] Enhanced thermal actuator timing matrix The nine channels are not independent of each other, but are related by the physical link of the thermal actuator. The control command drives the valve position feedback to change, the valve position feedback affects the downstream pressure, the winding current is affected by the control command and mechanical load, and the follow residual, residual difference, pressure reference deviation and current reference deviation reflect the abnormal deviation of the corresponding physical relationship.

[0138] This invention will enhance the timing matrix of thermal actuators. The nine channels are defined as nine nodes in the physical variable correlation graph. A physical prior adjacency matrix is ​​predefined based on the physical action relationship of the actuator. Then, trainable edge coefficients are introduced under the constraints of the physical prior adjacency matrix, so that the nine-dimensional enhanced features at each time step are first fused with node information on the physical variable correlation graph, and then enter the subsequent multi-scale temporal network along the time dimension. The specific steps are as follows:

[0139] 1> Establish a physical variable correlation graph containing 9 nodes. The physical variable correlation graph refers to the graph that enhances the timing matrix of the thermal actuator. The graph structure consists of 9 timing channels as nodes and the physical action relationship of the thermal actuator as edges. The first node corresponds to the control command channel, the second node corresponds to the valve position feedback channel, the third node corresponds to the valve downstream pressure channel, the fourth node corresponds to the winding current channel, the fifth node corresponds to the following residual channel, the sixth node corresponds to the residual differential channel, the seventh node corresponds to the pressure reference deviation channel, the eighth node corresponds to the current reference deviation channel, and the ninth node corresponds to the binary segmented mask channel.

[0140] The physical prior adjacency matrix is ​​denoted as , dimension , This refers to the pre-defined connection matrix of the physical variable correlation diagram, that is, Indicates the first The node to the first The physical prior connection state of each node. When When, it indicates the first The node and the first There is a clear physical connection between the nodes; when When, it indicates that the first is not preset. The node to the first The direct physical association of each node.

[0141] In one implementation, a control command node can be set to point to the valve position feedback node, winding current node, following residual node, and residual differential node to express the influence of control commands on actuator action and drive current; a valve position feedback node can be set to point to the downstream pressure node, following residual node, and pressure reference deviation node to express the influence of valve position change on pressure change and following deviation; a downstream pressure node can be set to point to the pressure reference deviation node to express the participation of measured pressure in pressure deviation construction; a winding current node can be set to point to the current reference deviation node to express the participation of measured current in current deviation construction; and a binary segmented mask node can be set to point to the following residual node, pressure reference deviation node, and current reference deviation node to express the influence of steady-state segment and transient segment on deviation interpretation.

[0142] In one implementation, the physical prior adjacency matrix It can be set to:

[0143] The control command is fed back to the valve position. The control command points to the winding current; The control command points to the residual; The valve position feedback indicates the downstream pressure. The valve position feedback direction follows the residual. The valve position feedback indicates a deviation from the pressure reference. The pressure after the valve points to the pressure reference deviation; The winding current deviates from the current reference. The binary segmentation mask points to the residual; The binary segmented mask points to the pressure reference deviation; The binary segmented mask points to the current reference deviation; the remaining elements are 0. This matrix encodes the physical causal relationships of the control, execution, feedback, and deviation signals within the thermal actuator.

[0144] It should be noted that the physical variable correlation graph is a graph structure consisting of a set of nodes and a set of edges. Specifically, the set of nodes consists of 9 nodes, each representing the timing matrix of the enhanced thermal actuator. One of the channels, the edge set is determined by the physical prior adjacency matrix Defined as an element in a matrix with a value of 1. This indicates the existence of a line from the first... The node points to the first The directed edges of a node represent physical interactions. For example, setting... In the middle, the control command node (1) points to the element of the valve position feedback node (2). This represents the physical relationship where a control command drives a change in valve position. During model calculation, information can propagate along this edge from node 1 to node 2.

[0145] 2> In the physical prior adjacency matrix Introducing a trainable edge coefficient matrix based on this And by adding self-loops, we obtain an adaptive physical adjacency matrix. , represented as:

[0146] ;

[0147] in, Let the matrix represent the trainable edge coefficients, with dimension 1. , is a trainable parameter used to adjust the strength of physical prior connections; This represents the Sigmoid activation function, which restricts the coefficients of trainable edges to between 0 and 1. This represents element-wise multiplication; This represents a 9th-order identity matrix, used to add self-connections to each node; This indicates a row-wise normalization operation, used to make the sum of the concatenation weights of each row equal to 1; Represents the adaptive physical adjacency matrix, with dimension . It is the node information propagation weight matrix obtained under the physical prior connection constraints.

[0148] It should be noted that, The term keeps positions without physical prior connections at 0, meaning the model can only adjust the strength of existing physical connections and cannot arbitrarily create connections without physical basis. This preserves the data's adaptive capability while limiting the propagation of spurious correlations in the physical variable association graph.

[0149] 3> For each time step From enhancing the timing matrix of thermal actuators Take the first one from the middle Line up and transpose to construct the node feature matrix. , dimension (regard it as) (dimensional matrix), node feature matrix It refers to the first The matrix is ​​formed by arranging the one-dimensional eigenvalues ​​of 9 physical variable nodes at each time step in the order of nodes. The first row corresponds to the control command eigenvalue, the second row corresponds to the valve position feedback eigenvalue, the third row corresponds to the valve downstream pressure eigenvalue, the fourth row corresponds to the winding current eigenvalue, the fifth row corresponds to the following residual eigenvalue, the sixth row corresponds to the residual differential eigenvalue, the seventh row corresponds to the pressure reference deviation eigenvalue, the eighth row corresponds to the current reference deviation eigenvalue, and the ninth row corresponds to the binary segmented mask eigenvalue.

[0150] 4> Node Feature Matrix Perform two layers of graph convolution to obtain the node-level graph convolution feature matrix. Specifically, the node features of the 0th layer graph are denoted as... ,in, Indicates the first The initial graph node feature matrix at the nth time step; for the nth time step Layered graph convolution, the inter-layer propagation method is represented as:

[0151] ;

[0152] in, This represents the graph convolutional layer index, which can be 0 or 1; Indicates the first The first time step The node feature matrix output by layer graph convolution; Indicates the first The trainable weight matrix of the layer graph convolution is used to map node features to a new feature space; Indicates the first Trainable bias vectors for layer graph convolution; This represents the Gaussian error linear unit activation function. If the output dimension of the first layer convolution is 32, and the output dimension of the second layer convolution is 32, then the... The node-level graph convolutional feature matrix is ​​finally obtained at each time step. , dimension Characterized in the first At each time step, the 9 nodes in the physical variable association graph are represented by 32 high-dimensional features after information transmission and fusion related to physical prior connections.

[0153] It should be noted that graph convolution calculation means first using an adaptive physical adjacency matrix. Information is transmitted and fused among the nine physical variable nodes, and then the fused node features are mapped to a high-dimensional feature space through a trainable weight matrix. Based on this, the physical relationships between control commands, valve position feedback, downstream pressure, winding current, and various deviation channels can be modeled at each time step.

[0154] 5> Convolution of the feature matrix at the node level graph Aggregate along the node dimension to obtain the first... The graph convolutional feature vector at each time step .

[0155] In one implementation, the feature vectors of 9 nodes (i.e., the node-level graph convolution feature matrix) are calculated simultaneously. The average pooling and max pooling results of the 9 row vectors are then concatenated along the feature dimension to obtain the graph convolution feature vector. , dimension The average pooling result is used to retain the overall association information of all physical variable nodes, while the max pooling result is used to retain the responses of key nodes with high anomaly intensity.

[0156] 6> The node feature matrix construction, graph convolution calculation, and node dimension aggregation are repeated for all 128 time steps to obtain the graph-enhanced temporal feature map. , dimension ,Graph Enhanced Temporal Feature Map It refers to the temporal feature map after physical variable correlation graph encoding, where each row corresponds to a graph convolution feature vector of a time step.

[0157] It should be noted that the physical variable correlation graph limits the range of node interactions through a predefined physical prior adjacency matrix and adjusts the strength of physical connections through a trainable edge coefficient matrix. This design differs from directly performing unconstrained fusion on all channels, and can reduce spurious correlations in small-sample fault diagnosis scenarios, improving the model's stability under different operating conditions and noise levels.

[0158] It should also be noted that the specific structure of the temporal graph convolutional coding operation is as follows:

[0159]

[0160] S4, Multi-scale Temporal Coding and Attention Pooling

[0161] Enhanced temporal feature graph While the information on the correlation of physical variables within the same time step has been integrated, thermal actuator failures also exhibit dynamic behaviors at different time scales. For example, jamming may manifest as local stagnation and sudden jumps in a short period of time, while pneumatic or hydraulic source anomalies may manifest as slow response over a longer period of time. Internal leakage may cause continuous pressure deviations in the steady state. Single-scale temporal convolution is difficult to cover both short-term abrupt changes and long-term trends simultaneously; ordinary average pooling can also easily dilute critical failure moments.

[0162] This invention utilizes global physical statistical feature vectors Multi-scale convolutional branch weights and channel modulation parameters are generated to enable the temporal coding process to adaptively adjust according to the macroscopic physical state of the current sample. Then, global physical statistical feature vectors are used. An attention query vector is generated, and the features from 128 time steps are weighted and pooled to obtain a fixed-length deep fault feature vector for global awareness. This ensures that the features used for classification can focus on both local key moments and the global operational state. The specific steps are as follows:

[0163] S401, Multi-scale dilated convolutional coding with global physical statistical feature modulation

[0164] Multi-scale dilated convolution is used to extract fault dynamic features under different temporal receptive fields. Global physical statistical feature modulation is used to adjust the response intensity of different time-scale branches and different channels according to the macroscopic physical state of the current sample. The specific steps are as follows:

[0165] 1> Enhance the graph with temporal feature maps As the input feature map of layer 0, denoted as Then, stack them. Temporal feature extraction is performed using dynamically modulated dilated convolutional blocks, where... This indicates the number of dynamically modulated dilated convolutional blocks, which can be 3; the... The input to each dynamically modulated dilated convolutional block is The output is , This represents the index of the dynamically modulated dilated convolution block, with a value ranging from 1 to... .

[0166] In the specific implementation, in the first... Within each dynamically modulated dilated convolutional block, set A parallel dilated convolution branch, Option 3 is acceptable. The dilation rate of each dilated convolution branch is denoted as , Used to control the interval between adjacent sampling points in the temporal convolution kernel. For example, the dilation rate of the three dilated convolution branches can be set to... , and The branch with an inflation rate of 1 is used to extract short-term local changes, the branch with an inflation rate of 2 is used to extract response hysteresis over a medium time range, and the branch with an inflation rate of 4 is used to extract slow trends over a longer time range.

[0167] In the specific implementation, each dilated convolution branch uses a 1D temporal convolution with a kernel length of up to 3. The convolution direction is along the time dimension, and the ends of the time series are filled with zeros or boundary copy padding to ensure that the time length of each branch output remains 128. In the dynamically modulated dilated convolution block, the first... The convolution kernel of each branch is denoted as... , Indicates the first The first dynamically modulated dilated convolutional block Trainable convolutional kernels with dilated convolutional branches are used to extract features from the input feature map. Extract dynamic features corresponding to the time scale. The output of each branch is denoted as , dimension , Indicates the first The number of output channels for a dynamically modulated dilated convolution block can be either 64 or 128.

[0168] It should be noted that the specific structure of the dynamically modulated dilated convolution block is as follows:

[0169]

[0170] 2> Utilize the global physical statistical feature vector Input the branch weights to generate the network, and obtain the first branch weights. Branch fusion weight vector of each dynamically modulated dilated convolutional block Branch fusion weight vector ,in, Indicates the first In the dynamically modulated dilated convolution block, the first... The fusion weights of the three dilated convolution branches are equal to 1.

[0171] In one implementation, the branch weight generation network consists of a single hidden layer multilayer perceptron and a softmax function. Specifically, the global physical statistical feature vector is first... The input linear layer is mapped to the hidden dimension (the hidden dimension can be 32), then the hidden state is obtained through the GELU activation function. Finally, the output linear layer outputs three branch scores, and the softmax function is used to obtain the branch fusion weight vector. When the duration of the transition segment accounts for a certain percentage When the inflation rate is high, the branch weight generation network can learn to increase the branch weights with an inflation rate of 4, making the model more concerned with long-term slow response; when the local mutations corresponding to the residual differences are obvious, the branch weight generation network can learn to increase the branch weights with an inflation rate of 1, making the model more concerned with short-term stuck jumps.

[0172] It should be noted that the dynamically modulated dilated convolutional blocks are stacked sequentially, and the input feature map of each block... The parameters of the convolutional kernels within each block differ. Therefore, each block requires its own independent branch weight generation network to generate weights based on global physical statistics. Dynamically adjust the fusion ratio of each branch within the current block. Although the network input is the same, the network's trainable parameters belong to different indices ( ) block.

[0173] 3> According to the branch fusion weight vector The outputs of multiple dilated convolution branches are weighted and fused to obtain the first... Multi-scale candidate feature maps of dynamically modulated dilated convolutional blocks , dimension Multi-scale candidate feature maps It is a candidate temporal feature map after fusing receptive field features from different times, and its first... The time step, the first The characteristic value of each channel is Through this weighted fusion, multi-scale candidate feature maps are generated. It can adaptively select short-term, medium-term, or long-term dynamic features based on the global physical state of the current sample.

[0174] in, Indicates the first The output of each dilated convolution branch The Middle The time step, the first Branch eigenvalues ​​of each channel; Indicates the first The fusion weights of the dilated convolution branches.

[0175] 4> Utilizing global physical statistical feature vectors Generate the first Channel modulation scaling vector of each dynamically modulated dilated convolution block and channel modulation translation vector Among them, the channel modulation scaling vector The dimension is This is used to control the amplification or suppression level of each output channel; channel modulation shift vector. The dimension is It is used to adjust the baseline activation level of each output channel.

[0176] In one implementation, the channel modulation scaling vector It is generated by a single hidden layer multilayer perceptron. Specifically, it generates the global physical statistical feature vector. The input linear layer is mapped to a hidden dimension (the hidden dimension can be 32 or 64), then passes through the GELU activation function, and finally passes through a linear layer for output. A dimensional vector is generated, and 1 is added to the output to make the initial modulation approximately unit scaled. Furthermore, the channel modulation translation vector... It is generated by another single-hidden-layer multilayer perceptron with the same structure but independent parameters. This is due to the global physical statistical feature vector. Includes the absolute value of the average residual of the transition section, the average pressure reference deviation of the transition section, the proportion of the transition section duration, and the average current reference deviation of the steady-state section, as well as the channel modulation scaling vector. and channel modulation translation vector It can dynamically adjust the convolution response based on the macroscopic state of the sample.

[0177] 5> Modulate the channel modulation scaling vector and channel modulation translation vector Broadcast along the time dimension to 128 time steps, and perform multi-scale candidate feature maps. Perform channel-level affine modulation. Specifically, for the first... The time step and the first Each output channel will display the candidate feature values. Multiply by the scaling parameter of the corresponding channel In addition, the corresponding channel translation parameters The modulated feature values ​​are obtained, i.e. .

[0178] in, Indicates the first In the dynamically modulated dilated convolution block, the first... The time step, the first Modulated characteristic values ​​of each channel; Represents the channel modulation scaling vector The first in One element; Represents the channel modulation shift vector The first in Each element.

[0179] 6> The modulated feature map (i.e., the feature map obtained after performing channel-level affine modulation on the multi-scale candidate feature map, its first...) Line 1 Column elements are Perform residual connections, layer normalization, and GELU activation to obtain the first... Output feature map of a dynamically modulated dilated convolution block .

[0180] In practical implementation, if the input feature map Number of channels and number of output channels If they are consistent, then directly input the feature map. As a residual term; if the number of channels is inconsistent, a 1×1 one-dimensional convolution or linear projection is used to transform the input feature map. Mapped to The dimension is used as a residual term. Among them, layer normalization is used to stabilize the numerical distribution between different samples and different channels, and the GELU activation function is used to enhance the nonlinear expressive power.

[0181] Complete in sequence After a dynamically modulated dilated convolutional block, a deep temporal feature map is obtained. , dimension Deep temporal feature map This refers to the temporal feature map after physical variable correlation graph encoding, multi-scale dilated convolution, and global physical statistical feature modulation; among which... This indicates the number of output channels of the last dynamically modulated dilated convolution block, which can be 128.

[0182] It should be noted that the multi-scale dilated convolution modulated by global physical statistical features does not apply a fixed time scale and fixed channel response to all samples. Instead, it transforms the global physical statistical features of each sample into branch fusion weights and channel modulation parameters. This enables the same diagnostic model to have different time scale focus and channel response focus when dealing with different fault states such as jamming, internal leakage, and abnormal pneumatic or hydraulic sources, thereby improving its adaptability to complex operating state differences.

[0183] S402, Global Query Weighted Attention Pooling

[0184] Deep temporal feature map While retaining features from 128 time steps, this invention uses a feature vector derived from global physical statistics to obtain a fixed-length feature vector for classification, while avoiding the weakening of critical fault moments by ordinary average pooling. The generated query vector is then weighted by attention at each time step, with the following specific steps:

[0185] 1> Global physical statistical feature vector Input a trainable linear projection layer to generate a global query vector. ,Right now .in, Represents the global query vector, with dimension . It is the time-step attention guidance vector obtained by mapping the global physical statistical features of the sample; This indicates a query for the projected weight matrix, with dimension 1. , are trainable parameters; This indicates a query for the projection bias vector, with dimension . , are trainable parameters; Representing deep temporal feature maps The number of channels.

[0186] It should be noted that the global query vector Its function is to transform macroscopic information such as the duration of the transition segment, the intensity of the transition segment deviation, and the current deviation of the steady-state segment into a reference direction for attention pooling, so that the pooling process pays more attention to the time step related to the current sample's fault mode.

[0187] 2> Calculate the global query vector With deep temporal feature map The similarity between feature vectors at each time step in the deep temporal feature map. In the The feature vectors at each time step are denoted as follows: , dimension Calculate the first by scaling the dot product The attention scores at each time step are calculated, and then the attention scores at 128 time steps are normalized using Softmax to obtain the temporal attention weights. , represented as:

[0188] ;

[0189] in, It refers to the first The time attention weights for each time step range from 0 to 1, and the sum of the time attention weights for all 128 time steps is 1. This represents the scaling factor, used to prevent the output of the Softmax function from becoming too concentrated due to excessively large dot product values. Indicates the first The deep temporal feature vector at each time step.

[0190] 3> According to time attention weight Deep temporal feature map The 128 time-step features are weighted and summed to obtain the deep fault feature vector. ,Right now .

[0191] in, Represents a deep fault feature vector with dimension . This refers to a fixed-length feature vector obtained after focusing attention on full-time fault information. Deep fault feature vector. It includes dynamic features extracted by multi-scale dilated convolution, and focuses on key moments most relevant to the global physical state, such as the moment of sudden change in the following residual caused by jamming, the moment of continuous pressure reference deviation caused by internal leakage, and the long transition moment caused by pneumatic or hydraulic source anomalies.

[0192] 4> Transfer deep fault feature vectors With global physical statistical eigenvectors By concatenating along the feature dimensions, a fused fault feature vector is obtained. fusion of fault feature vectors This refers to a classification input feature vector that simultaneously contains deep temporal features and sample-level physicostatistical features. This is achieved by preserving the global physicostatistical feature vector. The classification module can directly utilize sample-level indicators with clear physical meanings, such as the proportion of transition duration, the absolute value of the average residual of the transition, and the average current reference deviation of the steady-state section, to avoid these indicators being weakened during the propagation of deep networks.

[0193] It should be noted that global query weighted attention pooling differs from ordinary average pooling. Ordinary average pooling assigns the same weight to all 128 time steps, which can easily dilute short-term fault features; global query weighted attention pooling dynamically determines the weight of each time step based on global physical statistical features, which can adaptively highlight different time periods under different fault types and improve the discriminative power of classification features.

[0194] It should also be noted that the specific structure of multi-scale temporal coding and attention pooling operations is as follows:

[0195]

[0196] S5. Fault Diagnosis Model Training and Online Diagnosis

[0197] like Figure 4 As shown, after obtaining the fused fault feature vector Subsequently, this invention outputs the thermal actuator status categories through a classification layer and uses labeled samples for training. The four thermal actuator status categories are, in order, normal, jammed, internal leakage, and abnormal pneumatic or hydraulic power source. The specific steps are as follows:

[0198] 1> Fuse fault feature vectors Input the classification fully connected layer to obtain the state classification logical value vector. The dimension is 4, that is .in, The state classification logic value vector refers to the unnormalized score vector corresponding to the four thermal actuator state categories. The weight matrix of the fully connected classification layer represents trainable parameters. This represents the bias vector of the fully connected classification layer, which are trainable parameters. The state classification logistic value vector. The four elements correspond to normal, jamming, internal leakage, and abnormal pneumatic or hydraulic power source, respectively.

[0199] 2> Classify the state logical value vector Performing Softmax normalization yields the posterior probability vector of the state class. The dimension is 4, and it is the posterior probability vector of the state category. This refers to the probability distribution of the model predicting that the input sample belongs to one of the four thermal actuator state categories; the posterior probability vector of the state category. The first in The elements are denoted as , This indicates that the input sample belongs to the first... The posterior probability of each state category. This represents the state category index, with a value range from 1 to 4.

[0200] 3> During the training phase, labeled training samples are input into the fault diagnosis model in batches, and the class-weighted cross-entropy loss is used to measure the posterior probability vector of the state class. The difference between the label and the actual state category label. Among them, the first... The number of training samples for each state category is denoted as . The maximum number of samples across all state categories is denoted as , No. The class weights of each state category are denoted as: , can make Category weights This is used to improve the influence of minority class samples in the loss function and reduce the training bias when the number of normal samples is much greater than the number of faulty samples.

[0201] 4> Update the trainable parameters in the fault diagnosis model by minimizing the class-weighted cross-entropy loss.

[0202] In one implementation, the Adam optimizer is used to update trainable parameters. The initial learning rate can be 0.001, and the number of training epochs can be 100 to 200. During training, a learning rate decay strategy can be adopted based on the decrease in validation set loss. For example, if the validation set loss does not decrease for five consecutive training epochs, the learning rate is multiplied by 0.5. After each training epoch, the F1 scores for the four state categories are calculated using the validation set, and the arithmetic mean of the four F1 scores is taken to obtain the macro-average F1 score. The fault diagnosis model parameters corresponding to the highest macro-average F1 score are saved as the final fault diagnosis model parameters.

[0203] It should be noted that in the actuator nominal follower model , , and Pressure-valve position reference function, current-command reference function, and physical prior adjacency matrix All offline physical benchmarks are determined by the training set and are used consistently during neural network training. These offline physical benchmarks are all established using normal category training samples, without using test sample information, thus avoiding test data leakage and ensuring that the following residuals, pressure benchmark deviations, and current benchmark deviations are all based on the physical laws of normal actuators.

[0204] 5> During the online inference phase, the current diagnostic window is extracted from the online monitoring data stream of the thermal actuator to obtain the timing matrix of the thermal actuator to be diagnosed. Timing matrix of the thermal actuator to be diagnosed This refers to a sequence of control commands, valve position feedback, downstream pressure, and winding current timing data used for status determination at the current moment, with a size of [missing information]. The diagnostic window can slide in fixed steps, with the step size being 16, 32, or 64 time steps, to achieve continuous online diagnosis.

[0205] 6> Timing matrix of the thermal actuator to be diagnosed The process is exactly the same as in the training phase. Specifically, the normalized statistics saved during the training phase are first used to analyze the timing matrix of the thermal actuator to be diagnosed. Numerical normalization is performed; then, the actuator nominal following model is used to calculate the following residual and residual difference; next, the pressure reference deviation and current reference deviation are calculated using the pressure-valve position reference function and the current-command reference function; then, a binary segmented mask is generated based on the absolute value of the control command difference, and the global physical statistical feature vector is calculated; finally, the timing matrix of the enhanced thermal actuator to be diagnosed is obtained. and the global physical statistical feature vector to be diagnosed .

[0206] It should be noted that the online inference phase must use the normalized statistics, actuator nominal following model, pressure-valve position reference function, and current-command reference function saved during the training phase, and must not be based on the timing matrix of the thermal actuator to be diagnosed. The normalized range or physical baseline should be recalculated separately; otherwise, the relative strength of the fault deviation will be altered, affecting the stability of online diagnosis.

[0207] 7> Add the timing matrix of the enhanced thermal actuator to be diagnosed and the global physical statistical feature vector to be diagnosed The trained fault diagnosis model is input and sequentially processed through physical variable correlation graph encoding, multi-scale dilated convolutional encoding with global physical statistical feature modulation, global query weighted attention pooling, and a classification fully connected layer to obtain the posterior probability vector of the state to be diagnosed. posterior probability vector of the state category to be diagnosed This refers to the timing matrix of the thermal actuator to be diagnosed. The probability distributions for four states: normal, jammed, internal leakage, and abnormal pneumatic or hydraulic power source.

[0208] 8> Based on the posterior probability vector of the state to be diagnosed Output the diagnostic result. Specifically, if the maximum posterior probability is less than the confidence threshold... If the probability is not less than the confidence threshold, a low confidence level message will be output, and the subsequent diagnostic window will continue to be observed; if the maximum posterior probability is not less than the confidence threshold, the output will be lower. Then, the state category corresponding to the maximum posterior probability is used as the candidate diagnosis result for the current diagnosis window. Confidence threshold This represents the minimum probability threshold for a diagnosis to be accepted, and can be between 0.60 and 0.80.

[0209] In one embodiment, for example, if And confidence threshold If the maximum posterior probability is 0.72, corresponding to the second state category, the output candidate diagnosis result is "Kase"; if If the maximum posterior probability is 0.32, which is less than 0.60, a low confidence warning will be output, and further confirmation will be awaited in the subsequent diagnostic window.

[0210] 9. Perform time consistency verification on candidate diagnostic results across multiple consecutive diagnostic windows. The number of consecutive verification windows is denoted as... , This indicates the number of diagnostic windows that need to continuously exhibit the same fault category, and can be 3 to 5. When the same fault category appears consecutively... All diagnostic windows represent the maximum posterior probability category, and the corresponding maximum posterior probability is not less than the confidence threshold. If the candidate diagnostic results in an adjacent diagnostic window change frequently, the alarm will not be triggered, and subsequent diagnostic windows will continue to be collected.

[0211] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for diagnosing actuator faults in a thermal control system, characterized in that, Includes the following steps: S1. Collect timing signals of the thermal actuator under different operating conditions, label them, and construct a training sample set for the thermal actuator; the timing signals include control commands, valve position feedback, valve downstream pressure, and winding current; the labels are used to characterize the state category of the thermal actuator. S2. Construct a physical baseline model based on time-series signal samples of the normal category. Construct new time-series channels according to the deviation of the measured values ​​of the time-series signals from the physical baseline model. Extract global physical statistical features according to the change process of control commands to obtain the enhanced thermal actuator time-series matrix and global physical statistical feature vector. S3. Define multiple channels of the enhanced thermal actuator time series matrix as multiple nodes in the physical variable association graph. Predefine the physical prior adjacency matrix according to the physical action relationship of the actuator. Under the constraint of the physical prior adjacency matrix, introduce trainable edge coefficients so that the multidimensional enhanced features of each time step are first fused with node information on the physical variable association graph, and then enter the subsequent multi-scale time series network along the time dimension. S4. Generate modulation parameters and attention query vectors based on the global physical statistical feature vector. The multiple modulation parameters are used to adaptively adjust the encoding process of the multi-scale temporal network, and the attention query vector is used to perform weighted pooling on the time step features output by the multi-scale temporal network to obtain the fused fault feature vector. S5. Based on the fused fault feature vector, the classification layer outputs the state category of the thermal actuator.

2. The method for diagnosing actuator faults in a thermal control system according to claim 1, characterized in that, The labels include normal, stuck, internal leakage, and abnormal pneumatic or hydraulic power source.

3. The method for diagnosing actuator faults in a thermal control system according to claim 1, characterized in that, The physical reference model includes the actuator nominal follower model, the pressure-valve position reference function, and the current-... Instruction base function; The deviations include valve position feedback deviation, valve downstream pressure deviation, and winding current deviation. The multiple channels of the enhanced thermal actuator timing matrix include: control command, valve position feedback, valve downstream pressure, winding current, valve position feedback deviation, valve downstream pressure deviation, winding current deviation, valve position change rate, and pressure change rate.

4. The method for diagnosing actuator faults in a thermal control system according to claim 2, characterized in that, S2 specifically refers to: The nominal follower model of the actuator is identified using control commands and valve position feedback from normal category samples, and the follower residual is calculated step by step to ensure that the model input includes the actuator action deviation. Construct pressure-valve position reference function and current-command reference function, and calculate the deviation of measured values ​​from normal reference functions; A steady-state-transition segmented mask is generated based on the rate of change of the control command, and global physical statistical features are extracted under the constraints of the segmented mask.

5. The method for diagnosing actuator faults in a thermal control system according to claim 1, characterized in that, S3 specifically refers to: A physical variable association graph containing 9 nodes is established. The physical variable association graph is a graph structure with the 9 timing channels in the timing matrix of the enhanced thermal actuator as nodes and the physical interaction relationship of the thermal actuator as edges. Based on the physical prior adjacency matrix, a trainable edge coefficient matrix is ​​introduced, and self-loops are added to obtain an adaptive physical adjacency matrix; For each time step Take the first one from the timing matrix of the enhanced thermal actuator Line up and transpose to construct the node feature matrix; Perform two layers of graph convolution on the node feature matrix to obtain the node-level graph convolution feature matrix; The node-level graph convolutional feature matrices are converged along the node dimension to obtain the first... Graph convolutional feature vectors at each time step; The node feature matrix construction, graph convolution calculation, and node dimension aggregation are repeated for all time steps to obtain the graph-enhanced temporal feature map.

6. The method for diagnosing actuator faults in a thermal control system according to claim 5, characterized in that, The modulation parameters include multi-scale convolutional branch weights and channel modulation parameters, which are generated based on the global physical statistical feature vector, specifically: Using graph-enhanced temporal feature maps as input feature maps, multiple dynamically modulated dilated convolutional blocks are stacked to extract temporal features. The global physical statistical feature vector is input into a branch weight generation network consisting of a linear layer, a GELU activation function, a linear layer, and a Softmax function to obtain the branch fusion weight vector of the dynamically modulated dilated convolutional block. The outputs of multiple dilated convolution branches are weighted and fused according to the branch fusion weight vector to obtain the multi-scale candidate feature map of the dynamically modulated dilated convolution block. The multi-scale candidate feature map is a candidate temporal feature map after fusing receptive field features from different times. The channel modulation scaling vector and channel modulation translation vector of the dynamically modulated dilated convolutional block are generated using global physical statistical feature vectors. Multiply the feature value of each channel of the multi-scale candidate feature map by the scaling parameter of the corresponding channel, and add the translation parameter of the corresponding channel to obtain the modulated feature value; Residual connections, layer normalization, and GELU activation are performed on the modulated feature map to obtain the output feature map of the dynamically modulated dilated convolutional block. After performing multiple dynamically modulated dilated convolutional blocks in sequence, a deep temporal feature map is obtained.

7. The method for diagnosing actuator faults in a thermal control system according to claim 6, characterized in that, An attention query vector is generated based on the global physical statistical feature vector. This attention query vector is used to perform weighted pooling of the time-step features output by the multi-scale temporal network to obtain a fused fault feature vector. Specifically: Input the global physical statistical feature vector into the trainable linear projection layer to generate a global query vector; Calculate the similarity between the global query vector and the feature vector at each time step in the deep temporal feature map; The time step features of the deep temporal feature map are weighted and summed according to the time attention weight to obtain the deep fault feature vector. The deep fault feature vector and the global physical statistics feature vector are concatenated along the feature dimension to obtain the fused fault feature vector.

8. The method for diagnosing actuator faults in a thermal control system according to claim 1, characterized in that, Based on the deep fault feature vector and through a classification layer, the state category of the thermal actuator is output, specifically: The fused fault feature vector is input into the classification fully connected layer to obtain the state classification logical value vector; Perform Softmax normalization on the state classification logical value vector to obtain the posterior probability vector of the state category; The diagnostic result is output based on the posterior probability vector of the state category. If the maximum posterior probability is less than the confidence threshold, a low confidence warning is output, and the subsequent diagnostic window is observed. If the maximum posterior probability is not less than the confidence threshold, the state category corresponding to the maximum posterior probability is used as the candidate diagnostic result for the current diagnostic window, and the confidence threshold is set. This represents the minimum probability threshold for a diagnosis to be accepted.

9. A method for diagnosing actuator faults in a thermal control system according to claim 8, characterized in that, It also includes the consistency confirmation of the execution time of candidate diagnostic results in multiple consecutive diagnostic windows. When the same fault category is the maximum a posteriori probability category in multiple consecutive diagnostic windows, and the corresponding maximum a posteriori probability is not less than the confidence threshold, a formal fault alarm is output. If the candidate diagnostic results of adjacent diagnostic windows change frequently, no alarm is set off and subsequent diagnostic windows are collected.

Images