A valve super-high-speed action abnormality judgment method and system

Abstract

The application discloses a valve super-high-speed action abnormality judgment method and system, the method comprising: collecting original data of the whole motion process of the valve stem of a pneumatic valve actuator, obtaining an original event sequence and a triggering time; based on the original event sequence and the triggering time, performing event stream reconstruction and feature parameter extraction, obtaining a standardized feature vector; through a pre-determined lightweight fully connected neural network model, performing valve action abnormality detection on the standardized feature vector, obtaining a valve abnormality probability vector; based on the valve abnormality probability vector and a pre-set threshold, determining a final diagnosis result. The above scheme can improve the accuracy of valve super-high-speed action abnormality judgment, effectively reduce the blindness of fault troubleshooting, and shorten the downtime.

Description

Technical Field

[0001] This application relates to the field of valve anomaly judgment technology, specifically to a method and system for judging valve ultra-high speed action anomalies. Background Technology

[0002] Atomic Layer Deposition (ALD) technology, with its sub-single-layer thickness control precision and excellent conformal properties, has become an indispensable core process in advanced semiconductor manufacturing. The ALD process introduces precursor gas into the reaction chamber via alternating pulses, and within each pulse cycle, a pneumatic valve completes high-speed switching actions within milliseconds to achieve precise timing control and dosage of the reactant gas. Taking a typical ALD process as an example, the full stroke time of a single valve is typically required to be no more than 5ms. Any minute delay, speed fluctuation, or deviation in positioning can lead to errors in precursor injection volume, reaction window shift, and consequently, decreased film thickness uniformity, particulate contamination, or even the scrapping of an entire batch of wafers, directly impacting product yield and process stability.

[0003] Currently, monitoring the operational status of ALD pneumatic valves mainly relies on two types of methods: one is stroke completion detection based on valve position sensors (such as proximity switches and magnetostrictive displacement sensors). This method can only determine whether the valve is fully open or closed, but cannot capture dynamic behavior details such as speed changes, vibrations, and creeping of the valve stem throughout its movement, and lacks sensitivity to early fault signs; the other is random inspection using high-speed cameras or laser vibration meters during regular offline maintenance. Although this method can obtain complete motion curves, it requires stopping the machine and disassembling the equipment, which cannot achieve online real-time monitoring. Moreover, the detection cost is high and the efficiency is low, making it difficult to meet the maintenance requirements of semiconductor manufacturing for real-time knowledge of equipment status. Summary of the Invention

[0004] This application aims to provide a method and system for judging abnormal valve operation at ultra-high speed, which can improve the accuracy of judging abnormal valve operation at ultra-high speed, effectively reduce the blindness of fault diagnosis, and shorten downtime.

[0005] The technical solution of this application is implemented as follows: In a first aspect, embodiments of this application provide a method for judging abnormal ultra-high speed operation of a valve, the method comprising: Raw data was collected from the entire movement process of the valve stem of the pneumatic valve actuator to obtain the raw event sequence and trigger time; Based on the original event sequence and trigger time, event flow reconstruction and feature parameter extraction are performed to obtain a standardized feature vector; Using a pre-defined lightweight fully connected neural network model, valve action anomaly detection is performed on the standardized feature vector to obtain the probability vector of valve anomaly. The final diagnostic result is determined based on the probability vector of valve anomalies and a preset threshold. Specifically, based on the original event sequence and trigger time, event flow reconstruction and feature parameter extraction are performed to obtain a standardized feature vector, including: Based on the original event sequence, event stream temporal accumulation is performed to obtain an event count image sequence; The valve stem displacement time series is reconstructed from the event count image sequence to obtain the valve stem displacement time discrete sequence; Based on the discrete sequence of valve stem displacement time and the triggering time, action feature parameters are extracted to obtain a standardized feature vector.

[0006] Understandably, based on the original event sequence, event flow time-domain accumulation, valve stem displacement time-series reconstruction, and action feature parameter extraction are performed to obtain a standardized feature vector. Subsequently, a lightweight fully connected neural network model provides input data to perform valve action anomaly detection.

[0007] In the above scheme, based on the original event sequence, event stream temporal accumulation is performed to obtain an event counting image sequence, including: The original event sequence is divided into time slices by a pre-set cumulative time window, resulting in multiple time slices; Event statistics are performed on multiple time slices to obtain event subsets for each time slice. Based on the event subsets of each of the multiple time slices, generate event count images for each of the multiple time slices according to the preset region of interest; An event count image sequence is generated based on the event count images of multiple time slices.

[0008] Understandably, by using a pre-set cumulative time window, the original event sequence is divided into time slices, event statistics are performed, event count images of multiple time slices are generated, and further event count image sequences are generated, which facilitates subsequent reconstruction of the valve stem displacement time series.

[0009] In the above scheme, the valve stem displacement time series is reconstructed from the event counting image sequence to obtain the valve stem displacement time discrete sequence, including: For each time slice in the event count image sequence, edge response aggregation is performed to obtain the edge response value of each time slice; Based on the edge response value of each time slice, the edge position is accurately located to obtain the edge position of each time slice; Based on the edge position of each time slice, the displacement physical unit is converted to obtain the displacement of each time slice; Based on the displacement of each time slice, as well as the start and end times of each time slice, the discrete time sequence of valve stem displacement is determined.

[0010] It is understandable that by performing edge response aggregation, precise edge location positioning, and displacement physical unit conversion on the event count images of each time slice in the event count image sequence, the valve stem displacement time series can be reconstructed, thereby achieving complete capture and quantitative representation of the entire ultra-high-speed action.

[0011] In the above scheme, based on the discrete sequence of valve stem displacement time and the triggering time, action feature parameters are extracted to obtain a standardized feature vector, including: Determine the target displacement of the valve's rated stroke and the allowable error band for stability assessment; Based on the target displacement of the valve's rated stroke, the opening delay is extracted from the discrete sequence of valve stem displacement time. The maximum displacement is determined from the discrete time sequence of valve stem displacement; and the overshoot is determined based on the maximum displacement and the target displacement of the valve's rated stroke. Based on the target displacement of the valve's rated stroke and the allowable error band for stability determination, the stable time is extracted from the discrete sequence of valve stem displacement time. Based on the discrete time sequence of valve stem displacement, the instantaneous velocity of adjacent time slices is calculated; and the peak velocity of motion is determined from the instantaneous velocity of adjacent time slices. Calculate the velocity variation coefficient based on the instantaneous velocity of adjacent time slices; A standardized feature vector is generated based on the start-up delay, overshoot, settling time, peak velocity, and velocity variation coefficient.

[0012] Understandably, five key feature parameters—opening delay, overshoot, settling time, peak velocity, and velocity variation coefficient—are extracted from the reconstructed valve stem displacement time series. Standardized feature vectors are then generated using these parameters to provide input data for a subsequent lightweight fully connected neural network model, enabling abnormal valve action detection.

[0013] In the above scheme, a pre-determined lightweight fully connected neural network model is used to detect valve action anomalies on the standardized feature vector, resulting in a probability vector of valve anomalies, including: Intermediate features are obtained by mapping the standardized feature vectors through the first fully connected layer of a lightweight fully connected neural network model. By using the second fully connected layer of a lightweight fully connected neural network model, intermediate features are abstracted and compressed to obtain deep features; By mapping the deep features through the output layer of a lightweight fully connected neural network model, a probability vector of valve anomalies is obtained.

[0014] Understandably, by constructing a lightweight fully connected neural network model, valve action anomaly detection is performed on the standardized feature vector to obtain the probability vector of valve anomalies, which facilitates subsequent determination of the valve anomaly type based on the probability vector of valve anomalies.

[0015] In the above scheme, the preset thresholds include the threshold for stuck crawling, the threshold for tremor oscillation, the threshold for overshoot rebound, and the threshold for action timeout. Based on the probability vector of valve anomalies and preset thresholds, the final diagnostic result is determined, including: The probability of each type of valve abnormality in the probability vector is compared with the thresholds for stuck crawling, tremor oscillation, overshoot rebound, and action timeout to determine the type of valve abnormality. The final diagnostic result is generated based on the valve anomaly type.

[0016] Understandably, by using a lightweight fully connected neural network to output the probability vector of valve anomalies, and combining it with preset thresholds to determine the probability of various anomalies and the final diagnostic results, the problem of existing methods lacking systematic anomaly classification and judgment capabilities is solved. This achieves end-to-end automatic classification and identification of valve stem motion signals and four typical anomalies (stuck crawl, tremor oscillation, overshoot rebound, and action timeout), effectively reducing the blindness of fault diagnosis and shortening downtime.

[0017] Secondly, this application provides a valve ultra-high speed action anomaly judgment system. The valve ultra-high speed action anomaly judgment system includes: an acquisition module, an extraction module, a detection module, and a determination module; wherein, the acquisition module is used to collect raw data of the entire movement process of the valve stem of the pneumatic valve actuator to obtain the raw event sequence and trigger time; The extraction module is used to reconstruct the event stream and extract feature parameters based on the original event sequence and trigger time to obtain a standardized feature vector; The detection module is used to detect valve action anomalies by using a pre-determined lightweight fully connected neural network model to standardize feature vectors, and obtain a probability vector of valve anomalies. The determination module is used to determine the final diagnostic result based on the probability vector of valve anomalies and preset thresholds; The extraction module is also used to perform time-domain accumulation of the event stream based on the original event sequence to obtain an event count image sequence; to reconstruct the valve stem displacement time sequence from the event count image sequence to obtain a valve stem displacement time discrete sequence; and to extract action feature parameters based on the valve stem displacement time discrete sequence and the triggering time to obtain a standardized feature vector.

[0018] Thirdly, embodiments of this application provide a valve ultra-high speed action abnormality judgment device, including: a processor and a memory; wherein, the memory is used to store computer programs; A processor is used to retrieve and run computer programs from memory to perform methods such as those described in the first aspect.

[0019] Fourthly, embodiments of this application provide a computer-readable storage medium storing executable instructions for causing a processor to perform the method as described in the first aspect.

[0020] This application provides a method and system for judging abnormal ultra-high-speed valve action. The method includes: acquiring raw data of the entire motion process of the valve stem of a pneumatic valve actuator to obtain a raw event sequence and trigger time; reconstructing the event flow and extracting feature parameters based on the raw event sequence and trigger time to obtain a standardized feature vector; detecting valve action abnormalities using a pre-determined lightweight fully connected neural network model on the standardized feature vector to obtain a probability vector of valve abnormality; and determining the final diagnostic result based on the probability vector of valve abnormality and a preset threshold. The process of reconstructing the event flow and extracting feature parameters to obtain the standardized feature vector based on the raw event sequence and trigger time includes: performing time-domain accumulation of the event flow based on the raw event sequence to obtain an event count image sequence; reconstructing the valve stem displacement time sequence from the event count image sequence to obtain a valve stem displacement time discrete sequence; and extracting action feature parameters based on the valve stem displacement time discrete sequence and trigger time to obtain the standardized feature vector. In the above scheme, the original event sequence and trigger time of the entire valve stem movement process are collected. Based on the original event sequence and trigger time, event flow reconstruction and feature parameter extraction are performed to obtain a standardized feature vector. This can solve the problem that existing valve position sensors can only detect the switch position and cannot capture the dynamic details of the entire ultra-high speed movement. It provides a complete data foundation for the early identification of dynamic anomalies such as jamming, crawling and tremor. A lightweight fully connected neural network model is constructed and the final diagnostic result is output by combining preset thresholds. This solves the problem that existing methods lack the ability to classify and judge systematic anomalies, thereby improving the accuracy of valve ultra-high speed movement anomaly judgment, effectively reducing the blindness of fault diagnosis and shortening downtime. Attached Figure Description

[0021] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the specification, serve to explain the technical solutions of this application. Obviously, the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0022] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0023] Figure 1 A schematic flowchart of an optional method for judging abnormal valve operation at ultra-high speed, provided in an embodiment of this application; Figure 2 This is a schematic diagram of a valve ultra-high speed action abnormality judgment system provided in an embodiment of this application; Figure 3 This is a structural schematic diagram of a valve ultra-high speed action abnormality judgment device provided in an embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the specific technical solutions of this application will be further described in detail below with reference to the accompanying drawings of the embodiments of this application. The following embodiments are used to illustrate this application, but are not intended to limit the scope of this application.

[0025] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this application is for the purpose of describing embodiments of this application only and is not intended to be limiting of this application.

[0026] In the following description, references to "some embodiments," "this embodiment," "this application embodiment," and examples, etc., describe a subset of all possible embodiments. However, it is understood that "some embodiments" may be the same subset or different subset of all possible embodiments and may be combined with each other without conflict.

[0027] If the application documents contain similar descriptions such as "first / second", the following explanation shall be added: In the following description, the terms "first / second / third" are used only to distinguish similar objects and do not represent a specific order of objects. It is understood that "first / second / third" may be interchanged in a specific order or sequence where permitted, so that the embodiments of this application described herein can be implemented in an order other than that illustrated or described herein.

[0028] This application provides a method for judging abnormal ultra-high speed valve operation. Figure 1 This is an optional flowchart illustrating a method for judging abnormal ultra-high-speed valve operation provided in an embodiment of this application, which will be combined with... Figure 1 The steps shown are explained.

[0029] S101. Collect raw data of the entire movement process of the pneumatic valve actuator stem to obtain the raw event sequence and trigger time.

[0030] In some embodiments of this application, a dynamic vision sensor (DVS) is used to collect raw data of the entire movement process of the valve stem of the pneumatic valve actuator, and output the raw event sequence and trigger time with microsecond-level timestamps.

[0031] In some embodiments of this application, the DVS is fixed to the side of the valve stem of the valve actuator, ensuring that the optical axis is perpendicular to the direction of valve stem movement, so that the surface texture of the valve stem is clearly visible in the field of view. After installation, the target monitoring area in the image, i.e., the region of interest (ROI), is set as the rectangular pixel band covered by the expected movement trajectory of the valve stem, in order to reduce background event interference in non-interested areas. ROI parameters include: roi_u_start, roi_u_end: The start and end pixel coordinates of the ROI in the horizontal direction (U axis) of the image; roi_v_start, roi_v_end: The start and end pixel coordinates of the ROI in the vertical direction (V-axis) of the image; roi_width=roi_u_end-roi_u_start+1; roi_height=roi_v_end-roi_v_start+1.

[0032] Based on the typical valve actuation speed (full stroke time ≤ 5ms) in the ALD process, the DVS parameters are configured to ensure that the temporal resolution of the event stream is sufficient to fully capture the motion process. Configuration items include: 1. ev_time_resolution: Event timestamp resolution, fixed at 1μs; 2. `ev_contrast_threshold`: Contrast sensitivity threshold (unit: %, adjustable range 1%~100%), used to control event triggering sensitivity. The lower the threshold, the more sensitive it is to changes in brightness, and the more events are output; the higher the threshold, the more events with significant changes are triggered. An initial value of 15% is recommended, and it can be fine-tuned later according to the actual scenario. 3. ev_bias_refractory: Pixel non-response time (unit: μs), which is the shortest time to block subsequent events after a single pixel triggers an event. It is used to suppress noise bursts and is recommended to be set to 50μs.

[0033] After configuration, start DVS and enter the ready-to-acquire state, waiting for an external trigger signal.

[0034] Connect the trigger input pin of the DVS to the valve control signal output terminal of the ALD device to achieve hardware-level synchronization of event acquisition and valve action commands.

[0035] When the ALD device issues a control command to open or close the valve, it simultaneously sends a trigger pulse to the DVS. The system records the moment when this trigger pulse arrives at the DVS trigger input as the trigger time t_trigger (unit: μs).

[0036] After receiving the trigger signal, the camera continuously records the event stream within the set acquisition window. The parameters of the acquisition window are as follows: 1. acq_window_pre: Pre-acquisition time before triggering (unit: ms), used to record the static reference state of the valve stem before action. It is recommended to set it to 2ms. 2. acq_window_post: Continuous data collection time after triggering (unit: ms), used to fully record the action process and subsequent stabilization process. It is recommended to set it to 20ms (including the entire action process and the stabilization phase after reaching the target position). 3. acq_total_duration=acq_window_pre+acq_window_post: Total collection time.

[0037] Within the acquisition window, DVS outputs the event stream in real time. Each raw event e_k has a data structure of a four-tuple: e_k=(u_k,v_k,p_k,t_k); Where u_k and v_k are the pixel coordinates (integers, unit: pixels) where the event occurred, and their ranges are within [roi_u_start, roi_u_end] and [roi_v_start, roi_v_end] respectively; p_k is the event polarity, +1 indicates brightness enhancement (ON event), and -1 indicates brightness reduction (OFF event); t_k is the event timestamp (integer, unit: μs), which is the absolute time relative to the DVS internal clock.

[0038] After the acquisition window ends, all acquired events are sorted in ascending order by t_k to form the raw event sequence E_raw for a single action. E_raw = [e_1, e_2, e_3, ..., e_N], where N is the total number of events falling within the ROI in this acquisition.

[0039] S102. Based on the original event sequence and trigger time, perform event flow reconstruction and feature parameter extraction to obtain a standardized feature vector.

[0040] In some embodiments of this application, event stream time-domain accumulation is performed based on the original event sequence to obtain an event count image sequence; valve stem displacement time sequence is reconstructed from the event count image sequence to obtain a valve stem displacement time discrete sequence; and action feature parameters are extracted based on the valve stem displacement time discrete sequence and the triggering time to obtain a standardized feature vector.

[0041] In some embodiments of this application, S102 can be implemented by S1021-S1023, as follows: S1021. Based on the original event sequence, perform event stream time-domain accumulation to obtain an event count image sequence.

[0042] In some embodiments of this application, the original event sequence is divided into time slices by a pre-set cumulative time window to obtain multiple time slices; event statistics are performed on each of the multiple time slices to obtain event subsets for each of the multiple time slices; based on the event subsets of each of the multiple time slices, event counting images for each of the multiple time slices are generated according to a preset region of interest; and an event counting image sequence is generated based on the event counting images of each of the multiple time slices.

[0043] For example, since a single event only reflects a brightness change of a single pixel, events need to be accumulated within a spatiotemporal window to form a dense representation from which motion information can be extracted. An accumulation time window τ_acc (unit: μs) is set, and E_raw is divided into M consecutive time slices according to a fixed time window. The calculation formula is as follows: M=(t_N-t_1) / τ_acc; Where t_1 and t_N are the timestamps of the first and last events in E_raw, respectively.

[0044] The start time T_start[m] and end time T_end[m] of each time slice m (m=1,2,...,M) are: T_start[m]=t_1+(m-1)×τ_acc; T_end[m] = t_1 + m × τ_acc; All events falling within this time slice constitute the event subset E_slice[m]. It is recommended that τ_acc=50μs, which can achieve a balance between time resolution and signal-to-noise ratio.

[0045] For each time slice's event subset E_slice[m], an event counting image I_cnt[m] is generated within the ROI, thus obtaining an event counting image sequence. I_cnt[m] is a two-dimensional image with a height equal to roi_height and a width equal to roi_width. The value of each pixel position is calculated according to the following rules: Iterate through each event e_k=(u_k,v_k,p_k,t_k) in E_slice[m], and based on the pixel coordinates (u_k,v_k) of the event, accumulate its polarity value p_k to the corresponding pixel position in I_cnt[m].

[0046] For any pixel (u,v) within the ROI, its cumulative response value in the m-th time slice is equal to the sum of the polarities p_k of all events with coordinates (u,v) belonging to that slice. The specific calculation method is as follows: 1. Initialize I_cnt[m] as an all-zero matrix with size roi_height × roi_width; 2. For each event e_k in E_slice[m], execute: row=v_k-roi_v_start; col=u_k-roi_u_start; I_cnt[m][row][col]+=p_k; Where p_k=+1 represents an ON event (brightness increases) and p_k=-1 represents an OFF event (brightness decreases).

[0047] The result of this processing is that at the edge of the valve stem movement, due to the large number of events with the same polarity (both +1 or both -1), the cumulative response amplitude is significantly increased; while random noise events, due to their random positive and negative polarity and dispersed spatial location, cancel each other out during the summation process, and appear as a low amplitude background close to zero in I_cnt[m].

[0048] S1022. Reconstruct the valve stem displacement time series from the event count image sequence to obtain the valve stem displacement time discrete sequence.

[0049] In some embodiments of this application, edge response aggregation is performed on the event count image of each time slice in the event count image sequence to obtain the edge response value of each time slice; based on the edge response value of each time slice, the edge position is accurately located to obtain the edge position of each time slice; based on the edge position of each time slice, the displacement physical unit is converted to obtain the displacement of each time slice; based on the displacement of each time slice, as well as the start time and end time of each time slice, the valve stem displacement time discrete sequence is determined.

[0050] For example, the spatial position of the valve stem within each time slice is extracted from the event count image sequence to reconstruct the complete valve stem displacement time sequence D.

[0051] For each time slice I_cnt[m], the event counts of all pixels along the horizontal direction (U axis) of the ROI are accumulated and summed. The two-dimensional event count image is projected onto the vertical direction (V axis) to obtain the edge response value R[m](v). ​​That is, for each pixel coordinate v on the V axis, the event counts of all pixels in the U direction of that row are accumulated to form the corresponding edge response value R[m](v).

[0052] For the event count image I_cnt[m] corresponding to each time slice, the summation of all pixels along the horizontal direction (U axis) of the ROI is performed. The two-dimensional event count image is then projected onto the vertical direction (V axis) to obtain the edge response value R[m](v). ​​The specific calculation method of the edge response value R[m](v) is as follows: ; The centroid method is used to locate the edge position v_edge[m] (i.e., the row coordinate of the valve stem feature edge in the image) on the V-axis using R[m](v): ; Compared to the direct maximum value method, the centroid method offers sub-pixel-level positioning accuracy and effectively suppresses the interference of single-point noise on positioning. The valve stem feature edge refers to a pre-defined, clearly identifiable lateral edge on the valve stem (such as the valve stem end face or a scribe line).

[0053] The edge position v_edge[m] of the valve stem refers to the row index position of the valve stem feature edge in the V-axis direction of the image. When the valve stem is stationary, the edge remains at a fixed row; when the valve stem moves, the edge moves accordingly in the V-axis direction. The displacement is the offset of the edge position relative to its stationary state.

[0054] To obtain a reliable static reference position, the average edge position of all time slices within the pre-acquisition period before triggering is taken as the zero-position reference value v_zero. v_zero represents the average position of the valve stem's edge in the image row index when it is stationary before actuation. This value serves as the starting reference for subsequent displacement calculations.

[0055] Then the displacement d[m] (unit: pixel) of the m-th time slice is: d[m] = v_edge[m] - v_zero; Where d[m] is the row offset of the valve stem feature edge relative to its stationary position in the m-th time slice; d[m]>0 indicates that the valve stem moves along the positive V-axis, and d[m]<0 indicates that it moves along the negative V-axis.

[0056] In actual deployment, a conversion factor (scale, unit: mm / pixel) needs to be obtained through pre-executed pixel millimeter calibration to uniformly convert the D sequence to millimeter units before feature parameter extraction. The conversion formula is: d_mm[m] = d[m] × scale.

[0057] Thus, the discrete time sequence D of valve stem displacement is obtained: D=[(t_center[1],d_mm[1]),(t_center[2],d_mm[2]),...,(t_center[M],d_mm[M])]; Where t_center[m]=(T_start[m]+T_end[m]) / 2 is the center time of the m-th time slice.

[0058] S1023. Based on the discrete sequence of valve stem displacement time and the triggering time, the action feature parameters are extracted to obtain a standardized feature vector.

[0059] In some embodiments of this application, the following steps are taken: determining the target displacement of the valve's rated stroke and the allowable error band for stability determination; extracting the opening delay from the valve stem displacement time discrete sequence based on the target displacement of the valve's rated stroke; determining the maximum displacement from the valve stem displacement time discrete sequence; determining the overshoot based on the maximum displacement and the target displacement of the valve's rated stroke; extracting the settling time from the valve stem displacement time discrete sequence based on the target displacement of the valve's rated stroke and the allowable error band for stability determination; calculating the instantaneous velocity of adjacent time slices based on the valve stem displacement time discrete sequence; determining the peak velocity from the instantaneous velocity of adjacent time slices; calculating the velocity variation coefficient based on the instantaneous velocity of adjacent time slices; and generating a standardized feature vector based on the opening delay, overshoot, settling time, peak velocity, and velocity variation coefficient.

[0060] For example, based on the reconstructed valve stem displacement time series D, and combined with the valve action command type action_cmd (OPEN or CLOSE), five core feature parameters are extracted. Among them, action_cmd is used to determine the target movement direction of the valve stem. When action_cmd=OPEN, the opening direction displacement is taken as the positive direction. When action_cmd=CLOSE, the closing direction displacement is taken as the positive direction. Subsequent displacements and feature parameters are all calculated based on the unified positive direction.

[0061] Define the following intermediate variables: d_target: Target displacement of the valve's rated stroke (in mm), which is obtained from the valve specification and preset. t_trigger: Trigger time (in μs); ε_settle: The allowable error band for stability determination, which is set to ±1% of d_target by default.

[0062] Feature parameter 1: Enable delay t_delay (unit μs).

[0063] The opening delay is the time difference from the trigger moment t_trigger to the start of effective displacement of the valve stem, which directly reflects the response of the solenoid valve and the static friction resistance of the valve stem.

[0064] Find the moment t_start in D when the displacement first exceeds the starting threshold d_thresh_start (d_thresh_star (in mm), preset to 2% of the rated target displacement d_target): t_start=min{t_center[m]|d_mm[m]≥d_thresh_start}; t_delay = t_start - t_trigger; Corresponding anomaly: t_delay increases significantly, causing crawling to become stuck.

[0065] Characteristic parameter 2: Overshoot (%) Overshoot is the ratio of the maximum peak value of the valve stem exceeding the target displacement during its movement to the target displacement, reflecting the effectiveness of the buffer mechanism and the degree of impact upon arrival.

[0066] d_peak=max{d_mm[m]|m=1,2,...,M}, where M represents the total number of time slices; overshoot=max(0,(d_peak-d_target) / d_target×100%).

[0067] Corresponding anomaly: Significantly increased overshoot, leading to overshoot rebound, which may be accompanied by the risk of impact to the sealing surface.

[0068] Feature parameter 3: Settling time t_settle (unit: μs).

[0069] The settling time is the time interval from the trigger moment t_trigger until the valve stem enters the allowable error band of the target displacement and remains stable without exceeding the error band. It is used to reflect the convergence speed of residual oscillations after the valve stem is in place.

[0070] Find the steady-state moment t_settle_moment from D. This moment satisfies: Starting from t_settle_moment, the displacements of all subsequent sampling points lie within the interval [d_target-ε_settle, d_target+ε_settle] and do not exceed this error band. Therefore, the settling time is calculated as follows: t_settle=t_settle_moment-t_trigger; Corresponding anomaly: t_settle increases significantly, causing tremor oscillation, or the buffer mechanism fails.

[0071] Feature parameter four: Peak velocity v_peak (unit: mm / ms).

[0072] The peak velocity is the maximum value of the absolute value of the first-order difference of the valve stem displacement time series D, which represents the maximum instantaneous velocity during the valve stem movement and is used to assess whether the air source pressure is sufficient and whether the sealing resistance is abnormal.

[0073] Calculate the instantaneous velocity v_inst[m] and peak velocity v_peak of adjacent time slices: v_inst[m]=|d_mm[m+1]-d_mm[m]| / (t_center[m+1]-t_center[m])×1000; v_peak=max{v_inst[m]|m=1,2,...,M-1}; Corresponding anomalies: v_peak is significantly lower than the rated empirical value, resulting in jammed crawling or insufficient gas supply pressure; v_peak is significantly higher than the rated empirical value, resulting in buffer failure or excessively high gas source pressure.

[0074] Feature parameter 5: Velocity variation coefficient cv_v (unitless).

[0075] The coefficient of variation of velocity is the ratio of the standard deviation of instantaneous velocity to the mean. It is used to quantify the relative amplitude of velocity fluctuations during motion and is the most direct indicator of the severity of tremor / oscillation.

[0076] First, calculate the mean μ_v and standard deviation σ_v of the instantaneous velocity: Then the velocity variation coefficient is calculated as: cv_v = σ_v / μ_v.

[0077] The larger the cv_v, the more violent the speed fluctuation during the valve stem movement, and the worse the movement stability.

[0078] Corresponding anomaly: a significant increase in cv_v, leading to chattering or valve stem creep (alternating "jamming and slippage").

[0079] S103. Using a pre-determined lightweight fully connected neural network model, valve action anomaly detection is performed on the standardized feature vector to obtain the probability vector of valve anomaly.

[0080] In some embodiments of this application, a lightweight fully connected neural network model is first described. This model is designed to accurately learn the mapping relationship between standardized feature vectors and the probability vectors of valve anomalies. The lightweight fully connected neural network model employs a three-layer fully connected feedforward structure, the specific configuration of which is shown in Table 1.

[0081] Table 1 The standardized feature vectors of the input to the lightweight fully connected neural network model are shown in Table 2.

[0082] Table 2 The training method includes three parts: training dataset construction, network architecture and training objective design, and model training and optimization strategies.

[0083] First, the construction of the training dataset includes data sources, sample construction, and dataset partitioning. The data sources consist of a large number of valve action samples collected on a valve testing platform using the methods described above. Each valid sample contains input and a label. The input is a standardized feature vector F_norm with a dimension of 5, containing the opening delay t_delay, overshoot, settling time t_settle, peak velocity v_peak, and velocity variation coefficient cv_v. The label is an anomaly category label Y, a one-hot vector of length 4, corresponding to four anomaly types: stuck crawling, trembling oscillation, overshoot rebound, and action timeout.

[0084] Sample construction involves extracting a standardized feature vector F_norm for each valve action, following a valley flow pattern. This vector is then categorized by domain experts using synchronously acquired valve stem motion curves. The categorization rules are as follows: If a sample exhibits only a single anomaly, it is labeled with that anomaly category. If a sample exhibits multiple anomalies, it is labeled with the category that has the most significant feature, and the composite anomaly label is recorded for subsequent analysis. Normal samples are not used in model training but are only used for threshold statistics.

[0085] The dataset was divided according to the collection time sequence. It was divided into training set, validation set, and test set in a ratio of 70%, 15%, and 15%, respectively. During the division, it was ensured that the continuous operation of the same valve was not scattered in the training set and the validation set.

[0086] Second, the training objective is to minimize the cross-entropy loss, making the probability of the true class as close to 1 as possible in the 4-dimensional probability vector P output by the network, and the probabilities of the other classes as close to 0 as possible, thereby achieving accurate anomaly classification. The loss function used is the cross-entropy loss function. Since the output is a 4-class probability distribution, the cross-entropy loss function is as follows: Loss=-(Y[0]×log(P[0])+Y[1]×log(P[1])+Y[2]×log(P[2])+Y[3]×log(P[3])); Where Y[j] is the j-th dimension of the one-hot label vector (the true class is 1, and the rest are 0), and P[j] is the probability of the j-th class predicted by the network.

[0087] Third, model training and optimization strategies. The optimizer used is Adam, with an initial learning rate of 1×10⁻⁶. -3 The learning rate scheduling strategy uses ReduceLROnPlateau; the weight decay is 1×10. -4 Monitor the validation set loss, patience=10, factor=0.5, min_lr=1×10. -6 The batch size is 16; the maximum number of training epochs is 500; the training stops if the validation set loss does not decrease for 20 consecutive epochs, and the optimal weights are restored. The training process is as follows: A batch is randomly sampled from the training set (input feature shape (Batch, 5), label shape (Batch, 4)); forward propagation yields the predicted probability vector P, shape (Batch, 4); cross-entropy loss is calculated; backpropagation is performed to calculate the gradient; gradient clipping (norm 1.0) is applied to prevent gradient explosion; the optimizer updates the network parameters; after each epoch, the loss and classification accuracy are evaluated on the validation set; if the validation set loss does not decrease for 10 consecutive epochs, the learning rate is halved (minimum 1×10). -6 If the validation set loss does not decrease for 20 consecutive epochs, stop training and restore the optimal weights; after training is complete, save the model with the lowest validation set loss for online inference.

[0088] In some embodiments of this application, S103 can be implemented by S1031-S1033, as follows: S1031. The standardized feature vector is mapped through the first fully connected layer of the lightweight fully connected neural network model to obtain intermediate features.

[0089] S1032. Through the second fully connected layer of the lightweight fully connected neural network model, the intermediate features are abstracted and compressed to obtain deep features; S1033. The output layer of a lightweight fully connected neural network model is used to map deep features to obtain the probability vector of valve anomalies.

[0090] For example, fully connected layer 1 (i.e., the first fully connected layer) maps the input 5-dimensional normalized feature vector to 16-dimensional intermediate features. This dimensionality increase operation gives the network sufficient expressive power to cross-correlate the feature parameters in a unified semantic space, initially extracting clues to determine "what kind of abnormal mode the current action feature combination belongs to". For example, this layer can learn the cooperative relationship between t_delay and v_peak—if both deviate significantly from the rated empirical value at the same time, the probability of jamming / crawling increases. The ReLU activation function introduces non-linearity, enabling the network to fit the non-linear mapping relationship between feature parameters and abnormality types.

[0091] Fully connected layer 2 (i.e., the second fully connected layer) further abstracts the 16-dimensional intermediate features output by fully connected layer 1 into 8-dimensional deep features. This layer is responsible for extracting the following implicit relationships: the combined features of overshoot and settling time t_settle; and the joint distribution of velocity variation coefficient cv_v and peak velocity v_peak. Through the nonlinear superposition of 16-dimensional intermediate features to 8-dimensional deep features, the network can construct a continuous mapping relationship from feature parameters to anomaly types, providing the deep feature representation required for anomaly decision-making in the output layer.

[0092] The output layer maps the 8-dimensional deep features output from the fully connected layer 2 to a 4-dimensional output space, corresponding to the four anomaly types. The Softmax activation function is used to convert the original network output values ​​(logits) into a probability vector of valve anomalies, satisfying ΣP[j]=1 and each component located in the interval [0,1], so that the output results can be directly used for threshold determination.

[0093] The network output is a 4-dimensional probability vector P, with each component defined as follows: P[0] indicates stuck crawling, P[1] indicates tremor oscillation, P[2] indicates overshoot rebound, and P[3] indicates action timeout.

[0094] The output probability vector P satisfies the normalization constraint: P[0]+P[1]+P[2]+P[3]=1, and the values ​​of each component are between 0 and 1, which represents the probability estimate of the action belonging to the corresponding category.

[0095] S104. Based on the probability vector of valve abnormality and the preset threshold, determine the final diagnostic result.

[0096] In some embodiments of this application, the preset thresholds include a threshold for stuck crawling, a threshold for tremor oscillation, a threshold for overshoot rebound, and a threshold for action timeout.

[0097] In some embodiments of this application, the probability of various abnormalities in the probability vector of valve abnormalities is compared with the thresholds for stuck crawling, tremor oscillation, overshoot rebound, and action timeout to determine the valve abnormality type; based on the valve abnormality type, a final diagnostic result is generated.

[0098] Understandably, by collecting the original event sequence and trigger time of the entire valve stem movement process, and then reconstructing the event flow and extracting feature parameters based on the original event sequence and trigger time to obtain a standardized feature vector, the problem that existing valve position sensors can only detect the switch position and cannot capture the dynamic details of the entire ultra-high speed movement can be solved. This provides a complete data foundation for the early identification of dynamic anomalies such as jamming, crawling, and oscillation. By constructing a lightweight fully connected neural network model and combining it with preset thresholds to determine the final diagnostic result, the problem that existing methods lack the ability to classify and judge systematic anomalies is solved. This improves the accuracy of judging valve ultra-high speed movement anomalies, effectively reduces the blindness of fault diagnosis, and shortens downtime.

[0099] For example, the probabilities of various anomalies are compared with preset thresholds, and all anomaly types that exceed the thresholds are output.

[0100] The preset thresholds for each category are as follows: The threshold for kaser crawling, thr_stick, is 0.70; The threshold for tremor oscillation, thr_vibrate, is 0.75; The overshoot threshold thr_overshoot is 0.65; The threshold for action timeout, thr_timeout, is 0.80.

[0101] Judgment rule: For each anomaly category j, if P[j]>thr[j], then it is determined to be that anomaly type and an alarm is triggered.

[0102] The final diagnostic result (Result) is a collection of all abnormal types that exceed the threshold: Result={j|P[j]>thr[j],j=0,1,2,3}; If Result is an empty set (i.e., all P[j]≤thr[j]), then it is considered normal.

[0103] Where thr[j] represents the judgment threshold corresponding to the j-th type of anomaly, which corresponds to thr_stick, thr_vibrate, thr_overshoot and thr_timeout respectively.

[0104] This application has the following beneficial effects: First, a DVS is installed on the side of the valve stem and asynchronous event streams of the entire valve stem movement are acquired with microsecond-level time resolution. The complete valve stem displacement time series is reconstructed by generating event counting images and edge projection. This solves the problem that existing valve position sensors can only detect the switch position and cannot capture the dynamic details of the entire ultra-high speed movement. It provides a complete data foundation for the early identification of dynamic anomalies such as jamming, crawling, and tremor.

[0105] Secondly, five physically meaningful feature parameters—opening delay, overshoot, settling time, peak velocity, and velocity variation coefficient—are extracted from the reconstructed valve stem displacement time series. A lightweight fully connected neural network model is constructed, and combined with threshold judgment, the probability and diagnostic results of various anomalies are output. This solves the problem that existing methods lack the ability to classify and judge systematic anomalies, and realizes end-to-end automatic classification and identification of valve stem motion signals into four typical anomalies. This can effectively reduce the blindness of fault diagnosis and downtime.

[0106] Based on the valve ultra-high speed operation anomaly judgment method of the above embodiments, this application embodiment also provides a valve ultra-high speed operation anomaly judgment system, such as... Figure 2 As shown, Figure 2 This is a schematic diagram of a valve ultra-high speed action anomaly judgment system provided in an embodiment of this application. The valve ultra-high speed action anomaly judgment system 2 includes: an acquisition module 201, an extraction module 202, a detection module 203, and a determination module 204; wherein, the acquisition module 201 is used to collect raw data of the entire movement process of the valve stem of the pneumatic valve actuator to obtain the raw event sequence and trigger time; The extraction module 202 is used to reconstruct the event stream and extract feature parameters based on the original event sequence and trigger time to obtain a standardized feature vector; The detection module 203 is used to detect valve action anomalies by using a pre-determined lightweight fully connected neural network model to standardize the feature vector, and obtain the probability vector of valve anomalies. The determination module 204 is used to determine the final diagnostic result based on the probability vector of valve anomalies and preset thresholds; The extraction module 202 is also used to perform event stream time-domain accumulation based on the original event sequence to obtain an event count image sequence; to reconstruct the valve stem displacement time sequence from the event count image sequence to obtain a valve stem displacement time discrete sequence; and to extract action feature parameters based on the valve stem displacement time discrete sequence and the triggering time to obtain a standardized feature vector.

[0107] In this embodiment, the extraction module 202 is further configured to divide the original event sequence into time slices through a pre-set cumulative time window to obtain multiple time slices; perform event statistics on the multiple time slices to obtain event subsets for each of the multiple time slices; generate event counting images for each of the multiple time slices according to a preset region of interest based on the event subsets for each of the multiple time slices; and generate an event counting image sequence based on the event counting images for each of the multiple time slices.

[0108] In this embodiment, the extraction module 202 is further configured to perform edge response aggregation on the event count image of each time slice in the event count image sequence to obtain the edge response value of each time slice; perform precise edge location based on the edge response value of each time slice to obtain the edge position of each time slice; perform displacement physical unit conversion based on the edge position of each time slice to obtain the displacement of each time slice; and determine the valve stem displacement time discrete sequence based on the displacement of each time slice, as well as the start time and end time of each time slice.

[0109] In this embodiment, the extraction module 202 is further used to determine the target displacement of the valve's rated stroke and the allowable error band for stability determination; extract the opening delay from the valve stem displacement time discrete sequence based on the target displacement of the valve's rated stroke; determine the maximum displacement from the valve stem displacement time discrete sequence; and determine the overshoot based on the maximum displacement and the target displacement of the valve's rated stroke; extract the settling time from the valve stem displacement time discrete sequence based on the target displacement of the valve's rated stroke and the allowable error band for stability determination; calculate the instantaneous velocity of adjacent time slices based on the valve stem displacement time discrete sequence; determine the peak velocity from the instantaneous velocity of adjacent time slices; calculate the velocity variation coefficient based on the instantaneous velocity of adjacent time slices; and generate a standardized feature vector based on the opening delay, overshoot, settling time, peak velocity, and velocity variation coefficient.

[0110] In this embodiment, the detection module 203 is further configured to map the standardized feature vector through the first fully connected layer of the lightweight fully connected neural network model to obtain intermediate features; to abstract and compress the intermediate features through the second fully connected layer of the lightweight fully connected neural network model to obtain deep features; and to map the deep features through the output layer of the lightweight fully connected neural network model to obtain the probability vector of valve abnormality.

[0111] In this embodiment, the preset thresholds include the threshold for stuck crawling, the threshold for tremor oscillation, the threshold for overshoot rebound, and the threshold for action timeout. The determination module 204 is also used to compare the probability of various anomalies in the probability vector of valve anomalies with the thresholds for stuck crawling, tremor oscillation, overshoot rebound, and action timeout to determine the valve anomaly type; and generate the final diagnostic result based on the valve anomaly type.

[0112] Based on the valve ultra-high speed operation anomaly judgment method according to the above embodiments, this application embodiment also provides a valve ultra-high speed operation anomaly judgment device, such as... Figure 3 As shown, Figure 3 This is a schematic diagram of a valve ultra-high speed actuation anomaly judgment device provided in an embodiment of this application. The valve ultra-high speed actuation anomaly judgment device 3 includes: a processor 301 and a memory 302. The memory 302 is used to store computer programs; the processor 301 is used to call and run the computer programs from the memory to execute a valve ultra-high speed actuation anomaly judgment method as described in the above embodiment.

[0113] In the embodiments of this application, the processor 301 described above can be at least one of the following: Application-Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), Central Processing Unit (CPU), Controller, Microcontroller, and Microprocessor. It is understood that for different devices, the electronic device used to implement the above processor function can also be other types, and the embodiments of this application do not specifically limit it.

[0114] This application provides a computer-readable storage medium storing a computer program for implementing, when executed by a processor, a method for judging abnormal valve operation at ultra-high speed as described in any of the above embodiments.

[0115] For example, the program instructions corresponding to the valve ultra-high speed action anomaly judgment method in this embodiment can be stored on storage media such as optical disc, hard disk, and USB flash drive. When the program instructions corresponding to the valve ultra-high speed action anomaly judgment method in the storage medium are read or executed by an electronic device, the valve ultra-high speed action anomaly judgment method as described in any of the above embodiments can be implemented.

[0116] Furthermore, in the embodiments of this application, the functional modules can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional module.

[0117] If the integrated unit is implemented as a software functional module and is not sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this embodiment, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the method of this embodiment. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0118] It should be understood that the phrases "one embodiment," "an embodiment," or "some embodiments" mentioned throughout the specification mean that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "in one embodiment," "in one embodiment," or "in some embodiments" appearing throughout the specification do not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of this application, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application. The sequence numbers of the embodiments in this application are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. The descriptions of the various embodiments above tend to emphasize the differences between the various embodiments; their similarities or commonalities can be referred to mutually, and for the sake of brevity, these will not be repeated here.

[0119] The modules described above as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected to achieve the purpose of this embodiment according to actual needs.

[0120] In addition, each functional module in the various embodiments of this application can be integrated into one processing unit, or each module can be a separate unit, or two or more modules can be integrated into one unit; the integrated modules can be implemented in hardware or in the form of hardware plus software functional units.

[0121] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.

[0122] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0123] The features disclosed in the several product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0124] The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0125] The above description is merely an implementation method of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present application should be included within the protection scope of the present application.

Claims

1. A method for judging abnormal ultra-high speed operation of a valve, characterized in that, include: Raw data was collected from the entire movement process of the valve stem of the pneumatic valve actuator to obtain the raw event sequence and trigger time; Based on the original event sequence and the triggering time, event flow reconstruction and feature parameter extraction are performed to obtain a standardized feature vector; Using a pre-defined lightweight fully connected neural network model, valve action anomaly detection is performed on the standardized feature vector to obtain a probability vector of valve anomaly. Based on the probability vector of the valve malfunction and a preset threshold, the final diagnostic result is determined; The step of reconstructing the event stream and extracting feature parameters based on the original event sequence and the triggering time to obtain a standardized feature vector includes: Based on the original event sequence, event stream temporal accumulation is performed to obtain an event count image sequence; The event count image sequence is reconstructed using a valve stem displacement time sequence to obtain a discrete valve stem displacement time sequence; Based on the discrete sequence of valve stem displacement time and the triggering time, action feature parameters are extracted to obtain the standardized feature vector.

2. The method according to claim 1, characterized in that, The step of performing time-domain accumulation of the event stream based on the original event sequence to obtain an event count image sequence includes: The original event sequence is divided into time slices by a pre-set cumulative time window to obtain multiple time slices; Event statistics are performed on each of the multiple time slices to obtain a subset of events for each of the multiple time slices; Based on the event subsets of each of the multiple time slices, event count images of each of the multiple time slices are generated according to a preset region of interest. An event count image sequence is generated based on the event count images of each of the multiple time slices.

3. The method according to claim 1, characterized in that, The step of reconstructing the valve stem displacement time series from the event count image sequence to obtain a discrete valve stem displacement time series includes: For the event count image of each time slice in the event count image sequence, edge response aggregation is performed to obtain the edge response value of each time slice; Based on the edge response value of each time slice, the edge position is accurately located to obtain the edge position of each time slice; Based on the edge position of each time slice, the displacement physical unit is converted to obtain the displacement of each time slice; Based on the displacement of each time slice, and the start and end times of each time slice, the discrete time sequence of valve stem displacement is determined.

4. The method according to claim 1, characterized in that, The process of extracting action feature parameters based on the discrete sequence of valve stem displacement time and the triggering time to obtain the standardized feature vector includes: Determine the target displacement of the valve's rated stroke and the allowable error band for stability assessment; Based on the target displacement of the valve's rated stroke, the opening delay is extracted from the discrete sequence of valve stem displacement time. The maximum displacement is determined from the discrete time sequence of the valve stem displacement; and the overshoot is determined based on the maximum displacement and the target displacement of the valve's rated stroke. Based on the target displacement of the valve's rated stroke and the allowable error band for stability determination, the stable time is extracted from the discrete sequence of valve stem displacement time. Based on the discrete time sequence of valve stem displacement, the instantaneous velocity of adjacent time slices is calculated; and the peak velocity of motion is determined from the instantaneous velocity of adjacent time slices. The velocity variation coefficient is calculated based on the instantaneous velocity of the adjacent time slices; The standardized feature vector is generated based on the activation delay, the overshoot, the settling time, the peak motion speed, and the velocity variation coefficient.

5. The method according to claim 1, characterized in that, The step of using a pre-determined lightweight fully connected neural network model to perform valve action anomaly detection on the standardized feature vector, obtaining a probability vector of valve anomalies, includes: The standardized feature vector is mapped through the first fully connected layer of the lightweight fully connected neural network model to obtain intermediate features; The intermediate features are abstracted and compressed through the second fully connected layer of the lightweight fully connected neural network model to obtain deep features; The output layer of the lightweight fully connected neural network model is used to map the deep features to obtain the probability vector of the valve anomaly.

6. The method according to claim 1, characterized in that, The preset thresholds include the threshold for stuck crawling, the threshold for tremor oscillation, the threshold for overshoot rebound, and the threshold for action timeout. The determination of the final diagnostic result based on the probability vector of the valve anomaly and a preset threshold includes: The probability of each type of abnormality in the probability vector of the valve abnormality is compared with the threshold of jamming crawl, the threshold of tremor oscillation, the threshold of overshoot rebound, and the threshold of action timeout to determine the valve abnormality type. The final diagnostic result is generated based on the valve anomaly type.

7. A valve ultra-high speed actuation abnormality judgment system, characterized in that, include: The system includes an acquisition module, an extraction module, a detection module, and a determination module; wherein, the acquisition module is used to collect raw data of the entire movement process of the pneumatic valve actuator stem to obtain the raw event sequence and trigger time; The extraction module is used to reconstruct the event stream and extract feature parameters based on the original event sequence and the triggering time to obtain a standardized feature vector; The detection module is used to detect valve action anomalies by using a pre-determined lightweight fully connected neural network model to analyze the standardized feature vector, thereby obtaining a probability vector of valve anomalies. The determining module is used to determine the final diagnostic result based on the probability vector of the valve abnormality and a preset threshold. The extraction module is also used to perform event stream temporal accumulation based on the original event sequence to obtain an event count image sequence; The event count image sequence is reconstructed using a valve stem displacement time sequence to obtain a discrete valve stem displacement time sequence; Based on the discrete sequence of valve stem displacement time and the triggering time, action feature parameters are extracted to obtain the standardized feature vector.

8. A device for judging abnormal operation of a valve at ultra-high speed, characterized in that, include: A processor and a memory, wherein the memory is used to store computer programs; The processor is configured to call and run the computer program from the memory to perform the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores executable instructions for causing a processor to execute, thereby implementing the method of any one of claims 1 to 6.

Images