Abnormality detection method for electromagnetic valve pressure and flow distribution deviation

Abstract

The invention discloses an anomaly detection method for pressure and flow distribution deviation of an electromagnetic valve, which belongs to the technical field of anomaly detection and comprises the following steps of: constructing a sample corresponding to a window; calculating a quantile value, a smooth value, a normalized trend and a trend direction of each quantile of each window, and generating a non-zero direction set of each window; defining a splitting index of the window k; defining a state vector and a criterion parameter vector of the window k, defining an action space, and pre-training a strategy model; and sequentially marking each window, if the window k is normal, marking the next window, otherwise, judging that the window k enters a split state, and further judging based on a strategy model until the split is recovered or an abnormal alarm is triggered. According to the method, the defects of a traditional method in the aspect of progressive distribution offset recognition can be effectively overcome, the precision and stability of pressure and flow anomaly detection of the electromagnetic valve are remarkably improved, and an innovative solution with high robustness and engineering adaptation capacity is provided for an industrial fluid control system.

Description

Technical Field

[0001] This invention relates to the field of solenoid valve anomaly detection technology, and in particular to an anomaly detection method for pressure and flow distribution deviations in solenoid valves. Background Technology

[0002] With the rapid development of industrial automation and intelligent monitoring technologies, data-driven anomaly detection methods have shown great potential in areas such as industrial equipment operation monitoring, fluid control system safety assurance, and process control status assessment. As an intelligent detection method for industrial operation status analysis, anomaly identification technology based on distribution feature modeling, by introducing multi-scale statistical feature analysis and temporal behavior modeling, not only significantly improves the sensitivity of anomaly identification but also ensures a stable characterization of equipment operation status changes, thereby promoting the development of intelligent operation and maintenance and predictive maintenance technologies. In the scenario of solenoid valve operation monitoring, ensuring the accurate characterization of pressure and flow data distribution characteristics is crucial for achieving equipment health assessment and anomaly early warning. Practical applications often require detection methods to identify slowly evolving anomalies under complex operating conditions and possess strong resistance to transient interference. However, existing technologies often rely on fixed thresholds or single statistical indicators, making it difficult to effectively characterize the gradual changes in data distribution structure. This leads to problems such as identification lag or high false alarm rates in scenarios involving solenoid valve aging, blockage, or performance degradation. Summary of the Invention

[0003] The purpose of this invention is to provide an anomaly detection method for pressure and flow distribution deviation of solenoid valves, which addresses the shortcomings of traditional methods in identifying gradual distribution offsets, characterizes the gradual change features of data distribution structure, and thus improves the accuracy and stability of anomaly detection for solenoid valve pressure and flow.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: an anomaly detection method for the deviation of pressure and flow distribution in a solenoid valve, comprising the following steps: S1, construct samples using a sliding window on the sampling sequence X of the solenoid valve, wherein the sampled values ​​in X are flow rate values ​​or pressure values, and window k corresponds to sample X. k ; S2, select M quantiles, and obtain the i-th quantile q in window k. k,i quantile Processed into smooth values , 1≤i≤M; S3, q is calculated based on the smoothing value. k,i normalization trend and trend direction Construct a set of non-zero directions , ,and It contains m k One element; S4, a splitting index for window k based on normalized trend and trend direction, including directional splitting degree S. k Amplitude split degree V k Number of consecutive split windows L k Overall drift intensity D k ; S5, Define the state vector s of window k. k and criterion parameter vector θ k Define action space A and pre-train a policy model; s k =[S k V k ,L k D k ] T θ k =[W k ,τ s,k ,ϵ d,k ,L min,k ],A={a (1) ,…,a (6)}, where W k τ s,k ϵ d,k L min,k These are the current window length, the current direction split threshold, the current direction dead zone threshold, and the minimum duration required to confirm an anomaly, respectively, and action a. (1) ~a (6) Increase W respectively k Reduce W k Increase τ s,k Decrease τ s,k Increase ϵ d,k and trigger confirmation; The strategy model is used to determine based on s k Select action a k a k ∈A, perform action a k Then, update the criterion parameter vector θ of window k+1. k+1 ; S6, mark each window sequentially, where window k is marked using the following method: If S k ≤τ s,k If window k is marked as normal, otherwise it is determined that window k has entered a split state and S7 is executed; S7, update the criterion parameter vectors of the subsequent Tr windows using the policy model and observe the trend consistency of the corresponding windows, where the criterion parameter vector of the t-th window is θ. k+tThe trend consistency value for the corresponding window k+t is C. k+t , 1≤t≤Tr; If C k+t ≥1- If the condition is not met, it is considered a split recovery. Window k to window k+t-1 is marked as abnormal and window k+t as normal. The next judgment window jumps to window k+t+1. If C k+1 ~C k+Tr All less than 1- If the current split is determined to be an irreversible distribution offset anomaly, an anomaly alarm will be triggered.

[0005] As a preferred embodiment: S1 includes S11~S12; S11, the pressure data or flow data of the solenoid valve is sampled according to a preset period to obtain sampled values, and the sampled values ​​are arranged in time sequence to form a sampling sequence X, wherein the sampled values ​​are the pressure values ​​of the pressure data or the flow values ​​of the flow data. S12, use a window of length W and step size H to slide sampling on X, and form a sample by sampling points located in the window.

[0006] Preferably, step S2 includes S21 to S23; S21, select M quantiles in ascending order between (0,1), where the i-th quantile in window k is labeled q. k,i ; S22, obtain q using the quantile function. k,i quantile The quantile structure vector Q that constitutes window k k , T represents the transpose operation; S23, for Q k Perform first-order exponential smoothing to obtain the smoothed vector. , and that is , for The smoothed value.

[0007] As a preferred option: S3 specifically includes S31~S32; S31, Define q k,i normalization trend and trend direction ; , , In the formula, For sample X k-1 The smoothed value corresponding to the i-th quantile, ε r It is a local minimum value greater than 0. The threshold used to filter noise disturbances; S32, Construct the quantile trend direction vector d k Non-zero direction set , , ,and It contains m k Each element.

[0008] As a preferred option, S4 specifically refers to: , , In the formula C k This represents the trend consistency value. For indicator functions, for The j-th trend direction, for The corresponding m k indivual The mean, The preset direction splitting threshold.

[0009] As a preferred option: if an abnormal alarm is triggered in S7, proceed to step S8; S8, continuously monitor the trend consistency value of each subsequent window until the trend consistency value of a certain window B is ≥1- Window k to window B-1 are marked as abnormal, window B is marked as normal, and the next judgment window jumps to window B+1.

[0010] In this invention, regarding S2: based on quantiles at X k The quantile value corresponding to each quantile point is obtained, and to facilitate subsequent smoothing processing, the quantile values ​​are used to form the quantile structure vector Q of window k. k Then perform first-order smoothing to obtain the smoothed vector. , Inner element and Q k One-to-one correspondence, therefore and One-to-one correspondence. This invention uses M quantile values ​​to characterize the changing behavior of pressure and flow data in different statistical intervals, thereby effectively improving the ability to perceive changes in distribution patterns. The quantile structure vector is then processed into a smooth vector to make subsequent trend calculations more stable.

[0011] Regarding S3: The purpose is to obtain the trend direction at each quantile of window k and to find all non-zero trend directions of window k. This invention analyzes the smoothing vectors of two adjacent windows. and To construct the quantile trend direction vector of window k ,use This reflects the changing trend of the i-th quantile. For example, to calculate... You need to first follow the formula calculate , Data at quantile i for window k-1 and window k and The differences are related, and we obtain Then with By comparison, we can analyze the changing trend of quantile i from window k-1 to window k, and use the trend direction... This means that in the present invention The value is +1, -1, or 0, where +1 indicates an upward trend, -1 indicates a downward trend, and 0 indicates relative stability. It is considered valid when it is not 0, and is added to In the middle. And for If the system is in normal working condition, the different sample subgroups represented by each quantile should move in coordination and in roughly the same direction; if a gradual failure occurs, it will often affect a certain part of the interval first, thus causing directional splitting.

[0012] Regarding S4: Define the splitting index of window k, including the directional splitting degree S. k Amplitude split degree V k Number of consecutive split windows L k Overall drift intensity D k Among them, the directional splitting degree S k Used to reflect the consistency of the splitting direction, S k The larger the value, the more inconsistent the directions of the quantiles; the greater the amplitude splitting degree V. k This value reflects the degree of dispersion in the trend amplitude at each quantile; a larger value indicates a greater degree of divergence in amplitude. However, since start-ups, shutdowns, and upstream disturbances can cause temporary divergences, S should not be considered alone. k and V k Furthermore, criteria for persistence and recoverability are still needed. This invention introduces the number of consecutive split windows L. k ,when Only then will L k In addition, to avoid situations where the drift direction is consistent but the overall drift is slow, an overall drift intensity D is introduced. k This describes the degree of movement of the overall distribution. These parameters are used to construct the state vector s of window k. k .

[0013] Regarding S5: This invention is based on S4, using S... k V k L k D k Construct the state vector s of window k together kFurthermore, reinforcement learning is introduced, using a pre-trained policy model to select actions that change the criterion parameter vector θ for the next window. k+1 Thus changing θ k+1 The parameters in the table are used for subsequent anomaly detection.

[0014] Regarding S6: This invention divides the anomaly determination into two steps. The first step is to determine whether window k is normal or has entered a split state. If it is normal, no policy model intervention is required, and the next window k+1 is determined. If it has entered a split state, the second step of determination is performed according to step S7.

[0015] Regarding S7: The second step requires the introduction of a policy model, which updates the criterion parameter vectors for at most Tr windows and observes the trend consistency of the corresponding windows, including one C. k+t ≥1- If it is considered a split recovery, the window from window k to the window before recovery k+t-1 is marked as abnormal. The next loop starts judging from window k+t+1. If C k+1 ~C k+Tr All less than 1- If the current split is determined to be an irreversible distribution offset anomaly, an anomaly alarm will be triggered.

[0016] Compared with the prior art, the advantages of the present invention are as follows: (1) This invention introduces a multiquantile distribution structure modeling method. According to the S2 method, M quantile values ​​are used to construct the quantile structure vector of each window. The quantile values ​​are used to characterize the changing behavior of pressure and flow data in different statistical intervals, thereby effectively improving the ability to perceive changes in distribution patterns. This method not only improves the problem that traditional mean or variance is difficult to reflect the evolution of local distribution, but also ensures a stable description of the overall data structure changes, providing more comprehensive data support for the analysis of solenoid valve operating status.

[0017] (2) This invention introduces quantile change trend consistency analysis. First, based on the quantile structure vector and corresponding smoothing vector constructed in step S3, a quantile trend direction vector for each window is constructed to reflect the change trend of each quantile point. Then, in step S4, the splitting index of window k is defined, where the directional splitting degree S k Used to reflect the consistency of the splitting direction, amplitude splitting degree V k The number of consecutive split windows L is used to reflect the degree of tearing in amplitude. k To avoid temporary splitting caused by start-up / shutdown and upstream disturbances, the overall drift strength D k Used to characterize the degree of movement of the overall distribution. Then use S... k V k L k D k Construct state vector s kFurthermore, S5 introduces a reinforcement learning policy model that dynamically adjusts the criterion parameter vector θ based on the window state vector. k This method can identify gradual changes in pressure and flow data distribution without relying on fixed thresholds, improving the stability of anomaly detection. Furthermore, combined with a recoverability determination mechanism, it effectively distinguishes between short-term disturbances and genuine anomalies by observing the recovery of trend consistency, significantly reducing the false alarm rate. Simultaneously, through continuous analysis of the quantile trend structure, it can capture distribution changes caused by local performance degradation of the solenoid valve earlier, achieving early detection of potential anomalies and enhancing the system's early warning capability. This method accurately identifies the phenomenon of internal structural splitting in the distribution by characterizing the collaborative changes of different quantile intervals during the time evolution process. This method not only focuses on the overall trend but also effectively captures the phenomenon of premature degradation in local intervals, significantly improving the accuracy of early identification of gradual anomalies in solenoid valves, especially demonstrating high practical application value in long-term operation monitoring scenarios.

[0018] In summary, this invention introduces multiquantile distribution modeling, trend consistency analysis, and reinforcement learning, which can effectively solve the shortcomings of traditional methods in progressive distribution offset identification, significantly improve the accuracy and stability of solenoid valve pressure and flow anomaly detection, and provide an innovative solution with high robustness and engineering adaptability for industrial fluid control systems. Attached Figure Description

[0019] Figure 1 This is a flowchart of the present invention; Figure 2 The flowchart shows the inference process for anomaly detection in window k. Detailed Implementation

[0020] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0021] Example 1: See Figure 1 and Figure 2 An anomaly detection method for pressure and flow distribution deviation in solenoid valves includes the following steps: S1, construct samples using a sliding window on the sampling sequence X of the solenoid valve, wherein the sampled values ​​in X are flow rate values ​​or pressure values, and window k corresponds to sample X. k ; S2, select M quantiles, and obtain the i-th quantile q in window k. k,i quantile Processed into smooth values , 1≤i≤M; S3, q is calculated based on the smoothing value. k,i normalization trend and trend direction Construct a set of non-zero directions , ,and It contains m k One element; S4, a splitting index for window k based on normalized trend and trend direction, including directional splitting degree S. k Amplitude split degree V k Number of consecutive split windows L k Overall drift intensity D k : , , In the formula C k This represents the trend consistency value. For indicator functions, for The j-th trend direction, for The corresponding m k indivual The mean, The preset direction splitting threshold; S5, Define the state vector s of window k. k and criterion parameter vector θ k Define action space A and pre-train a policy model; s k =[S k V k ,L k D k ] T θ k =[W k ,τ s,k ,ϵ d,k ,L min,k ],A={a (1) ,…,a (6)}, where W k τ s,k ϵ d,k L min,k These are the current window length, the current direction split threshold, the current direction dead zone threshold, and the minimum duration required to confirm an anomaly, respectively, and action a. (1) ~a (6) Increase W respectively k Reduce W k Increase τ s,k Decrease τ s,k Increase ϵ d,k and trigger confirmation; The strategy model is used to determine based on s k Select action a k ak ∈A, perform action a k Then, update the criterion parameter vector θ of window k+1. k+1 ; S6, mark each window sequentially, where window k is marked using the following method: If S k ≤τ s,k If window k is marked as normal, otherwise it is determined that window k has entered a split state and S7 is executed; S7, update the criterion parameter vectors of the subsequent Tr windows using the policy model and observe the trend consistency of the corresponding windows, where the criterion parameter vector of the t-th window is θ. k+t The trend consistency value for the corresponding window k+t is C. k+t , 1≤t≤Tr; If C k+t ≥1- If the condition is not met, it is considered a split recovery. Window k to window k+t-1 is marked as abnormal and window k+t as normal. The next judgment window jumps to window k+t+1. If C k+1 ~C k+Tr All less than 1- If the current split is determined to be an irreversible distribution shift anomaly, an anomaly alarm is triggered. In this embodiment, The value is 0.4.

[0022] Example 2: See Figures 1-2 Based on Example 1, S1 includes S11~S12; S11, the pressure data or flow data of the solenoid valve is sampled according to a preset period to obtain sampled values, and a sampling sequence X is formed according to the time sequence. The sampled values ​​are the pressure values ​​of the pressure data or the flow values ​​of the flow data. If it is pressure data, the obtained sampling sequence X is a sequence of pressure values ​​formed according to the time sequence. If it is flow data, the obtained sampling sequence X is a sequence of flow values ​​formed according to the time sequence. S12, use a window of length W and step size H to slide and sample on X, and define the sample X of window k. k , x s For X k The s-th sampling point.

[0023] Step S2 includes S21 to S23; S21, select M quantiles in ascending order between (0,1), where the i-th quantile in window k is labeled q. k,i For example, if M=5, the five quantiles are 0.1, 0.25, 0.5, 0.75, and 0.9. 0.25 is labeled as q in window k. k,2 ; S22, according to formula Get q k,i quantile , where Quantile(∙,q k,i () is the quantile function, used to return X k In q k,i The quantile values, 1≤i≤M, are used to form Q by combining the M quantile values ​​for smoothing. k , T represents the transpose operation; S23, for Q k Perform first-order exponential smoothing to obtain the smoothed vector. Specifically, you can refer to the formula: The processing involves α, where α is the smoothing coefficient, 0 ≤ α < 1. The smoothed vector for window k-1 is obtained after processing. Structure and Q k It is the same, and also contains M elements, so it is denoted as and will Recorded as The smoothed value.

[0024] S3 specifically includes S31~S32; S31, Define q k,i normalization trend and trend direction ; , , In the formula, For sample X k-1 The smoothed value corresponding to the i-th quantile, ε r It is a local minimum value greater than 0. The threshold used to filter noise disturbances; S32, Construct the quantile trend direction vector d k Non-zero direction set , , ,and It contains m k Each element.

[0025] In the splitting index of S4, the directional splitting degree S k Amplitude split degree V k Number of consecutive split windows L k Overall drift intensity D k Specifically, it is defined according to the following formula: , , In the formula Ck This represents the trend consistency value. For indicator functions, for The j-th trend direction, for The corresponding m k indivual The mean, The preset direction splitting threshold.

[0026] Example 3: See Figures 1-2 Based on Example 1 or Example 2, a specific method for defining the action space A and a method for pre-training the strategy model are given.

[0027] The method for defining the action space A is: A = {a (1) ,…,a (6)}, action a (1) ~a (6) Increase W respectively k Reduce W k Increase τ s,k Decrease τ s,k Increase ϵ d,k And trigger confirmation; action a (1) and a (2) Increasing and decreasing W k The change is ΔW, and the action a (3) and a (4) Increasing and decreasing τ s,k The change is Action a (5) Increasing ϵ d,k The change in ε is Δϵ; W min <W k <W max , 0 < τ s,k <1, 1≤ϵ d,k ≤ϵ max Among them, W min W max These are the preset lower and upper bounds of the window length, respectively. max This is the upper limit of the preset directional dead zone threshold.

[0028] The pre-trained policy model is as follows: Sa1, taking k as the current time, the policy model is based on s k Select action a k , and according to s k and a k The cost of calculating time k ; , In the formula, λ1~λ4 are the weights of the corresponding multiplication terms. As an indicator function, ΔW k For action a k The resulting window change; Sa2, the optimization objective is defined as minimizing the expected cumulative discount cost J(θ); , In the formula, γ∈(0,1) is the discount factor, and u is the future step offset. Let k+u be the cost. For the expectation under strategy π, To find a strategy π The value of γ is the smallest; in this implementation, the value of γ is 0.99. Sa3 samples trajectories by interacting with the environment through the policy model and updates the policy model parameters using the policy gradient algorithm.

[0029] Example 4: Based on Example 1, if an abnormal alarm is triggered in step S7, proceed to step S8; S8, continuously monitor the trend consistency value of each subsequent window until the trend consistency value of a certain window B is ≥1- Window k to window B-1 are marked as abnormal, window B is marked as normal, and the next judgment window jumps to window B+1.

[0030] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting anomalies in the pressure and flow distribution deviation of a solenoid valve, characterized in that, Includes the following steps: S1, construct samples using a sliding window on the sampling sequence X of the solenoid valve, wherein the sampled values ​​in X are flow rate values ​​or pressure values, and window k corresponds to sample X. k ; S2, select M quantiles, and obtain the i-th quantile q in window k. k,i quantile Processed into smooth values , 1≤i≤M; S3, Calculate q based on smoothing value k,i normalization trend and trend direction Construct a set of non-zero directions , ,and It contains m k One element; S4, a splitting index for window k based on normalized trend and trend direction, including directional splitting degree S. k Amplitude split degree V k Number of consecutive split windows L k Overall drift intensity D k ; S5, Define the state vector s of window k. k and criterion parameter vector θ k Define action space A and pre-train a policy model; s k =[S k V k ,L k D k ] T θ k =[W k ,τ s,k ,ϵ d,k ,L min,k ],A={a (1) ,…,a (6) }, where W k τ s,k ϵ d,k L min,k These are the current window length, the current direction split threshold, the current direction dead zone threshold, and the minimum duration required to confirm an anomaly, respectively, and action a. (1) ~a (6) Increase W respectively k Reduce W k Increase τ s,k Decrease τ s,k Increase ϵ d,k and trigger confirmation; The strategy model is used to determine based on s k Select action a k a k ∈A, perform action a k Then, update the criterion parameter vector θ of window k+1. k+1 ; S6, mark each window sequentially, where window k is marked using the following method: If S k ≤τ s,k If window k is marked as normal, otherwise it is determined that window k has entered a split state and S7 is executed; S7, update the criterion parameter vectors of the subsequent Tr windows using the policy model and observe the trend consistency of the corresponding windows, where the criterion parameter vector of the t-th window is θ. k+t The trend consistency value for the corresponding window k+t is C. k+t , 1≤t≤Tr; If C k+t ≥1- If the condition is not met, it is considered a split recovery. Window k to window k+t-1 is marked as abnormal and window k+t as normal. The next judgment window jumps to window k+t+1. If C k+1 ~C k+Tr All less than 1- If the current split is determined to be an irreversible distribution offset anomaly, an anomaly alarm will be triggered.

2. The method for detecting anomalies in the pressure and flow distribution of a solenoid valve according to claim 1, characterized in that: S1 includes S11~S12; S11, the pressure data or flow data of the solenoid valve is sampled according to a preset period to obtain sampled values, and the sampled values ​​are arranged in time sequence to form a sampling sequence X, wherein the sampled values ​​are the pressure values ​​of the pressure data or the flow values ​​of the flow data. S12, use a window of length W and step size H to slide sampling on X, and form a sample by sampling points located in the window.

3. The method for detecting anomalies in the pressure and flow distribution deviation of a solenoid valve according to claim 1, characterized in that: Step S2 includes S21 to S23; S21, select M quantiles in ascending order between (0,1), where the i-th quantile in window k is labeled q. k,i ; S22, obtain q using the quantile function. k,i quantile The quantile structure vector Q that constitutes window k k , T represents the transpose operation; S23, for Q k Perform first-order exponential smoothing to obtain the smoothed vector. , and that is , for The smoothed value.

4. The method for detecting anomalies in the pressure and flow distribution deviation of a solenoid valve according to claim 1, characterized in that: S3 specifically includes S31~S32; S31, Define q k,i normalization trend and trend direction ; ， ， In the formula, For sample X k-1 The smoothed value corresponding to the i-th quantile, ε r It is a local minimum value greater than 0. This is the threshold used to filter out noise disturbances; S32, Construct the quantile trend direction vector d k Non-zero direction set , , ,and It contains m k Each element.

5. The method for detecting anomalies in the pressure and flow distribution of a solenoid valve according to claim 1, characterized in that: S4 specifically refers to: ， ， In the formula C k This represents the trend consistency value. For indicator functions, for The j-th trend direction, for The corresponding m k indivual The mean, The preset direction splitting threshold.

6. The method for detecting anomalies in the pressure and flow distribution deviation of a solenoid valve according to claim 1, characterized in that: In S7, if an abnormal alarm is triggered, proceed to step S8; S8, continuously monitor the trend consistency value of each subsequent window until the trend consistency value of a certain window B is ≥1- Window k to window B-1 are marked as abnormal, window B is marked as normal, and the next judgment window jumps to window B+1.

Images