A deep learning-based fault early warning method and system

Abstract

The application discloses a kind of based on deep learning's fault early warning method and system, it is related to deep learning technical field, by the systematic modeling of hydraulic response relation between valve operation event and pressure time series Pre, flow time series Flo, realize the early identification and quantitative early warning of valve implicit anomaly.Combining target valve normal operation historical data constructs normal hysteresis baseline set Bas, so that hysteresis identification model Mdl can accurately depict the hysteresis behavior distribution under normal operating conditions.Hysteresis feature set Lag generated in real time is input into the hysteresis identification model Mdl, and a single, comparable hysteresis risk score Rsk is obtained by averaging the dimensionless multi-hysteresis feature deviation degree, and an early warning trigger set Trg is formed when the hysteresis risk score Rsk is continuously abnormal, and a structured early warning result set Alr is output.The early warning can be issued when the valve has not completely failed, but only shows action delay or execution not in place.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of deep learning, in particular to a fault early warning method and system based on deep learning. BACKGROUND

[0002] In the production and operation process of water companies, the water supply system is usually in a continuous and high-load operation state, and various pipe network facilities, pump station equipment and regulating valves jointly constitute a complex hydraulic system. With the improvement of informatization and automation level, water service enterprises gradually introduce operation management mode based on data analysis and intelligent algorithm, hoping to realize early perception and risk warning of system operation state through comprehensive analysis of a large amount of operation data.

[0003] In the actual water supply system, the valve is an important executive component for regulating flow and pressure, and its operation state directly affects the hydraulic balance of the pipe network and the reliability of water supply. At present, the monitoring of the operation state of the valve in the water service system mainly depends on the valve position feedback signal, manual inspection or simple rule judgment based on experience. However, in the long-term operation process, the valve is prone to hidden abnormal conditions such as jamming, incomplete opening and closing, performance degradation of the actuator or distortion of the feedback signal. Such abnormalities often do not immediately reflect as obvious valve position abnormalities or alarm signals, making it difficult for existing monitoring methods to discover problems in time and difficult to meet the needs of fine operation management. SUMMARY

[0004] In view of the deficiencies of the prior art, the present application provides a fault early warning method and system based on deep learning, which solves the problems mentioned in the background art.

[0005] To achieve the above purpose, the present application is realized by the following technical scheme: a fault early warning method based on deep learning, comprising the following steps: S1, collecting valve operation records related to the target valve to form a valve operation event set Eve, and synchronously collecting pressure time series of valve associated measuring points to form pressure time series Pre and flow time series to form flow time series Flo, constructing an event window based on the valve operation event set Eve, and preprocessing the pressure time series Pre and the flow time series Flo to obtain an event segment set Seg; S2, based on the event segment set Seg, extracting the hysteresis characteristics between pressure change and flow change to form a hysteresis feature set Lag; S3, based on the historical data of the normal operation of the target valve, constructing a normal hysteresis baseline set Bas from the hysteresis feature set Lag, and training a deep learning model based on the normal hysteresis baseline set Bas to obtain a hysteresis identification model Mdl; S4. Input the real-time generated hysteresis feature set Lag into the hysteresis recognition model Mdl for processing, output the risk assessment result of valve response hysteresis, form a hysteresis risk score Rsk, and generate an early warning trigger set Trg when the hysteresis risk score Rsk meets the preset conditions. S5. Based on the warning results of the valve hidden abnormality output by the warning trigger set Trg, a warning result set Alr is formed and sent down for storage.

[0006] Preferably, S1 includes S11; S11. Read the operation records of the target valve from the water company's existing SCADA data interface, historical database interface, or operation data bus; The operation record includes valve identifier, operation command type, command issuance time, and execution status code; the operation record is processed by event parsing: using the same valve identifier as the primary key, multiple operation records with adjacent time intervals less than the preset merging time are merged into one operation event, and the start time of one operation event is recorded as the event start point, and the end time of one operation event is recorded as the event end point, thus constructing a valve operation event set Eve; Based on the valve operation event set Eve, the range of measurement points that have a hydraulic coupling relationship with the target valve is determined, and pressure and flow sampling values ​​within the range of measurement points are retrieved from the existing measurement point data interface; the pressure sampling values ​​are arranged in chronological order to form a pressure time series Pre, and the flow sampling values ​​are arranged in chronological order to form a flow time series Flo; when the sampling periods of the pressure time series Pre and the flow time series Flo are inconsistent, a time rasterization alignment method is used to map them to a unified time raster; The time granularity of the unified time grid is determined by the least common multiple of the sampling periods of the two, so that the pressure time series Pre and the flow time series Flo have one-to-one corresponding sampling points on the same time axis.

[0007] Preferably, S1 further includes S12; S12. Based on each valve operation event in the valve operation event set Eve, extend a preset pre-time before the event start point and extend a preset post-time after the event end point to form an event window corresponding to the valve operation event, and combine all event windows into an event window set Win; In the event window set Win, corresponding subsequences of the pressure time series Pre and the flow time series Flo, which have the same time range as each event window, are extracted to obtain candidate segments; preprocessing operations are performed on the candidate segments, including missing data repair and abnormal spike suppression, wherein: The missing data repair process employs an interpolation method based on the consistency of adjacent time points, so that the value of the missing point is determined by the effective sampling points on both sides of the missing point; The abnormal spike suppression process employs a suppression method based on sliding window robust statistics, which brings sampling points that exceed the robust range back to the robust range. In the abnormal peak suppression process, the robust center value of the candidate segment within the sliding window is defined as Med, and the robust dispersion of the candidate segment within the sliding window is defined as Mad. Abnormal peaks are suppressed according to the following rules: When the deviation of any sampling point from Med is greater than the product of a preset threshold and Mad, the sampling point is replaced with Med. After completing the preprocessing operation, the candidate fragments are combined into event-level hydraulic response fragments according to time synchronization, forming the event fragment set Seg.

[0008] Preferably, S2 includes S21; S21. Perform time-series analysis on each event-level hydraulic response segment in the event segment set Seg to analyze the dynamic response relationship between pressure changes and flow rate changes before and after valve operation; In each event-level hydraulic response segment, the onset time, process, and stable interval of the pressure and flow rate changes are determined, and hysteresis features describing the pressure change relative to the flow rate change are extracted. These hysteresis features include the time difference between the onset time of the pressure change and the onset time of the flow rate change, the stable duration required for the pressure change to reach the stable interval, the stable duration required for the flow rate change to reach the stable interval, and the relative displacement of the pressure and flow rate changes on the time axis. The hysteresis features extracted from each event-level hydraulic response segment are combined and vectorized, and the hysteresis features corresponding to all event-level hydraulic response segments are aggregated to form a hysteresis feature set Lag for describing the hysteresis behavior of the valve hydraulic response.

[0009] Preferably, S3 includes S31; S31. Based on the historical operating data of the target valve, the continuity of the target valve's operating status in the historical time dimension is determined to identify the historical time period in which the target valve is continuously in normal operating condition; wherein, the historical time period in which the target valve is continuously in normal operating condition is determined in the following manner: Obtain the historical operation and maintenance records corresponding to the target valve, and divide the historical time into multiple continuous time segments based on the time axis; within each time segment, determine whether the target valve meets the preset normal operation judgment conditions; The normal operation determination conditions include: no valve maintenance operation or valve fault handling operation is recorded within the time segment, and no manual forced switching or emergency dispatch operation occurs; when a preset number of adjacent time segments meet the normal operation determination conditions, multiple adjacent time segments are merged to form a historical time segment in which the target valve is continuously in a normal operation state. After determining the historical time period, the historical time period is used as the time filtering condition. Hysteresis features formed within the historical time period are extracted from the obtained hysteresis feature set Lag and used as reference feature samples to represent the hydraulic response hysteresis behavior of the target valve under normal operating conditions. Statistical modeling is performed on the reference feature sample. The statistical modeling process includes: calculating the value distribution, central tendency and fluctuation range of each type of hysteresis feature in the reference feature sample to obtain statistical description results that describe the overall distribution of hysteresis features of the target valve under normal operating conditions. Based on the statistical description results, a normal hysteresis baseline set Bas is constructed to represent the overall distribution characteristics of the pressure and flow response hysteresis behavior of the target valve under normal operating conditions.

[0010] Preferably, S3 further includes S32; S32. After completing the construction of the normal hysteresis baseline set Bas, the normal hysteresis baseline set Bas is used as the source of training samples for model training. A deep learning model is constructed based on autoencoder deep technology. The deep learning model includes an encoding network and a decoding network, wherein: the encoding network is used to map the hysteresis features of the input to a low-dimensional latent representation space, and the decoding network is used to reconstruct the hysteresis features of the input based on the low-dimensional latent representation. Using the hysteresis features in the normal hysteresis baseline set Bas as training samples, the deep learning model is iteratively trained so that the deep learning model learns the distribution pattern and correlation of the hysteresis features of the target valve under normal operating conditions by minimizing the difference between the input hysteresis features and the reconstructed hysteresis features. After completing the iterative training and satisfying the preset convergence condition, the model parameters of the deep learning model are solidified to obtain a hysteresis recognition model Mdl, which is used to evaluate the deviation of the input hysteresis features from the distribution represented by the normal hysteresis baseline set Bas.

[0011] Preferably, S4 includes S41; S41. Based on the newly formed event-level hydraulic response fragment, generate and update the hysteresis feature set Lag in real time; From the real-time updated hysteresis feature set Lag, select the hysteresis feature corresponding to the current valve operating state, use the hysteresis feature as input, and substitute it into the obtained hysteresis recognition model Mdl. The hysteresis recognition model Mdl, based on the autoencoder structure, encodes and reconstructs the hysteresis feature one by one, and calculates the deviation between the input value and the corresponding reconstructed value for each hysteresis feature, thus obtaining a feature deviation set corresponding to each hysteresis feature. Dimensionless processing is performed on each feature deviation in the feature deviation set. After the dimensionless processing is completed, the dimensionless feature deviations are summed with equal weights. The hysteresis risk score Rsk, which represents the degree of deviation of the current valve hydraulic response hysteresis behavior from the normal operating state, is calculated by summing the dimensionless feature deviations and dividing by the number of hysteresis features.

[0012] Preferably, S4 further includes S42; S42. After obtaining the hysteresis risk score Rsk, record and update the hysteresis risk score Rsk in chronological order. Based on the current value of the hysteresis risk score Rsk and its change over time, risk assessment processing is performed on the hysteresis risk score Rsk to determine whether the current hydraulic response hysteresis behavior of the valve meets the preset early warning triggering conditions. When the hysteresis risk score Rsk meets the preset warning triggering condition, a warning triggering record containing the target valve identifier, warning triggering time and corresponding hysteresis risk score Rsk is generated, and the warning triggering records are collected to form a warning triggering set Trg for indicating abnormal valve response hysteresis. No action is taken when the hysteresis risk score Rsk does not meet the preset warning triggering conditions; The warning triggering conditions include any of the following determination methods: a determination method based on the hysteresis risk score Rsk exceeding a preset risk threshold, or a determination method based on the hysteresis risk score Rsk continuously increasing over multiple consecutive assessment periods.

[0013] Preferably, S5 includes S51; S51. Based on the warning trigger set Trg obtained in step S42, organize and output the warning trigger records included: collect the warning trigger records according to the target valve identifier and trigger time to form a warning result set Alr used to characterize the hidden operational abnormality risk of the target valve; The warning result set Alr is used to indicate the potential risks of valve jamming, inadequate execution, or abnormal feedback in the future operation of the target valve, and is output and stored as a warning result of valve operation status.

[0014] A fault early warning system based on deep learning includes a valve operation data acquisition module, a hysteresis feature extraction module, a deep model construction module, an early warning risk assessment module, and an early warning result generation module. The valve operation data acquisition module collects valve operation records related to the target valve to form a valve operation event set Eve, and simultaneously collects the pressure time series of valve-related measuring points to form a pressure time series Pre and the flow time series to form a flow time series Flo. Based on the valve operation event set Eve, an event window is constructed, and the pressure time series Pre and the flow time series Flo are preprocessed to obtain an event fragment set Seg. The hysteresis feature extraction module extracts hysteresis features between pressure changes and flow rate changes based on the event fragment set Seg, forming a hysteresis feature set Lag. The deep model building module constructs a normal hysteresis baseline set Bas from the hysteresis feature set Lag based on the historical data of the normal operation of the target valve, and trains a deep learning model on the normal hysteresis baseline set Bas to obtain the hysteresis recognition model Mdl. The early warning risk assessment module inputs the real-time generated hysteresis feature set Lag into the hysteresis recognition model Mdl for processing, outputs the risk assessment result of valve response hysteresis, forms a hysteresis risk score Rsk, and generates an early warning trigger set Trg when the hysteresis risk score Rsk meets the preset conditions. The warning result generation module outputs warning results for hidden valve anomalies based on the warning trigger set Trg, forms a warning result set Alr, and sends it down for storage.

[0015] This invention provides a fault early warning method and system based on deep learning, which has the following beneficial effects: (1) By systematically modeling the hydraulic response relationship between valve operation events and pressure time series Pre and flow time series Flo, early identification and quantitative warning of hidden valve anomalies are achieved. This method does not rely on valve position feedback hardware, but instead uses the event fragment set Seg as a basis to extract the hysteresis feature set Lag that reflects the dynamic response relationship between pressure and flow, and combines it with the historical data of the target valve's normal operation to construct the normal hysteresis baseline set Bas, so that the hysteresis identification model Mdl can accurately characterize the hysteresis behavior distribution under normal operating conditions. In actual operation, the real-time generated hysteresis feature set Lag is input into the hysteresis identification model Mdl, and a single, comparable hysteresis risk score Rsk is obtained by averaging the deviation of multiple hysteresis features after dimensionless transformation, and an early warning trigger set Trg is formed when it is continuously abnormal, and finally a structured early warning result set Alr is output. Compared with the existing methods that rely on valve position signals or manual inspection, this method can issue an early warning when the valve has not completely failed and only shows slow action or incomplete execution.

[0016] (2) By precisely defining the start time, process, and stable interval of pressure and flow changes in the event-level hydraulic response segment, the time difference feature, pressure stability duration feature, flow stability duration feature, and relative displacement feature can respectively characterize the dynamic response differences caused by valve operation from different dimensions, avoiding misjudgment caused by relying solely on a single time indicator; at the same time, by using the historical time segment of the target valve continuously in normal operation, the hysteresis feature set Lag is time-constrained and screened to construct the normal hysteresis baseline set Bas, so that the normal state learned by the model comes directly from the real operating conditions, rather than an ideal setting. On this basis, the hysteresis recognition model Mdl, trained based on the autoencoder structure, can stably reconstruct the normal hysteresis features. When the valve's pressure and flow response relationship gradually deviates from the normal distribution due to wear, scaling, or aging of the actuator, even if a single operation has not yet triggered an obvious alarm, the reconstruction deviation of the model output will continue to increase, thus providing a sensitive signal for subsequent risk assessment.

[0017] (3) At the level of risk assessment and early warning output, the system achieves quantification, continuity, and practical output of hidden valve anomalies. Its core technical advantages are reflected in the following: by inputting the real-time updated hysteresis feature set Lag into the trained hysteresis recognition model Mdl, and by performing dimensionless and weighted averaging on the feature deviations corresponding to each hysteresis feature, a single, stable, and physically meaningful hysteresis risk score Rsk is finally formed, thus avoiding the problem of multiple indicators being difficult to comprehensively judge in traditional methods. On this basis, by recording the hysteresis risk score Rsk in a time series and determining its trend, the early warning trigger no longer depends on the instantaneous anomaly of a single operation, but can identify the continuous accumulation process of risk. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the steps of a fault early warning method based on deep learning according to the present invention. Figure 2 This is a schematic diagram of a fault early warning system based on deep learning according to the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention. Example 1

[0020] This invention provides a fault early warning method based on deep learning. Please refer to [link / reference]. Figure 1 This includes the following steps: S1. Collect valve operation records related to the target valve to form a valve operation event set Eve, and simultaneously collect the pressure time series of valve-related measuring points to form a pressure time series Pre and the flow time series to form a flow time series Flo. Based on the valve operation event set Eve, construct an event window, and preprocess the pressure time series Pre and the flow time series Flo to obtain an event fragment set Seg. S2. Based on the event fragment set Seg, extract the hysteresis features between pressure changes and flow rate changes to form a hysteresis feature set Lag. S3. Based on the historical data of the normal operation of the target valve, construct a normal hysteresis baseline set Bas from the hysteresis feature set Lag, and train a deep learning model on the normal hysteresis baseline set Bas to obtain the hysteresis recognition model Mdl. S4. Input the real-time generated hysteresis feature set Lag into the hysteresis recognition model Mdl for processing, output the risk assessment result of valve response hysteresis, form a hysteresis risk score Rsk, and generate an early warning trigger set Trg when the hysteresis risk score Rsk meets the preset conditions. S5. Based on the warning results of the valve hidden abnormality output by the warning trigger set Trg, a warning result set Alr is formed and sent down for storage.

[0021] In this embodiment, a deep learning-based fault early warning method constructed based on steps S1 to S5 above achieves early identification and quantitative early warning of hidden valve anomalies by systematically modeling the hydraulic response relationship between valve operation events and pressure time series Pre and flow time series Flo. This method does not rely on valve position feedback hardware, but instead uses the event fragment set Seg as a basis to extract a hysteresis feature set Lag reflecting the dynamic response relationship between pressure and flow. It then combines this with historical normal operation data of the target valve to construct a normal hysteresis baseline set Bas, enabling the hysteresis identification model Mdl to accurately characterize the hysteresis behavior distribution under normal operating conditions. In actual operation, the real-time generated hysteresis feature set Lag is input into the hysteresis identification model Mdl. A single, comparable hysteresis risk score Rsk is obtained by averaging the deviations of multiple hysteresis features after dimensionless scaling. When this score remains abnormal, an early warning trigger set Trag is formed, ultimately outputting a structured early warning result set Alr. Compared to existing methods that rely on valve position signals or manual inspections, this method can issue early warnings even before the valve has completely failed, when it only exhibits slow action or incomplete execution. For example, if a valve experiences a slight jamming during pipeline operation, causing a significant lag in pressure changes while the valve position feedback still shows normal, this method can still identify risks in a timely manner based on the increase in the hysteresis risk score Rsk. This allows water maintenance personnel to gain time to respond in advance, reducing the probability of sudden water outages, scheduling failures, or pipeline operation anomalies, and significantly improving valve operation safety and the level of precision in pipeline management. Example 2

[0022] Specifically: S1 includes S11; S11. Read the operation records of the target valve from the water company's existing SCADA data interface, historical database interface, or operation data bus; The operation record includes valve identifier, operation command type, command issuance time, and execution status code; the operation record is processed by event parsing: using the same valve identifier as the primary key, multiple operation records with adjacent time intervals less than the preset merging time are merged into one operation event, and the start time of one operation event is recorded as the event start point, and the end time of one operation event is recorded as the event end point, thus constructing a valve operation event set Eve; Based on the valve operation event set Eve, the range of measurement points that have a hydraulic coupling relationship with the target valve is determined, and pressure and flow sampling values ​​within the range of measurement points are retrieved from the existing measurement point data interface; the pressure sampling values ​​are arranged in chronological order to form a pressure time series Pre, and the flow sampling values ​​are arranged in chronological order to form a flow time series Flo; when the sampling periods of the pressure time series Pre and the flow time series Flo are inconsistent, a time rasterization alignment method is used to map them to a unified time raster; The time granularity of the unified time grid is determined by the least common multiple of the sampling periods of the two, so that the pressure time series Pre and the flow time series Flo have one-to-one corresponding sampling points on the same time axis; It should be noted that: Valve operation event set Eve: A set of events obtained by parsing valve opening, closing or regulation commands under the same valve identifier. Each event contains at least an event start point and an event end point, which is used to divide continuous operation data into traceable segments around valve actions. Pressure Time Series Pre: A sequence of pressure sampling values ​​from hydraulically associated measurement points of the target valve arranged in a uniform time grid, used to reflect the dynamic process of valve action in response to pipeline pressure. Flow time series Flo: A sequence of flow sampling values ​​from hydraulically associated measurement points of the target valve arranged in a uniform time grid, used to reflect the dynamic process of valve action in response to flow distribution. The technical advantage of the time rasterization alignment method is that it avoids misalignment of pressure and flow rate changes on the time axis due to inconsistent sampling periods, thereby providing a comparable data basis for subsequent hysteresis analysis.

[0023] S1 further includes S12; S12. Based on each valve operation event in the valve operation event set Eve, extend a preset pre-time before the event start point and extend a preset post-time after the event end point to form an event window corresponding to the valve operation event, and combine all event windows into an event window set Win; In the event window set Win, corresponding subsequences of the pressure time series Pre and the flow time series Flo, which have the same time range as each event window, are extracted to obtain candidate segments; preprocessing operations are performed on the candidate segments, including missing data repair and abnormal spike suppression, wherein: The missing data repair process employs an interpolation method based on the consistency of adjacent time points, so that the value of the missing point is determined by the effective sampling points on both sides of the missing point; The abnormal spike suppression process employs a suppression method based on sliding window robust statistics, which brings sampling points that exceed the robust range back to the robust range. In the abnormal peak suppression process, the robust center value of the candidate segment within the sliding window is defined as Med, and the robust dispersion of the candidate segment within the sliding window is defined as Mad. Abnormal peaks are suppressed according to the following rules: When the deviation of any sampling point from Med is greater than the product of a preset threshold and Mad, the sampling point is replaced with Med. After completing the preprocessing operation, the candidate fragments are combined into event-level hydraulic response fragments according to time synchronization, forming the event fragment set Seg; It should be noted that: The event window set Win is a collection of time ranges built around a single valve operation event. Each event window covers at least the baseline operating state before the valve operation and the hydraulic response process after the valve operation. By extending the preset pre-time before the start of the event and the preset post-time after the end of the event, the complete impact of the valve operation on pressure and flow changes is included in the same event window, thus providing a complete time context for subsequent response hysteresis analysis. The abnormal spike suppression process is used to weaken non-physical spikes in candidate segments caused by communication jitter, transient interference, or acquisition anomalies. Here, Med represents the robust center value of the candidate segment within the sliding window, reflecting the typical level of pressure or flow within that time window; Mad represents the robust dispersion of the candidate segment within the sliding window, reflecting the robust fluctuation range of pressure or flow within that time window. By replacing sampling points that deviate from Med and exceed the Mad adjustment threshold with Med, the impact of abnormal spikes on response hysteresis analysis can be suppressed while maintaining the overall trend.

[0024] In this embodiment, based on the collaborative processing of S11 and S12, a highly consistent and traceable reconstruction of valve operation behavior and hydraulic response process is achieved at the data level. Its direct technical effect is that it transforms the operation records, pressure sampling values, and flow sampling values ​​originally scattered in the water company's SCADA system into a set of event segments (Seg) organized around valve actions, thereby significantly improving the data reliability and interpretation accuracy of subsequent hysteresis analysis. By constructing a valve operation event set (Eve) and further forming an event window set (Win), this method can accurately distinguish between the stable operating state before valve action and the hydraulic response process after valve action, avoiding the response truncation or background interference problems commonly found in traditional fixed-time interception methods. Simultaneously, by using time-raster alignment processing of the pressure time series (Pre) and the flow time series (Flo), even when pressure sampling is at the second level and flow sampling is at the minute level in actual engineering, it can still ensure a one-to-one correspondence between the two on the same time axis, providing a strict temporal basis for subsequent hysteresis feature extraction. Furthermore, by introducing anomaly spike suppression processing based on Med and Mad into the candidate segments, non-physical fluctuations introduced by communication jitter or instantaneous sensor anomalies can be effectively eliminated. For example, instantaneous flow spikes caused by signal jitter under low-flow conditions at night will not be misjudged as valve response changes, thus ensuring that the resulting event segment set Seg truly reflects the actual impact of valve operation on the hydraulic state of the pipeline network. This data organization method centered on valve events allows subsequent analysis to be built on a high-quality, physically consistent data foundation, which is significantly better than the method of directly modeling based solely on raw time-series data. It is especially suitable for real-world operating scenarios in water pipeline networks with complex operating conditions and inconsistent sampling frequencies. Example 3

[0025] Specifically: S2 includes S21; S21. Perform time-series analysis on each event-level hydraulic response segment in the event segment set Seg to analyze the dynamic response relationship between pressure changes and flow rate changes before and after valve operation; In each event-level hydraulic response segment, the onset time, process, and stable interval of the pressure and flow rate changes are determined, and hysteresis features describing the pressure change relative to the flow rate change are extracted. These hysteresis features include the time difference between the onset time of the pressure change and the onset time of the flow rate change, the stable duration required for the pressure change to reach the stable interval, the stable duration required for the flow rate change to reach the stable interval, and the relative displacement of the pressure and flow rate changes on the time axis. The hysteresis features extracted from each event-level hydraulic response segment are combined and vectorized, and the hysteresis features corresponding to all event-level hydraulic response segments are aggregated to form a hysteresis feature set Lag for describing the hysteresis behavior of the valve hydraulic response. It should be noted that: The change initiation time refers to the point in time when the pressure sampling value sequence or flow sampling value sequence first shows a sustained change relative to the baseline state before valve operation; wherein, the baseline state before valve operation is obtained by statistically analyzing the sampling values ​​within the event window and within a preset time range before the valve operation occurs, and the statistical results are used to represent the stable operating level before the valve operation occurs; When the offset of a sampling point in the pressure sampling value sequence or the flow sampling value sequence relative to the baseline state continuously exceeds the preset change judgment threshold, and the offset maintains the same trend of change in multiple consecutive sampling points, the timestamp corresponding to the sampling point is determined as the change start time. In the above way, the pressure change start time and the flow change start time are determined respectively. This can effectively distinguish between the real hydraulic response change caused by valve operation and short-term noise fluctuations, thereby improving the stability and reliability of the change start time determination. The change process refers to the time interval from the start of the change to before the pressure or flow sampling value sequence enters the stable range. During the change process, the pressure or flow sampling value sequence exhibits a trend of continuous increase, continuous decrease, or step change, which is used to reflect the dynamic adjustment stage of the impact of valve operation on the hydraulic system. By clarifying the time range of the change process, the dynamic response caused by valve operation can be distinguished from the stable state before and after the operation, providing a clear time boundary for the subsequent hysteresis characteristic quantification. The stable interval refers to the time period after the change process ends when the pressure or flow sampling value sequence enters a period where the fluctuation amplitude is less than a preset stability threshold. When the change amplitude of the pressure or flow sampling value sequence is always within the stability threshold range in multiple consecutive sampling points, and the overall change trend no longer shows a monotonically increasing or monotonically decreasing characteristic, this time period is determined as the stable interval. By introducing the determination of the stable interval, it is possible to quantify whether the system has completed hydraulic adjustment after valve operation, providing a basis for the subsequent extraction of stable duration characteristics. The time difference characteristic between the start time of pressure change and the start time of flow rate change is calculated by subtracting the timestamps corresponding to the start time of pressure change and the start time of flow rate change. The time difference characteristic is used to represent the order and degree of delay of the response of pressure change relative to flow rate change. When the time difference characteristic increases or decreases abnormally, it indicates that there is a risk of slow, stuck or incomplete execution in the hydraulic response process caused by valve operation. The pressure stabilization duration characteristic refers to the length of time from the start of the pressure change to the first entry of the pressure sampling value sequence into the stable pressure change interval. The pressure stabilization duration characteristic is calculated by the difference between the timestamps of the start of the pressure change and the start of the stable pressure change interval. The pressure stabilization duration characteristic is used to reflect the recovery speed of the pressure adjustment process after the valve operation is completed. Its abnormal increase is usually related to problems such as valve jamming and valve incomplete opening. The flow stability duration feature refers to the length of time from the start of the flow change to the first entry of the flow sampling value sequence into the flow stability interval. The flow stability duration feature is obtained by calculating the difference between the timestamps of the start of the flow change and the start of the flow stability interval. The flow stability duration feature is used to reflect the efficiency of valve operation in regulating the actual flow rate. Abnormal changes in the feature can be used to identify situations where valve execution is slow or the regulation capacity is reduced. The relative displacement feature is used to describe the overall alignment of the pressure change process and the flow rate change process on the time axis. It is obtained by performing time axis translation analysis on the pressure sample value sequence change process and the flow rate sample value sequence change process. When it is necessary to achieve the maximum similarity in the shape of the two change processes, one of the change processes is translated on the time axis, and the amount of translation time required is used as the relative displacement feature. The relative displacement feature can comprehensively reflect the overall response hysteresis relationship between pressure change and flow rate change. Compared with the single time difference feature, it has stronger robustness and discrimination ability for hidden valve anomalies.

[0026] S3 includes S31; S31. Based on the historical operating data of the target valve, the continuity of the target valve's operating status in the historical time dimension is determined to identify the historical time period in which the target valve is continuously in normal operating condition; wherein, the historical time period in which the target valve is continuously in normal operating condition is determined in the following manner: Obtain the historical operation and maintenance records corresponding to the target valve, and divide the historical time into multiple continuous time segments based on the time axis; within each time segment, determine whether the target valve meets the preset normal operation judgment conditions; The normal operation determination conditions include: no valve maintenance operation or valve fault handling operation is recorded within the time segment, and no manual forced switching or emergency dispatch operation occurs; when a preset number of adjacent time segments meet the normal operation determination conditions, multiple adjacent time segments are merged to form a historical time segment in which the target valve is continuously in a normal operation state. After determining the historical time period, the historical time period is used as the time filtering condition. Hysteresis features formed within the historical time period are extracted from the obtained hysteresis feature set Lag and used as reference feature samples to represent the hydraulic response hysteresis behavior of the target valve under normal operating conditions. Statistical modeling is performed on the reference feature sample. The statistical modeling process includes: calculating the value distribution, central tendency and fluctuation range of each type of hysteresis feature in the reference feature sample to obtain statistical description results that describe the overall distribution of hysteresis features of the target valve under normal operating conditions. Based on the statistical description results, a normal hysteresis baseline set Bas is constructed to represent the overall distribution characteristics of the pressure and flow response hysteresis behavior of the target valve under normal operating conditions; It should be noted that: The historical time intervals that are continuously in normal operating condition are represented as a continuous time range on a time axis. Each time interval satisfies the condition that the target valve has not undergone maintenance intervention, has not performed maintenance operations, and has not been recorded as an abnormal operating state within that time range. By discretizing the historical operating time into multiple time segments and determining the operating state of each segment, and then merging adjacent time segments that meet the conditions, the historical time intervals can truly reflect the continuous period of stable operation of the target valve in actual operation. It is possible to obtain a reproducible and interpretable normal operating time reference range solely through existing operating and maintenance records without relying on valve position feedback signals. Each hysteresis feature in the hysteresis feature set Lag is accompanied by its formation time information. By matching the formation time information with the historical time segment, hysteresis features formed within the historical time segment can be selected from the hysteresis feature set Lag, thereby forming a reference feature sample set. This ensures that the selection of hysteresis features is based solely on time segment constraints, without modifying or removing the hysteresis features themselves, thus avoiding the masking of potential hysteresis anomalies. The statistical modeling process is a distribution modeling process performed on each type of hysteresis feature in the reference feature sample. Its output includes the typical value range and variation scale of each type of hysteresis feature under normal operating conditions. The statistical description results are used to characterize the overall distribution pattern of hysteresis features under normal operating conditions, rather than to classify and judge individual hysteresis features. By performing statistical modeling on the reference feature sample, normal hysteresis behavior can be clearly described in a distributional form, providing a foundation for subsequent model training based on the degree of distribution deviation. The normal hysteresis baseline set Bas is a hysteresis feature reference set composed of the statistical description results. It is used to describe the typical distribution of the pressure-flow response hysteresis behavior of the target valve under normal operating conditions. The normal hysteresis baseline set Bas serves as the reference input for the deep learning model, enabling the model to learn the distribution characteristics of normal hysteresis behavior, thereby providing a basis for identifying hysteresis anomalies.

[0027] S3 further includes S32; S32. After completing the construction of the normal hysteresis baseline set Bas, the normal hysteresis baseline set Bas is used as the source of training samples for model training. A deep learning model is constructed based on autoencoder deep technology. The deep learning model includes an encoding network and a decoding network, wherein: the encoding network is used to map the hysteresis features of the input to a low-dimensional latent representation space, and the decoding network is used to reconstruct the hysteresis features of the input based on the low-dimensional latent representation. Using the hysteresis features in the normal hysteresis baseline set Bas as training samples, the deep learning model is iteratively trained so that the deep learning model learns the distribution pattern and correlation of the hysteresis features of the target valve under normal operating conditions by minimizing the difference between the input hysteresis features and the reconstructed hysteresis features. After completing the iterative training and satisfying the preset convergence condition, the model parameters of the deep learning model are solidified to obtain a hysteresis recognition model Mdl for evaluating the deviation of the input hysteresis features from the distribution represented by the normal hysteresis baseline set Bas. It should be noted that: The autoencoder is an existing deep learning representation learning model that uses a combination of an encoding network and a decoding network to achieve compressed representation and reconstruction learning of input features. The encoding network maps high-dimensional input features to a low-dimensional latent representation space, and the decoding network reconstructs the original input features based on the latent representation. By training using only samples under normal operating conditions, the autoencoder can learn the inherent distribution structure of normal features, resulting in smaller reconstruction errors for normal samples during the reconstruction process. Technical effect description: Self encoders have been widely used in the field of industrial equipment condition monitoring and anomaly detection. They are particularly suitable for scenarios where anomaly samples are scarce or anomaly forms are diverse, and can meet the engineering requirements for valve response hysteresis anomaly identification. Each sample in the normal hysteresis baseline set Bas is a hysteresis feature formed under normal operating conditions, and its feature dimension and structure are consistent with the hysteresis features defined in step S2. During training, the hysteresis features in the normal hysteresis baseline set Bas are directly used as input samples of the autoencoder, so that the model learns how the hysteresis features should be effectively represented and reconstructed under normal operating conditions. By maintaining the structural consistency between the training samples and the real-time input features, it is ensured that the trained model can be directly used for online evaluation of hysteresis features. During model training, the difference between the input hysteresis features and the reconstructed hysteresis features output by the decoding network is used as the training error. The model parameters are updated through multiple iterations to gradually reduce the training error. When the training error no longer decreases significantly in multiple iterations, the model training process is considered to have reached convergence. Through the training mechanism, the autoencoder can form a stable reconstruction capability for the normal hysteresis feature distribution, thereby providing a reliable basis for subsequent hysteresis anomaly assessment based on the degree of reconstruction deviation. After training is completed, the parameters of the encoding and decoding networks of the autoencoder model are solidified to form the hysteresis recognition model Mdl. When subsequent hysteresis features are input into the hysteresis recognition model Mdl, the model reconstructs the input hysteresis features and evaluates the degree of deviation of the hysteresis feature from the normal hysteresis baseline set Bas based on the difference between the input hysteresis features and the reconstructed hysteresis features. Provided that the normal hysteresis distribution is learned through reconstruction and the deviation evaluation result is output, the autoencoder can be implemented in a variety of existing structural forms, including but not limited to fully connected autoencoders, sparse autoencoders, or deep autoencoders. Different structures have different implementation methods, but they are all trained based on the normal hysteresis baseline set Bas, and their hysteresis recognition principle and technical effect are consistent.

[0028] In this embodiment, based on the collaborative design of S21, S31, and S32, interpretable modeling and robust identification of valve hydraulic response hysteresis behavior are achieved at three levels: feature expression layer, normal baseline layer, and model learning layer. Its key technical effect lies in transforming the previously difficult-to-quantify and experience-dependent problem of valve latent anomalies into a problem of identifying distribution deviations based on the hysteresis feature set Lag and the normal hysteresis baseline set Bas. By precisely defining the start time, process, and stable interval of pressure and flow changes in event-level hydraulic response segments, time difference features, pressure stability duration features, flow stability duration features, and relative displacement features can characterize the dynamic response differences caused by valve operation from different dimensions, avoiding misjudgments caused by relying solely on a single time indicator. Simultaneously, by utilizing historical time periods when the target valve is continuously in normal operating condition, the hysteresis feature set Lag is time-constrained and filtered to construct the normal hysteresis baseline set Bas, ensuring that the normal state learned by the model directly originates from real operating conditions, rather than idealized settings. Based on this, the hysteresis identification model Mdl, trained using an autoencoder structure, can stably reconstruct normal hysteresis characteristics. When the pressure and flow response relationship of a valve gradually deviates from the normal distribution due to wear, scaling, or aging of the actuator, even if a single operation has not triggered a significant alarm, the reconstruction deviation of the model output will continue to increase, thus providing a sensitive signal for subsequent risk assessment. For example, in actual pipeline operation, after multiple small-scale opening adjustments, a certain regulating valve exhibits a gradually lengthening duration of stable flow, while the pressure change remains within a reasonable range. Traditional monitoring methods struggle to detect anomalies in a timely manner, but the hysteresis identification model Mdl can identify the overall deviation of such composite hysteresis characteristics, providing a reliable basis for maintenance personnel to arrange maintenance in advance, significantly improving the foresight and engineering practicality of valve operating status identification. Example 4

[0029] Specifically: S4 includes S41; S41. Based on the newly formed event-level hydraulic response fragment, generate and update the hysteresis feature set Lag in real time; From the real-time updated hysteresis feature set Lag, select the hysteresis feature corresponding to the current valve operating state, use the hysteresis feature as input, and substitute it into the obtained hysteresis recognition model Mdl. The hysteresis recognition model Mdl, based on the autoencoder structure, encodes and reconstructs the hysteresis feature one by one, and calculates the deviation between the input value and the corresponding reconstructed value for each hysteresis feature, thus obtaining a feature deviation set corresponding to each hysteresis feature. Dimensionless processing is performed on each feature deviation in the feature deviation set. After the dimensionless processing is completed, the dimensionless feature deviations are summed with equal weights. The hysteresis risk score Rsk, which represents the degree of deviation of the current valve hydraulic response hysteresis behavior from the normal operating state, is calculated by summing the dimensionless feature deviations and dividing by the number of hysteresis features. It should be noted that: The feature deviation refers to: for each hysteresis feature, the deviation measure calculated based on the difference between the input hysteresis feature and its reconstructed hysteresis feature using the hysteresis recognition model Mdl; each hysteresis feature corresponds to a feature deviation measure, which reflects the degree of deviation of the hysteresis feature relative to the normal hysteresis baseline set Bas.

[0030] S4 also includes S42; S42. After obtaining the hysteresis risk score Rsk, record and update the hysteresis risk score Rsk in chronological order. Based on the current value of the hysteresis risk score Rsk and its change over time, risk assessment processing is performed on the hysteresis risk score Rsk to determine whether the current hydraulic response hysteresis behavior of the valve meets the preset early warning triggering conditions. When the hysteresis risk score Rsk meets the preset warning triggering condition, a warning triggering record containing the target valve identifier, warning triggering time and corresponding hysteresis risk score Rsk is generated, and the warning triggering records are collected to form a warning triggering set Trg for indicating abnormal valve response hysteresis. No action is taken when the hysteresis risk score Rsk does not meet the preset warning triggering conditions; The warning triggering conditions include any of the following determination methods: a determination method based on the hysteresis risk score Rsk exceeding a preset risk threshold, or a determination method based on the hysteresis risk score Rsk continuously increasing over multiple consecutive assessment periods. By combining numerical magnitude and trend analysis, false triggers caused by short-term fluctuations can be avoided, while ensuring effective identification of persistent lag risks.

[0031] S5 includes S51; S51. Based on the warning trigger set Trg obtained in step S42, organize and output the warning trigger records included: collect the warning trigger records according to the target valve identifier and trigger time to form a warning result set Alr used to characterize the hidden operational abnormality risk of the target valve; The warning result set Alr is used to indicate the potential risks of valve jamming, inadequate execution, or abnormal feedback during the future operation of the target valve. It is output and stored as a valve operation status warning result, thereby achieving early warning of hidden valve abnormalities without relying on valve position feedback hardware.

[0032] In this embodiment, based on the continuous processing flow of S41, S42, and S51 described above, the quantification, continuity, and practical output of hidden valve anomalies are achieved at the risk assessment and early warning output level. Its core technical advantage lies in: by inputting the real-time updated hysteresis feature set Lag into the trained hysteresis recognition model Mdl, and performing dimensionless and weighted averaging on the feature deviations corresponding to each hysteresis feature, a single, stable, and physically meaningful hysteresis risk score Rsk is ultimately formed, thus avoiding the problem of difficulty in comprehensively judging multiple indicators in traditional methods. Furthermore, by recording the hysteresis risk score Rsk in a time-series format and determining its trend, the early warning trigger no longer depends on the instantaneous anomaly of a single operation, but can identify the continuous accumulation process of risk. For example, in actual operation, when the flow recovery speed of a valve gradually slows down after multiple adjustment operations due to actuator wear, a single operation may still be within an acceptable range, but the corresponding hysteresis risk score Rsk will continuously increase over multiple assessment periods, thereby triggering the early warning trigger set Trg and ultimately forming the early warning result set Alr. This early warning mechanism, which is based on continuous risk evolution, enables maintenance personnel to know the potential risks of valve jamming, inadequate execution, or abnormal feedback before the valves have experienced obvious failures or accidents. Furthermore, without the need for additional valve position feedback hardware, the early warning results can be directly stored and applied to the maintenance decision-making system, significantly improving the foresight, reliability, and engineering practical value of valve operation status monitoring in water pipe networks. Example 5

[0033] A fault early warning system based on deep learning, please refer to... Figure 2 Specifically, it includes a valve operation data acquisition module, a hysteresis feature extraction module, a deep model construction module, an early warning risk assessment module, and an early warning result generation module; The valve operation data acquisition module collects valve operation records related to the target valve to form a valve operation event set Eve, and simultaneously collects the pressure time series of valve-related measuring points to form a pressure time series Pre and the flow time series to form a flow time series Flo. Based on the valve operation event set Eve, an event window is constructed, and the pressure time series Pre and the flow time series Flo are preprocessed to obtain an event fragment set Seg. The hysteresis feature extraction module extracts hysteresis features between pressure changes and flow rate changes based on the event fragment set Seg, forming a hysteresis feature set Lag. The deep model building module constructs a normal hysteresis baseline set Bas from the hysteresis feature set Lag based on the historical data of the normal operation of the target valve, and trains a deep learning model on the normal hysteresis baseline set Bas to obtain the hysteresis recognition model Mdl. The early warning risk assessment module inputs the real-time generated hysteresis feature set Lag into the hysteresis recognition model Mdl for processing, outputs the risk assessment result of valve response hysteresis, forms a hysteresis risk score Rsk, and generates an early warning trigger set Trg when the hysteresis risk score Rsk meets the preset conditions. The warning result generation module outputs warning results for hidden valve anomalies based on the warning trigger set Trg, forms a warning result set Alr, and sends it down for storage.

[0034] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A fault early warning method based on deep learning, characterized in that: Includes the following steps: S1. Collect valve operation records related to the target valve to form a valve operation event set Eve, and simultaneously collect the pressure time series of valve-related measuring points to form a pressure time series Pre and the flow time series to form a flow time series Flo. Based on the valve operation event set Eve, construct an event window, and preprocess the pressure time series Pre and the flow time series Flo to obtain an event fragment set Seg. S2. Based on the event fragment set Seg, extract the hysteresis features between pressure changes and flow rate changes to form a hysteresis feature set Lag. S3. Based on the historical data of the normal operation of the target valve, construct a normal hysteresis baseline set Bas from the hysteresis feature set Lag, and train a deep learning model on the normal hysteresis baseline set Bas to obtain the hysteresis recognition model Mdl. S4. Input the real-time generated hysteresis feature set Lag into the hysteresis recognition model Mdl for processing, output the risk assessment result of valve response hysteresis, form a hysteresis risk score Rsk, and generate an early warning trigger set Trg when the hysteresis risk score Rsk meets the preset conditions. S5. Based on the warning results of the valve hidden abnormality output by the warning trigger set Trg, a warning result set Alr is formed and sent down for storage.

2. The fault early warning method based on deep learning according to claim 1, characterized in that: S1 includes S11; S11. Read the operation records of the target valve from the water company's existing SCADA data interface, historical database interface, or operation data bus; The operation record includes valve identifier, operation command type, command issuance time, and execution status code; the operation record is processed by event parsing: using the same valve identifier as the primary key, multiple operation records with adjacent time intervals less than the preset merging time are merged into one operation event, and the start time of one operation event is recorded as the event start point, and the end time of one operation event is recorded as the event end point, thus constructing a valve operation event set Eve; Based on the valve operation event set Eve, the range of measurement points that have a hydraulic coupling relationship with the target valve is determined, and pressure and flow sampling values ​​within the range of measurement points are retrieved from the existing measurement point data interface; the pressure sampling values ​​are arranged in chronological order to form a pressure time series Pre, and the flow sampling values ​​are arranged in chronological order to form a flow time series Flo; when the sampling periods of the pressure time series Pre and the flow time series Flo are inconsistent, a time rasterization alignment method is used to map them to a unified time raster; The time granularity of the unified time grid is determined by the least common multiple of the sampling periods of the two, so that the pressure time series Pre and the flow time series Flo have one-to-one corresponding sampling points on the same time axis.

3. The fault early warning method based on deep learning according to claim 2, characterized in that: S1 further includes S12; S12. Based on each valve operation event in the valve operation event set Eve, extend a preset pre-time before the event start point and extend a preset post-time after the event end point to form an event window corresponding to the valve operation event, and combine all event windows into an event window set Win; In the event window set Win, corresponding subsequences of the pressure time series Pre and the flow time series Flo, which have the same time range as each event window, are extracted to obtain candidate segments; preprocessing operations are performed on the candidate segments, including missing data repair and abnormal spike suppression, wherein: The missing data repair process employs an interpolation method based on the consistency of adjacent time points, so that the value of the missing point is determined by the effective sampling points on both sides of the missing point; The abnormal spike suppression process employs a suppression method based on sliding window robust statistics, which brings sampling points that exceed the robust range back to the robust range. In the abnormal peak suppression process, the robust center value of the candidate segment within the sliding window is defined as Med, and the robust dispersion of the candidate segment within the sliding window is defined as Mad. Abnormal peaks are suppressed according to the following rules: When the deviation of any sampling point from Med is greater than the product of a preset threshold and Mad, the sampling point is replaced with Med. After completing the preprocessing operation, the candidate fragments are combined into event-level hydraulic response fragments according to time synchronization, forming the event fragment set Seg.

4. The fault early warning method based on deep learning according to claim 3, characterized in that: S2 includes S21; S21. Perform time-series analysis on each event-level hydraulic response segment in the event segment set Seg to analyze the dynamic response relationship between pressure changes and flow rate changes before and after valve operation; In each event-level hydraulic response segment, the onset time, process, and stable interval of the pressure and flow rate changes are determined, and hysteresis features describing the pressure change relative to the flow rate change are extracted. These hysteresis features include the time difference between the onset time of the pressure change and the onset time of the flow rate change, the stable duration required for the pressure change to reach the stable interval, the stable duration required for the flow rate change to reach the stable interval, and the relative displacement of the pressure and flow rate changes on the time axis. The hysteresis features extracted from each event-level hydraulic response segment are combined and vectorized, and the hysteresis features corresponding to all event-level hydraulic response segments are aggregated to form a hysteresis feature set Lag for describing the hysteresis behavior of the valve hydraulic response.

5. The fault early warning method based on deep learning according to claim 4, characterized in that: S3 includes S31; S31. Based on the historical operating data of the target valve, the continuity of the target valve's operating status in the historical time dimension is determined to identify the historical time period in which the target valve is continuously in normal operating condition; wherein, the historical time period in which the target valve is continuously in normal operating condition is determined in the following manner: Obtain the historical operation and maintenance records corresponding to the target valve, and divide the historical time into multiple continuous time segments based on the time axis; within each time segment, determine whether the target valve meets the preset normal operation judgment conditions; The normal operation determination conditions include: no valve maintenance operation or valve fault handling operation is recorded within the time segment, and no manual forced switching or emergency dispatch operation occurs; when a preset number of adjacent time segments meet the normal operation determination conditions, multiple adjacent time segments are merged to form a historical time segment in which the target valve is continuously in a normal operation state. After determining the historical time period, the historical time period is used as the time filtering condition. Hysteresis features formed within the historical time period are extracted from the obtained hysteresis feature set Lag and used as reference feature samples to represent the hydraulic response hysteresis behavior of the target valve under normal operating conditions. Statistical modeling is performed on the reference feature sample. The statistical modeling process includes: calculating the value distribution, central tendency and fluctuation range of each type of hysteresis feature in the reference feature sample to obtain statistical description results that describe the overall distribution of hysteresis features of the target valve under normal operating conditions. Based on the statistical description results, a normal hysteresis baseline set Bas is constructed to represent the overall distribution characteristics of the pressure and flow response hysteresis behavior of the target valve under normal operating conditions.

6. The fault early warning method based on deep learning according to claim 5, characterized in that: S3 further includes S32; S32. After completing the construction of the normal hysteresis baseline set Bas, the normal hysteresis baseline set Bas is used as the source of training samples for model training. A deep learning model is constructed based on autoencoder deep technology. The deep learning model includes an encoding network and a decoding network, wherein: the encoding network is used to map the hysteresis features of the input to a low-dimensional latent representation space, and the decoding network is used to reconstruct the hysteresis features of the input based on the low-dimensional latent representation. Using the hysteresis features in the normal hysteresis baseline set Bas as training samples, the deep learning model is iteratively trained so that the deep learning model learns the distribution pattern and correlation of the hysteresis features of the target valve under normal operating conditions by minimizing the difference between the input hysteresis features and the reconstructed hysteresis features. After completing the iterative training and satisfying the preset convergence condition, the model parameters of the deep learning model are solidified to obtain a hysteresis recognition model Mdl, which is used to evaluate the deviation of the input hysteresis features from the distribution represented by the normal hysteresis baseline set Bas.

7. The fault early warning method based on deep learning according to claim 6, characterized in that: S4 includes S41; S41. Based on the newly formed event-level hydraulic response fragment, generate and update the hysteresis feature set Lag in real time; From the real-time updated hysteresis feature set Lag, select the hysteresis feature corresponding to the current valve operating state, use the hysteresis feature as input, and substitute it into the obtained hysteresis recognition model Mdl. The hysteresis recognition model Mdl, based on the autoencoder structure, encodes and reconstructs the hysteresis feature one by one, and calculates the deviation between the input value and the corresponding reconstructed value for each hysteresis feature, thus obtaining a feature deviation set corresponding to each hysteresis feature. Dimensionless processing is performed on each feature deviation in the feature deviation set. After the dimensionless processing is completed, the dimensionless feature deviations are summed with equal weights. The hysteresis risk score Rsk, which represents the degree of deviation of the current valve hydraulic response hysteresis behavior from the normal operating state, is calculated by summing the dimensionless feature deviations and dividing by the number of hysteresis features.

8. The fault early warning method based on deep learning according to claim 7, characterized in that: S4 also includes S42; S42. After obtaining the hysteresis risk score Rsk, record and update the hysteresis risk score Rsk in chronological order. Based on the current value of the hysteresis risk score Rsk and its change over time, risk assessment processing is performed on the hysteresis risk score Rsk to determine whether the current hydraulic response hysteresis behavior of the valve meets the preset early warning triggering conditions. When the hysteresis risk score Rsk meets the preset warning triggering condition, a warning triggering record containing the target valve identifier, warning triggering time and corresponding hysteresis risk score Rsk is generated, and the warning triggering records are collected to form a warning triggering set Trg for indicating abnormal valve response hysteresis. No action is taken when the hysteresis risk score Rsk does not meet the preset warning triggering conditions; The warning triggering conditions include any of the following determination methods: a determination method based on the hysteresis risk score Rsk exceeding a preset risk threshold, or a determination method based on the hysteresis risk score Rsk continuously increasing over multiple consecutive assessment periods.

9. A fault early warning method based on deep learning according to claim 8, characterized in that: S5 includes S51; S51. Based on the warning trigger set Trg obtained in step S42, organize and output the warning trigger records included: collect the warning trigger records according to the target valve identifier and trigger time to form a warning result set Alr used to characterize the hidden operational abnormality risk of the target valve; The warning result set Alr is used to indicate the potential risks of valve jamming, inadequate execution, or abnormal feedback in the future operation of the target valve, and is output and stored as a warning result of valve operation status.

10. A fault early warning system based on deep learning, applied to the fault early warning method based on deep learning as described in any one of claims 1 to 9, characterized in that: It includes a valve operation data acquisition module, a hysteresis feature extraction module, a deep model construction module, an early warning risk assessment module, and an early warning result generation module; The valve operation data acquisition module collects valve operation records related to the target valve to form a valve operation event set Eve, and simultaneously collects the pressure time series of valve-related measuring points to form a pressure time series Pre and the flow time series to form a flow time series Flo. Based on the valve operation event set Eve, an event window is constructed, and the pressure time series Pre and the flow time series Flo are preprocessed to obtain an event fragment set Seg. The hysteresis feature extraction module extracts hysteresis features between pressure changes and flow rate changes based on the event fragment set Seg, forming a hysteresis feature set Lag. The deep model building module constructs a normal hysteresis baseline set Bas from the hysteresis feature set Lag based on the historical data of the normal operation of the target valve, and trains a deep learning model on the normal hysteresis baseline set Bas to obtain the hysteresis recognition model Mdl. The early warning risk assessment module inputs the real-time generated hysteresis feature set Lag into the hysteresis recognition model Mdl for processing, outputs the risk assessment result of valve response hysteresis, forms a hysteresis risk score Rsk, and generates an early warning trigger set Trg when the hysteresis risk score Rsk meets the preset conditions. The warning result generation module outputs warning results for hidden valve anomalies based on the warning trigger set Trg, forms a warning result set Alr, and sends it down for storage.

Images