Safety valve leakage intelligent alarm method and system

Abstract

The invention relates to the technical field of safety monitoring, and discloses an intelligent alarm method and system for safety valve leakage, and the method comprises the steps: obtaining original vibration data and original pressure data of a safety valve, and carrying out the synchronous processing and denoising, thereby obtaining a vibration data sequence and a pressure data sequence; calculating signal energy distribution based on the data sequence and analyzing a frequency component to obtain a frequency domain feature vector; carrying out peak value labeling, noise filtering and risk assessment on the frequency domain feature vector, and identifying an initial leakage sign; and fusing the multi-source features to generate a comprehensive risk index and triggering a corresponding alarm. According to the method, weak leakage signals caused by early-stage tiny abrasion of the sealing surface of the safety valve can be accurately recognized under the complex noise background, and accurate early warning of leakage risks is achieved.

Description

Technical Field

[0001] This invention relates to the field of safety monitoring technology, and in particular to an intelligent alarm method and system for safety valve leakage. Background Technology

[0002] Currently, safety valves are key components for ensuring the safety of industrial systems. Real-time monitoring and early warning of their operating status are crucial for preventing leakage accidents and ensuring production safety. Utilizing industrial Internet of Things (IoT) technology to monitor the status of safety valves online has become an important development direction.

[0003] In one existing technology, vibration and pressure sensors with integrated sensor network chips are deployed on a safety valve to collect signals in real time. Anomaly detection is then performed using methods such as fixed threshold comparison or basic spectrum template matching. An alarm is triggered when the monitored data exceeds a preset peak threshold or matches a known fault spectrum. However, industrial operating conditions are complex and variable. The leakage characteristics representing early, minor wear on the sealing surface in the signals collected by the sensor network chip are extremely weak and easily masked by strong background noise. Existing judgment methods based on fixed rules struggle to dynamically adapt to changing operating conditions and cannot effectively filter out noise interference, resulting in insufficient ability to identify early, weak leakage characteristics.

[0004] Therefore, existing technologies suffer from low accuracy in early warning of safety valve leaks due to the difficulty in accurately identifying early, subtle leak characteristics of safety valves in complex noise environments. Summary of the Invention

[0005] This invention provides a method and system for an intelligent alarm system for safety valve leakage, which solves the problem in the prior art where the early warning accuracy of safety valve leakage is low because it is difficult to accurately identify the early weak leakage characteristics of safety valves in complex noise backgrounds.

[0006] Firstly, in order to solve the above-mentioned technical problems, the present invention provides an intelligent alarm method for safety valve leakage, comprising: The raw vibration data and raw pressure data of the operating environment of the safety valve are acquired, and the raw vibration data and raw pressure data are synchronously processed and denoised to obtain the vibration data sequence and pressure data sequence. Based on the vibration data sequence and the pressure data sequence, the signal energy distribution is calculated and the frequency components are analyzed to obtain the frequency domain feature vector; Peak labeling and noise filtering are performed on the frequency domain feature vector. After noise filtering, secondary peaks are extracted as peak components. It is determined whether the peak components exceed the preset peak benchmark. If so, the logical relationship of high-risk signals is mapped and structured data is constructed. If not, structured data is not constructed and the anomaly detection process ends. Based on the structured data, deviation features are extracted, the rate of change of deviation features is calculated, and cross-compared with historical benchmark data to obtain the comparison difference value. Based on the changing trend and frequency offset range of the comparison difference value, it is determined whether the increasing trend and the preset frequency offset range are satisfied. If yes, the result of the initial leakage sign determination is that it exists; otherwise, the result of the initial leakage sign determination is that it does not exist. When the initial signs of leakage are determined to exist, a comprehensive risk index is generated by fusing the deviation change rate with the difference value of the pressure data sequence. The comprehensive risk index is then compared with historical data to obtain the risk range result. Based on the risk range results, an alarm signal sequence is generated, and the alarm signal sequence is adaptively smoothed to obtain a confidence segment; An iterative verification process is performed on the confidence level segment to determine the final leakage risk level of the safety valve, trigger the corresponding alarm, and obtain the risk rating result.

[0007] Secondly, the present invention provides an intelligent alarm system for safety valve leakage, comprising: The data acquisition and preprocessing module is used to acquire the raw vibration data and raw pressure data of the operating environment of the safety valve, and to perform synchronous processing and noise reduction on the raw vibration data and raw pressure data to obtain vibration data sequence and pressure data sequence. The frequency domain feature extraction module is used to calculate the signal energy distribution and analyze the frequency components based on the vibration data sequence and the pressure data sequence to obtain the frequency domain feature vector; The high-risk signal screening and structuring module is used to perform peak labeling and noise filtering on the frequency domain feature vector. After noise filtering, secondary peaks are extracted as peak components. It is determined whether the peak components exceed the preset peak benchmark. If so, the logical relationship of the high-risk signal is mapped and structured data is constructed. If not, structured data is not constructed and the current anomaly detection process ends. The deviation feature extraction and rate comparison module is used to extract deviation features based on the structured data, calculate the deviation change rate of the deviation features and cross-compare them with historical benchmark data to obtain the comparison difference value. The initial leakage determination module is used to determine whether the increasing trend and preset frequency offset range are satisfied based on the changing trend and frequency offset range of the comparison difference value. If yes, the module outputs the initial leakage indication determination result as existing; otherwise, the module outputs the initial leakage indication determination result as not existing. The comprehensive risk assessment module is used to generate a comprehensive risk index by fusing the deviation change rate and the difference value of the pressure data sequence when the initial signs of leakage are determined to exist. The comprehensive risk index is then compared with historical data to obtain the risk range result. The alarm processing and filtering module is used to generate an alarm signal sequence based on the risk range result, and adaptively smooth the alarm signal sequence to obtain a confidence segment; The risk rating alarm module is used to perform an iterative verification process on the confidence level segment, determine the final leakage risk level of the safety valve, trigger the corresponding alarm, and obtain the risk rating result.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention uses a multi-channel sensor array to synchronously collect vibration and pressure data, and uses wavelet transform, cross wavelet transform and wavelet packet transform for joint time-frequency domain analysis and denoising, which effectively removes strong background noise and significantly improves the signal-to-noise ratio, providing a high-quality data foundation for solving the problem of early weak leakage signals being submerged.

[0009] (2) This invention constructs a two-dimensional signal intensity distribution matrix, uses dynamic thresholds to screen the region of interest, and finely separates the fundamental frequency and residual sideband components to identify weak signal feature points, thereby constructing a frequency domain feature vector with high discrimination, which greatly enhances the ability to capture and detect extremely weak frequency domain modulation features induced by early wear of the safety valve sealing surface.

[0010] (3) This invention constructs the identified high-risk frequency domain features into structured data containing logical mapping relationships, extracts deviation feature sequences on this basis, analyzes the trend of their change rate using the sliding window method, and then cross-compares them with historical benchmarks, thus realizing the leap from static threshold judgment to dynamic trend intelligent analysis, effectively reducing false alarms caused by normal fluctuations in operating conditions, and accurately identifying initial signs of leakage.

[0011] (4) This invention generates a comprehensive risk index by integrating the deviation change characteristics of the characterization of state evolution rate with the real-time pressure difference characteristics, and introduces a dynamic safety envelope based on historical data statistics for comparison, thereby realizing a comprehensive risk assessment of multiple physical quantities and multi-dimensional information, making the risk judgment more comprehensive and more adaptable to the actual operating conditions of the equipment. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of a safety valve leakage intelligent alarm method provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a safety valve leakage intelligent alarm system provided in the second embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Reference Figure 1 The first embodiment of the present invention provides a smart alarm method for safety valve leakage, comprising the following steps: S11, acquire the original vibration data and original pressure data of the operating environment of the safety valve, and perform synchronous processing and noise reduction on the original vibration data and the original pressure data to obtain the vibration data sequence and the pressure data sequence; S12, based on the vibration data sequence and the pressure data sequence, calculate the signal energy distribution and analyze the frequency components to obtain the frequency domain feature vector; S13, perform peak labeling and noise filtering on the frequency domain feature vector, extract secondary peaks as peak components after noise filtering, determine whether the peak components exceed the preset peak reference, if so, map the logical relationship of high-risk signals and construct structured data, if not, do not construct structured data and end the current anomaly detection process. S14, extract deviation features based on the structured data, calculate the deviation change rate of the deviation features and cross-compare them with historical benchmark data to obtain the comparison difference value; S15. Based on the changing trend and frequency offset range of the comparison difference value, determine whether the increasing trend and preset frequency offset range are satisfied. If yes, the result of the initial leakage sign determination is that it exists; otherwise, the result of the initial leakage sign determination is that it does not exist. S16, when the initial signs of leakage are determined to exist, a comprehensive risk index is generated by fusing the difference between the deviation change rate and the pressure data sequence, and the comprehensive risk index is compared with historical data to obtain the risk range result; S17, Based on the risk range result, generate an alarm signal sequence, and adaptively smooth the alarm signal sequence to obtain a confidence segment; S18, perform an iterative verification process on the confidence level segment to determine the final leakage risk level of the safety valve, trigger the corresponding alarm, and obtain the risk rating result.

[0015] In step S11, it is necessary to acquire the raw vibration data and raw pressure data of the safety valve's operating environment, and to perform synchronous processing and noise reduction on the raw vibration data and raw pressure data to obtain a vibration data sequence and a pressure data sequence, including: The raw vibration data and raw pressure data of the safety valve's operating environment are synchronously collected by a multi-channel sensor array and the timestamps are recorded to obtain the raw dataset; Based on the original dataset, the original vibration data is decomposed into sub-band coefficients at multiple scales, and the pressure data is synchronously segmented to obtain vibration sub-band coefficients and pressure sub-band sequences. The time-frequency domain coherence spectra of the vibration sub-band coefficients and the pressure sub-band sequence are calculated by cross wavelet transform, the phase difference sequence is determined, and the coupling feature vector is obtained. The joint energy distribution features are extracted from the coupled feature vectors and the abnormal features are fused to obtain vibration data sequences and pressure data sequences.

[0016] It should be noted that the timestamp accuracy needs to reach the millisecond level; feature correlation specifically refers to the correlation between the amplitude fluctuation of the vibration signal and the change amplitude of the pressure signal at the same time, and the correlation between the coordinated changes of vibration components in different frequency bands and pressure fluctuations, etc.

[0017] A continuous Morlet wavelet is used as the mother wavelet for multi-scale decomposition. By adjusting the scale parameters, the analysis focuses on the core frequency range of 10Hz to 500Hz, which contains the operating characteristics and potential fault signals of safety valves in industrial scenarios. Among these, the frequencies of periodic vibrations generated by the normal opening and closing of the valve disc, vibrations caused by friction of the sealing surface, stable fluctuations in medium pressure, and abnormal pressure changes caused by leakage are concentrated in the operating characteristic frequency range of industrial equipment from 10Hz to 500Hz.

[0018] Specifically, the cross-wavelet transform generates a time-frequency domain coherence spectrum by calculating the intensity of the coordinated changes of the vibration subband and the pressure subband at different times and frequencies. The spectrum ranges from 0 to 1; the closer the value is to 1, the higher the coupling degree between the vibration and pressure signals at the same time and frequency. For example, during normal operation, the two signals change coordinatedly, resulting in a higher coherence spectrum value. A value closer to 0 indicates a lower coupling degree. During a fault, the coordinated relationship is disrupted, and the coherence spectrum value decreases. This is the quantification of the time-frequency coupling relationship. The phase difference sequence records the phase difference between the vibration and pressure signals at the same frequency. During normal operation, energy transfer is timely, and the phase difference remains stable within a small range, such as ±0.1π. When the safety valve experiences faults such as wear on the sealing surface, the transmission path of vibration energy to the pressure signal is blocked, leading to a delay in the pressure signal response and a significant increase in the phase difference, such as exceeding 0.2π. This energy transfer delay can be captured by observing abnormal fluctuations in the sequence.

[0019] It is worth noting that low-frequency energy, typically associated with the main structure of the safety valve and major flow-induced vibrations, such as the 10Hz to 100Hz range, dominates the total energy. Meanwhile, mid-to-high-frequency energy, associated with more refined dynamic processes such as sealing surface friction, fluid turbulence, and leakage jetting, such as the 100Hz to 500Hz range, accounts for a smaller proportion during normal operation. When a leakage fault occurs, due to increased sealing surface friction and the impact of the leaking fluid, the vibration and pressure signal energy in the mid-to-high-frequency bands often shows a significant increase, while the proportion of low-frequency dominant energy may decrease accordingly. This shift in energy distribution characteristics can serve as an important basis for identifying anomalies.

[0020] For ease of understanding, the following is an illustrative example: a sensor array includes accelerometers and pressure transmitters mounted near the valve body sealing surface and adjacent pipelines, synchronously acquiring data at a sampling rate of 1000Hz, with timestamp accuracy controlled within 1 millisecond. Continuous acquisition for 10 seconds yields 10,000 synchronous data points. The useful signals in the vibration data are concentrated between 10Hz and 500Hz, while noise signals are distributed above 500Hz. The pressure data is then divided into 8 sub-band sequences according to the same time scale using 8-scale continuous Morlet wavelet decomposition.

[0021] Calculations using cross-wavelet transform showed that during normal operation, the coherence spectrum of the 10Hz-100Hz low-frequency vibration sub-band and its corresponding pressure sub-band was 0.85, with a stable phase difference of 0.05π, indicating tight coupling. When a minute wear of 0.1mm appeared on the sealing surface, the coherence spectrum dropped to 0.48, and the phase difference lagged to 0.2π, reflecting the disruption of the coupling relationship and energy transfer delay caused by the fault. Based on this coupling feature vector, wavelet packet decomposition was performed to extract the energy distribution of each node. Under normal conditions, the energy proportion of the 10Hz-100Hz low-frequency nodes reached 72%, the energy proportion of the 100Hz-500Hz mid-high frequency nodes was 18%, and the energy proportion of the noise band above 500Hz was 10%. During wear and leakage, the energy proportion of the 10Hz-100Hz low-frequency nodes dropped to 55%, the energy proportion of the 100Hz-500Hz mid-high frequency nodes rose to 42%, and the energy proportion of the noise band above 500Hz was compressed to 3%. After fusing this abnormal feature, vibration and pressure data sequences with a signal-to-noise ratio improved by 2.5 times were generated.

[0022] In step S12, based on the vibration data sequence and the pressure data sequence, the signal energy distribution is calculated and the frequency components are analyzed to obtain a frequency domain feature vector, including: Based on the vibration data sequence and the pressure data sequence, the signal intensity distribution matrix is ​​obtained by segmenting the data into preset time windows and calculating the root mean square value of each slice. Based on the two-dimensional signal intensity distribution matrix, a dynamic threshold is set to determine the range of signals of interest. Perform frequency domain transformation on the signal range of interest and separate the core component to obtain the fundamental frequency component and the residual sideband component; Based on the fundamental frequency component and the residual sideband component, weak signal feature points are identified and combined to obtain a frequency domain feature vector.

[0023] It should be noted that the preset time window length is set to 0.5 seconds, with a step size of 0.2 seconds. A window length that is too short will result in unstable data statistics within a single window, while a window that is too long will smooth out rapid transient changes in the signal. The overlap of the 0.5-second window length and the 0.2-second step size aims to effectively capture early fault transient characteristics that may last for hundreds of milliseconds, while ensuring that each window has enough sampling points (e.g., 500 points) for stable energy calculation. The 0.5-second window length, combined with a 1000Hz sampling rate, allows each window to contain 500 data points, ensuring sufficient data for accurate energy calculation without masking short-term signal fluctuations due to an excessively long window. The 0.2-second step size means that there is a 0.3-second overlap between adjacent time windows. This overlay sampling design is to avoid missing transient energy changes, which refer to sudden and short-duration energy fluctuations during the operation of the safety valve, such as slight valve disc sticking or momentary friction of the sealing surface. These changes are often critical signals in the early stages of a fault, and their duration may only be 0.1 to 0.2 seconds. The 0.2-second step size ensures that these short-term signals are completely covered by at least one time window and are not missed due to falling between two windows.

[0024] The dynamic threshold equals the matrix mean plus 1.5 times the matrix standard deviation. The matrix mean reflects the overall signal strength level under the current operating conditions, while the standard deviation reflects the dispersion of signal strength. Choosing the combination of the matrix mean plus 1.5 times the matrix standard deviation is a reasonable approach based on extensive statistical analysis of industrial operating conditions. The 1.5 times standard deviation coefficient can both cover signal fluctuations during normal operation, avoiding misjudging normal fluctuations as abnormalities, and accurately capture abnormal signals exceeding the normal range, avoiding missing true fault signals. In industrial scenarios, signal strength varies significantly under different operating conditions, and a fixed threshold cannot adapt to such fluctuations; therefore, a dynamic threshold is used.

[0025] The frequency domain transformation employs Fast Fourier Transform (FFT). The fundamental frequency component is the dominant frequency component during normal operation of the safety valve, generated by stable factors such as inherent equipment vibration and pipeline excitation; its frequency and amplitude are fixed. The residual sideband component consists of frequency components distributed around the fundamental frequency, generated by signal modulation effects; during normal operation, its amplitude is small and its distribution is symmetrical. Wear on the sealing surface leads to uneven contact, generating additional frictional vibrations. This vibrational energy is superimposed on the fundamental frequency as a modulated signal, directly causing changes in the residual sideband component. In the early stages of wear, the amplitude of the sideband component increases slowly, and the increase is positively correlated with the degree of wear; as wear intensifies, the sideband component expands towards both sides of the fundamental frequency, increasing its frequency distribution range.

[0026] It is worth noting that power spectral density analysis of the residual sideband components is crucial for identifying weak leakage signals. Weak leakage signals refer to those caused by early faults such as minor wear or cracks on the sealing surface of safety valves. These signals have low energy, indistinct features, and are mixed in with background noise, making them difficult to identify using conventional time-domain analysis methods. Power spectral density analysis, by calculating the signal power within a unit frequency range, can highlight weak frequency features masked by noise in the time domain in the frequency domain. Weak leakage caused by a fault will produce vibrations at specific frequencies, corresponding to an abnormally high power density in a certain frequency band of the residual sideband components. A power density at the fault feature point that is more than twice the background noise level is the optimal threshold choice for balancing identification sensitivity and reliability in engineering practice. If the ratio is less than twice, random noise fluctuations are easily misjudged as fault signals, leading to false detections; if it is more than twice, it ensures that the feature point is a signal generated by a real fault, rather than noise interference, thus guaranteeing the accuracy of feature extraction.

[0027] For example, the vibration and pressure data sequences are segmented into 0.5-second time windows with 0.2-second steps, sampled at a 1000Hz rate, with each window containing 500 data points and adjacent windows overlapping by 0.3 seconds. The root mean square value of each slice is calculated, constructing a two-dimensional signal intensity distribution matrix with rows corresponding to time slices and columns representing vibration and pressure intensities. Assuming the mean vibration intensity of the matrix is ​​0.3 m / s² with a standard deviation of 0.08 m / s², and the mean pressure intensity is 0.1 MPa with a standard deviation of 0.02 MPa, the dynamic threshold is determined by adding 1.5 times the standard deviation to the mean. That is, the vibration intensity threshold is 0.3 + 0.12 = 0.42 m / s², and the pressure intensity threshold is 0.1 + 0.03 = 0.13 MPa. Regions with intensities exceeding the thresholds are marked as signal intervals of interest. An FFT transformation is performed on the joint vibration and pressure sequence within this interval to separate the 50Hz fundamental frequency component. This frequency is the pipeline excitation frequency and belongs to the dominant frequency for normal operation of the safety valve. In a scenario where there is a minute wear of 0.1 mm on the sealing surface, the frictional vibration caused by the wear modulates the fundamental frequency, forming a weak peak at ±10 Hz from the fundamental frequency, which is the residual sideband component. Power spectral density analysis revealed that the power density of this weak peak is 2.3 times that of the background noise, meeting the identification criterion of more than 2 times. By combining the frequency and amplitude parameters of the fundamental frequency with the position ±10 Hz and amplitude of the weak peak, a multi-dimensional frequency domain feature vector containing multi-dimensional fault characteristics is formed.

[0028] In step S13, peak labeling and noise filtering are performed on the frequency domain feature vector. After noise filtering, secondary peaks are extracted as peak components. It is determined whether the peak components exceed a preset peak benchmark. If so, the logical relationship of high-risk signals is mapped, and structured data is constructed. If not, structured data is not constructed, and the anomaly detection process ends. This includes: Based on the frequency domain feature vector, the peak position and noise component are extracted and noise filtering is performed to obtain the filtered spectrum features. Based on the filtered spectral features and the peak position, secondary peaks are extracted as peak components; The amplitude ratio of the peak component relative to the base frequency main peak is calculated. If the amplitude ratio exceeds a preset peak reference, it is identified as a high-risk signal source. Structured data is constructed based on key-value pairs. The structured data includes a logical mapping relationship between the frequency offset, amplitude ratio, and risk level of the high-risk signal source. If the amplitude ratio does not exceed the preset peak reference, it is determined to be a normal signal component, no structured data is constructed, and the anomaly detection process ends.

[0029] In one implementation, noise filtering employs a moving average filtering method with a window width of 5 frequency points. Moving average filtering smooths the amplitude spectrum of the frequency domain feature vector through a sliding window. Specifically, the amplitude of each frequency point is obtained by taking the arithmetic mean of its own amplitude and the amplitudes of its two preceding and following frequency points.

[0030] The true peak profile is a core signal feature that reflects the operating status of the equipment. Its amplitude changes smoothly and is concentrated in a specific frequency range. The window width is controlled at 5 frequency points, which can filter high-frequency noise without causing peak profile distortion.

[0031] The preset peak threshold is set to 0.15, meaning that a signal source is considered high-risk when the amplitude ratio of the secondary peak to the fundamental frequency peak exceeds 0.15. This threshold is determined based on signal-to-noise ratio (SNR) quantization analysis. SNR is the ratio of the target signal amplitude to the background noise amplitude. In engineering practice, an SNR ≥ 3 is typically used as the critical standard for effective signal identification. Statistical analysis of extensive historical fault data reveals that under normal operating conditions, the SNR of the secondary peak is generally below 2, corresponding to an amplitude ratio below 0.1. During slight leakage, the SNR of the secondary peak is ≥ 3, with the amplitude ratio stabilizing between 0.15 and 0.25. During moderate and severe leakage, the SNR further increases, and the amplitude ratio also increases accordingly. Setting 0.15 as the threshold essentially corresponds to an effective signal threshold of SNR 3. This avoids misjudging normal fluctuations with an SNR less than 2 as faults, ensuring reliability, while accurately capturing weak signals in the early stages of leakage with an SNR ≥ 3, ensuring sensitivity and achieving a balance between sensitivity and reliability.

[0032] It should be noted that structured data is stored in key-value pair format, with a one-to-one correspondence between "keys" and "values." The "key" identifies the data type, and the "value" is the corresponding data content. The logical mapping relationship of risk levels is determined based on the combination of frequency offset and amplitude ratio. The specific characteristic parameters include frequency offset and amplitude ratio. Frequency offset is the frequency difference between the secondary peak and the fundamental frequency peak, and amplitude ratio is the ratio of the amplitude of the secondary peak to the amplitude of the fundamental frequency peak. The mapping relationship is established through statistical analysis of a large amount of historical fault data. For example, it can be assumed that when the frequency offset is between ±5Hz and ±10Hz and the amplitude ratio is between 0.15 and 0.25, it corresponds to slight leakage; when the frequency offset is between ±10Hz and ±20Hz and the amplitude ratio is between 0.25 and 0.4, it corresponds to moderate leakage; and when the frequency offset exceeds ±20Hz and the amplitude ratio is greater than 0.4, it corresponds to severe leakage.

[0033] It should be noted that this implementation method is suitable for industrial scenarios with complex noise components, including chemical production workshops, metallurgical plants, and thermal power plants. Chemical production workshops experience combined vibration interference from various equipment such as pumps and compressors, along with noise from media flow. Metallurgical plants suffer from electromagnetic interference and signal distortion caused by high temperatures, resulting in a chaotic noise frequency distribution. In thermal power plants, vibration noise from large equipment such as fans and pumps is superimposed with safety valve signals, leading to complex noise components. This method uses a dynamic weighted moving average filtering algorithm, rather than fixed-rule filtering. Its core principle is to adaptively adjust the noise suppression strength by dynamically adjusting the weights of the filtering coefficients. High-frequency components refer to frequency components above the upper limit of the equipment's normal operating characteristic frequency by 50Hz, and the amplitude standard deviation is an indicator of the severity of high-frequency component fluctuations. The filter coefficient weights are weighted coefficients of the amplitude at each frequency point within the sliding window. They are set based on the standard deviation of the amplitude of the high-frequency components. When the standard deviation of the amplitude is large, i.e., when the noise is severe, the weights are concentrated at the center frequency point of the window, and the weights of the edge frequencies are reduced. For example, the weight of the center frequency point is 0.4, the weights of the two adjacent frequencies are 0.2 each, and the weights of the two outermost frequencies are 0.1 each, which enhances the suppression effect on high-frequency noise. When the standard deviation of the amplitude is small, i.e. when the noise is gentle, the weights are distributed more evenly. For example, the weight of each frequency point is 0.2, which avoids excessive smoothing and loss of useful signal details.

[0034] It is worth noting that when multiple high-risk signal sources exist, their logical relationships need to be aggregated to improve the comprehensiveness of risk identification. Aggregating logical relationships refers to statistically analyzing the key characteristics of multiple high-risk signal sources and their inherent correlations, including the symmetry of frequency offset, the synchronicity of amplitude ratio, and the overlap of occurrence times. Specifically, the quantitative criterion for synchronously increasing amplitude ratios is that within three consecutive time units, the difference in amplitude ratio changes of multiple high-risk signal sources is within ±0.02, and the overall trend is monotonically upward. The quantitative criterion for overlapping times is that the detection times of multiple high-risk signal sources fall within the same 0.5-second analysis time window, or the time difference between detection times is ≤0.1 seconds; this value is the set time tolerance. Asymmetric sideband sources refer to sideband sources distributed around both sides of the fundamental frequency, with asymmetrical frequency offset directions. For example, one side is concentrated at +5Hz and +10Hz, while the other side only has +15Hz, without corresponding -5Hz and -10Hz, or the amplitude ratio is asymmetrical, for example, the amplitude ratio of sideband sources on one side is generally above 0.2, while the other side is only around 0.15. When the sealing surface wears more intensely, it leads to an imbalance in the valve's vibration characteristics. The originally symmetrical sideband distribution is disrupted, and the frictional vibration generated by the wear excites multiple sideband sources of different frequencies, with the amplitude ratios of these sideband sources increasing synchronously. Therefore, when multiple asymmetrical sideband sources appear simultaneously, with their amplitude ratios increasing synchronously and overlapping in time, it can be determined as a trend of intensified wear, avoiding the bias caused by relying solely on a single high-risk signal source.

[0035] For example, by scanning the amplitude spectrum of the frequency domain feature vector, local peaks with amplitudes higher than twice the overall mean are identified as peak positions, and the remaining low-amplitude regions are considered as noise components. A moving average filter with a window width of 5 frequency points is used. The amplitude of each frequency point is obtained by averaging its own amplitude and the amplitudes of the two frequency points before and after it. For example, if the original amplitude of a frequency point is 0.3, and the amplitudes of the adjacent frequency points are 0.28, 0.29, 0.31, and 0.32 respectively, the average amplitude is 0.3.

[0036] Secondary peaks within ±15Hz of the peak position are extracted as peak components. The amplitude ratio of this component to the fundamental frequency main peak is calculated to be 0.18, exceeding the preset peak reference of 0.15, thus identifying it as a high-risk signal source. Structured data is constructed based on key-value pairs as {"Frequency Offset": +10Hz, "Amplitude Ratio": 0.18, "Risk Level": "Slight Leakage"}. If another high-risk signal source is detected simultaneously, with a frequency offset of +12Hz and an amplitude ratio of 0.2, the frequency offsets of the two signal sources are asymmetrical, there are no corresponding -10Hz and -12Hz sideband sources, and they are detected within the same 0.5-second analysis time window, meeting the criteria for time overlap. The difference in the amplitude ratio changes of the two sources within three consecutive time units is 0.02, meeting the quantitative requirement of synchronously increasing amplitude ratios. After aggregating their logical relationships, it is determined to be a trend of accelerated wear, and proceeds to the subsequent anomaly analysis process.

[0037] In step S14, deviation features are extracted based on the structured data, the rate of change of the deviation features is calculated, and cross-compared with historical benchmark data to obtain the comparison difference value, including: The amplitude deviation values ​​of the structured data are extracted and sorted by time to obtain the deviation feature sequence; The rate of change of the deviation characteristic sequence is calculated using the sliding window method to obtain the deviation change rate. Based on the rate of change of the deviation, a cross-comparison is performed with historical benchmark data to obtain the comparison difference value.

[0038] It should be noted that the deviation feature sequence is an ordered set of amplitude deviation values ​​extracted from structured data and related to high-risk signal sources. The amplitude deviation value refers to the difference between the amplitude ratio recorded in the structured data and the baseline amplitude ratio during normal operation of the safety valve, directly reflecting the degree of deviation of the current amplitude parameter from the normal state. The core parameters such as the amplitude ratio stored in the structured data dynamically change with the operating state of the safety valve, and the deviation values ​​of these parameters are directly related to the severity of faults such as sealing surface wear. Sort the amplitude deviation values ​​according to the chronological order of data acquisition to form a continuous time-series data chain. Fault development is gradual; for example, as sealing surface wear progresses from slight to severe, the corresponding amplitude deviation value gradually increases. The time-sorted sequence can intuitively present this dynamic evolution process from nothing to something, from weak to strong, providing a foundation for subsequently capturing fault change patterns.

[0039] The sliding window method is used to calculate the rate of change of the deviation characteristic sequence. The window width is set to 3 seconds (i.e., covering 3 consecutive data analysis units), and each sliding step is 1 second. The 3 time points represent a 3-second time series data segment, ensuring that the data within the window reflects a short-term and stable trend. Historical benchmark data is a reference dataset of deviation change rates obtained through continuous data acquisition and statistical analysis under long-term normal operation of the safety valve. The data source must cover at least 6 months of trouble-free operation to ensure that it covers the normal deviation change patterns under different operating conditions, such as high load, low load, and temperature and humidity fluctuations. The data acquisition frequency is consistent with the current detection frequency, such as a 1000Hz sampling rate. The amplitude deviation values ​​are extracted from the acquired normal operation data, sorted by time to form a normal deviation sequence, and then the same sliding window method as the current one is used, with the window width set to 3 time points, to calculate the deviation change rate under normal conditions. Finally, the mean, standard deviation, maximum value, and other core statistical parameters of this rate are obtained through statistical analysis.

[0040] It is worth noting that the cross-comparison uses the difference between the current rate of change of deviation and the average of historical baseline data to quantify the degree of deviation between the current state and the normal state. The sign of the difference reflects the direction of deviation; a positive difference indicates that the current rate of change of deviation is higher than the normal level, suggesting the possibility of fault development; a negative difference indicates that the current rate of change is lower than the normal level, mostly due to stable operating conditions or minor disturbances. The absolute value of the difference directly quantifies the degree of deviation; the larger the absolute value, the more significant the difference between the current state and the normal state, and the higher the probability of fault development; the smaller the absolute value, the closer the current state is to the normal operating level. This comparison difference value accurately quantifies the dynamic trend of fault development and is the core basis for subsequent judgment of initial signs of leakage. Only when the difference value shows a stable increasing trend can it be a sign of initial leakage caused by faults such as wear of the sealing surface.

[0041] For example, the amplitude deviation values ​​of five consecutive time slices are extracted from structured data, namely 0.16, 0.17, 0.18, 0.19, and 0.20. Each value corresponds to the statistical result of one second of running data. These values ​​are sorted according to the chronological order of data acquisition to form a deviation characteristic sequence. This sequence shows a gradually increasing trend, reflecting that the wear on the sealing surface may be continuously developing. The sliding window width is set to three time points, i.e., three consecutive seconds. The first window contains data of 0.16, 0.17, and 0.18, and the differences between adjacent data are calculated to be 0.01 and 0.01, respectively. The average value is taken to obtain the deviation change rate of this window as 0.01 / second. The second window contains data of 0.17, 0.18, and 0.19, with the same change rate of 0.01 / second. The third window contains data of 0.18, 0.19, and 0.20, with the same change rate of 0.01 / second. Historical baseline data is derived from six months of trouble-free operation data of the safety valve. Amplitude deviation values ​​are extracted from the normal operation data to form a normal deviation sequence. The rate of change of deviation under normal conditions is calculated using the same three-time-point sliding window method. Statistically, the mean is 0.005 / second and the standard deviation is 0.001 / second. Cross-comparison calculations show that the difference is 0.01 minus 0.005, which equals 0.005 / second. This difference is positive and its absolute value is significantly greater than the standard deviation of the historical baseline data (i.e., the absolute value of the difference is greater than twice or more than the standard deviation of the historical baseline data). The current absolute value of the difference, 0.005, is five times the standard deviation of 0.001, meeting the criteria of "significantly greater than," indicating that the current rate of change of deviation is significantly higher than normal, suggesting signs of a potential fault development.

[0042] S15. Based on the changing trend and frequency offset range of the comparison difference value, determine whether the increasing trend and preset frequency offset range are satisfied. If yes, the result of the initial leakage indication determination is that it exists; otherwise, the result of the initial leakage indication determination is that it does not exist.

[0043] It should be noted that the criterion for determining an increasing trend is a monotonically increasing comparison difference value for three or more consecutive time units. The time unit is a fixed time interval between data acquisition and analysis, consistent with the time slices used in the data processing described earlier. Setting three or more time units as the criterion is based on the gradual nature of fault development and the signal characteristics of industrial scenarios. Leakage-related faults such as wear on sealing surfaces gradually worsen, and the corresponding comparison difference value will show a stable upward trend. An increase of only one or two time units is likely due to accidental factors such as momentary sensor interference or temporary fluctuations in operating conditions, and is not a true fault signal. A monotonically increasing value for three or more consecutive time units effectively filters out such accidental fluctuations, ensuring that a stable fault evolution trend is captured.

[0044] The preset frequency offset range is set to the fundamental frequency ±20Hz. The fundamental frequency is the dominant frequency component during normal operation of the safety valve, determined by stable factors such as inherent equipment vibration and pipeline excitation, and serves as the core frequency reference for the signal. Frequency offset is the frequency difference between the secondary peak of a high-risk signal source and the fundamental frequency, reflecting the impact of a fault on the signal frequency characteristics. Setting the preset range to the fundamental frequency ±20Hz is beneficial because initial leakage faults caused by minor wear on the sealing surface will lead to specific changes in the frequency characteristics of vibration and pressure signals, with frequency offsets mostly concentrated within the fundamental frequency ±20Hz range. Frequency offsets exceeding this range are more often caused by non-leakage factors such as loose equipment installation, external impacts, and environmental vibrations, and are unrelated to wear on the sealing surface. Therefore, the setting is ±20Hz.

[0045] It is worth noting that both conditions must be met simultaneously to determine the presence of initial signs of leakage. Meeting only one condition may lead to a misjudgment; an increase in the difference value alone could be due to temporary factors such as adjustments to operating conditions or fluctuations in medium parameters; a frequency deviation within a preset range alone could be an accidental phenomenon caused by random noise interference. Combining both conditions significantly improves the rigor of the judgment, effectively distinguishing normal fluctuations from genuine fault signals, and ensuring the accuracy of identifying initial signs of leakage.

[0046] For example, the time unit for data acquisition and analysis is set to 1 second. The comparison difference values ​​are 0.003, 0.004, 0.005, 0.006, and 0.007 per time unit for five consecutive time units, respectively. The value of each subsequent time unit is strictly greater than the previous one, showing a clear monotonically increasing trend, which meets the judgment criterion of three or more consecutive time units. Assuming that the fundamental frequency of the safety valve is 50Hz during normal operation and the secondary peak frequency of the high-risk signal source is 58Hz, the calculated frequency offset is +8Hz, which falls within the preset fundamental frequency ±20Hz range. Both conditions are met, therefore, the initial signs of leakage are determined to exist. If, in a certain case, the comparison difference value increases for three consecutive time units, but the frequency offset of the high-risk signal source is +25Hz, exceeding the fundamental frequency ±20Hz range, then the initial signs of leakage are determined to not exist.

[0047] In step S16, when the initial signs of leakage are determined to exist, a comprehensive risk index is generated by fusing the deviation change rate with the difference value of the pressure data sequence. The comprehensive risk index is then compared with historical data to obtain a risk range result, including: When the initial signs of leakage are determined to be present, a discretization difference operation is performed based on the pressure data sequence to obtain a pressure difference feature sequence. Based on the pressure differential characteristic sequence and the deviation change rate, a weighted summation operation is performed to obtain a comprehensive risk index; Collect historical operating data of the safety valve, calculate the mean and standard deviation of the historical operating data to obtain statistical distribution characteristics, construct a safety boundary based on the statistical distribution characteristics, and obtain a dynamic safety envelope. By comparing the comprehensive risk index with the dynamic safety envelope, if the comprehensive risk index falls outside the range defined by the dynamic safety envelope, the risk range result is obtained.

[0048] It should be noted that discretization differential operation directly obtains the pressure change rate within each time unit by calculating the pressure difference between two adjacent time units in the pressure data sequence. This difference is the instantaneous rate of pressure change within that time unit. For example, if the pressures of two adjacent time units are 2.0 MPa and 2.1 MPa, the difference of 0.1 MPa is the instantaneous rate of pressure change within that time period. In a leakage state, the leakage of the medium disrupts the pressure balance within the pipeline. The pressure will fluctuate continuously and gradually decrease as the leakage increases. The corresponding pressure differential characteristic sequence will show an abnormal increasing trend. In the early stages of leakage, the differential value increases slightly, and as the leakage intensifies, the differential value increases rapidly, directly reflecting the degree of abnormality in pressure change.

[0049] The pressure differential feature sequence and the deviation change rate form complementary features. The deviation change rate, derived from vibration data processing, primarily reflects vibration deviation changes caused by mechanical faults such as sealing surface wear, belonging to the mechanical state dimension. The pressure differential feature sequence, derived from pressure data processing, primarily reflects medium pressure fluctuations caused by leakage, belonging to the medium state dimension. Complementary features mean that the two features describe leakage risk from different dimensions. A single feature may lead to misjudgment due to factors such as operating condition fluctuations, while the combination of the two can complement and verify each other, covering multiple aspects of leakage faults and improving the comprehensiveness and accuracy of risk assessment.

[0050] In the weighted summation calculation, the weight of the deviation change rate is set to 0.6, and the weight of the pressure differential characteristic is set to 0.4. This weight allocation is determined based on statistical analysis of a large amount of historical fault data and analysis of the degree of fault impact. The deviation change rate is directly related to the rate of wear development of the sealing surface and is more critical for predicting fault development, therefore it is given a higher weight. The pressure differential characteristic is an indirect result of leakage, and its changes may be affected by fluctuations in operating parameters such as medium temperature and flow rate. Its direct indication of leakage risk is slightly weaker than that of the deviation change rate, therefore it is given a relatively lower weight.

[0051] The process of constructing the dynamic safety envelope is as follows: First, collect at least 6 months of historical normal operation data of the safety valve, covering operation data under different working conditions such as high load, low load, and fluctuations in ambient temperature and humidity, to ensure the comprehensiveness and representativeness of the data; then, perform statistical analysis on the basic data related to comprehensive risk indicators in these historical data, namely vibration deviation data and pressure data, to calculate the mean and standard deviation. The mean reflects the baseline level of characteristic data during normal operation, and the standard deviation reflects the amplitude of fluctuations under normal working conditions; finally, expand the range of the dynamic safety envelope by 1.5 times above and below the mean.

[0052] Among them, the characteristic data under normal operating conditions are mostly approximately normally distributed, and an interval of 1.5 times the standard deviation can cover about 93% of the normal operating data, which can accommodate reasonable fluctuations under different operating conditions and avoid masking the real fault signals due to the excessively wide interval.

[0053] For example, the pressure data sequence is 2.0, 2.1, 2.3, and 2.6 MPa. When performing discretization difference operation, the difference between two adjacent data points is calculated sequentially, resulting in a pressure difference feature sequence of 0.1, 0.2, and 0.3 MPa. This sequence shows a clear increasing trend, reflecting a continuously accelerating instantaneous pressure change rate, consistent with the characteristics of a leakage state. The mean of this pressure difference feature sequence is calculated to be 0.2 MPa, and the deviation change rate obtained from vibration data processing is currently 0.01 / second. Min-max normalization is performed on these two feature parameters, mapping them uniformly to the 0-1 interval. Assuming statistical analysis of historical data, the historical range of the deviation change rate is 0-0.02 / second, the current value is 0.01 / second, and after min-max normalization, it is 0.5; the historical range of the mean of the pressure difference feature is 0-0.5 MPa, the current value is 0.2 MPa, and after min-max normalization, it is 0.4. Subsequently, according to the preset weighting rules, the deviation change rate was assigned a weight of 0.6, and the mean of the pressure difference characteristic was assigned a weight of 0.4. The two normalized characteristic parameters were then weighted and summed to obtain a comprehensive risk index of 0.46. Historical normal operation data of the safety valve over the past six months was collected, covering different operating conditions such as medium pressure between 3.0 and 4.0 MPa under high load and medium pressure between 1.0 and 2.0 MPa under low load. The normalized comprehensive risk index had a mean of 0.3 and a standard deviation of 0.08. A dynamic safety envelope constructed based on the mean ± 1.5 times the standard deviation ranged from 0.3 minus 1.5 multiplied by 0.08 to 0.3 plus 1.5 multiplied by 0.08, which is 0.18 to 0.42. The current comprehensive risk index of 0.46 falls outside this range, indicating that it exceeds the reasonable range of fluctuations under normal operating conditions, suggesting a potential leakage risk.

[0054] In step S17, based on the risk range result, an alarm signal sequence is generated, and the alarm signal sequence is adaptively smoothed to obtain a confidence segment, including: Based on the comprehensive risk indicators and the risk range results, a comprehensive risk data stream is obtained; If the comprehensive risk data stream exceeds the preset security baseline, an alarm signal sequence is generated; The adaptive sliding window width is determined based on the local flip frequency of the alarm signal sequence, and a weighted average processing is performed on the alarm signal sequence based on the adaptive sliding window width to obtain a smooth alarm data stream; Identify the numerical troughs in the smoothed alarm data stream, use the numerical troughs as dividing boundaries, and segment the smoothed alarm data stream to obtain alarm data segments; Calculate the integral of the amplitude and duration of each alarm data segment to obtain a confidence value. If the confidence value is greater than a preset confidence threshold, then output the confidence segment.

[0055] It should be noted that the preset safety baseline is set as the upper limit of the dynamic safety envelope. This setting is derived from the judgment logic of the risk range results. The upper limit of the dynamic safety envelope is the maximum safe boundary of the comprehensive risk index under normal operating conditions. The core judgment criterion for the risk range results is whether the comprehensive risk index exceeds this upper limit. If it exceeds, it indicates a leakage risk; if it does not exceed, it is within the safe range. Keeping the preset safety baseline consistent with this upper limit ensures that the generation logic of the alarm signal sequence is completely consistent with the risk range results. An alarm signal will only be triggered when the value in the comprehensive risk data stream exceeds the upper limit of the dynamic safety envelope, that is, when the preset safety baseline is reached. This avoids logical contradictions such as "the risk range is judged to be safe but the alarm is triggered" or "the risk range is judged to be risky but the alarm is not triggered," thus achieving consistency between risk judgment and alarm generation.

[0056] The local flip frequency refers to the frequency at which "0" and "1" switch between each other in an alarm signal sequence. "0" represents a risk-free alarm, and "1" represents a risky alarm. The local flip frequency is calculated by first traversing the entire alarm signal sequence and counting the total number of signal state switches between adjacent time units (i.e., the previous time unit is "0" and the next is "1", or vice versa). Then, the total number of switches is divided by the total number of time units in the sequence minus one. Since N time units correspond to N-1 adjacent pairs, the switching frequency per unit time is the local flip frequency. The adaptive sliding window width is determined based on the local flip frequency. A higher frequency indicates more significant noise interference to the signal, requiring a smaller window width, with a minimum of 3 time units and a maximum of 10 time units. The time unit setting should be consistent with the time slices used in the data processing described earlier, such as 1 second per time unit, to ensure logical consistency in the analysis. The window width is limited to 3-10 time units to avoid insufficient data within the window, which could affect smoothing due to single-point fluctuations, or an excessively wide window coverage that could obscure the short-term trend of the signal.

[0057] The confidence level is calculated by integrating the amplitude and duration. The amplitude reflects the intensity of the alarm, i.e., the severity of the leakage risk; the duration reflects the persistence of the anomaly, i.e., the continuous state of the leakage risk; one time unit corresponds to 1 second. The higher the reliability of the alarm, the better. The preset confidence threshold is set to 1.0, which is the optimal value verified by a large amount of historical fault data and simulation experiments. Confidence levels below 1.0 are mostly false alarms caused by noise interference, temporary fluctuations in operating conditions, etc. For example, an amplitude of 0.1 and an integral result of 0.8 after 8 time units are likely invalid signals. Confidence levels above 1.0 are basically valid alarms caused by real faults such as wear on the sealing surface and leakage.

[0058] It's worth noting that the numerical trough is the point with the smallest value in the smoothed alarm data stream, with values ​​on both sides exceeding this point. Using the numerical trough as a dividing boundary—that is, the intensity of the preceding alarm event drops from its peak to a trough, while the intensity of the following alarm event rises from a trough to its peak—the trough becomes a natural boundary between the two events. This approach splits the continuous smoothed alarm data stream into multiple independent alarm data segments, each corresponding to a complete anomaly event. This facilitates the individual calculation of the confidence value for each event and the assessment of its reliability.

[0059] For example, the time unit is set to 1 second, and the comprehensive risk data stream is 0.07, 0.08, 0.09, 0.085, 0.095, 0.10, 0.098. The upper limit of the dynamic safety envelope is 0.08, so the preset safety baseline is set to 0.08. Traversing the comprehensive risk data stream, values ​​greater than 0.08 are marked as "1", and values ​​less than or equal to 0.08 are marked as "0", generating the original alarm signal sequence 0, 1, 1, 1, 1, 1, 1. The local flip frequency of the sequence was statistically analyzed. A switch from "0" to "1" occurred only once between the first and second time units, indicating an extremely low flip frequency. Therefore, the adaptive sliding window width was determined to be 5 time units, and weighting coefficients were set to 0.1, 0.2, 0.4, 0.2, and 0.1, with the highest weight at the center of the window decreasing towards both sides. A weighted average was applied to the comprehensive risk data corresponding to the original alarm signal sequence, resulting in a smoothed alarm data stream of 0.082, 0.081, 0.088, 0.091, 0.093, 0.096, and 0.094.

[0060] The identified trough in the smoothed alarm data stream is 0.081. Only when the values ​​on both sides of this point are greater than themselves is a complete alarm data segment obtained using this point as the dividing boundary. This segment covers the data stream for the following 6 time units, and its average amplitude is calculated to be 0.091. The duration of this segment is 6 time units, therefore the confidence score is the integral of the average amplitude and the duration, resulting in 0.546. Since this confidence score is less than the preset confidence threshold of 1.0, it indicates that the effective strength of this alarm segment does not meet the criteria for a true anomaly, and it is highly likely a minor anomaly caused by fluctuations in operating conditions. Therefore, this segment is not output as a confidence segment. If the subsequent comprehensive risk data stream continues to rise, causing the integral of the average amplitude and duration of the alarm data segment to exceed 1.0, then the corresponding confidence segment will be output, and the process will proceed to subsequent iterative verification.

[0061] In step S18, an iterative verification process is performed on the confidence level segment to determine the final leakage risk level of the safety valve, trigger the corresponding alarm, and obtain the risk rating result, including: The data from the confidence level segment is input into a preset simulation model to obtain a simulated feature sequence; Calculate the matching degree between the confidence segment and the simulated feature sequence. If the matching degree is greater than a preset matching benchmark, generate abnormal interval labeling information. Based on the abnormal interval labeling information, a result set is constructed, and the average risk value, volatility variance and duration are extracted from the result set as core features to obtain the deterioration feature vector. The degradation feature vector is mapped to a preset wear evolution trend map to obtain a wear severity value; The wear severity value is analyzed based on a preset risk range. When the wear severity value is within the preset risk range, a risk alarm corresponding to the risk level is triggered, and a risk rating result is obtained.

[0062] It is worth noting that the confidence level segment is a time-series risk data sequence containing the comprehensive risk value at each time point. To facilitate simulation model processing, the statistical characteristics of this segment are extracted, including the average risk value, volatility variance, and duration.

[0063] It should be noted that the training data for the preset simulation model is the full lifecycle failure data of the safety valve. This full lifecycle failure data covers the complete data chain of the safety valve from its brand-new condition upon leaving the factory, normal operation, to minor leakage, moderate leakage, severe leakage, and finally, shutdown due to failure. This includes vibration data, pressure data, operating parameters, and corresponding leakage degree labels for each stage. The construction process consists of three steps: First, data preprocessing involves denoising, aligning, and standardizing the collected full lifecycle data to remove abnormal interference data and ensure data quality. Second, feature extraction extracts features strongly correlated with leakage degree from the processed data, such as frequency domain feature vectors, deviation change rates, and pressure difference features, forming a feature dataset. Third, model training: since severe failure data is difficult to obtain in reality, numerical simulations are constructed based on the fluid-structure interaction physical mechanism of the safety valve. This generates simulated failure data with different operating conditions and leakage degrees. This simulated failure data is then fused with real full lifecycle data to expand the training set (7:3 data ratio). The XGBoost gradient boosting tree algorithm is then used, with the fused feature dataset as input and leakage degree labels as output, to train and generate the simulation model.

[0064] The preset matching benchmark was set to 0.85. By analyzing the matching degree distribution of 1,000 sets of fault data with different leakage levels and simulated feature sequences, it was found that when the matching degree was set to 0.85, cases of atypical signals caused by operating condition fluctuations or complex faults could be filtered out. The obtained cases could match the core characteristics of known fault modes, with only minor differences in environmental interference.

[0065] The matching degree is calculated using a dynamic time warping algorithm, which uses dynamic programming to find the optimal alignment path between two sequences and calculates the cumulative distance after alignment. To obtain a matching degree in the range of 0-1, a min-max normalization method is used to map the sum of distances to the 0-1 interval. The sum of distances is then calculated and divided by the maximum possible cumulative distance under this algorithm. The closer the value is to 1, the better the temporal characteristics match.

[0066] The high degree of consistency refers to the fact that the signal characteristics of the confidence segment, such as frequency distribution, amplitude change, and continuity pattern, are highly consistent with the simulated feature sequence of a certain leakage level output by the preset simulation model. The key feature points and change trends of the two are not significantly different, indicating that the evolution path and manifestation of the current anomaly are consistent with a certain type of leakage fault, and the anomaly type can be identified as that type of leakage fault.

[0067] The process of constructing the wear evolution trend map is as follows: First, a large number of historical failure cases at different wear stages are collected. Each case contains a complete deterioration feature vector, namely the average risk value, variance of fluctuation, duration, and corresponding actual wear degree detection data, such as the wear depth of the sealing surface and the leakage amount. Then, K-means algorithm is used to perform cluster analysis on these case data. The cases are divided into three stages according to the actual wear degree: slight wear, moderate wear, and severe wear. The corresponding cluster center is 3. The vector space boundary of the stage is defined as the range of the 5th quantile to the 95th quantile of the feature values ​​of all cases in that stage, so as to exclude the influence of outliers. Feature weights were determined by analyzing historical case data using principal component analysis. The contribution rates of average risk value, volatility variance, and duration were 40%, 30%, and 30%, respectively, and their weights were set to 0.4, 0.3, and 0.3. Next, for each stage, the distribution range of its degradation feature vectors was statistically analyzed to divide the vector space for each stage. Specifically, the vector space for the slight wear stage corresponds to low average risk value, small volatility variance, and short duration; the vector space for the moderate wear stage corresponds to medium-level feature parameters; and the vector space for the severe wear stage corresponds to high average risk value, large volatility variance, and long duration. Finally, the vector-to-numerical mapping was achieved by weighted summation combined with interval interpolation to form a complete wear evolution trend map.

[0068] The minor wear stage corresponds to a wear depth of 0 to 0.1 mm on the sealing surface and a very small leakage; the moderate wear stage corresponds to a wear depth of 0.1 to 0.3 mm and a moderate leakage; and the severe wear stage corresponds to a wear depth exceeding 0.3 mm and a large leakage. The vector space boundary of each stage is determined by statistically analyzing the mean and standard deviation of each characteristic parameter to ensure the accuracy of the division.

[0069] The specific mapping logic from vector to numerical value is as follows: First, the average risk value, volatility variance, and duration are assigned weights of 0.4, 0.3, and 0.3, respectively; then, the min-max normalization method is used for each component of the degradation feature vector, and the values ​​are calculated according to the minimum and maximum values ​​of the feature vector space in the stage, mapping them to the interval of 0 to 3, corresponding to 3 numerical levels for each stage; finally, the weighted sum is calculated to obtain the wear severity values ​​from 1 to 10.

[0070] It is worth noting that the preset risk range is divided into three levels based on the severity of wear: 1 to 3, 4 to 7, and 8 to 10 refer to the range of wear severity values. This classification is based on historical failure consequence analysis and maintenance cost optimization results. Statistical analysis of the accident rate and maintenance difficulty corresponding to different wear severity levels reveals that when the wear severity value is 1 to 3, the sealing surface wear is slight, leakage is minimal, and the accident rate is less than 0.5%, requiring no emergency treatment; therefore, it is classified as low risk. When the value is 4 to 7, wear is moderate, leakage gradually increases, and the accident rate rises to 5% to 10%, requiring timely inspection and intervention; therefore, it is classified as medium risk. When the value is 8 to 10, wear is severe, leakage is significant, and the accident rate exceeds 30%, easily leading to safety accidents; therefore, it is classified as high risk.

[0071] Different risk levels correspond to different alarm mechanisms. Low risk triggers an early warning, which is only recorded in the system background and a light attention notification is pushed. No downtime is required, and regular maintenance can be performed. Medium risk triggers an inspection alarm, and the system automatically sends an inspection work order to the maintenance personnel, prompting them to complete the equipment inspection and targeted maintenance within 24 hours. High risk triggers an emergency shutdown alarm, and the system immediately issues an audible and visual alarm. At the same time, it links with the industrial control system to automatically cut off relevant pipeline valves, stop equipment operation, prevent leakage from expanding and causing accidents, and ensure the targeted and effective maintenance measures.

[0072] For example, the confidence level segment data is a signal sequence with a mean amplitude of 0.16, a duration of 8 time units, and a variance of 0.002, which is input into a preset simulation model. This model is built based on 5 years of full life cycle failure data of a certain type of safety valve and has learned the feature mapping relationship of three types of leakage: minor, moderate, and severe. After inputting the current data, it outputs the simulated feature sequence corresponding to moderate leakage. The matching degree between the two is calculated to be 0.92, which is greater than the preset matching benchmark of 0.85, indicating that the frequency distribution and amplitude variation pattern of the current anomaly are highly consistent with the known moderate leakage failure mode. Anomaly interval labeling information is generated, covering the entire segment, starting at the 10th time unit and lasting for 8 time units.

[0073] Based on the labeled information, a result set was constructed, and the average risk value (0.16), variance (0.002), and duration (8) were extracted as core features, resulting in degradation feature vectors of 0.16, 0.002, and 8. These vectors were mapped to a pre-defined wear evolution trend map, constructed through clustering of 2000 historical cases. This map has been divided into vector spaces for three wear stages: slight, moderate, and severe. The vector space for the slight stage is: average risk value 0.05 to 0.1, variance (0.0005 to 0.001), and duration (2 to 5); for the moderate stage, it is: average risk value 0.1 to 0.2, variance (0.001 to 0.003), and duration (5 to 10); and for the severe stage, it is: average risk value above 0.2 and variance (0.003). For values ​​above 10 and a duration of 10 or more, the current degradation feature vector falls within the medium wear vector space. A normalized weighted sum is calculated using the weighted average risk value (0.4), fluctuation variance (0.3), and duration (0.3), all using the min-max normalization method. The normalized average risk value is 0.6; the normalized fluctuation variance is 0.002 minus 0.001 divided by 0.003 minus 0.001, resulting in 0.5; the normalized duration is 8 minus 5 divided by 10 minus 5, resulting in 0.6; the weighted sum is 0.6 multiplied by 0.4 plus 0.5 multiplied by 0.3 plus 0.6 multiplied by 0.3, resulting in 0.57. This corresponds to a medium stage value of 4 plus 0.57 multiplied by 3, resulting in 5.71. After rounding, the wear severity value is 5.7. The value falls within the medium-risk range of 4 to 7. The system triggers an inspection alarm, automatically sends an inspection work order to the maintenance personnel, and prompts them to complete the sealing surface inspection and prepare for spare parts replacement within 24 hours, thus obtaining the risk rating result.

[0074] In summary, this invention discloses an intelligent alarm method for safety valve leakage. Through multi-source data fusion, dynamic feature extraction, and adaptive verification, this invention achieves accurate identification of early and weak leakage of safety valves under complex operating conditions, effectively reducing false alarm and missed alarm rates and improving industrial production safety.

[0075] Reference Figure 2 The second embodiment of the present invention provides an intelligent alarm system for safety valve leakage, comprising: The data acquisition and preprocessing module is used to acquire the raw vibration data and raw pressure data of the operating environment of the safety valve, and to perform synchronous processing and noise reduction on the raw vibration data and raw pressure data to obtain vibration data sequence and pressure data sequence. The frequency domain feature extraction module is used to calculate the signal energy distribution and analyze the frequency components based on the vibration data sequence and the pressure data sequence to obtain the frequency domain feature vector; The high-risk signal screening and structuring module is used to perform peak labeling and noise filtering on the frequency domain feature vector. After noise filtering, secondary peaks are extracted as peak components. It is determined whether the peak components exceed the preset peak benchmark. If so, the logical relationship of the high-risk signal is mapped and structured data is constructed. If not, structured data is not constructed and the current anomaly detection process ends. The deviation feature extraction and rate comparison module is used to extract deviation features based on the structured data, calculate the deviation change rate of the deviation features and cross-compare them with historical benchmark data to obtain the comparison difference value. The initial leakage determination module is used to determine whether the increasing trend and preset frequency offset range are satisfied based on the changing trend and frequency offset range of the comparison difference value. If yes, the module outputs the initial leakage indication determination result as existing; otherwise, the module outputs the initial leakage indication determination result as not existing. The comprehensive risk assessment module is used to generate a comprehensive risk index by fusing the deviation change rate and the difference value of the pressure data sequence when the initial signs of leakage are determined to exist. The comprehensive risk index is then compared with historical data to obtain the risk range result. The alarm processing and filtering module is used to generate an alarm signal sequence based on the risk range result, and adaptively smooth the alarm signal sequence to obtain a confidence segment; The risk rating alarm module is used to perform an iterative verification process on the confidence level segment, determine the final leakage risk level of the safety valve, trigger the corresponding alarm, and obtain the risk rating result.

[0076] It should be noted that the intelligent alarm system for safety valve leakage provided in this embodiment of the invention is used to execute all the process steps of the intelligent alarm method for safety valve leakage in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0077] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0078] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for intelligent alarm of safety valve leakage, characterized in that, include: The raw vibration data and raw pressure data of the operating environment of the safety valve are acquired, and the raw vibration data and raw pressure data are synchronously processed and denoised to obtain the vibration data sequence and pressure data sequence. Based on the vibration data sequence and the pressure data sequence, the signal energy distribution is calculated and the frequency components are analyzed to obtain the frequency domain feature vector; Peak labeling and noise filtering are performed on the frequency domain feature vector. After noise filtering, secondary peaks are extracted as peak components. It is determined whether the peak components exceed the preset peak benchmark. If so, the logical relationship of high-risk signals is mapped and structured data is constructed. If not, structured data is not constructed and the anomaly detection process ends. Based on the structured data, deviation features are extracted, the rate of change of deviation features is calculated, and cross-compared with historical benchmark data to obtain the comparison difference value. Based on the changing trend and frequency offset range of the comparison difference value, it is determined whether the increasing trend and the preset frequency offset range are satisfied. If yes, the result of the initial leakage sign determination is that it exists; otherwise, the result of the initial leakage sign determination is that it does not exist. When the initial signs of leakage are determined to exist, a comprehensive risk index is generated by combining the deviation change rate with the difference value of the pressure data sequence. The comprehensive risk index is then compared with historical data to obtain the risk range result. Based on the risk range results, an alarm signal sequence is generated, and the alarm signal sequence is adaptively smoothed to obtain a confidence segment; An iterative verification process is performed on the confidence level segment to determine the final leakage risk level of the safety valve, trigger the corresponding alarm, and obtain the risk rating result.

2. The intelligent alarm method for safety valve leakage according to claim 1, characterized in that, The process involves acquiring raw vibration and pressure data of the safety valve's operating environment, synchronously processing and denoising the raw vibration and pressure data to obtain vibration and pressure data sequences, including: The raw vibration data and raw pressure data of the safety valve's operating environment are synchronously collected by a multi-channel sensor array and the timestamps are recorded to obtain the raw dataset; Based on the original dataset, the original vibration data is decomposed into sub-band coefficients at multiple scales, and the original pressure data is synchronously segmented to obtain vibration sub-band coefficients and pressure sub-band sequences. The time-frequency domain coherence spectra of the vibration sub-band coefficients and the pressure sub-band sequence are calculated by cross wavelet transform, the phase difference sequence is determined, and the coupling feature vector is obtained. The joint energy distribution features are extracted from the coupled feature vectors and the abnormal features are fused to obtain vibration data sequences and pressure data sequences.

3. The intelligent alarm method for safety valve leakage according to claim 1, characterized in that, The step of calculating the signal energy distribution and analyzing the frequency components based on the vibration data sequence and the pressure data sequence to obtain the frequency domain feature vector includes: Based on the vibration data sequence and the pressure data sequence, the signal intensity distribution matrix is ​​obtained by segmenting the data into preset time windows and calculating the root mean square value of each slice. Based on the two-dimensional signal intensity distribution matrix, a dynamic threshold is set to determine the range of signals of interest. Perform frequency domain transformation on the signal range of interest and separate the core component to obtain the fundamental frequency component and the residual sideband component; Based on the fundamental frequency component and the residual sideband component, weak signal feature points are identified and combined to obtain a frequency domain feature vector.

4. The intelligent alarm method for safety valve leakage according to claim 1, characterized in that, The process of peak labeling and noise filtering of the frequency domain feature vector, extracting secondary peaks as peak components after noise filtering, and determining whether the peak components exceed a preset peak benchmark are all performed. If so, the logical relationship of high-risk signals is mapped, and structured data is constructed. If not, structured data is not constructed, and the anomaly detection process ends. This includes: Based on the frequency domain feature vector, the peak position and noise component are extracted and noise filtering is performed to obtain the filtered spectrum features. Based on the filtered spectral features and the peak position, secondary peaks are extracted as peak components; Calculate the amplitude ratio of the peak component relative to the base frequency main peak. If the amplitude ratio exceeds a preset peak reference, it is identified as a high-risk signal source. Construct structured data based on key-value pairs. The structured data includes a logical mapping relationship between the frequency offset, amplitude ratio, and risk level of the high-risk signal source. If the amplitude ratio does not exceed the preset peak reference, it is determined to be a normal signal component, no structured data is constructed, and the anomaly detection process ends.

5. The intelligent alarm method for safety valve leakage according to claim 1, characterized in that, The step of extracting deviation features from the structured data, calculating the rate of change of the deviation features, and cross-comparing them with historical benchmark data to obtain a comparison difference value includes: The amplitude deviation values ​​of the structured data are extracted and sorted by time to obtain the deviation feature sequence; The rate of change of the deviation characteristic sequence is calculated using the sliding window method to obtain the deviation change rate. Based on the rate of change of the deviation, a cross-comparison is performed with historical benchmark data to obtain the comparison difference value.

6. The intelligent alarm method for safety valve leakage according to claim 1, characterized in that, When the initial signs of leakage are determined to exist, a comprehensive risk index is generated by fusing the deviation change rate with the difference value of the pressure data sequence. This comprehensive risk index is then compared with historical data to obtain a risk range result, including: When the initial signs of leakage are determined to be present, a discretization difference operation is performed based on the pressure data sequence to obtain a pressure difference feature sequence. Based on the pressure difference characteristic sequence and the deviation change rate, a weighted summation operation is performed to obtain a comprehensive risk index; Collect historical operating data of the safety valve, calculate the mean and standard deviation of the historical operating data to obtain statistical distribution characteristics, construct a safety boundary based on the statistical distribution characteristics, and obtain a dynamic safety envelope. By comparing the comprehensive risk index with the dynamic safety envelope, if the comprehensive risk index falls outside the range defined by the dynamic safety envelope, the risk range result is obtained.

7. The intelligent alarm method for safety valve leakage according to claim 1, characterized in that, Based on the risk range result, an alarm signal sequence is generated, and the alarm signal sequence is adaptively smoothed to obtain a confidence segment, including: Based on the comprehensive risk indicators and the risk range results, a comprehensive risk data stream is obtained; If the comprehensive risk data stream exceeds the preset security baseline, an alarm signal sequence is generated; The adaptive sliding window width is determined based on the local flip frequency of the alarm signal sequence, and a weighted average processing is performed on the alarm signal sequence based on the adaptive sliding window width to obtain a smooth alarm data stream; Identify the numerical troughs in the smoothed alarm data stream, use the numerical troughs as dividing boundaries, and segment the smoothed alarm data stream to obtain alarm data segments; Calculate the integral of the amplitude and duration of each alarm data segment to obtain a confidence value. If the confidence value is greater than a preset confidence threshold, then output the confidence segment.

8. The intelligent alarm method for safety valve leakage according to claim 1, characterized in that, The iterative verification process performed on the confidence level segment determines the final leakage risk level of the safety valve, triggers the corresponding alarm, and obtains the risk rating result, including: The data from the confidence level segment is input into a preset simulation model to obtain a simulated feature sequence; Calculate the matching degree between the confidence segment and the simulated feature sequence. If the matching degree is greater than a preset matching benchmark, generate abnormal interval labeling information. Based on the abnormal interval labeling information, a result set is constructed, and the average risk value, volatility variance and duration are extracted from the result set as core features to obtain the deterioration feature vector. The degradation feature vector is mapped to a preset wear evolution trend map to obtain a wear severity value; The wear severity value is analyzed based on a preset risk range. When the wear severity value is within the preset risk range, a risk alarm corresponding to the risk level is triggered, and a risk rating result is obtained.

9. A smart alarm system for safety valve leakage, characterized in that, include: The data acquisition and preprocessing module is used to acquire the raw vibration data and raw pressure data of the operating environment of the safety valve, and to perform synchronous processing and noise reduction on the raw vibration data and raw pressure data to obtain vibration data sequence and pressure data sequence. The frequency domain feature extraction module is used to calculate the signal energy distribution and analyze the frequency components based on the vibration data sequence and the pressure data sequence to obtain the frequency domain feature vector; The high-risk signal screening and structuring module is used to perform peak labeling and noise filtering on the frequency domain feature vector. After noise filtering, secondary peaks are extracted as peak components. It is determined whether the peak components exceed the preset peak benchmark. If so, the logical relationship of the high-risk signal is mapped and structured data is constructed. If not, structured data is not constructed and the current anomaly detection process ends. The deviation feature extraction and rate comparison module is used to extract deviation features based on the structured data, calculate the deviation change rate of the deviation features and cross-compare them with historical benchmark data to obtain the comparison difference value. The initial leakage determination module is used to determine whether the increasing trend and preset frequency offset range are satisfied based on the changing trend and frequency offset range of the comparison difference value. If so, the initial leakage sign determination result is output as "existence". If not, the output result for the initial leakage indication determination is that it does not exist; The comprehensive risk assessment module is used to generate a comprehensive risk index by fusing the deviation change rate and the difference value of the pressure data sequence when the initial signs of leakage are determined to exist. The comprehensive risk index is then compared with historical data to obtain the risk range result. The alarm processing and filtering module is used to generate an alarm signal sequence based on the risk range result, and adaptively smooth the alarm signal sequence to obtain a confidence segment; The risk rating alarm module is used to perform an iterative verification process on the confidence level segment, determine the final leakage risk level of the safety valve, trigger the corresponding alarm, and obtain the risk rating result.

Images