A Real-Time Anomaly Diagnosis Method for Hydropower Plant Inspection and Monitoring Data Based on Edge Computing

By performing streaming reconstruction and frequency domain phase analysis on hydropower plant inspection data at edge computing nodes, and combining the covariance matrix condition number to assess anomaly confidence, the problems of delayed diagnostic results and unutilized correlation of multi-source signals in hydropower plant inspection and monitoring systems are solved, enabling timely capture and efficient storage of early anomalies.

CN123115696APending Publication Date: 2026-11-17HUADIAN SICHUAN POWER GENERATION CO LTD WAWUSHAN BRANCH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611544653.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-08
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing hydropower plant inspection and monitoring systems, centralized data processing platforms result in high communication bandwidth consumption, delayed diagnostic results, difficulty in capturing frequency-phase relationship changes caused by early wear or loosening of equipment, and insufficient utilization of multi-source signal correlation structures, leading to limited confidence and discrimination in anomaly assessment.

Method used

On edge computing nodes, multi-source heterogeneous monitoring data are streamed and reconstructed and subjected to frequency domain phase analysis. The abrupt time profile is located by phase difference. The anomaly confidence is evaluated by the condition number of the covariance matrix of vibration and pressure data. Consistent hashing is used to achieve balanced and persistent storage of scoring data.

Benefits of technology

It enables timely capture and hierarchical storage of abnormal signs in hydropower plant inspection and monitoring data, reducing the probability of missed reports and improving the storage scalability and data retrieval efficiency of edge nodes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115696A_ABST
    Figure CN123115696A_ABST
Patent Text Reader

Abstract

This invention discloses a real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing, belonging to the technical field of edge anomaly diagnosis for hydropower plant inspection data. The method deploys an inspection data acquisition queue and reassembles multi-source heterogeneous data streams to generate structured time-series data blocks; it uses Fast Fourier Transform to extract the frequency domain main peak sequence, calculates the phase difference between adjacent frequency points, and locates abrupt change time sections based on the continuous rate of change; it extracts the time-domain signal at the abrupt change point for empirical mode decomposition, extracting the density of the first intrinsic mode function zero-crossing point as the initial anomaly criterion; when the density exceeds a threshold, it calculates the covariance matrix condition number by operating on the covariance of vibration and pressure data, using the condition number logarithmic value as the anomaly confidence score; after mapping to hash codes, it uses consistent hashing to distribute the data to multiple processing units for persistent storage. This invention achieves real-time anomaly diagnosis and distributed storage of hydropower plant inspection data.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of edge anomaly diagnosis technology for hydropower plant inspection data, specifically a real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing. Background Technology

[0002] Hydropower plant inspection and monitoring systems collect heterogeneous data from multiple sources, including vibration, pressure, temperature, and displacement, for equipment operation status monitoring and anomaly detection. Current real-time anomaly diagnosis largely relies on centralized data processing platforms, aggregating inspection data to a central control center for threshold discrimination, statistical trend analysis, or spectrum analysis. However, the large number of edge measuring points, high sampling frequency, communication protocol instability, and data arrival time contribute to jitter. Centralized uploads consume significant communication bandwidth, resulting in noticeable time lags in diagnostic results. Threshold-based discrimination methods, which judge amplitude exceeding limits at single measuring points, fail to reflect early anomalies caused by high-frequency transient components and phase changes. Frequency domain analysis typically extracts only the peak amplitude spectrum, ignoring the continuous change in phase difference between adjacent frequency points over time, leading to insufficient accuracy in locating abrupt changes.

[0003] When early wear or loosening occurs in localized areas of equipment, the signal does not immediately manifest as amplitude exceeding limits. Instead, the phase relationship between frequencies changes, and the zero-crossing density of high-frequency intrinsic mode components becomes abnormal within a short time window, which is difficult to capture effectively using conventional methods. Existing anomaly confidence assessments often employ single-parameter thresholds or simple feature fusion, failing to fully exploit the correlation structure changes between vibration and pressure signals, resulting in limited anomaly scoring discrimination. Edge computing nodes possess multiple processing units, and conventional storage methods do not consider the balanced distribution of scoring data, leading to significant data redistribution overhead when nodes are expanded or processing units are changed. Further improvements are needed in real-time location of frequency domain abrupt change time sections at the edge and the formation of initial criteria, as well as in quantifying and reasonably persisting anomaly confidence using the correlation structure of multi-source signals. Summary of the Invention

[0004] To address the issues of insufficient real-time processing, inadequate utilization of phase information, and ineffective quantification of the correlation structure of multi-source signals in existing real-time anomaly diagnosis of hydropower plant inspection and monitoring data, this invention aims to provide a real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing. This method performs streaming reconstruction and frequency domain phase analysis on multi-source heterogeneous monitoring data at edge computing nodes to locate abrupt time segments. The covariance matrix condition number of vibration and pressure data is used as the anomaly confidence score, and consistent hashing is used to achieve balanced persistence of the score data.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing, comprising the following steps: deploying an inspection data acquisition queue on an edge computing node; performing streaming reconstruction of multi-source heterogeneous monitoring data according to the timestamp and source device identifier of the data packets to generate structured time-series data blocks; performing fast Fourier transform on the structured time-series data blocks to extract the frequency domain main peak sequence; calculating the phase difference between adjacent frequency points in the main peak sequence; locating abrupt change time segments in the frequency domain based on the continuous rate of change of the phase difference; extracting time-domain signal segments at the abrupt change time segments; and analyzing the time-domain signal segments at the abrupt change time segments. Empirical Mode Decomposition (EMD) is performed on the signal segment to obtain the intrinsic mode function (EMF) set. The zero-crossing density of the first EMF in the EMF set is extracted as the initial anomaly criterion. When the zero-crossing density exceeds a preset density threshold, the vibration data and pressure data in the structured time-series data block are covariance calculated to obtain the covariance matrix. The condition number of the covariance matrix is ​​then calculated, and the logarithm of the condition number is used as the anomaly confidence score. The anomaly confidence score is mapped to a hash code, and a consistent hashing algorithm is used to distribute the hash code to multiple processing units on the edge computing node. Each processing unit then performs persistent storage of the anomaly confidence score. This method completes data reconstruction, frequency domain localization, time domain verification, and confidence scoring at the edge, enabling timely capture and hierarchical storage of anomalies in hydropower plant inspection and monitoring data.

[0006] Preferably, when generating structured time-series data blocks, an independent data acquisition queue is created in the memory of the edge computing node for each source device identifier. When a data packet arrives, its timestamp and source device identifier are extracted, and the data packet is appended to the tail of the corresponding data acquisition queue according to the source device identifier. Data packets are taken from the head of each data acquisition queue, and the timestamp of the taken data packet is compared with the current system time. Data packets with timestamp lag exceeding the allowable jitter are filtered out, marked as delayed data packets, and removed from the queue. The remaining data packets are arranged in ascending order of timestamp, and data packets with the same timestamp are grouped together. The sensor values ​​of all data packets in the same group are concatenated into a row vector according to the preset measurement point order to generate structured time-series data blocks. By eliminating delayed data and aligning timestamps, the multi-source data maintains time consistency before entering the frequency domain analysis, reducing the interference of time misalignment on subsequent abrupt change detection.

[0007] Preferably, when performing Fast Fourier Transform (FFT) on structured time-series data blocks and locating abrupt time segments, the structured time-series data blocks are divided into multiple subframes with a fixed time window length. FFT is performed on each subframe to obtain its frequency domain amplitude sequence and frequency domain phase sequence. Frequency points corresponding to local maxima are extracted from the frequency domain amplitude sequences of each subframe to form the main peak sequence of that subframe. The frequency points in the main peak sequence are arranged in ascending order of frequency. The phase difference between two adjacent frequency points is calculated to obtain a phase difference sequence. The phase difference sequence of the current subframe is then differentially analyzed point-by-point with the phase difference sequence of the previous subframe to obtain a phase difference change rate sequence. When the absolute values ​​of three or more consecutive values ​​in the phase difference change rate sequence exceed a set threshold, the start time of the corresponding subframe is recorded as the abrupt time segment. By judging continuous exceedances of the phase difference change rate, the time position of abrupt changes in the frequency domain main peak structure can be accurately captured, avoiding misjudgments caused by single-point phase fluctuations.

[0008] Preferably, when extracting a time-domain signal segment at the abrupt change time section and performing empirical mode decomposition (EMD), the time-domain signal segment is extracted by extending sampling points forward and backward by a preset length, centered on the abrupt change time section. EMD is then performed on this time-domain signal segment, extracting a finite number of intrinsic mode functions (EMFs) through a screening process. These EMFs are arranged in descending order of frequency to form a set. The first EMF in the set is selected, and all sampling points of this EMF are iterated through. The number of times the product of the amplitudes of two adjacent sampling points is negative is counted to obtain the total number of zero-crossings. The total number of zero-crossings is divided by the duration of the time-domain signal segment to obtain the zero-crossing density. The zero-crossing density is compared with a preset density threshold. If the zero-crossing density exceeds the preset density threshold, it is determined that there is an initial abnormality at the abrupt change time section. The first EMF centrally reflects the high-frequency transient components in the signal, and its zero-crossing density can serve as an initial criterion for anomalies, exhibiting good sensitivity even under low signal-to-noise ratio conditions.

[0009] Preferably, when the zero-crossing density exceeds a preset density threshold, vibration amplitude sequences and pressure amplitude sequences aligned with the time of abrupt change cross-sections are extracted from the structured time-series data block. These sequences are then paired one-to-one according to sampling time to form a two-dimensional observation vector sequence. The mean vector of the two-dimensional observation vector sequence is calculated, and the mean vector is subtracted from each two-dimensional observation vector to obtain a decentralized vector sequence. The decentralized vector sequence is then multiplied by its transpose and summed, and divided by the total number of observation vectors minus one to obtain the covariance matrix. Singular value decomposition is performed on the covariance matrix to extract the maximum and minimum singular values. The maximum singular value is divided by the minimum singular value to obtain the condition number. The logarithm of the condition number, with the natural constant as the base, is taken to obtain the anomaly confidence score. The covariance condition number of the vibration and pressure data can reflect the degree of coupling anomaly between unit vibration and pressure, and the logarithmic score maintains a stable quantitative expression even when the condition number varies widely.

[0010] Preferably, when performing hash mapping and persistent storage of the anomaly confidence score, the anomaly confidence score is multiplied by a preset scaling factor, and the integer part is converted into a fixed-length binary string. Cyclic redundancy check (CRC) is performed on this binary string, and a check bit is appended to the end of the binary string to form a hash code. The unique identifiers of all processing units in the edge computing node are obtained, and the MD5 value of each unique identifier is calculated. The MD5 value is mapped to a position on the hash ring, and the MD5 value of the hash code is calculated and mapped to a position on the hash ring. The processing unit identifier closest to the hash code mapping position is found clockwise on the hash ring, and the anomaly confidence score and its corresponding hash code are sent to that processing unit, which then writes the data to its local storage medium. This consistent hashing allocation mechanism ensures a balanced distribution of anomaly score data across different processing units, reduces local write hotspots, and improves the throughput of persistent storage on edge nodes.

[0011] Further preferably, the phase difference sequence of the main peak sequence is obtained as follows: Adjacent frequency points are sequentially extracted from the main peak sequence in ascending order of frequency. The phase values ​​of adjacent frequency points are extracted from the frequency domain phase sequence. The phase difference value is obtained by subtracting the previous phase value from the subsequent phase value. This process is repeated for all adjacent frequency points to form the phase difference sequence. For each phase difference value in the phase difference sequence, the difference between it and the phase difference value at the same index in the phase difference sequence of the previous subframe is calculated to obtain a point-by-point difference value. The absolute values ​​of all point-by-point difference values ​​are taken and averaged to obtain the average phase difference change rate of the subframe. A weighted moving average is then applied to the average phase difference change rate and the phase difference change rate of the previous subframe to obtain a smoothed phase difference change rate sequence. Through averaging and moving average processing, the phase difference change rate sequence is smoother, reducing random fluctuations caused by local phase estimation errors.

[0012] A further preferred method for empirical mode decomposition (EMD) screening includes: taking the sampling point index corresponding to the abrupt change time section as the center, taking a first preset number of sampling points forward and a second preset number of sampling points backward, and extracting a time-domain signal segment with a total length equal to the sum of the first and second preset numbers plus one; identifying all local maxima and local minima in the time-domain signal segment, constructing upper and lower envelopes respectively using cubic spline interpolation, and calculating the mean envelope of the upper and lower envelopes; subtracting the mean envelope from the time-domain signal segment to obtain candidate intrinsic mode functions (EMFs); determining whether the absolute value of the difference between the number of extrema and the number of zero-crossings of the candidate EMFs is less than or equal to one; if so, outputting the candidate EMF as the first EMF; otherwise, repeating the screening process based on the candidate EMFs. This screening method can adaptively separate high-frequency components in the signal and is suitable for processing non-stationary transient fluctuations in hydropower plant inspection and monitoring data.

[0013] Further preferably, the two-dimensional observation vector sequence is formed as follows: Based on the absolute time of the abrupt change time section, all data rows within the range of the third preset number of sampling points before the abrupt change time section to the fourth preset number of sampling points after the section are selected from the structured time-series data block; from each selected data row, the amplitude of the vibration sensor is extracted as the first component, and the amplitude of the pressure sensor is extracted as the second component. Using the sampling time as the row index, the first and second components are combined into a two-dimensional observation vector. All two-dimensional observation vectors are arranged in ascending order by sampling time to form a two-dimensional observation vector sequence. Using a window extraction method aligned with the abrupt change time section ensures that the vibration and pressure data used for covariance calculation are synchronized with the abrupt change location, improving the accuracy of the anomaly confidence score.

[0014] A further preferred embodiment of the anomaly confidence score calculation includes: performing singular value decomposition on the two-dimensional covariance matrix to obtain the left singular matrix, the singular value diagonal matrix, and the transpose of the right singular matrix; extracting all singular values ​​from the diagonal of the singular value diagonal matrix; comparing the magnitudes of all singular values ​​to identify the largest and smallest singular values; dividing the largest singular value by the smallest singular value, using the quotient as the condition number of the covariance matrix; adding a preset regularization factor to the condition number; and taking the logarithm with the natural constant as the base. The result of this logarithmic operation is used as the anomaly confidence score at the time point of the abrupt change. The introduction of the regularization factor avoids the condition number becoming too large or incalculable when the singular values ​​are close to zero, thus improving the stability of the score.

[0015] The technical effects and advantages provided by the present invention in the above technical solution are as follows: After performing a Fast Fourier Transform on the structured time-series data block, the main peak sequence in the frequency domain is extracted, and the phase difference between adjacent frequency points is calculated. The abrupt change time segment in the frequency domain is located based on the continuous rate of change of the phase difference. The continuous rate of change of the phase difference reflects the non-stationary changes in the phase structure between frequency components, and is more sensitive to early anomalies where the amplitude has not yet significantly exceeded its limit, providing the specific time location of the abrupt change in the frequency domain. A time-domain signal segment is extracted at the abrupt change time segment for empirical mode decomposition, and the zero-crossing density of the first intrinsic mode function is extracted as the initial criterion for anomaly detection. The first intrinsic mode function corresponds to the highest-frequency oscillation component in the signal, and the zero-crossing density characterizes the oscillation speed of the waveform within a short time window. When a transient change occurs in the local state of the device, this density becomes abnormal, enabling the identification of the abrupt change time segment in the time domain. This method no longer relies on a single amplitude threshold, but integrates frequency domain phase changes and high-frequency zero-crossing information in the time domain, making the initial judgment of early anomaly symptoms more timely and reducing the probability of missed detection.

[0016] When the zero-crossing density exceeds a preset threshold, vibration and pressure amplitude sequences aligned with the time of abrupt change are extracted to form a two-dimensional observation vector sequence. This sequence is then centered and subjected to outer product operations to generate a covariance matrix. Singular value decomposition (SVD) is performed on the covariance matrix, and the ratio of the maximum to minimum singular value is used as the condition number. The natural logarithm is then taken to obtain the anomaly confidence score. Vibration and pressure data exhibit a relatively stable correlation structure under normal operating conditions. When equipment malfunctions, this correlation structure changes, increasing the condition number of the covariance matrix. Logarithmic transformation can compress the numerical span and characterize the degree of anomaly, forming a continuous confidence score rather than a simple limit-crossing indicator. This score is mapped to a hash code, and a consistent hashing algorithm is used to distribute the hash code to multiple processing units on the edge computing node. The corresponding processing unit performs persistent storage. The score data is evenly distributed across multiple processing units. When a processing unit is added or removed, only a small amount of adjacent hash ring data needs to be migrated. Edge storage exhibits good scalability and data locality, resulting in high efficiency for writing and subsequent retrieval of anomaly records. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0018] Figure 1 This is a flowchart of a real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing. Figure 2 This is a flowchart for calculating anomaly confidence scores based on the correlation between vibration and pressure data. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0020] See Figure 1 This invention provides a real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing, comprising the following steps: A data collection queue for inspection is deployed on the edge computing node. Multi-source heterogeneous monitoring data enters the edge computing node in the form of data packets, which carry timestamps and source device identifiers. Based on the timestamps and source device identifiers of the data packets, the multi-source heterogeneous monitoring data is reassembled in a streaming manner to generate structured time-series data blocks.

[0021] Perform a Fast Fourier Transform on the structured time-series data block, extract the frequency domain main peak sequence from the Fast Fourier Transform result, calculate the phase difference between adjacent frequency points in the main peak sequence, and locate the abrupt time section in the frequency domain based on the continuous rate of change of the phase difference.

[0022] A time-domain signal segment is extracted at the abrupt change time section, and empirical mode decomposition is performed on the time-domain signal segment to obtain the intrinsic mode function set. The zero-crossing density of the first intrinsic mode function in the intrinsic mode function set is extracted as the initial criterion for anomaly.

[0023] When the density at the zero-crossing point exceeds the preset density threshold, the vibration data and pressure data in the structured time-series data block are subjected to covariance calculation to obtain the covariance matrix. Then, the condition number of the covariance matrix is ​​calculated, and the logarithm of the condition number is used as the anomaly confidence score.

[0024] The anomaly confidence score is mapped to a hash code, and the hash code is distributed to multiple processing units on the edge computing node using a consistent hashing algorithm. Each processing unit then performs the persistent storage of the anomaly confidence score.

[0025] In practical implementation, edge computing nodes allocate a data acquisition queue management area in memory. This area contains a mapping relationship between source device identifiers and data acquisition queue structures. The source device identifier can be the equipment code, equipment serial number, or monitoring point identifier from a hydropower plant inspection and monitoring system. The edge computing node reads the header field of the inspection data packet, which contains a timestamp and the source device identifier. The data acquisition queue management area searches for the corresponding data acquisition queue based on the source device identifier. If the corresponding data acquisition queue does not exist, the data acquisition queue management area creates a new data acquisition queue and writes the mapping relationship between the source device identifier and the new data acquisition queue into the data acquisition queue management area. The data acquisition queue is implemented using a circular buffer. The circular buffer pre-allocates a fixed-size memory block, and inspection data packets are appended to the end of the circular buffer. When the circular buffer is full, it overwrites the oldest inspection data packet.

[0026] In some embodiments, edge computing nodes acquire inspection data packets via interruption or polling when they arrive. After the inspection data packet enters the edge computing node, the node parses it to obtain a timestamp and source device identifier. Using the source device identifier as the key, the edge computing node searches for the corresponding data acquisition queue in the data acquisition queue management area. The edge computing node then appends the entire inspection data packet to the end of the corresponding data acquisition queue, maintaining the original byte order within the queue.

[0027] In practice, edge computing nodes retrieve inspection data packets from the head of each data acquisition queue. The retrieval process is executed periodically, with the edge computing node triggering the retrieval action according to a preset inspection data processing cycle. The edge computing node reads the timestamp of the retrieved inspection data packet and its current system time. The edge computing node calculates the timestamp lag, which is equal to its current system time minus the timestamp of the retrieved inspection data packet. The edge computing node compares the timestamp lag with the allowable jitter. The allowable jitter is a pre-configured time threshold, determined by adding the difference between the historical maximum and minimum network transmission delay between the edge computing node and the inspection data source to the edge computing node's local clock drift. The edge computing node filters out inspection data packets with timestamp lag exceeding the allowable jitter, marking them as delayed data packets and removing them from the data acquisition queue. Delayed data packets removed from the data acquisition queue do not participate in the subsequent structured time-series data block generation process.

[0028] In the remaining inspection data packets, the edge computing nodes arrange the remaining inspection data packets in ascending order of timestamp. The sorting uses either heap sort or merge sort. The sorted remaining inspection data packets form an ascending timestamp sequence. The edge computing nodes group inspection data packets with the same timestamp together, grouping the remaining inspection data packets with the same timestamp into the same group. The edge computing nodes read the preset measurement point order, which is a list of ordered measurement point identifiers loaded from the configuration file. For each group of inspection data packets with the same timestamp, the edge computing nodes extract the sensor values ​​of the corresponding measurement points from each inspection data packet according to the preset measurement point order, and concatenate the sensor values ​​of all inspection data packets in the same group into a row vector according to the preset measurement point order. The edge computing nodes combine the timestamps and row vectors to generate structured time-series data blocks. The edge computing nodes store the structured time-series data blocks in a two-dimensional array structure in memory. Each row of the two-dimensional array structure corresponds to a timestamp, the first column of the two-dimensional array structure stores the timestamp, and the remaining columns store the sensor values ​​from the row vectors.

[0029] In practice, edge computing nodes read structured time-series data blocks and determine a fixed time window length. The fixed time window length is read from a configuration file by the edge computing node, and its unit is the number of sampling points. The edge computing node divides the structured time-series data block into multiple subframes according to the fixed time window length. During the division process, the edge computing node starts with the sampling point corresponding to the beginning timestamp of the structured time-series data block and sequentially extracts sampling points corresponding to the fixed time window length in the direction of increasing time, forming a subframe. Adjacent subframes do not overlap. For the last subframe, if the number of remaining sampling points is less than the fixed time window length, the edge computing node adds zero-value sampling points to the end of the remaining sampling points to ensure that the number of sampling points in the last subframe reaches the fixed time window length.

[0030] In some embodiments, the edge computing nodes perform a Fast Fourier Transform (FFT) on each subframe. The FFT uses a radix-2 decimation-time algorithm. When the number of sampling points in a subframe is not a power of 2, the edge computing nodes add zero-value sampling points at the end of the subframe until the number of sampling points is a power of 2. The FFT outputs a sequence of complex numbers in the frequency domain, where each complex number corresponds to a frequency point. The edge computing nodes calculate the modulus of each complex number in the frequency domain sequence to obtain a frequency domain amplitude sequence. The edge computing nodes also calculate the argument of each complex number in the frequency domain sequence to obtain a frequency domain phase sequence. The lengths of the frequency domain amplitude and phase sequences are the same as the number of points in the FFT.

[0031] Optionally, the edge computing nodes extract the frequency points corresponding to local maxima from the frequency domain amplitude sequence of each subframe. For a frequency point in the frequency domain amplitude sequence, if the amplitude of the frequency point is greater than the amplitude of the immediately preceding frequency point and the amplitude of the immediately following frequency point, the frequency point is identified as a local maximum. The edge computing nodes traverse all frequency points in the frequency domain amplitude sequence and extract the frequency points corresponding to all local maxima. The edge computing nodes form a main peak sequence by combining all the frequency points corresponding to the local maxima extracted from the same subframe. The edge computing nodes arrange the frequency points in the main peak sequence in ascending order of frequency. The frequency value of a frequency point is obtained by multiplying the frequency point index by the frequency resolution.

[0032] It can be understood that the edge computing node extracts the r-th and (r+1)-th frequency points sequentially in ascending frequency order from the arranged main peak sequence. The edge computing node extracts the phase value of the r-th frequency point from the frequency domain phase sequence, and also extracts the phase value of the (r+1)-th frequency point. The edge computing node subtracts the phase value of the r-th frequency point from the phase value of the (r+1)-th frequency point to obtain the r-th phase difference value. When the r-th phase difference value is less than... At that time, the edge computing node adds the r-th phase difference value. When the r-th phase difference is greater than At that time, the edge computing node subtracts the r-th phase difference value. The edge computing nodes traverse all adjacent frequency points in the main peak sequence to obtain the phase difference sequence of the current subframe.

[0033] In practice, for each phase difference value in the current subframe phase difference sequence, the edge computing node reads the phase difference value at the same index in the previous subframe phase difference sequence. When the length of the current subframe phase difference sequence differs from the length of the previous subframe phase difference sequence, the edge computing node uses the shorter of the two phase difference sequences as the length for point-by-point difference processing. The edge computing node subtracts the phase difference value at the same index in the previous subframe phase difference sequence from the phase difference value in the current subframe phase difference sequence to obtain the point-by-point difference value. The edge computing node performs the same operation on each phase difference value to obtain a sequence of point-by-point difference values.

[0034] Optionally, the edge computing node takes the absolute value of all point-by-point difference values, then calculates the sum of all absolute values, and divides by the total number of point-by-point difference values ​​to obtain the average phase difference change rate of the current subframe. The edge computing node uses the average phase difference change rate of the current subframe as the phase difference change rate of the current subframe. The edge computing node reads the smoothed phase difference change rate of the previous subframe, and performs a weighted moving average between the average phase difference change rate of the current subframe and the smoothed phase difference change rate of the previous subframe to obtain the smoothed phase difference change rate of the current subframe. The expression for the weighted moving average is: in, Indicates the first The smoothed phase difference change rate of each subframe; The subframe number indicates the position of the currently processed subframe in the time series; Indicates the first Average phase difference change rate of each subframe; Indicates the first The smoothed phase difference change rate of each subframe; Indicates the weighting coefficient. Weighting coefficients The method for obtaining the weighting coefficients is as follows: the edge computing node reads the weighting coefficients from the configuration parameters. Weighting coefficients in configuration parameters The weighting coefficients are set based on the magnitude of the high-frequency fluctuation components in the phase difference change rate sequence. When the high-frequency fluctuation components are large, the weighting coefficients are adjusted accordingly. Pick to The values ​​between; when the high-frequency fluctuation component is small, the weighting coefficient Pick to The values ​​between [a certain value]. In this embodiment, the weighting coefficients [are used]. Set as Set weighting coefficients for The basis for this is that the phase difference change rate sequence in the hydropower plant inspection and monitoring data contains non-stationary abrupt changes and random noise. The weighting coefficients can suppress random noise while preserving the trend of sudden changes.

[0035] In some embodiments, edge computing nodes generate a smoothed phase difference rate of change sequence in chronological order. The edge computing nodes sequentially determine the absolute value of each value in the smoothed phase difference rate of change sequence. When an edge computing node detects that the absolute values ​​of three or more consecutive values ​​in the smoothed phase difference rate of change sequence all exceed a set threshold, the edge computing node records the start time of the subframe corresponding to the first value among these consecutive values ​​as a sudden change time section.

[0036] The threshold is obtained as follows: Edge computing nodes collect smoothed historical samples of the phase difference change rate sequence under normal operating conditions. The mean and standard deviation of the absolute values ​​of these smoothed historical samples are calculated. The threshold is set to the mean of the absolute values ​​of these smoothed historical samples plus three times the standard deviation. In this embodiment, the threshold is set to... Set the threshold to The basis for this is that, under normal operating conditions, the mean of the absolute values ​​of the smoothed phase difference change rate sequence is close to... The standard deviation is close to Three standard deviations plus the mean is Therefore, the threshold is set to .

[0037] In practical implementation, edge computing nodes acquire abrupt change time segments and align the time of these segments with the timestamp sequence in the structured time-series data block. When the time of the abrupt change time segment perfectly matches a sampling time in the structured time-series data block, the edge computing node uses the sampling point index corresponding to that sampling time as the center sampling point index. When the time of the abrupt change time segment does not perfectly match any sampling time in the structured time-series data block, the edge computing node selects the sampling point index corresponding to the sampling time closest to the abrupt change time segment as the center sampling point index. The edge computing node reads a first preset quantity and a second preset quantity from the configuration parameters. The first preset quantity represents the number of sampling points extending from the center sampling point index in the direction of decreasing time, and the second preset quantity represents the number of sampling points extending from the center sampling point index in the direction of increasing time. The first and second preset quantities are set based on the following: the time-domain signal segment needs to completely cover the duration of the abrupt transient process, and the total number of sampling points in the time-domain signal segment needs to meet the minimum requirement of Empirical Mode Decomposition (EMD) for the number of sampling points, which is 64 sampling points. The edge computing node uses the index of the center sampling point as a reference, takes a first preset number of sampling points forward and a second preset number of sampling points backward, and extracts a time-domain signal segment with a total length equal to the sum of the first and second preset numbers plus one. The time-domain signal segment includes the center sampling point, all sampling points before the center sampling point, and all sampling points after the center sampling point.

[0038] In some embodiments, edge computing nodes perform an empirical mode decomposition (EMD) screening process on time-domain signal segments. The edge computing nodes identify all local maxima and local minima within the time-domain signal segment. For a sample point in the time-domain signal segment, the edge computing node identifies it as a local maximum when its amplitude is greater than that of the preceding and following sample points. For a sample point in the time-domain signal segment, the edge computing node identifies it as a local minimum when its amplitude is less than that of the preceding and following sample points. The edge computing nodes construct an upper envelope using cubic spline interpolation, which passes through all local maxima. The cubic spline interpolation uses a cubic polynomial between adjacent local maxima, and the cubic polynomial satisfies second-derivative continuity at the local maxima. Edge computing nodes construct a lower envelope using cubic spline interpolation. This lower envelope passes through all local minima. Cubic spline interpolation uses a cubic polynomial between adjacent local minima, and this polynomial satisfies second-derivative continuity at the local minima. Edge computing nodes then calculate the mean envelope of the upper and lower envelopes. The magnitude of the mean envelope at each sampling point is equal to the sum of the magnitudes of the upper and lower envelopes at the same sampling point, divided by two.

[0039] Optionally, the edge computing node subtracts the mean envelope from the time-domain signal segment to obtain candidate intrinsic mode functions (IMFs). The edge computing node iterates through all sampling points of the candidate IMFs, counting the number of extreme points by identifying local maxima and minima. The edge computing node iterates through all adjacent sampling points of the candidate IMFs, counting the number of times the product of the amplitudes of two adjacent sampling points is negative, thus obtaining the number of zero-crossing points of the candidate IMFs. The edge computing node calculates the absolute value of the difference between the number of extreme points and the number of zero-crossing points of the candidate IMFs. When the absolute value of the difference between the number of extreme points and the number of zero-crossing points of the candidate IMFs is less than or equal to one, the edge computing node outputs the candidate IMF as the first IMF. When the absolute value of the difference between the number of extrema and the number of zero-crossings of a candidate intrinsic mode function (EMF) is greater than one, the edge computing node treats the candidate EMF as a new time-domain signal segment and repeatedly performs the process of identifying local maxima and minima, constructing the upper envelope, constructing the lower envelope, calculating the mean envelope, and subtracting the mean envelope to obtain a new candidate EMF. The edge computing node is set to a maximum of 10 iterations. When the number of iterations reaches 10, regardless of whether the absolute value of the difference between the number of extrema and the number of zero-crossings of the candidate EMF is less than or equal to one, the edge computing node outputs the current candidate EMF as the first EMF. The reason for setting the maximum number of iterations to 10 is that excessive iterations in empirical mode decomposition can lead to over-smoothing of the EMF and loss of detailed information; 10 iterations preserve anomalous features while avoiding over-decomposition.

[0040] It can be understood that edge computing nodes extract a finite number of intrinsic mode functions (EMFs) through a filtering process. After obtaining the first EMF, the edge computing node subtracts the first EMF from the original time-domain signal segment to obtain the remaining signal. The edge computing node repeats the Empirical Mode Decomposition (EMD) filtering process on the remaining signal to obtain the second EMF. The edge computing node repeats the above process until the remaining signal becomes a monotonic function or a constant, at which point the extraction stops, resulting in a finite number of EMFs. The EMFs extracted by the EMF filtering process are arranged in descending order of frequency, with the first EMF being the most frequent. The edge computing node retrieves the first EMF from the set of EMFs, iterates through all sampling points of the first EMF, and counts the number of times the product of the amplitudes of two adjacent sampling points of the first EMF is negative, thus obtaining the total number of zero-crossings.

[0041] In practical implementation, the edge computing nodes calculate the zero-crossing density, and the formula for calculating the zero-crossing density is as follows: in, This represents the density at the zero-crossing point, expressed in units of particles per second. This represents the total number of zero-crossings, and its value is the number of times the product of the amplitudes of two adjacent sampling points in the first eigenmode function is negative. This represents the duration of a time-domain signal segment, measured in seconds. The acquisition method is as follows: the edge computing node reads the total number of sampling points of the time domain signal segment. The total number of sampling points of the time domain signal segment is equal to the sum of the first preset number and the second preset number plus one. The edge computing node divides the total number of sampling points of the time domain signal segment by the sampling rate to obtain the duration of the time domain signal segment.

[0042] The edge computing node reads a preset density threshold from the configuration parameters. The preset density threshold is set based on the following: under normal operating conditions, the zero-crossing density of the intrinsic mode function of the hydropower unit's vibration signal remains within a stable range; the preset density threshold is the upper limit of the zero-crossing density under normal operating conditions. The edge computing node compares the zero-crossing density with the preset density threshold. When the zero-crossing density is greater than the preset density threshold, the edge computing node determines that there is an abnormal initial sign at the abrupt change time segment. When the zero-crossing density is less than or equal to the preset density threshold, the edge computing node determines that there is no abnormal initial sign at the abrupt change time segment.

[0043] In specific implementation, please refer to Figure 2The edge computing node reads the absolute time of the abrupt change time segment. The absolute time of the abrupt change time segment is the recorded start time of the subframe used as the abrupt change time segment. The edge computing node reads the third and fourth preset quantities from the configuration parameters. The third preset quantity represents the number of sampling points sampled from the absolute time of the abrupt change time segment in the direction of decreasing time, and the fourth preset quantity represents the number of sampling points sampled from the absolute time of the abrupt change time segment in the direction of increasing time. The third and fourth preset quantities are set based on the following: the two-dimensional observation vector sequence needs to cover the correlation range of vibration and pressure data during the anomaly development process; simultaneously, the total number of observation vectors in the two-dimensional observation vector sequence needs to meet the minimum requirement of the singular value decomposition of the covariance matrix for the total number of observation vectors, which is that the total number of observation vectors is not less than 2. Using the absolute time of the abrupt change time segment as a reference, the edge computing node selects all data rows from the structured time-series data block within the range of the third preset quantity of sampling points before the absolute time of the abrupt change time segment to the fourth preset quantity of sampling points after the absolute time of the abrupt change time segment. During data row filtering, edge computing nodes compare the sampling time of each data row in the structured time-series data block with the filtering range. Data rows are included in the filtering range when their sampling time equals the absolute time of the abrupt change timeline. Data rows are also included in the filtering range when their sampling time falls between the times corresponding to the third preset number of sampling points before the absolute time of the abrupt change timeline and the times corresponding to the fourth preset number of sampling points after the absolute time of the abrupt change timeline. The filtering range includes the sampling times at both ends.

[0044] In some embodiments, the edge computing node extracts the amplitude of the vibration sensor from each selected data row, and the extracted vibration sensor amplitude is used as the first component. The edge computing node also extracts the amplitude of the pressure sensor from each selected data row, and the extracted pressure sensor amplitude is used as the second component. Both the vibration sensor amplitude and the pressure sensor amplitude are sensor values ​​from the corresponding data rows in the structured time-series data block. Using the sampling time as the row index, the edge computing node combines the first and second components of each data row into a two-dimensional observation vector. The two-dimensional observation vector is a column vector containing two components; the first row elements are the first component, and the second row elements are the second component. The edge computing node arranges all two-dimensional observation vectors in ascending order of sampling time, forming a two-dimensional observation vector sequence. The order of the two-dimensional observation vectors in the sequence is consistent with the ascending order of sampling time of the data rows in the structured time-series data block.

[0045] Optionally, the edge computing node calculates the mean vector of the two-dimensional observation vector sequence. The mean vector is a two-dimensional column vector obtained by averaging all two-dimensional observation vectors in the sequence according to their components. The edge computing node subtracts the mean vector from each two-dimensional observation vector in the sequence to obtain a decentralized vector. The first row of the decentralized vector is equal to the first component of the original two-dimensional observation vector minus the first row of the mean vector, and the second row of the decentralized vector is equal to the second component of the original two-dimensional observation vector minus the second row of the mean vector. The edge computing node performs an outer product operation on each decentralized vector in the sequence with its transpose to obtain an outer product matrix. The outer product matrix is ​​a 2x2 square matrix. The edge computing node sums all the outer product matrices to obtain the sum of the outer product matrices. The edge computing node divides the sum of the outer product matrices by the total number of observation vectors minus one to obtain the covariance matrix. The total number of observation vectors is equal to the number of two-dimensional observation vectors in the sequence.

[0046] Understanding this process, edge computing nodes perform singular value decomposition (SVD) on the covariance matrix. SVD decomposes the covariance matrix into a left singular matrix, a singular value diagonal matrix, and the transpose of the right singular matrix. The left singular matrix is ​​a 2x2 orthogonal matrix, the singular value diagonal matrix is ​​a 2x2 diagonal matrix, and the transpose of the right singular matrix is ​​also a 2x2 orthogonal matrix. The diagonal elements of the singular value diagonal matrix are the singular values ​​of the covariance matrix. Edge computing nodes extract all singular values ​​from the diagonal of the singular value diagonal matrix. The edge computing nodes then compare the magnitudes of all extracted singular values, identifying the largest and smallest singular values.

[0047] In practice, the edge computing node divides the largest singular value by the smallest singular value, and the quotient is used as the condition number of the covariance matrix. The condition number reflects the numerical stability of the covariance matrix; the larger the condition number, the less stable the correlation between vibration and pressure data. The edge computing node reads a preset regularization factor. This preset regularization factor is used to prevent the logarithmic result from approaching negative infinity due to an excessively small condition number. The preset regularization factor is set based on the following: under normal operating conditions in a hydropower plant, the condition number of the covariance matrix is ​​usually in the range of 1 to 10. When the condition number is close to 1, the impact of the regularization factor on the anomaly confidence score needs to be less than the perceptual resolution of the anomaly confidence score. The perceptual resolution is set to 0.001, and the regularization factor is set to 0.001. The reason for setting the regularization factor to 0.001 is that when the regularization factor is less than or equal to the perception resolution, the regularization factor will not mask the difference in the number of conditions between normal and abnormal operating conditions. At the same time, when the number of conditions is close to zero, the input value of the logarithmic operation is 0.001, the result of the logarithmic operation is -6.907, and the anomaly confidence score is still a finite value.

[0048] After adding a preset small amount of regularization to the condition number of the covariance matrix of the edge computing nodes, the logarithm with base to the natural constant is taken, and the result of the logarithmic operation is used as the anomaly confidence score at the time section of the abrupt change. The formula for calculating the anomaly confidence score is: in, Indicates the confidence score for anomalies; The condition number represents the covariance matrix, and the condition number is the quotient of the largest singular value and the smallest singular value. This represents the preset regularization minima, which is set to... ; This represents a logarithmic function with the natural constant as the base. Anomaly confidence score. The dimension is one, and the higher the anomaly confidence score, the higher the degree of anomaly in the correlation between vibration data and pressure data.

[0049] In practice, edge computing nodes acquire anomaly confidence scores. The edge computing nodes then read a preset scaling factor. This preset scaling factor is used to convert the anomaly confidence score into an integer representation, and its setting is based on the resolution requirements of the anomaly confidence score. The decimal part of the anomaly confidence score contains information about the degree of anomaly; the preset scaling factor is set to... Set the preset scaling factor to The basis for this is that the resolution accuracy of the anomaly confidence score needs to reach [a certain level]. Multiply the abnormal confidence score by Rounding down to the integer part preserves three decimal places in the anomaly confidence score while avoiding rounding errors inherent in floating-point representation. The edge computing node multiplies the anomaly confidence score by a preset scaling factor to obtain a scaled anomaly confidence score. The edge computing node then rounds down the integer part of the scaled anomaly confidence score. The formula for rounding down to the integer part is: in, This represents the integer value obtained by multiplying the anomaly confidence score by a preset scaling factor and then rounding it down. This represents the anomaly confidence score, with values ​​derived from logarithmic calculations. This represents the preset scaling factor, with a value of [value missing]. ; This represents the floor function, which rounds the input value down to the nearest integer in the direction of negative infinity.

[0050] In some embodiments, the edge computing node converts the rounded integer into a fixed-length binary string. The fixed length is read by the edge computing node from configuration parameters and is set to a specific value. Position. Fixed length set to The basis for the number of bits is: the local storage medium of the processing unit is... The hash code is stored in an unsigned integer format. A binary string can cover the possible range of values ​​for anomaly confidence scores after scaling, and maintains an appropriate hash code length after being merged with the check bits generated by cyclic redundancy check. Edge computing nodes determine whether the rounded integer exceeds... The range of unsigned integers. When the rounded integer is less than zero, the edge computing node sets the rounded integer to zero. When the rounded integer is greater than... At that time, the edge computing node sets the rounded integer to... Edge computing nodes convert the rounded integer to... A binary string, Insufficient binary string When the position is filled on the left, add it. .

[0051] Optionally, the edge computing node pairs are transformed Cyclic redundancy check (CRC) is performed on the binary string. The CRC-32 algorithm is used for CRC-32. The CRC-32 algorithm uses a generator polynomial. Edge computing nodes will A binary string of bits is used as input data for the CRC-32 algorithm. The CRC-32 algorithm calculation process is as follows: the edge computing node will... The left side of the binary string is supplemented. indivual ,get Bit-extended binary strings. Edge computing nodes use generator polynomials corresponding to... binary divisor right Bit-extended binary string binary modulo Division. Binary modulo Division is a bitwise XOR operation. Edge computing nodes will perform binary modulo operations. The result of division The remainder is used as a check bit. Edge computing nodes append the check bit to... The last two bits of the binary string are combined to form Bit hash code. The bit hash code is derived from the original Bit binary string and The check digits are composed of bits.

[0052] It can be understood that edge computing nodes obtain unique identifiers for all processing units within the edge computing node. A processing unit is a computing core or module within the edge computing node. The unique identifier is either a hardware identifier or an identifier assigned by the operating system. The edge computing node calculates the MD5 value for the unique identifier of each processing unit. The MD5 calculation process is as follows: the edge computing node takes the unique identifier as input and pads the input message by adding a unique identifier to the end of the input message. Position and several The bit makes the length of the padded message equal to the length of the padded message. The remainder when modulo is Then add The message length is specified in bits. Edge computing nodes will then populate the message according to... Bit-grouping, each group performs four rounds of compression function operations, including nonlinear functions and modulo operations. Addition and circular left shift operations. The edge computing nodes will perform four rounds of compression function operations on the resulting four nodes. The bit state variables are concatenated in order to obtain The MD5 value is a bit of the MD5 hash. A binary number with a value range of 1. arrive Edge computing nodes map the MD5 value corresponding to the unique identifier of each processing unit to a position on a hash ring. The hash ring is... arrive A ring-shaped address space. Edge computing nodes pair MD5 values... The modulo operation is performed, and the result serves as the unique identifier of the processing unit's position on the hash ring.

[0053] In practice, edge computing nodes calculate the MD5 hash value of the hash code. The process of calculating the MD5 hash value is the same as the process of calculating the MD5 hash value of the unique identifier of the processing unit. The edge computing nodes will... The bit hash code is used as the input message for the MD5 algorithm. The bit hash code is padded, grouped, and compressed using four rounds of functions to obtain... The edge computing node will use the MD5 hash value to... Take the modulo operation, and use the result as the position of the hash code on the hash ring.

[0054] In some embodiments, the edge computing node searches for the processing unit identifier closest to the hash code mapping position in a clockwise direction on the hash ring. Specifically, the edge computing node arranges the positions of all unique processing unit identifiers on the hash ring in ascending order, forming an ordered position list. The edge computing node compares the hash code mapping position with each position in the ordered position list. When there is a position in the ordered position list that is greater than the hash code mapping position, the edge computing node selects the first position in the ordered position list that is greater than the hash code mapping position, and the unique processing unit identifier corresponding to this position is the found processing unit identifier. When there is no position in the ordered position list that is greater than the hash code mapping position, the edge computing node selects the first position in the ordered position list, and the unique processing unit identifier corresponding to this position is the found processing unit identifier. The edge computing node sends the anomaly confidence score and the corresponding hash code to the found processing unit. The sending method is either bus transmission within the edge computing node or shared memory writing. The edge computing node encapsulates the anomaly confidence score and the corresponding hash code into a persistent message, which contains the anomaly confidence score, hash code, unique processing unit identifier, and timestamp. Edge computing nodes write persistent messages to the receiving queue of the processing unit.

[0055] Optionally, the processing unit reads persistent messages from the receive queue. The processing unit parses the persistent messages to obtain the anomaly confidence score and the corresponding hash code. The processing unit writes the anomaly confidence score and the corresponding hash code to the local storage medium. The local storage medium is the non-volatile memory corresponding to the processing unit. The processing unit writes the anomaly confidence score and the corresponding hash code to the local storage medium in key-value pair format, where the key is the hash code and the value is the anomaly confidence score. Before writing to the local storage medium, the processing unit performs a write verification. The write verification includes the processing unit writing data to the local storage medium, reading data from the local storage medium, and comparing it with the written data. When the read data matches the written data, the write verification passes. When the read data does not match the written data, the processing unit re-executes the write operation. After the write operation is completed, the processing unit releases the memory space occupied by the persistent messages.

[0056] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing, characterized in that, Includes the following steps: Deploy inspection data collection queues on edge computing nodes, and perform streaming reassembly of multi-source heterogeneous monitoring data based on the timestamps of data packets and source device identifiers to generate structured time-series data blocks; Perform a fast Fourier transform on the structured time-series data block to extract the main peak sequence in the frequency domain, calculate the phase difference between adjacent frequency points in the main peak sequence, and locate the abrupt time section in the frequency domain based on the continuous rate of change of the phase difference. A time-domain signal segment is extracted at the abrupt change time section, and empirical mode decomposition is performed on the time-domain signal segment to obtain the intrinsic mode function set. The zero-crossing density of the first intrinsic mode function in the intrinsic mode function set is extracted as the initial criterion for anomaly. When the density at the zero point exceeds the preset density threshold, the vibration data and pressure data in the structured time series data block are subjected to covariance calculation to obtain the covariance matrix. Then, the condition number of the covariance matrix is ​​calculated, and the logarithm of the condition number is used as the anomaly confidence score. The anomaly confidence score is mapped to a hash code, and the hash code is distributed to multiple processing units on the edge computing node using a consistent hashing algorithm. Each processing unit then performs the persistent storage of the anomaly confidence score.

2. The real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing according to claim 1, characterized in that, Perform a Fast Fourier Transform on the structured time-series data block to extract the main peak sequence in the frequency domain. Calculate the phase difference between adjacent frequency points in the main peak sequence. Based on the continuous rate of change of the phase difference, locate the abrupt time segment in the frequency domain. Specifically, this includes: The structured time-series data block is divided into multiple subframes with a fixed time window length. A fast Fourier transform is performed on each subframe to obtain the frequency domain amplitude sequence and frequency domain phase sequence of each subframe. Extract the frequency points corresponding to the local maxima from the frequency domain amplitude sequence of each subframe to form the main peak sequence of that subframe, and arrange the frequency points in the main peak sequence in ascending order of frequency. Take two adjacent frequency points, calculate the difference between their phase values ​​to obtain a phase difference sequence, and perform point-by-point difference between the phase difference sequence of the current subframe and the phase difference sequence of the previous subframe to obtain a phase difference change rate sequence. When the absolute values ​​of three or more consecutive values ​​in the phase difference change rate sequence all exceed the set threshold, the start time of the corresponding subframe is recorded as a sudden change time section.

3. The real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing according to claim 2, characterized in that, A time-domain signal segment is extracted at the abrupt change time section. Empirical mode decomposition (EMD) is performed on this time-domain signal segment to obtain the intrinsic mode function (IMF) set. The zero-crossing density of the first IMF in the IMF set is extracted as the initial anomaly criterion, specifically including: Using the abrupt change time section as the center point, sampling points are extended forward and backward by a preset length to extract time-domain signal segments. Empirical mode decomposition is performed on these time-domain signal segments, and a finite number of intrinsic mode functions are extracted through a screening process. These functions are then arranged in descending order of frequency to form a set. Take the first intrinsic mode function in the set, iterate through all the sampling points of the intrinsic mode function, count the number of times the product of the amplitudes of two adjacent sampling points is negative, and obtain the total number of zero crossings; The total number of zero-crossings is divided by the duration of the time-domain signal segment to obtain the zero-crossing density. The zero-crossing density is compared with a preset density threshold. If the zero-crossing density exceeds the preset density threshold, it is determined that there is an abnormal initial sign at the time section of the abrupt change.

4. The real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing according to claim 3, characterized in that, When the density at the zero-crossing point exceeds a preset density threshold, the vibration data and pressure data in the structured time-series data block are subjected to covariance calculation to obtain a covariance matrix. The condition number of the covariance matrix is ​​then calculated, and the logarithm of the condition number is used as the anomaly confidence score. Specifically, this includes: The vibration amplitude sequence and pressure amplitude sequence aligned with the time of abrupt change cross section are extracted from the structured time series data block. The vibration amplitude sequence and pressure amplitude sequence are paired one by one according to the sampling time to form a two-dimensional observation vector sequence. Calculate the mean vector of the two-dimensional observation vector sequence, subtract the mean vector from each two-dimensional observation vector to obtain a decentralized vector sequence, sum the outer product of the decentralized vector sequence and its transpose, and then divide by the total number of observation vectors minus one to obtain the covariance matrix. Singular value decomposition is performed on the covariance matrix to extract the maximum and minimum singular values. The maximum singular value is divided by the minimum singular value to obtain the condition number. The logarithm of the condition number with the natural constant as the base is then taken to obtain the anomaly confidence score.

5. The real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing according to claim 4, characterized in that, The anomaly confidence score is mapped to a hash code, and a consistent hashing algorithm is used to distribute the hash codes to multiple processing units on the edge computing node. Each processing unit then performs the persistent storage of the anomaly confidence score, specifically including: The anomaly confidence score is multiplied by a preset scaling factor and the integer part is taken. This integer part is converted into a fixed-length binary string. Then, a cyclic redundancy check is performed on the binary string, and a check bit is generated and appended to the end of the binary string to form a hash code. Obtain the unique identifier of all processing units in the edge computing node, calculate the MD5 value of each unique identifier, map the MD5 value to the position on the hash ring, and then calculate the MD5 value of the hash code and map it to the position on the hash ring. The system searches for the processing unit identifier closest to the hash code mapping position in a clockwise direction on the hash ring, sends the anomaly confidence score and the corresponding hash code to that processing unit, and then writes the data to the local storage medium.

6. The real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing according to claim 1, characterized in that, Generate structured time-series data blocks, specifically including: An independent data acquisition queue is created in the memory of the edge computing node for each source device identifier. When a data packet arrives, its timestamp and source device identifier are extracted, and the data packet is appended to the tail of the corresponding data acquisition queue according to the source device identifier. Data packets are retrieved from the head of each data acquisition queue. The timestamp of the retrieved data packets is compared with the current system time. Data packets whose timestamp lag exceeds the allowable jitter are filtered out, marked as delayed data packets, and removed from the queue. The remaining data packets are arranged in ascending order of timestamp. Data packets with the same timestamp are grouped together, and the sensor values ​​of all data packets in the same group are concatenated into a row vector according to the preset measurement point order to generate a structured time-series data block.

7. The real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing according to claim 2, characterized in that, The phase difference change rate sequence was obtained, specifically including: In the main peak sequence, the i-th frequency point and the (i+1)-th frequency point are taken out in ascending order of frequency. The phase value of the i-th frequency point and the phase value of the (i+1)-th frequency point are extracted from the frequency domain phase sequence respectively. The latter is subtracted from the former to obtain the i-th phase difference value. The phase difference sequence is obtained by traversing all adjacent frequency points. For each phase difference value in the phase difference sequence, calculate the difference between it and the phase difference value at the same index in the phase difference sequence of the previous subframe to obtain the point-by-point difference value. Take the absolute value of all point-by-point difference values ​​and then average them to obtain the average phase difference change rate of the subframe. The average phase difference change rate is used as the phase difference change rate of this subframe. This value is then weighted and averaged with the phase difference change rate of the previous subframe to obtain a smoothed phase difference change rate sequence.

8. The method for real-time anomaly diagnosis of hydropower plant inspection and monitoring data based on edge computing according to claim 3, characterized in that, Using the abrupt change time segment as the center point, sampling points are extended forward and backward by a predetermined length to extract time-domain signal segments. Empirical mode decomposition (EMD) is performed on these time-domain signal segments, and a finite number of intrinsic mode functions (EMFs) are extracted through a screening process. These EMFs are then arranged in descending order of frequency to form a set, specifically including: Centered on the sampling point index corresponding to the abrupt change time section, take a first preset number of sampling points forward and a second preset number of sampling points backward, and extract a time domain signal segment with a total length equal to the sum of the first preset number and the second preset number plus one. The empirical mode decomposition process is performed on the time-domain signal segment to identify all local maxima and local minima in the time-domain signal segment. The upper and lower envelopes are constructed by cubic spline interpolation, and the mean envelopes of the upper and lower envelopes are calculated. Subtract the mean envelope from the time-domain signal segment to obtain candidate intrinsic mode functions. Determine whether the absolute value of the difference between the number of extreme points and the number of zero crossings of the candidate intrinsic mode function is less than or equal to one. If so, output the candidate intrinsic mode function as the first intrinsic mode function; otherwise, repeat the selection process.

9. The real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing according to claim 4, characterized in that, The formation of a two-dimensional observation vector sequence specifically includes: Based on the absolute time of the abrupt change time section, select all data rows from the structured time series data block whose sampling time is within the range of the third preset number of sampling points before the abrupt change time section to the fourth preset number of sampling points after the abrupt change time section. The amplitude of the vibration sensor is extracted as the first component from each row of data selected, and the amplitude of the pressure sensor is extracted as the second component. The sampling time is used as the row index, and the first and second components are combined into a two-dimensional observation vector. All two-dimensional observation vectors are arranged in ascending order according to the sampling time to form a two-dimensional observation vector sequence.

10. The real-time anomaly diagnosis method for hydropower plant inspection and monitoring data based on edge computing according to claim 4, characterized in that, The abnormal confidence score is obtained, specifically including: Perform singular value decomposition on the two-dimensional covariance matrix to obtain the left singular matrix, the singular value diagonal matrix, and the transpose of the right singular matrix. Extract all singular values ​​from the diagonal of the singular value diagonal matrix, compare the magnitudes of all singular values, and find the singular value with the largest and smallest values. Divide the largest singular value by the smallest singular value, and use the quotient as the condition number of the covariance matrix. Add a preset regularization factor to the condition number, and take the logarithm with the natural constant as the base. Use the result of the logarithmic operation as the anomaly confidence score at the time section of the mutation.