A feature recognition-based hydroelectric equipment state monitoring system and method

Abstract

The application discloses a kind of water and electricity equipment condition monitoring system and method based on feature recognition, it is related to water and electricity equipment monitoring technical field, including data acquisition module, data preprocessing module, working condition feature library and partition module, double-layer coupling quantitative analysis module, adaptive iteration calibration and safety locking module;Data acquisition module is used to collect unit historical archival data and real-time operation data;Data preprocessing module filters effective sample meeting steady-state constraint, as model iteration update benchmark;Working condition feature library and partition module are used to build multi-working condition characteristic fingerprint library and rely on cosine similarity quantification sample feature matching degree;Double-layer coupling quantitative analysis module is used to decouple equipment natural aging component and fault abnormal mutation component;Adaptive iteration calibration and safety locking module only complete corresponding interval submodel local update after sample standard, realize multi-working condition benchmark model adaptive calibration, complete water and electricity equipment condition monitoring.

Description

Technical Field

[0001] This invention relates to the field of hydropower equipment monitoring technology, specifically a hydropower equipment condition monitoring system and method based on feature recognition. Background Technology

[0002] Hydropower units operate under complex hydraulic conditions with varying head and load for extended periods. Complex factors such as hydraulic disturbances, cavitation vortices, and load fluctuations continuously affect operating parameters such as unit vibration and pressure pulsation, resulting in significant differences in the steady-state characteristics of hydropower equipment. This places extremely high demands on the accuracy and stability of equipment condition monitoring.

[0003] Currently, most existing hydropower equipment condition monitoring technologies adopt a single fixed steady-state characteristic baseline and a unified model iteration update mechanism. They do not distinguish the operating condition disturbance characteristics of different head ranges and cannot adapt to different operating condition scenarios such as high-frequency stability, medium-frequency disturbance, and low-frequency strong disturbance. Under stable operating conditions, baseline updates are prone to lag and the inability to track minor aging and deterioration of equipment. Under strong disturbance scarce operating conditions, they are easily affected by hydraulic noise and operating condition fluctuations, resulting in model baseline shift and fault characteristics being masked by normal operating condition disturbances, which in turn leads to monitoring false alarms and missed alarms.

[0004] Meanwhile, traditional monitoring technologies generally use fixed sample size and fixed update cycle to carry out model iteration. They do not adaptively adjust the sample entry threshold and iteration rhythm according to the intensity of operating condition disturbance, data signal-to-noise ratio, and the scarcity of operating conditions. This results in defects such as the inability to quantify the quality of steady-state samples and the mixing of effective and invalid samples, making it difficult to ensure the consistency and reliability of model iteration under multiple operating conditions.

[0005] Furthermore, existing technologies cannot accurately distinguish between the slow evolution of equipment due to long-term natural aging and the abrupt change of state due to instantaneous fault impact. The two types of state components are mixed in the model iteration, which easily leads to technical problems such as normal aging of equipment being judged as abnormal and instantaneous faults being assimilated and smoothed out by the baseline. This makes it impossible to achieve refined and high-precision state monitoring and deterioration identification of hydropower equipment under complex and variable working conditions, and it is difficult to meet the actual application needs of long-term stable monitoring of current smart hydropower equipment. Summary of the Invention

[0006] The purpose of this invention is to provide a feature-based hydropower equipment condition monitoring system and method to solve the problems raised in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for monitoring the condition of hydropower equipment based on feature recognition, the method comprising: Step S1: Synchronously collect historical full-condition archived data and real-time operation data of the hydropower plant. The data includes hydraulic, mechanical vibration, electrical and equipment auxiliary status parameters. All data are equipped with GPS timestamps to achieve global time series alignment. Historical archived data is used for model pre-training and parameter statistics, while real-time operation data is used for online sample accumulation and equipment status monitoring. Step S2: Preprocess and clean the raw data, remove non-steady and dirty data, and select effective steady-state data with head fluctuation ≤ ±0.5m, load fluctuation ≤ ±2%, and continuous steady-state duration ≥ 5min as the benchmark sample for online model update; Step S3: Construct a standard dataset using historical data of the unit's healthy steady state without defects after overhaul, pre-train and build a multi-head interval operating condition feature fingerprint database, and rely on four quantitative indicators: head deviation rate, pressure fluctuation variance, data signal-to-noise ratio, and operating condition probability. Adaptively divide the operating head into three types of operating condition intervals: high-frequency stable, medium-frequency disturbance, and low-frequency strong disturbance through an unsupervised clustering algorithm, and extract the effective data time-domain and frequency-domain features. Use the feature vector cosine similarity algorithm to calculate the sample feature matching degree of the corresponding interval. Step S4: Construct a dual-layer coupled logical architecture of working condition confidence quantification and time series state evolution, adaptively solve the optimal sample capacity for each head interval, combine the sample feature matching degree to complete the steady-state sample confidence weighting conversion, and obtain the standardized effective sample duration; and analyze the time series evolution law through the residual time series second-order differential algorithm, and decouple the equipment natural aging component and fault abnormal component by mechanism. Step S5: Set the confidence trigger coefficient to construct the model iteration admission threshold. After the effective sample duration reaches the target, only the local update of the corresponding head interval sub-model is triggered. The gradient is quantized and the model update cycle is adaptively tuned according to the interval operating condition disturbance intensity coefficient. Combined with the time residual time sequence analysis in step S4, the iteration permissions are distinguished. Natural aging components are allowed to participate in baseline update and the iteration behavior of blocked fault components are allowed. With the help of a multi-dimensional safety blocking mechanism, abnormal iterations are intercepted, and the adaptive calibration update of the benchmark model of each head interval is completed.

[0008] Furthermore, the historical full-condition archived data includes the unit's full-condition operation data spanning the flood season and dry season over the past three years, including vibration swing time-domain data, pressure pulsation high-frequency waveform data, active and reactive load data, net water head, flow rate through the unit, tailwater aeration parameters, bearing temperature, guide vane opening, and unit speed parameters; the real-time operation data consists of hydraulic, mechanical, electrical, and auxiliary system status data collected in milliseconds by the LCU and high-frequency acquisition devices, and automatically distinguishes between power generation, phase adjustment, and no-load operation modes.

[0009] Furthermore, the non-steady-state dirty data removed during preprocessing specifically includes data from start-up and shutdown transitions, load step adjustments, high-frequency fluctuations during primary frequency modulation, gas injection switching moments, load shedding tests, sensor zero-point drift, and communication packet loss / distortion.

[0010] Furthermore, step S3 includes the following specific steps: Step S31: Perform steady-state screening and denoising on the standard dataset to remove abnormal samples and retain clean and healthy steady-state data. Then, for the pre-processed steady-state data of each head interval, extract the time-domain feature mean, time-domain variance, steady-state amplitude, frequency domain main frequency, vortex band characteristic frequency, and spectral energy ratio of each sample to form a multi-dimensional original feature vector. Then, perform mean clustering convergence calculation on the massive original feature vectors under the same head condition to remove discrete abnormal feature samples that deviate from the cluster distribution. After iterative convergence, solidify the standardized steady-state feature fingerprint vector that uniquely corresponds to each head interval. Associate and bind the standardized steady-state feature fingerprint vector corresponding to each head interval with the corresponding working condition interval label, working condition disturbance attribute, and operating parameter range, and store it in a structured database to form a multi-head interval exclusive working condition feature fingerprint database that is partitioned independently, has unique features, and can be matched and compared, thus completing the overall construction of the fingerprint database. Step S32: Obtain the real-time net head and rated head data from the real-time operation data record, calculate the head deviation rate, head deviation rate = (real-time net head - rated head) / rated head * 100%; obtain the interval pressure pulsation variance as the second-order statistical variance of the single-interval pressure pulsation time series data; obtain the data signal-to-noise ratio as the ratio of effective steady-state signal energy to noise signal energy, and the operating condition probability as the percentage of the steady-state operating time of a single head interval to the total steady-state operating time of the unit; after normalizing the above four quantitative indicators, construct a multi-dimensional operating condition disturbance feature matrix, and input the feature matrix as a whole into an unsupervised clustering algorithm; during the algorithm execution, the preset number of cluster categories is three, and the iterative convergence optimization objective is to minimize the overall variance of samples within a category and maximize the feature difference between categories. When the change in cluster centers is less than the preset convergence threshold, the algorithm iteration is determined to end, and the adaptive unlabeled grouping of all head operating condition samples is completed; Step S33: The algorithm automatically outputs three groups of characteristic-differentiated working condition sample clusters. Based on the inherent working condition characteristics of the four quantitative parameters corresponding to each group, the interval classification is completed. Among them, the sample cluster with the smallest pressure fluctuation variance, the highest signal-to-noise ratio, the small head deviation rate, and the highest working condition probability is classified as the high-frequency stable working condition interval; the sample cluster with the pressure fluctuation variance, signal-to-noise ratio, head deviation rate, and working condition probability at a medium level is classified as the medium-frequency disturbance working condition interval; the sample cluster with the largest pressure fluctuation variance, the lowest signal-to-noise ratio, the large head deviation rate, and the low working condition probability is classified as the low-frequency strong disturbance working condition interval. Step S34: Predefine the standardized steady-state feature fingerprint of each head interval as the reference vector A, and the same-dimensional working condition feature extracted from real-time effective steady-state data as the test vector B, setting the total number of feature dimensions to n; where the reference vector A = [A1, A2, ..., An], A1, A2, ..., An correspond to the standard feature values ​​of the interval standard in terms of time domain mean, time domain variance, steady-state amplitude, frequency domain dominant frequency, eddy band characteristic frequency, and spectral energy ratio, respectively; the test vector B = [B1, B2, ..., Bn] B1, B2, ..., Bn correspond to the measured values ​​of the working condition features of each dimension extracted synchronously from the real-time samples. The sample feature matching degree η is calculated using the cosine similarity formula: η = (A·B) / (||A||×||B||), where A·B is the dot product of the reference vector and the vector to be tested, obtained by summing the product of the corresponding dimension feature values ​​pairwise; ||A|| is the arithmetic square root of the sum of squares of the feature values ​​of each dimension of the reference vector; and ||B|| is the arithmetic square root of the sum of squares of the feature values ​​of each dimension of the vector to be tested. The value of η ranges from 0 to 1. The closer η is to 1, the higher the overlap between the multidimensional working condition features of the real-time sample and the corresponding head interval healthy steady-state standard fingerprint, and the stronger the representativeness and effectiveness of the sample working condition. The closer η is to 0, the lower the effective value of the sample features deviating from the steady-state working condition.

[0011] Furthermore, step S4 includes the following: Step S41: The dual-layer coupled logic architecture includes a working condition confidence quantization logic layer and a time-series state evolution logic layer. The two layers operate and are coupled and linked at each level. The working condition confidence quantization logic layer is used to solve for the optimal sample size of the working condition and to perform confidence weighted purification of the samples. The time-series state evolution logic layer is used to decouple the mechanism of equipment natural aging and instantaneous failure. Step S42: The working condition confidence quantization logic layer calculates the adaptive optimal effective sample capacity Ni corresponding to the i-th type of head interval through the optimal sample size adaptive algorithm. Ni=Nmin×(σi / σ0)×(1 / SNRi)×(1 / Pi 1 / 2 ); Where Nmin represents the minimum basic sample size under the healthy and stable baseline operating condition of the unit; σi represents the comprehensive standard deviation calculated by fusing multi-dimensional data of pressure pulsation and vibration amplitude of the unit within the i-th target head interval, used to characterize the severity of overall hydraulic and mechanical disturbances in the current interval; σ0 represents the fixed baseline value of the comprehensive standard deviation corresponding to the high-frequency stable baseline head interval, serving as a reference scale for the degree of disturbance under all operating conditions; SNRi is the signal-to-noise ratio of steady-state data in the i-th target head interval, reflecting the energy proportion of effective operating condition signals and noise interference; Pi is the proportion of annual steady-state operating time in the i-th target head interval, characterizing the scarcity of samples under this operating condition; based on the baseline... Based on the minimum reliable sample size for each working condition, the algorithm performs linked corrections by considering the disturbance ratio, noise ratio, and scarcity of working conditions relative to the benchmark in the current section. When the disturbance in the target head section is stronger, the noise is higher, and the working conditions are scarcer, the algorithm automatically increases the required optimal sample size to offset the data errors caused by strong disturbances, low quality, and low sample ratio. When the working conditions in the target head section are stable, the noise is low, and the operating ratio is high, the algorithm automatically reduces the required sample size to avoid redundant sample calculations. Ultimately, the algorithm achieves intelligent adaptive matching of the sample entry threshold under different disturbance intensities, ensuring that all samples participating in the model iteration in all head sections have the same accuracy and reliability. Step S43: After determining the optimal effective sample capacity Ni for the i-th head interval, further perform real-time sample quality correction calculation to obtain the standardized effective sample duration Ti that can be used for threshold comparison. The conversion formula is Ti = Tr × ηi; where Tr is the duration of the original continuous steady-state sample obtained in real-time screening under the i-th working condition, and ηi is the feature matching degree of the corresponding sample under the i-th working condition. This step uses the feature matching degree as the quality weight to perform quality purification and correction on the original samples that only represent the quantity of duration, eliminating invalid sample weights due to feature mismatch or working condition mismatch; the entire set of operations forms a complete logical link of fixed threshold solution + real-time sample correction.

[0012] Furthermore, the timing state evolution logic layer in step S4 also includes the following: Collect the difference between real-time monitoring parameters and the corresponding benchmark parameters of the working condition interval, and construct a continuous time-series residual sequence with uniform sampling at equal intervals; perform first-order difference operation and second-order differential operation on the time-series residual sequence frame by frame. The first-order difference is used to characterize the temporal change slope of the residual sequence, and the second-order differential is used to characterize the change curvature and instantaneous change intensity of the residual sequence. A fixed first-order difference stationarity threshold range and a second-order derivative zero reference threshold are pre-configured. The residual difference data of a consecutive preset number of frames are statistically judged. If all first-order difference calculation results in the frame segment fall into the preset first-order difference stationarity threshold range, and the absolute value of all second-order derivative calculation results in the corresponding frame segment is not greater than the preset second-order derivative zero reference threshold, the residual sequence as a whole shows a uniform unidirectional shift change pattern without abrupt inflection points or shock fluctuations. The time series component of this segment is quantitatively determined to be the natural aging component of the device, and the baseline iteration update permission for the corresponding time period sample is opened. If the first-order difference calculation result of a single frame or a few consecutive frames exceeds the first-order difference stationarity threshold range, and the absolute value of the second-order differential calculation result of the corresponding frame point is greater than the preset second-order differential zero reference threshold, and the residual sequence exhibits numerical step jumps, local abrupt inflection points, or instantaneous shock fluctuations, the time series component in that segment is quantitatively determined to be an abnormal component of equipment defects or early faults, and the model iteration permission for samples in that segment is locked. This accurately achieves the decoupling and discrimination of the quantization mechanism for three types of states: normal operating condition disturbances, natural equipment aging, and abnormal equipment faults.

[0013] Furthermore, the model iteration admission threshold is K*Ni, where K is a 0.95 confidence trigger coefficient. When the interval weighted effective sample duration Ti meets the standard, which means that the standardized effective sample duration Ti calculated in real time for the i-th head interval is greater than or equal to the iteration admission threshold K×Ni, only the local update of the corresponding interval sub-model is triggered, without interfering with the global model. The interval sub-model refers to a pre-constructed exclusive working condition feature fingerprint benchmark model that is independently matched to each type of head working condition interval, including independent sub-models corresponding to the three types of working condition intervals: high-frequency stable, medium-frequency disturbance, and low-frequency strong disturbance. Define the disturbance intensity coefficient Kdi, Kdi=(σi / σ0)×(1 / SNRi), and construct the update cycle quantification formula Li, Li=Tb×Kdi×ai, where Tb represents the baseline update cycle and ai represents the correction coefficient for the operating probability of the i-th type of head interval.

[0014] A feature-based hydropower equipment condition monitoring system includes a data acquisition module, a data preprocessing module, a working condition feature database and partitioning module, a two-layer coupled quantitative analysis module, and an adaptive iterative calibration and safety interlocking module. The data acquisition module is used to connect to various monitoring systems in the power plant, and synchronously collect historical archived data and real-time operating data of the unit, covering hydraulic, mechanical, electrical and auxiliary equipment status parameters. After GPS time-series alignment, the data is used for model pre-training statistics and online sample monitoring accumulation. The data preprocessing module cleans the raw data, removes non-steady-state data, and selects valid samples that meet the steady-state constraints as the benchmark for model iteration and update. The module for building a database and partitioning the operating condition features relies on the unit's overhaul health steady-state data to construct a standard dataset. After purification, time-frequency domain features are extracted, and the standardized steady-state fingerprint vectors of each head interval are solidified by mean clustering and added to the database to build a multi-operating condition feature fingerprint database. The perturbation feature matrix is ​​constructed by combining multiple types of operating condition quantitative indicators. Multiple head operating condition intervals are adaptively divided by constrained unsupervised clustering, and the sample feature matching degree is quantified by cosine similarity. The dual-layer coupled quantization analysis module has a dual-layer logical architecture, which can adaptively solve the optimal sample capacity for each working condition, obtain the standardized effective sample duration by combining feature matching degree correction, and decouple the natural aging component of the equipment and the fault abnormal mutation component through the first and second order differential quantization discrimination mechanism of the time residual. The adaptive iterative calibration and safety interlocking module sets the iterative admission threshold by setting a fixed confidence coefficient. After the sample meets the standard, only the corresponding interval sub-model is locally updated. The update cycle is adaptively adjusted according to the working condition disturbance gradient. It differentiates the control of iteration permissions, allowing the aging component to update the baseline and the interlocking fault component to iterate. With the help of the multi-dimensional safety interlocking mechanism, it realizes the adaptive calibration of the multi-working condition benchmark model and completes the status monitoring of hydropower equipment.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention utilizes four quantitative indicators—head deviation rate, pressure pulsation variance, data signal-to-noise ratio, and operating condition probability—to adaptively partition head conditions using an unsupervised clustering algorithm. It automatically divides the data into three differentiated operating condition intervals based on the inherent distribution characteristics of the data: high-frequency stable, mid-frequency disturbed, and low-frequency strongly disturbed, thus avoiding the drawbacks of traditional manual threshold partitioning and a single global baseline. Simultaneously, it independently constructs dedicated feature fingerprint sub-models for operating condition intervals with different disturbance characteristics, enabling refined modeling of equipment steady-state characteristics under various hydraulic disturbance scenarios. This eliminates monitoring interference caused by variable operating condition disturbances at the source, significantly improving the accuracy of identifying the state characteristics of hydropower equipment under complex operating conditions and effectively avoiding false alarms and missed alarms caused by operating condition fluctuations.

[0016] This invention introduces an innovative adaptive algorithm for confidence quantization based on operating conditions. It adaptively solves for the optimal sample size based on the disturbance intensity, noise level, and scarcity of operating conditions across different operating condition intervals. This achieves differentiated intelligent adaptation of the sample entry threshold for each operating condition interval, offsetting data errors from strong disturbances, low signal-to-noise ratios, and scarce operating conditions by dynamically adjusting the sample base. Simultaneously, it combines feature vector cosine similarity metric to quantify sample matching quality, performing confidence-weighted purification of the original steady-state samples to eliminate invalid samples with mismatched operating conditions or discrete features. This achieves dual quantitative verification of sample quantity and quality, solving the problems of traditional fixed sample sizes being unable to adapt to multiple operating conditions and inconsistent sample reliability. This ensures consistent accuracy and stable reliability in model iterations across all operating condition intervals.

[0017] This invention employs a residual time-series second-order differential quantization discrimination algorithm, replacing traditional fuzzy judgment logic with standardized numerical thresholds. This accurately distinguishes between the uniform, gradual time-series components of natural equipment aging and the instantaneous, abrupt time-series components of faults and anomalies. It differentiates model iteration permissions, allowing only normal aging components to participate in baseline updates and strictly blocking iteration of fault and anomaly components. Simultaneously, it incorporates an adaptive update frequency mechanism based on operating condition disturbance gradients and a multi-dimensional safety interlocking strategy. High-frequency iterations adapt to aging evolution under stable operating conditions, while low-frequency, cautious iterations under strong disturbance conditions avoid baseline drift. This completely resolves the shortcomings of traditional technologies where fault characteristics are assimilated by the baseline and normal aging is misjudged as anomalies, achieving long-term, high-precision, and high-reliability adaptive monitoring of hydropower equipment status. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating a feature-based hydropower equipment status monitoring method according to the present invention. Detailed Implementation

[0019] 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] Example: Figure 1 As shown, the present invention provides a method for monitoring the status of hydropower equipment based on feature recognition, the method comprising: Step S1: Synchronously collect historical full-condition archived data and real-time operation data of the hydropower plant. The data includes hydraulic, mechanical vibration, electrical and equipment auxiliary status parameters. All data are equipped with GPS timestamps to achieve global time series alignment. Historical archived data is used for model pre-training and parameter statistics, while real-time operation data is used for online sample accumulation and equipment status monitoring. Step S2: Preprocess and clean the raw data, remove non-steady and dirty data, and select effective steady-state data with head fluctuation ≤ ±0.5m, load fluctuation ≤ ±2%, and continuous steady-state duration ≥ 5min as the benchmark sample for online model update; Step S3: Construct a standard dataset using historical data of the unit's healthy steady state without defects after overhaul, pre-train and build a multi-head interval operating condition feature fingerprint database, and rely on four quantitative indicators: head deviation rate, pressure fluctuation variance, data signal-to-noise ratio, and operating condition probability. Adaptively divide the operating head into three types of operating condition intervals: high-frequency stable, medium-frequency disturbance, and low-frequency strong disturbance through an unsupervised clustering algorithm, and extract the effective data time-domain and frequency-domain features. Use the feature vector cosine similarity algorithm to calculate the sample feature matching degree of the corresponding interval. Step S4: Construct a dual-layer coupled logical architecture of working condition confidence quantification and time series state evolution, adaptively solve the optimal sample capacity for each head interval, combine the sample feature matching degree to complete the steady-state sample confidence weighting conversion, and obtain the standardized effective sample duration; and analyze the time series evolution law through the residual time series second-order differential algorithm, and decouple the equipment natural aging component and fault abnormal component by mechanism. Step S5: Set the confidence trigger coefficient to construct the model iteration admission threshold. After the effective sample duration reaches the target, only the local update of the corresponding head interval sub-model is triggered. The gradient is quantized and the model update cycle is adaptively tuned according to the interval operating condition disturbance intensity coefficient. Combined with the time residual time sequence analysis in step S4, the iteration permissions are distinguished. Natural aging components are allowed to participate in baseline update and the iteration behavior of blocked fault components are allowed. With the help of a multi-dimensional safety blocking mechanism, abnormal iterations are intercepted, and the adaptive calibration update of the benchmark model of each head interval is completed.

[0021] The historical full-condition archive data consists of the unit's full-condition operation data spanning the flood season and dry season over the past three years, including vibration swing time domain data, pressure pulsation high-frequency waveform data, active and reactive load data, net water head, flow rate through the unit, tailwater aeration parameters, bearing temperature, guide vane opening, and unit speed parameters; the real-time operation data consists of hydraulic, mechanical, electrical, and auxiliary system status data collected in milliseconds by the LCU and high-frequency acquisition devices, and automatically distinguishes between power generation, phase adjustment, and no-load operation modes.

[0022] The non-steady-state dirty data removed during preprocessing specifically includes data from start-up and shutdown transitions, load step adjustments, high-frequency fluctuations during primary frequency modulation, gas injection switching moments, load shedding tests, sensor zero-point drift, and communication packet loss / distortion.

[0023] Step S3 includes the following specific steps: Step S31: Perform steady-state screening and denoising on the standard dataset to remove abnormal samples and retain clean and healthy steady-state data. Then, for the pre-processed steady-state data of each head interval, extract the time-domain feature mean, time-domain variance, steady-state amplitude, frequency domain main frequency, vortex band characteristic frequency, and spectral energy ratio of each sample to form a multi-dimensional original feature vector. Then, perform mean clustering convergence calculation on the massive original feature vectors under the same head condition to remove discrete abnormal feature samples that deviate from the cluster distribution. After iterative convergence, solidify the standardized steady-state feature fingerprint vector that uniquely corresponds to each head interval. Associate and bind the standardized steady-state feature fingerprint vector corresponding to each head interval with the corresponding working condition interval label, working condition disturbance attribute, and operating parameter range, and store it in a structured database to form a multi-head interval exclusive working condition feature fingerprint database that is partitioned independently, has unique features, and can be matched and compared, thus completing the overall construction of the fingerprint database. As shown in the example: Six-dimensional core operating condition features are extracted from steady-state data frame by frame. The sampling time for each frame is 1 second, and a unified 6-dimensional feature vector is constructed. The specific calculation methods for each dimension are as follows: The first dimension, the time-domain mean, is the arithmetic mean of the pressure pulsation and vibration amplitude parameters within a single frame, representing the steady-state basic amplitude level; the second dimension, the time-domain variance, is the discrete statistic of the relative mean of the parameters in a single frame, representing the degree of steady-state fluctuation; the third dimension, the steady-state amplitude, is the peak steady-state mean of the effective signal in a single frame; the fourth dimension, the frequency domain dominant frequency, is the frequency corresponding to the maximum spectral energy after the fast Fourier transform; the fifth dimension, the vortex band characteristic frequency, is the center value of the low-frequency characteristic band corresponding to the tailrace vortex band of the mixed-flow turbine; and the sixth dimension, the spectral energy ratio, is the ratio of the energy of the effective characteristic band to the total energy of the entire frequency band. Through the above calculation methods, features are extracted frame by frame from all steady-state samples of each type of head interval, generating a massive set of original 6-dimensional feature vectors. For massive original feature vectors within a single interval, mean clustering convergence is used to purify features. The mean and dispersion of all original vectors are statistically analyzed dimension by dimension, and discrete outlier feature samples with single-dimensional feature deviations greater than twice the interval standard deviation are removed to avoid feature interference caused by accidental steady-state fluctuations. The interval feature cluster centers are continuously updated iteratively, with the iteration convergence precision set to ≤10 for changes in cluster centers. -4 After iterative convergence, the final cluster center vector is used as the standardized steady-state fingerprint vector that uniquely corresponds to the head interval. Step S32: Obtain the real-time net head and rated head data from the real-time operation data record, calculate the head deviation rate, head deviation rate = (real-time net head - rated head) / rated head * 100%; obtain the interval pressure pulsation variance as the second-order statistical variance of the single-interval pressure pulsation time series data; obtain the data signal-to-noise ratio as the ratio of effective steady-state signal energy to noise signal energy, and the operating condition probability as the percentage of the steady-state operating time of a single head interval to the total steady-state operating time of the unit; after normalizing the above four quantitative indicators, construct a multi-dimensional operating condition disturbance feature matrix, and input the feature matrix as a whole into an unsupervised clustering algorithm; during the algorithm execution, the preset number of cluster categories is three, and the iterative convergence optimization objective is to minimize the overall variance of samples within a category and maximize the feature difference between categories. When the change in cluster centers is less than the preset convergence threshold, the algorithm iteration is determined to end, and the adaptive unlabeled grouping of all head operating condition samples is completed; Step S33: The algorithm automatically outputs three groups of characteristic-differentiated working condition sample clusters. Based on the inherent working condition characteristics of the four quantitative parameters corresponding to each group, the interval classification is completed. Among them, the sample cluster with the smallest pressure fluctuation variance, the highest signal-to-noise ratio, the small head deviation rate, and the highest working condition probability is classified as the high-frequency stable working condition interval; the sample cluster with the pressure fluctuation variance, signal-to-noise ratio, head deviation rate, and working condition probability at a medium level is classified as the medium-frequency disturbance working condition interval; the sample cluster with the largest pressure fluctuation variance, the lowest signal-to-noise ratio, the large head deviation rate, and the low working condition probability is classified as the low-frequency strong disturbance working condition interval. Step S34: Predefine the standardized steady-state feature fingerprint of each head interval as the reference vector A, and the same-dimensional working condition feature extracted from real-time effective steady-state data as the test vector B, setting the total number of feature dimensions to n; where the reference vector A = [A1, A2, ..., An], A1, A2, ..., An correspond to the standard feature values ​​of the interval standard in terms of time domain mean, time domain variance, steady-state amplitude, frequency domain dominant frequency, eddy band characteristic frequency, and spectral energy ratio, respectively; the test vector B = [B1, B2, ..., Bn] B1, B2, ..., Bn correspond to the measured values ​​of the working condition features of each dimension extracted synchronously from the real-time samples. The sample feature matching degree η is calculated using the cosine similarity formula: η = (A·B) / (||A||×||B||), where A·B is the dot product of the reference vector and the vector to be tested, obtained by summing the product of the corresponding dimension feature values ​​pairwise; ||A|| is the arithmetic square root of the sum of squares of the feature values ​​of each dimension of the reference vector; and ||B|| is the arithmetic square root of the sum of squares of the feature values ​​of each dimension of the vector to be tested. The value of η ranges from 0 to 1. The closer η is to 1, the higher the overlap between the multidimensional working condition features of the real-time sample and the corresponding head interval healthy steady-state standard fingerprint, and the stronger the representativeness and effectiveness of the sample working condition. The closer η is to 0, the lower the effective value of the sample features deviating from the steady-state working condition. When the similarity η between the real-time sample and the corresponding interval baseline vector is ≥0.92, it is determined to be a highly matched valid sample; when η<0.92, it is determined to be a feature mismatch sample and will not participate in the model iteration.

[0024] Step S4 includes the following: Step S41: The dual-layer coupled logic architecture includes a working condition confidence quantization logic layer and a time-series state evolution logic layer. The two layers operate and are coupled and linked at each level. The working condition confidence quantization logic layer is used to solve for the optimal sample size of the working condition and to perform confidence weighted purification of the samples. The time-series state evolution logic layer is used to decouple the mechanism of equipment natural aging and instantaneous failure. Step S42: The working condition confidence quantization logic layer calculates the adaptive optimal effective sample capacity Ni corresponding to the i-th type of head interval through the optimal sample size adaptive algorithm. Ni=Nmin×(σi / σ0)×(1 / SNRi)×(1 / Pi 1 / 2 ); Where Nmin represents the minimum basic sample size under the healthy and stable baseline operating condition of the unit; σi represents the comprehensive standard deviation calculated by fusing multi-dimensional data of pressure pulsation and vibration amplitude of the unit within the i-th target head interval, used to characterize the severity of overall hydraulic and mechanical disturbances in the current interval; σ0 represents the fixed baseline value of the comprehensive standard deviation corresponding to the high-frequency stable baseline head interval, serving as a reference scale for the degree of disturbance under all operating conditions; SNRi is the signal-to-noise ratio of steady-state data in the i-th target head interval, reflecting the energy proportion of effective operating condition signals and noise interference; Pi is the proportion of annual steady-state operating time in the i-th target head interval, characterizing the scarcity of samples under this operating condition; based on the baseline... Based on the minimum reliable sample size for each working condition, the algorithm performs linked corrections by considering the disturbance ratio, noise ratio, and scarcity of working conditions relative to the benchmark in the current section. When the disturbance in the target head section is stronger, the noise is higher, and the working conditions are scarcer, the algorithm automatically increases the required optimal sample size to offset the data errors caused by strong disturbances, low quality, and low sample ratio. When the working conditions in the target head section are stable, the noise is low, and the operating ratio is high, the algorithm automatically reduces the required sample size to avoid redundant sample calculations. Ultimately, the algorithm achieves intelligent adaptive matching of the sample entry threshold under different disturbance intensities, ensuring that all samples participating in the model iteration in all head sections have the same accuracy and reliability. Step S43: After determining the optimal effective sample capacity Ni for the i-th head interval, further perform real-time sample quality correction calculation to obtain the standardized effective sample duration Ti that can be used for threshold comparison. The conversion formula is Ti = Tr × ηi; where Tr is the duration of the original continuous steady-state sample obtained in real-time screening under the i-th working condition, and ηi is the feature matching degree of the corresponding sample under the i-th working condition. This step uses the feature matching degree as the quality weight to perform quality purification and correction on the original samples that only represent the quantity of duration, eliminating invalid sample weights due to feature mismatch or working condition mismatch; the entire set of operations forms a complete logical link of fixed threshold solution + real-time sample correction.

[0025] The timing state evolution logic layer in step S4 also includes the following: Collect the difference between real-time monitoring parameters and the corresponding benchmark parameters of the working condition interval, and construct a continuous time-series residual sequence with uniform sampling at equal intervals; perform first-order difference operation and second-order differential operation on the time-series residual sequence frame by frame. The first-order difference is used to characterize the temporal change slope of the residual sequence, and the second-order differential is used to characterize the change curvature and instantaneous change intensity of the residual sequence. A fixed first-order difference stationarity threshold range and a second-order derivative zero reference threshold are pre-configured. The residual difference data of a consecutive preset number of frames are statistically judged. If all first-order difference calculation results in the frame segment fall into the preset first-order difference stationarity threshold range, and the absolute value of all second-order derivative calculation results in the corresponding frame segment is not greater than the preset second-order derivative zero reference threshold, the residual sequence as a whole shows a uniform unidirectional shift change pattern without abrupt inflection points or shock fluctuations. The time series component of this segment is quantitatively determined to be the natural aging component of the device, and the baseline iteration update permission for the corresponding time period sample is opened. If the first-order difference calculation result of a single frame or a few consecutive frames exceeds the first-order difference stationarity threshold range, and the absolute value of the second-order differential calculation result of the corresponding frame point is greater than the preset second-order differential zero reference threshold, and the residual sequence exhibits numerical step jumps, local abrupt inflection points, or instantaneous shock fluctuations, the time series component in that segment is quantitatively determined to be an abnormal component of equipment defects or early faults, and the model iteration permission for samples in that segment is locked. This accurately achieves the decoupling and discrimination of the quantization mechanism for three types of states: normal operating condition disturbances, natural equipment aging, and abnormal equipment faults.

[0026] As shown in the example: First, a time-series residual sequence is constructed. A fixed sampling time interval Δt is set. Deviations between real-time monitoring feature parameters and corresponding standardized steady-state fingerprint benchmark parameters are collected within the same operating condition range. A discrete time-series residual sequence R(t) = [R1, R2, R3, ..., Rn] is constructed, where each residual point in the sequence corresponds to a real-time feature deviation within a 1-second sampling interval. Based on this, point-by-point differential operations are performed: the first-order difference is used to characterize the real-time slope of the residual sequence, calculated as ΔRk = R k+1 -R k The residual is the difference between two adjacent sampling times. The sign of the first-order difference result indicates the direction of residual shift, and the magnitude indicates the rate of residual change. A second-order difference operation is performed on the first-order difference result to obtain the second-order differential value, which is used to characterize the curvature and abrupt change intensity of the residual change. The calculation formula is Δ. 2 Rk=ΔR k+1 -ΔR k The second-order derivative is used to identify the essential difference between a steady, gradual change in a sequence and a sudden, abrupt change. In this embodiment, a fixed quantization threshold is uniformly configured: the first-order differential stationary threshold range is set to [-0.03, 0.03], the second-order derivative zero-reference threshold is set to 0.005, and the number of steady-state determination statistical frames is fixed at 120 frames.

[0027] When all first-order difference calculation results of 120 consecutive sampling points fall within the range of [-0.03, 0.03], and the absolute value of the second-order derivative calculation result of each corresponding frame is ≤0.005, it proves that the slope of the residual sequence is constant, the curvature of the change is close to zero, the sequence as a whole shows a uniform unidirectional smooth shift, without abrupt changes or shock fluctuations. It is quantitatively determined that the time series component of this segment is a component of long-term slow natural aging of the equipment, and the baseline iteration update permission of the sample in this period is opened. When, within any 10 consecutive sampling points, there is a first-order difference calculation result that exceeds the threshold range of [-0.03, 0.03], and the absolute value of the second-order derivative calculation result of the corresponding frame point is >0.005, showing significant numerical jumps and curvature abrupt changes, it proves that there is an instantaneous step inflection point or shock fluctuation in the residual sequence. It is quantitatively determined that the time series component of this segment is an abnormal component of equipment defects or early faults, and the model iteration permission of the sample in this period is immediately locked.

[0028] The model iteration admission threshold is K*Ni, where K is a 0.95 confidence trigger coefficient. When the interval weighted effective sample duration Ti meets the standard, which means that the standardized effective sample duration Ti calculated in real time for the i-th head interval is greater than or equal to the iteration admission threshold K×Ni, only the local update of the corresponding interval sub-model is triggered, without interfering with the global model. The interval sub-model refers to a pre-constructed exclusive working condition feature fingerprint benchmark model that is independently matched to each type of head working condition interval, including independent sub-models corresponding to three types of working condition intervals: high-frequency stable, medium-frequency disturbance, and low-frequency strong disturbance. A disturbance intensity coefficient Kdi is defined as Kdi = (σi / σ0) × (1 / SNRi), and an update cycle quantification formula Li is constructed as Li = Tb × Kdi × ai, where Tb represents the baseline update cycle and ai represents the correction coefficient for the operating probability of the i-th head interval. In the high-frequency stable interval, Kdi and ai approach 1, with an update cycle of 7–15 days. In the mid-frequency disturbance interval, the parameters are moderate, with an update cycle of 15–30 days. In the low-frequency strong disturbance interval, Kdi is significantly larger and ai is smaller, with an update cycle of 90–120 days, achieving an adaptive gradient change in model update rhythm according to the disturbance intensity of the operating condition.

[0029] The multi-dimensional security interlocking mechanism includes residual mutation interlocking, baseline offset interlocking, feature distortion interlocking, and feature consistency interlocking, which are used to intercept abnormal iteration scenarios such as numerical step anomalies, benchmark deviations, feature mismatches, and sample feature discrepancies exceeding the standard.

[0030] A feature-based hydropower equipment condition monitoring system includes a data acquisition module, a data preprocessing module, a working condition feature database and partitioning module, a two-layer coupled quantitative analysis module, and an adaptive iterative calibration and safety interlocking module. The data acquisition module is used to connect to various monitoring systems in the power plant, and synchronously collect historical archived data and real-time operating data of the unit, covering hydraulic, mechanical, electrical and auxiliary equipment status parameters. After GPS time-series alignment, the data is used for model pre-training statistics and online sample monitoring accumulation. The data preprocessing module cleans the raw data, removes non-steady-state data, and selects valid samples that meet the steady-state constraints as the benchmark for model iteration and update. The module for building a database and partitioning the operating condition features relies on the unit's overhaul health steady-state data to construct a standard dataset. After purification, time-frequency domain features are extracted, and the standardized steady-state fingerprint vectors of each head interval are solidified by mean clustering and added to the database to build a multi-operating condition feature fingerprint database. The perturbation feature matrix is ​​constructed by combining multiple types of operating condition quantitative indicators. Multiple head operating condition intervals are adaptively divided by constrained unsupervised clustering, and the sample feature matching degree is quantified by cosine similarity. The dual-layer coupled quantization analysis module has a dual-layer logical architecture, which can adaptively solve the optimal sample capacity for each working condition, obtain the standardized effective sample duration by combining feature matching degree correction, and decouple the natural aging component of the equipment and the fault abnormal mutation component through the first and second order differential quantization discrimination mechanism of the time residual. The adaptive iterative calibration and safety interlocking module sets the iterative admission threshold by setting a fixed confidence coefficient. After the sample meets the standard, only the corresponding interval sub-model is locally updated. The update cycle is adaptively adjusted according to the working condition disturbance gradient. It differentiates the control of iteration permissions, allowing the aging component to update the baseline and the interlocking fault component to iterate. With the help of the multi-dimensional safety interlocking mechanism, it realizes the adaptive calibration of the multi-working condition benchmark model and completes the status monitoring of hydropower equipment.

[0031] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring the condition of hydropower equipment based on feature recognition, characterized in that: The method includes: Step S1: Synchronously collect historical full-condition archived data and real-time operation data of the hydropower plant. The data includes hydraulic, mechanical vibration, electrical and equipment auxiliary status parameters. All data are equipped with GPS timestamps to achieve global time sequence alignment. Step S2: Preprocess and clean the raw data, remove non-steady and dirty data, and select effective steady-state data with head fluctuation ≤ ±0.5m, load fluctuation ≤ ±2%, and continuous steady-state duration ≥ 5min as the benchmark sample for online model update; Step S3: Construct a standard dataset using historical data of the unit's healthy steady state without defects after overhaul, pre-train and build a multi-head interval operating condition feature fingerprint database, and rely on four quantitative indicators: head deviation rate, pressure fluctuation variance, data signal-to-noise ratio, and operating condition probability. Adaptively divide the operating head into three types of operating condition intervals: high-frequency stable, medium-frequency disturbance, and low-frequency strong disturbance through an unsupervised clustering algorithm, and extract the effective data time-domain and frequency-domain features. Use the feature vector cosine similarity algorithm to calculate the sample feature matching degree of the corresponding interval. Step S4: Construct a dual-layer coupled logical architecture of working condition confidence quantification and time series state evolution, adaptively solve the optimal sample capacity for each head interval, combine the sample feature matching degree to complete the steady-state sample confidence weighting conversion, and obtain the standardized effective sample duration; and analyze the time series evolution law through the residual time series second-order differential algorithm, and decouple the equipment natural aging component and fault abnormal component by mechanism. Step S5: Set the confidence trigger coefficient to construct the model iteration admission threshold. After the effective sample duration reaches the target, only the local update of the corresponding head interval sub-model is triggered. The gradient is quantized and the model update cycle is adaptively tuned according to the interval operating condition disturbance intensity coefficient. Combined with the time residual time sequence analysis in step S4, the iteration permissions are distinguished. Natural aging components are allowed to participate in baseline update and the iteration behavior of blocked fault components are allowed. With the help of a multi-dimensional safety blocking mechanism, abnormal iterations are intercepted, and the adaptive calibration update of the benchmark model of each head interval is completed.

2. The method for monitoring the condition of hydropower equipment based on feature recognition according to claim 1, characterized in that: The historical full-condition archived data consists of the unit's full-condition operation data spanning the flood season and dry season over the past three years, including vibration swing time domain data, pressure pulsation high-frequency waveform data, active and reactive load data, net water head, flow rate through the unit, tailwater aeration parameters, bearing temperature, guide vane opening, and unit speed parameters; the real-time operation data consists of hydraulic, mechanical, electrical, and auxiliary system status data collected in milliseconds by the LCU and high-frequency acquisition device, and automatically distinguishes between power generation, phase adjustment, and no-load operation modes.

3. The method for monitoring the condition of hydropower equipment based on feature recognition according to claim 1, characterized in that: The non-steady-state dirty data removed during preprocessing specifically includes data from start-up and shutdown transitions, load step adjustments, high-frequency fluctuations during primary frequency modulation, gas injection switching moments, load shedding tests, sensor zero-point drift, and communication packet loss / distortion.

4. The method for monitoring the condition of hydropower equipment based on feature recognition according to claim 1, characterized in that: Step S3 includes the following specific steps: Step S31: Perform steady-state screening and denoising on the standard dataset to remove abnormal samples and retain clean and healthy steady-state data. Then, for the pre-processed steady-state data of each head interval, extract the time-domain feature mean, time-domain variance, steady-state amplitude, frequency domain main frequency, vortex band characteristic frequency, and spectral energy ratio of each sample to form a multi-dimensional original feature vector. Then, perform mean clustering convergence calculation on the massive original feature vectors under the same head condition to remove discrete abnormal feature samples that deviate from the cluster distribution. After iterative convergence, solidify the standardized steady-state feature fingerprint vector that uniquely corresponds to each head interval. Associate and bind the standardized steady-state feature fingerprint vector corresponding to each head interval with the corresponding working condition interval label, working condition disturbance attribute, and operating parameter range, and store it in a structured database to form a multi-head interval exclusive working condition feature fingerprint database that is partitioned independently, has unique features, and can be matched and compared, thus completing the overall construction of the fingerprint database. Step S32: Obtain the real-time net head and rated head data from the real-time operation data record, calculate the head deviation rate, head deviation rate = (real-time net head - rated head) / rated head * 100%; obtain the interval pressure pulsation variance as the second-order statistical variance of the single-interval pressure pulsation time series data; obtain the data signal-to-noise ratio as the ratio of effective steady-state signal energy to noise signal energy, and the operating condition probability as the percentage of the steady-state operating time of a single head interval to the total steady-state operating time of the unit; after normalizing the above four quantitative indicators, construct a multi-dimensional operating condition disturbance feature matrix, and input the feature matrix as a whole into an unsupervised clustering algorithm; During the algorithm execution, the preset number of cluster categories is three. The iterative convergence optimization objective is to minimize the overall variance of samples within each category and maximize the difference in features between categories. When the change in cluster centers is less than the preset convergence threshold, the algorithm iteration is determined to end, and the adaptive unlabeled grouping of all water head condition samples is completed. Step S33: The algorithm automatically outputs three groups of characteristic differential working condition sample clusters. Based on the inherent working condition characteristics of the four quantitative parameters corresponding to each group of clusters, the interval classification is completed. Among them, the sample cluster with the smallest pressure fluctuation variance, the highest signal-to-noise ratio, the small head deviation rate, and the highest working condition probability is classified as the high-frequency stable working condition interval. The sample cluster with moderate pressure fluctuation variance, signal-to-noise ratio, head deviation rate, and operating probability is classified as the medium-frequency disturbance operating condition range; the sample cluster with the largest pressure fluctuation variance, lowest signal-to-noise ratio, large head deviation rate, and low operating probability is classified as the low-frequency strong disturbance operating condition range. Step S34: Predefine the standardized steady-state feature fingerprint of each head interval as the reference vector A, and extract the same dimension working condition feature from the real-time effective steady-state data as the test vector B, and set the total number of feature dimensions to n; where the reference vector A=[A1,A2,…,An], A1,A2,…,An correspond to the standard feature values ​​of the interval standard time domain mean, time domain variance, steady-state amplitude, frequency domain dominant frequency, eddy band characteristic frequency, and spectral energy ratio, respectively; The vector to be tested is B = [B1, B2, ..., Bn], where B1, B2, ..., Bn correspond to the measured values ​​of the working condition features of each dimension extracted synchronously from the real-time samples. The feature matching degree η of the samples is calculated using the cosine similarity formula, η = (A·B) / (||A||×||B||), where A·B is the dot product of the reference vector and the vector to be tested, which is obtained by multiplying the corresponding feature values ​​of each dimension and summing them. ||A|| is the arithmetic square root of the sum of squares of the feature values ​​of each dimension of the reference vector. ||B|| is the arithmetic square root of the sum of squares of the feature values ​​of each dimension of the vector to be tested.

5. The method for monitoring the condition of hydropower equipment based on feature recognition according to claim 1, characterized in that: Step S4 includes the following: Step S41: The dual-layer coupled logic architecture includes a working condition confidence quantization logic layer and a time-series state evolution logic layer. The two layers operate and are coupled and linked at each level. The working condition confidence quantization logic layer is used to solve for the optimal sample size of the working condition and to perform confidence weighted purification of the samples. The time-series state evolution logic layer is used to decouple the mechanism of equipment natural aging and instantaneous failure. Step S42: The working condition confidence quantization logic layer calculates the adaptive optimal effective sample capacity Ni corresponding to the i-th type of head interval through the optimal sample size adaptive algorithm. Ni=Nmin×(σi / σ0)×(1 / SNRi)×(1 / Pi 1 / 2 ); Where Nmin represents the minimum basic sample size under the healthy and stable baseline operating condition of the unit, σi represents the comprehensive standard deviation calculated by fusing multi-dimensional data on pressure pulsation and vibration amplitude of the unit within the i-th target head interval, σ0 represents the fixed baseline value of the comprehensive standard deviation corresponding to the high-frequency stable baseline head interval, SNRi is the signal-to-noise ratio of the steady-state data in the i-th target head interval, and Pi is the proportion of annual steady-state operating time in the i-th target head interval. Step S43: After determining the optimal effective sample capacity Ni for the i-th type of head interval, further complete the real-time sample quality correction calculation to obtain the standardized effective sample duration Ti that can be used for threshold comparison. The conversion formula is Ti = Tr × ηi; where Tr is the original continuous steady-state sample duration obtained by real-time screening under the i-th type of working condition, and ηi is the feature matching degree of the corresponding sample under the i-th type of working condition.

6. The method for monitoring the condition of hydropower equipment based on feature recognition according to claim 1, characterized in that: The timing state evolution logic layer in step S4 also includes the following: Collect the difference between real-time monitoring parameters and the corresponding benchmark parameters of the working condition interval, and construct a continuous time-series residual sequence with uniform sampling at equal intervals; perform first-order difference operation and second-order differential operation on the time-series residual sequence frame by frame. The first-order difference is used to characterize the temporal change slope of the residual sequence, and the second-order differential is used to characterize the change curvature and instantaneous change intensity of the residual sequence. A fixed first-order difference stationarity threshold range and a second-order derivative zero reference threshold are pre-configured. The residual difference data of a consecutive preset number of frames are statistically judged. If all first-order difference calculation results in the frame segment fall into the preset first-order difference stationarity threshold range, and the absolute value of all second-order derivative calculation results in the corresponding frame segment is not greater than the preset second-order derivative zero reference threshold, the residual sequence as a whole shows a uniform unidirectional shift change pattern without abrupt inflection points or shock fluctuations. The time series component of this segment is quantitatively determined to be the natural aging component of the device, and the baseline iteration update permission for the corresponding time period sample is opened. If the first-order difference calculation result of a single frame or a few consecutive frames exceeds the first-order difference stationarity threshold range, and the absolute value of the second-order differential calculation result of the corresponding frame point is greater than the preset second-order differential zero reference threshold, and the residual sequence shows a numerical step jump, local abrupt change inflection point or instantaneous shock fluctuation pattern, the time series component of this segment is quantitatively determined to be an abnormal component of equipment defect or early fault, and the model iteration permission of the sample in this segment is locked.

7. The method for monitoring the condition of hydropower equipment based on feature recognition according to claim 1, characterized in that: Step S5 includes the following: The model iteration admission threshold is K*Ni, where K is a 0.95 confidence trigger coefficient. When the interval weighted effective sample duration Ti meets the standard, the standard means that the standardized effective sample duration Ti calculated in real time for the i-th type of head interval is greater than or equal to the iteration admission threshold K×Ni. Only the local update of the corresponding interval sub-model is triggered, without interfering with the global model. The interval sub-model refers to a pre-constructed exclusive working condition feature fingerprint benchmark model that is independently matched to each type of head working condition interval, including independent sub-models corresponding to three types of working condition intervals: high-frequency stable, medium-frequency disturbance, and low-frequency strong disturbance. Define the disturbance intensity coefficient Kdi, Kdi=(σi / σ0)×(1 / SNRi), and construct the update cycle quantification formula Li, Li=Tb×Kdi×ai, where Tb represents the baseline update cycle and ai represents the correction coefficient for the operating probability of the i-th type of head interval.

8. A feature-based hydropower equipment condition monitoring system, using the feature-based hydropower equipment condition monitoring method according to any one of claims 1-7, characterized in that: The system includes a data acquisition module, a data preprocessing module, a working condition feature database and partitioning module, a two-layer coupled quantization analysis module, and an adaptive iterative calibration and safety interlocking module. The data acquisition module is used to connect to various monitoring systems in the power plant, and synchronously collect historical archived data and real-time operating data of the unit, covering hydraulic, mechanical, electrical and auxiliary equipment status parameters. After GPS time-series alignment, the data is used for model pre-training statistics and online sample monitoring accumulation, respectively. The data preprocessing module cleans the raw data, removes non-steady-state data, and selects valid samples that meet steady-state constraints as the benchmark for model iteration and update. The operating condition feature database construction and partitioning module builds a standard dataset based on the unit overhaul health steady-state data, extracts time-frequency domain features after purification, and solidifies the standardized steady-state fingerprint vectors of each head interval through mean clustering and stores them in the database to build a multi-operating condition feature fingerprint database. It also constructs a perturbation feature matrix by combining multiple types of operating condition quantitative indicators, adaptively divides multiple head operating condition intervals through constrained unsupervised clustering, and quantifies the sample feature matching degree based on cosine similarity. The dual-layer coupled quantization analysis module has a dual-layer logical architecture, which can adaptively solve the optimal sample capacity for each working condition, obtain the standardized effective sample duration by combining feature matching degree correction, and decouple the equipment natural aging component and the fault abnormal mutation component through the first and second order differential quantization discrimination mechanism of time residual. The adaptive iterative calibration and safety interlocking module sets the iterative admission threshold by setting a fixed confidence coefficient. After the sample meets the standard, only the corresponding interval sub-model is locally updated. The update cycle is adaptively adjusted according to the working condition disturbance gradient. It differentiates the control of iteration permissions, allowing the aging component to update the baseline and the interlocking fault component to iterate. With the help of the multi-dimensional safety interlocking mechanism, it realizes the adaptive calibration of the multi-working condition benchmark model and completes the status monitoring of hydropower equipment.

Images