Real-time monitoring system for operating state of hydraulic unit based on vibration signals of multiple measuring points

Abstract

This invention relates to the field of hydropower equipment condition monitoring technology, specifically a real-time monitoring system for the operating status of hydropower units based on multi-point vibration signals. The system includes modules for multi-point vibration signal acquisition, data preprocessing, extraction of spatiotemporal coupling features from multiple measurement points, dynamic updating of adaptive thresholds for operating conditions, status identification, and real-time prediction and early warning of vibration degradation trends. This invention deploys a multi-point acquisition architecture along the vibration transmission path, constructs spatial weight coefficients based on vibration transmission characteristics to extract spatiotemporal coupling features, dynamically updates early warning thresholds by associating with unit operating parameters, achieves advanced prediction of degradation trends through a time-series model, and employs an edge-end collaborative deployment architecture. This improves the accuracy of early anomaly identification and adaptability to changing operating conditions, achieving millisecond-level low-latency real-time monitoring, and is suitable for the safe management and control of the operating status of various hydropower units.

Description

Technical Field

[0001] This invention relates to the field of hydropower equipment condition monitoring technology, specifically a real-time monitoring system for the operating status of hydropower units based on multi-point vibration signals. Background Technology

[0002] With the rapid development of my country's hydropower industry, hydraulic generating units are continuously evolving towards larger capacity, higher head, and greater intelligence. Their operational stability directly affects the safe and reliable power supply of the power system. Statistical data shows that approximately 80% of faults and accidents in hydraulic generating units are characterized by vibration signals. Therefore, vibration signal monitoring is a core technical means for the operation status control, fault early warning, and safety protection of hydraulic generating units. It is of great significance for ensuring the long-term safe and stable operation of the units, reducing the risk of unplanned shutdowns, and improving the level of intelligent operation and maintenance of power plants.

[0003] Currently, technologies related to vibration monitoring of hydraulic turbine units have been widely applied. Existing monitoring schemes mostly involve deploying vibration sensors at key parts of the unit, collecting vibration signals, and uploading them to a monitoring platform for analysis and processing, thereby achieving monitoring of the unit's basic operating status and early warning of faults. However, existing technologies still have many shortcomings in practical applications and are difficult to adapt to the high-precision real-time monitoring requirements of today's large-capacity, variable-operating-condition hydraulic turbine units.

[0004] As prior art related to this invention, publication number CN102183295A discloses a sound and vibration monitoring system for hydropower units. In the technical solution of this prior art, the system as a whole consists of at least one monitoring unit and at least one monitoring center distributed within the same area. Each monitoring unit includes at least one copolymer piezoelectric sound sensing unit, at least one copolymer piezoelectric vibration sensing unit, and a processing unit. The sensing units can collect sound and vibration signals from hydropower units in the frequency band from 0.01Hz to 10kHz. The processing unit preprocesses the received signals or data from the sensing units and then transmits them to the monitoring center via wired or wireless means. The monitoring center is responsible for receiving and analyzing data from multiple monitoring units within the area. Multiple monitoring centers can exchange data to obtain sound and vibration information from all hydropower units within the area, ultimately achieving fault prediction for the hydropower units. This solution aims to achieve vibration and sound monitoring of a group of hydropower units within an area through a low-cost sensing architecture, solving some of the problems of high cost and poor scalability in traditional monitoring systems.

[0005] However, the technical solutions in the aforementioned comparative documents still have significant technical shortcomings in the actual operation monitoring of hydraulic units. Firstly, these solutions only achieve basic acquisition, preprocessing, and transmission of vibration and sound signals, without specifically addressing the multi-point spatiotemporal coupling characteristics of hydraulic unit vibration signals. Each measuring point's signal is acquired and processed independently, failing to consider the spatial correlation characteristics of the vibration transmission path. This makes it impossible to extract the combined spatiotemporal coupling characteristics of multiple measuring points, hindering accurate vibration source localization and identification of early, subtle anomalies. Furthermore, these solutions are susceptible to fluctuations in operating conditions and environmental noise interference, resulting in insufficient monitoring accuracy and anomaly identification sensitivity. Secondly, these solutions lack an adaptive early warning mechanism designed for the wide range of variable operating conditions of hydraulic units. They do not correlate core operating parameters such as unit head, speed, and output, making dynamic updates of early warning thresholds impossible. This leads to false or missed early warnings when unit operating conditions frequently change, and the solutions are unsuitable for monitoring the full operating conditions of various types of hydraulic units, including mixed-flow, impulse, and pumped-storage units. Third, this scheme can only achieve basic monitoring of the unit's current operating status, lacking a module for advanced prediction of vibration degradation trends. It cannot predict unit degradation trends based on time-series vibration characteristic sequences, making early warning of faults difficult and failing to provide forward-looking guidance for unit operation and maintenance. Fourth, this scheme adopts an architecture of sensor unit acquisition - processing unit preprocessing - centralized analysis at the monitoring center, without employing a collaborative deployment method between the edge and the host computer. A large amount of data needs to be uploaded to the monitoring center for centralized analysis, resulting in large transmission delays and insufficient real-time performance, failing to meet the application requirements for millisecond-level real-time monitoring and emergency protection of hydraulic turbine units. Summary of the Invention

[0006] The purpose of this invention is to provide a real-time monitoring system for the operating status of hydraulic turbine units based on multi-point vibration signals, so as to solve the problems mentioned in the background art, such as the lack of spatial correlation of multi-point signals, insufficient accuracy of early warning for changing operating conditions, poor real-time performance, and lack of ability to predict deterioration trends.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The real-time monitoring system for the operating status of hydraulic turbine units based on multi-point vibration signals includes a multi-point vibration signal acquisition module, a data preprocessing module, a status identification module, and a real-time early warning module. It also includes a multi-point spatiotemporal coupling feature extraction module, an operating condition adaptive threshold dynamic update module, and a vibration deterioration trend real-time prediction module. The multi-point vibration signal acquisition module deploys M triaxial vibration acceleration measurement points along the vibration transmission path of the hydraulic turbine unit, and synchronously acquires the triaxial vibration time-domain signal and the key phase pulse signal of the unit at each measurement point during unit operation. The sampling frequency is synchronously locked with the real-time rotation speed of the unit. The data preprocessing module completes the time-domain synchronous alignment of multi-channel vibration signals based on the key phase pulse signal, and sequentially completes signal detrending, power frequency harmonic filtering and outlier removal, outputting a standardized multi-measurement point synchronous vibration signal. The multi-measurement point spatiotemporal coupling feature extraction module constructs a multi-measurement point spatiotemporal feature matrix based on standardized signals, and extracts a set of spatiotemporal coupling features from multiple measurement points by combining the spatial weight coefficients of the vibration transmission path. The adaptive threshold dynamic update module for operating conditions associates the real-time head, speed, and output operating condition parameters of the unit, and dynamically updates the warning thresholds of each spatiotemporal coupling feature based on the historical normal operating condition dataset. The status recognition module completes the real-time discrimination of the unit's operating status based on the spatiotemporal coupling feature set and the real-time early warning threshold. The vibration deterioration trend real-time prediction module completes the advanced prediction of the unit's vibration deterioration trend based on the time-series feature sequence. The real-time early warning module outputs graded early warning signals according to the status discrimination results and trend prediction results.

[0008] Preferably, the M triaxial vibration acceleration measurement points of the multi-point vibration signal acquisition module are respectively arranged at vibration-sensitive locations on the upper frame (X / Y direction), lower frame (X / Y direction), top cover (vertical direction), turbine main shaft (radial direction), and tailrace inlet wall of the hydraulic turbine unit, where M ≥ 8; the sampling frequency is synchronously locked with the real-time rotational speed of the unit to satisfy the following: In the formula, For real-time sampling frequency, This is the harmonic factor, which can be 128, 256, or 512. This represents the real-time rotational speed of the unit, expressed in r / min.

[0009] Preferably, the multi-channel vibration signals in the data preprocessing module are synchronized in the time domain using a cross-correlation delay estimation method based on key phase pulses. After alignment, the delay deviation of each channel signal is no greater than one sampling period. The cross-correlation delay estimation satisfies the following: In the formula, Let be the cross-correlation function between the signal at the i-th measurement point and the key phase signal of the reference channel. The original vibration signal at the i-th measuring point. This is the key phase reference signal. The time delay offset is N, where N is the number of signal sampling points, and the value corresponding to the peak value of the cross-correlation function is taken. This is the delay compensation value for this channel.

[0010] Preferably, the spatial weighting coefficients of the multi-point spatiotemporal coupling feature extraction module are constructed based on the vibration transmission path attenuation characteristics between the measuring points and the vibration source. satisfy: In the formula, The vibration transmission attenuation coefficient is determined by taking specific values ​​for both the steel structure of the hydraulic turbine unit and the water medium. Let be the linear distance in space from the i-th measuring point to the vibration source of the unit's main shaft. Let be the total energy of the vibration signal at the i-th measuring point. This represents the total number of measurement points.

[0011] Preferably, the multi-point spatiotemporal coupling feature extraction module first performs adaptive variational mode decomposition on the standardized vibration signal of each measuring point to obtain K intrinsic mode function (IMF) components. It then extracts the effective value, peak value, kurtosis, spectral kurtosis, and 1-10 octave amplitude values ​​of each IMF component as a single-point time-domain and frequency-domain feature set. Finally, it combines spatial weighting coefficients... Construct a multi-point spatiotemporal feature matrix Where F is the feature dimension of a single measurement point, and the spatiotemporal coupled feature set is obtained through weighted fusion. ,in Let be the single-point feature vector of the i-th measurement point.

[0012] Preferably, the adaptive threshold dynamic update module first divides the historical normal operating condition dataset of the unit into multiple operating condition clusters using a density peak clustering algorithm. For each operating condition cluster, a kernel density estimation method is used to calculate the probability density distribution of each spatiotemporal coupling feature, and the upper boundary corresponding to the 99.7% confidence interval is taken as the warning threshold for the feature under that operating condition. The kernel density estimation satisfies: In the formula, The kernel density estimation function for the eigenvalues. The number of samples within the operating condition cluster. To determine the optimal bandwidth, cross-validation is used. For Gaussian kernel function, Let be the eigenvalues ​​to be estimated. It is the i-th feature sample within the operating condition cluster.

[0013] Preferably, the real-time vibration degradation trend prediction module employs a gated recurrent unit network with a spatiotemporal attention mechanism to perform temporal modeling of the spatiotemporally coupled feature set within a continuous time window, predicting the feature change trend in the next T time steps in advance; the core computation of the gated recurrent unit satisfies: In the formula, To update the gate output, To reset the gate output, This represents the hidden layer state from the previous time step. The spatiotemporal coupling features input at the current moment, It is the sigmoid activation function. This is the weight matrix. For bias terms, This represents the candidate hidden layer state. This represents the current hidden layer state.

[0014] Preferably, the status recognition module, based on a real-time spatiotemporal coupled feature set and a warning threshold corresponding to the operating condition, determines the anomaly index of the computer group's operating status. This anomaly index is calculated using weighted Mahalanobis distance and satisfies the following: In the formula, To weighted Mahalanobis distance outlier, This is a real-time spatiotemporal coupled feature vector. This represents the mean vector of features within the corresponding working condition cluster. This is a feature weight diagonal matrix, where the diagonal elements are the sensitivity coefficients of each feature to the vibration degradation of the unit. is the inverse of the characteristic covariance matrix.

[0015] Preferably, the real-time early warning module incorporates a four-level hierarchical early warning logic, corresponding to four levels: normal, early warning, alarm, and emergency shutdown. Based on the single-feature early warning threshold output by the operating condition adaptive threshold dynamic update module, the corresponding level's anomaly thresholds are sequentially defined: the early warning threshold is the baseline threshold, the alarm threshold is 1.2 times the baseline threshold, and the emergency threshold is 1.5 times the baseline threshold. Early warning triggering employs a dual-factor coupled criterion of real-time anomaly degree and degradation trend: the normal level corresponds to an anomaly degree below the early warning threshold and no degradation trend, only outputting the unit's normal operating status identifier; the early warning level corresponds to an anomaly degree exceeding the early warning threshold but not reaching the alarm threshold, and the degradation trend prediction result shows no continuous upward trend, outputting an early warning signal while simultaneously triggering high-density data acquisition and storage and event timestamp marking. The alarm level corresponds to an anomaly exceeding the alarm threshold but not reaching the emergency threshold, or a degradation trend prediction indicating that the characteristic value will exceed the alarm threshold within the next T time steps. An alarm signal is output and linked with local audible and visual prompts, and the location of the abnormal measurement point and details of the characteristic anomaly are pushed to the host computer simultaneously. The emergency shutdown warning level corresponds to an anomaly exceeding the emergency threshold, or a degradation trend prediction indicating that the characteristic value will exceed the emergency threshold within the next T time steps. An emergency shutdown warning signal is output and a hard-wired linkage interface with the unit speed control system and emergency shutdown circuit is reserved.

[0016] Preferably, the state initial judgment unit of the multi-point vibration signal acquisition module, the data preprocessing module, the multi-point spatiotemporal coupling feature extraction module, and the state recognition module are all deployed on the edge computing terminal at the unit site. The edge computing terminal adopts an IP65 protection level design, is suitable for the strong electromagnetic and high humidity operating environment of the hydraulic power plant, and has a built-in IEEE1588 precision clock synchronization unit to achieve co-source clock synchronization with the key phase pulse signal of the multi-point vibration signal acquisition module, with a synchronization error of no more than 1μs. The edge computing terminal allocates computing power and imposes latency constraints on each module. The preprocessing latency of a single frame vibration signal is ≤1ms, the latency of spatiotemporal coupling feature extraction is ≤2ms, the latency of initial state judgment is ≤1ms, and the total end-to-end processing latency is ≤5ms. After the edge terminal completes the processing, it uses a sparse coding algorithm to compress the spatiotemporal coupling feature set with a compression ratio ≥5:1. After encryption using the national cryptographic SM4 algorithm, only the compressed feature set, state judgment results, trend prediction results, and abnormal event data are uploaded to the host computer monitoring platform. At the same time, it locally redundantly stores no less than 30 days of full feature data and original event data. It has a built-in network outage resume mechanism, which automatically retransmits the monitoring data during the network outage period after the network is restored.

[0017] Compared with the prior art, the beneficial effects of the present invention are: (1) This invention addresses the technical shortcomings of existing hydraulic turbine vibration monitoring technologies, such as independent processing of signals from multiple measuring points, lack of spatial correlation, and insufficient sensitivity for early anomaly identification. Through the core design of multi-measuring point spatiotemporal coupling feature extraction, this invention achieves high-precision feature extraction and anomaly characterization of hydraulic turbine vibration signals. This invention deploys a multi-measuring point acquisition architecture along the vibration transmission path, and combines the cross-correlation delay estimation of key phase pulses to achieve accurate synchronous alignment of multi-channel signals. With the sampling mechanism of speed frequency doubling synchronization locking, the temporal synchronization and consistency of multi-measuring point data are guaranteed from the source of signal acquisition. At the same time, based on the attenuation characteristics of the vibration transmission path, a spatial weighting coefficient is constructed, and the time-domain and frequency-domain features of multiple measuring points are integrated to form a spatiotemporal coupling feature set. This breaks the technical limitations of isolated analysis of signals from each measuring point in the existing technology, effectively weakens the interference of operating condition fluctuations and environmental noise on the monitoring results, accurately strengthens the characteristic information related to the vibration source, and significantly improves the identification accuracy of early weak vibration anomalies of the unit. It solves the technical problems of the inability to accurately capture early fault characterization of the unit and weak anti-interference ability in the comparison file.

[0018] (2) This invention addresses the technical problems of existing monitoring schemes, such as fixed thresholds being unable to adapt to the wide range of variable operating conditions of the unit, poor early warning accuracy, and lack of ability to predict deterioration trends. Through the collaborative design of dynamic updating of operating condition adaptive thresholds and real-time prediction of vibration deterioration trends, it achieves accurate early warning and forward-looking control of the unit under all operating conditions. This invention uses a density peak clustering algorithm to divide historical normal operating data into operating condition clusters, and combines this with a kernel density estimation method to achieve adaptive dynamic updates of warning thresholds for each feature. This breaks through the technical constraints of fixed thresholds in existing technologies, effectively avoiding false and missed warnings caused by frequent switching of unit operating conditions. It can fully adapt to the full-condition operation monitoring needs of various types of hydraulic units, such as mixed-flow, impulse, and pumped storage units. At the same time, it uses a gated cyclic unit network with a spatiotemporal attention mechanism to construct a time-series prediction model, realizing the advanced prediction of unit vibration deterioration trends. This breaks through the technical barrier of existing technologies that can only achieve passive monitoring of the current operating state, providing forward-looking technical support for unit operation and maintenance, fundamentally improving the safety redundancy of unit operation, and filling the technical gaps in the comparative documents that lack adaptive capabilities for changing operating conditions and advanced warning functions.

[0019] (3) In view of the technical defects of the existing monitoring system’s centralized processing architecture, such as large transmission delay, insufficient real-time performance and poor field adaptability, the present invention achieves low-latency and high-reliability real-time monitoring of the vibration status of hydraulic units through the deployment architecture design of edge terminal and host computer collaboration. This invention deploys signal acquisition, data preprocessing, spatiotemporal coupling feature extraction, and initial state judgment units on the edge computing terminal at the unit's site. Through refined computing power allocation and latency constraints, it achieves millisecond-level end-to-end processing of single-frame signals, meeting the low-latency application requirements for real-time monitoring and emergency protection of hydraulic units. Simultaneously, it uses a sparse coding algorithm to compress feature data at a high rate, combined with encrypted transmission using the national cryptographic SM4 algorithm, significantly reducing data transmission bandwidth consumption and ensuring the security of monitoring data transmission. The edge terminal adopts a high-protection, strong anti-interference design, adapting to the complex operating environment of strong electromagnetic fields and high humidity in hydraulic power plants. Combined with local redundant storage and a network interruption resumption mechanism, it comprehensively ensures the integrity of monitoring data and the stability of system operation, giving the system strong engineering practicality and field adaptability. This solves the technical problems of insufficient real-time performance and poor adaptability to the field environment in the centralized analysis architecture described in the comparison documents. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are explained in detail together with the embodiments of the invention, but do not constitute a limitation thereof.

[0021] Figure 1 This is a diagram illustrating the architecture of the real-time monitoring system for the operating status of a hydraulic turbine unit based on multi-point vibration signals, as described in this invention. Figure 2 This is a flowchart of the data acquisition and preprocessing process of the present invention; Figure 3 This is a flowchart of the feature extraction and threshold update process of the present invention; Figure 4 This is a flowchart of the state recognition and trend prediction process of the present invention; Figure 5 This is a diagram illustrating the hierarchical early warning logic and system deployment of this invention. Detailed Implementation

[0022] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figures 1-5 As shown, the real-time monitoring system for the operating status of a hydraulic turbine unit based on multi-point vibration signals of the present invention is used for real-time monitoring of vibration status, anomaly identification, and prediction of deterioration trend during the full-condition operation of the hydraulic turbine unit. The system consists of a multi-point vibration signal acquisition module, a data preprocessing module, a multi-point spatiotemporal coupling feature extraction module, an operating condition adaptive threshold dynamic update module, a status identification module, a vibration deterioration trend real-time prediction module, and a real-time early warning module.

[0024] The multi-point vibration signal acquisition module is deployed along the vibration transmission path of the hydraulic turbine unit. Three-axis vibration acceleration measuring points, covering sensitive locations throughout the entire vibration transmission chain of the hydraulic turbine unit, specifically including the upper frame. Up and up rack Up and down racks Up and down racks The core sensitive locations are: the vertical direction of the top cover, the radial direction of the turbine main shaft, and the inlet wall of the tailrace pipe. The value of is not less than 8. The module synchronously acquires the three-dimensional vibration time-domain signals and the key phase pulse signals of the unit at each measuring point during operation. The sampling frequency is synchronized with the real-time speed of the unit through frequency doubling. The calculation method for frequency doubling synchronization is as follows: In the formula, For real-time sampling frequency, This is the frequency doubling factor, which can be selected as 128, 256, or 512 depending on the monitoring requirements. This refers to the real-time speed of the generator unit, expressed in r / min. When the generator unit is at its rated speed... r / min, octave coefficient When, it can be calculated Hz, to complete the synchronous locking of the sampling frequency.

[0025] The data preprocessing module performs time-domain synchronization alignment of multi-channel vibration signals based on key-phase pulse signals, sequentially performing signal detrending, power frequency harmonic filtering, and outlier removal, outputting standardized multi-point synchronized vibration signals. The time-domain synchronization alignment of multi-channel vibration signals employs a cross-correlation delay estimation method based on key-phase pulses. After alignment, the delay deviation of each channel signal is no greater than one sampling period. The calculation method for the cross-correlation delay estimation is as follows: In the formula, For the first The cross-correlation function between the signal at each measurement point and the key phase signal of the reference channel. For the first The original vibration signal at each measuring point This is the key phase reference signal. This is the time delay offset. This represents the number of signal sampling points. After calculation, the peak value of the cross-correlation function is taken. This is the delay compensation value for this channel, used to achieve time-domain synchronization and alignment of multiple channel signals. When When, the peak value of the cross-correlation function at a certain measuring point is calculated. At each sampling point, the delay compensation value for that channel is 3 sampling cycles, and signal alignment is completed.

[0026] The multi-measurement-point spatiotemporal coupling feature extraction module constructs a multi-measurement-point spatiotemporal feature matrix based on standardized signals. Combined with spatial weighting coefficients of the vibration transmission path, it extracts a joint spatiotemporal coupling feature set from the multiple measurement points. The spatial weighting coefficients are constructed based on the vibration transmission path attenuation characteristics between the measurement points and the vibration source, and are calculated as follows: In the formula, The vibration transmission attenuation coefficient is set to 0.02 for the steel structure medium of the hydraulic turbine unit and 0.08 for the water medium. For the first The straight-line distance between each measuring point and the vibration source of the unit's main shaft. For the first The total energy of the vibration signal at each measuring point This represents the total number of measurement points. When... A certain measuring point of m, steel structure medium The total energy of the signal at this measuring point The total energy at all measuring points is When, it can be calculated .

[0027] The specific implementation process of feature extraction is as follows: first, adaptive variational mode decomposition is performed on the standardized vibration signal of each measuring point to obtain... One Intrinsic Mode Function (IMF) component The value ranges from 6 to 10. The effective value, peak value, kurtosis, spectral kurtosis, and 1-10 octave amplitude values ​​of each IMF component are extracted as a single-point time-domain and frequency-domain feature set. The single-point feature dimension... Not less than 20. This is then combined with the spatial weighting coefficient. Construct a multi-point spatiotemporal feature matrix The spatiotemporal coupled feature set is obtained through weighted fusion. ,in For the first Single-point feature vector of each measurement point.

[0028] The adaptive threshold dynamic update module associates real-time head, speed, and output operating parameters of the unit, and dynamically updates the warning thresholds for each spatiotemporal coupling feature based on historical normal operating condition datasets. The specific implementation process is as follows: First, the historical normal operating condition dataset is divided into multiple operating condition clusters using a density peak clustering algorithm. For each operating condition cluster, the probability density distribution of each spatiotemporal coupling feature is calculated using a kernel density estimation method. The upper boundary corresponding to the 99.7% confidence interval is taken as the warning threshold for that feature under that operating condition. The kernel density estimation is calculated as follows: In the formula, The kernel density estimation function for the eigenvalues. The number of samples within the operating condition cluster. To determine the optimal bandwidth, cross-validation is used. For Gaussian kernel function, Let be the eigenvalues ​​to be estimated. The first in the working condition cluster A feature sample. Within a certain operating condition cluster... 1 feature sample, optimal bandwidth Gaussian kernel function When the kernel density estimate corresponding to the feature value is calculated, the upper boundary of the 99.7% confidence interval is determined to be 3.2, which serves as the warning threshold for the feature.

[0029] The status identification module, based on a spatiotemporal coupled feature set and a real-time early warning threshold, completes real-time identification of the unit's operating status. It uses a weighted Mahalanobis distance to calculate the anomaly index of the unit's operating status, and the calculation method is as follows: In the formula, To weighted Mahalanobis distance outlier, This is a real-time spatiotemporal coupled feature vector. This represents the mean vector of features within the corresponding working condition cluster. This is a feature weight diagonal matrix, where the diagonal elements are the sensitivity coefficients of each feature to the vibration degradation of the unit. is the inverse of the characteristic covariance matrix.

[0030] The real-time vibration degradation trend prediction module predicts the vibration degradation trend of the unit based on time-series feature sequences. It adopts a gated recurrent unit network with a spatiotemporal attention mechanism to perform time-series modeling of the spatiotemporally coupled feature set within a continuous time window, and predicts the future in advance. The characteristic change trend of each time step The value ranges from 10 to 100. The core calculation method of the gated loop unit is... In the formula, To update the gate output, To reset the gate output, This represents the hidden layer state from the previous time step. The spatiotemporal coupling features input at the current moment, It is the sigmoid activation function. This is the weight matrix. For bias terms, This represents the candidate hidden layer state. This represents the current hidden layer state.

[0031] The real-time early warning module outputs tiered early warning signals based on state judgment and trend prediction results. It incorporates a four-level tiered early warning logic, corresponding to four levels: normal, early warning, alarm, and emergency shutdown. Using the single-feature early warning threshold output by the adaptive threshold dynamic update module as a benchmark, corresponding anomaly thresholds are sequentially defined for each level: the early warning threshold is the benchmark threshold, the alarm threshold is 1.2 times the benchmark threshold, and the emergency threshold is 1.5 times the benchmark threshold. Early warning triggering employs a dual-factor coupled criterion of real-time anomaly degree and degradation trend. The normal level corresponds to an anomaly degree below the early warning threshold and no degradation trend, outputting only a normal unit operation status indicator. The early warning level corresponds to an anomaly degree exceeding the early warning threshold but not reaching the alarm threshold, and the degradation trend prediction result shows no continuous upward trend; simultaneously, an early warning signal is output, triggering high-density data acquisition and storage, and event timestamp marking. The alarm level corresponds to an anomaly degree exceeding the alarm threshold but not reaching the emergency threshold, or the degradation trend prediction shows future... If the feature value exceeds the alarm threshold within a given time step, an alarm signal will be output and a local audio-visual alert will be triggered. Simultaneously, the location of the abnormal measurement point and details of the anomaly will be pushed to the host computer. An emergency shutdown warning level corresponds to an anomaly exceeding the emergency threshold, or a deterioration trend prediction will be displayed for future... Within a certain time step, the characteristic value will exceed the emergency threshold, outputting an emergency shutdown warning signal and reserving a hard-wired linkage interface with the unit speed control system and emergency shutdown circuit.

[0032] The system adopts a collaborative deployment architecture between the edge terminal and the host computer. The multi-point vibration signal acquisition module, data preprocessing module, multi-point spatiotemporal coupling feature extraction module, and state identification module's initial state judgment unit are all deployed on the edge computing terminal at the unit's site. The edge computing terminal is designed with IP65 protection, adapting to the strong electromagnetic and high humidity operating environment of the hydraulic power plant. It has a built-in IEEE1588 precision clock synchronization unit, achieving clock synchronization with the key-phase pulse signal of the multi-point vibration signal acquisition module, with a synchronization error not exceeding 1μs. The edge computing terminal allocates computing power and imposes latency constraints on each module. The preprocessing latency for a single frame of vibration signal does not exceed 1ms, the spatiotemporal coupling feature extraction latency does not exceed 2ms, the initial state judgment latency does not exceed 1ms, and the total end-to-end processing latency does not exceed 5ms. After the edge terminal completes the processing, it uses a sparse coding algorithm to compress the spatiotemporal coupled feature set with a compression ratio of no less than 5:1. After encryption using the national cryptographic SM4 algorithm, it only uploads the compressed feature set, status discrimination results, trend prediction results, and abnormal event data to the host computer monitoring platform. At the same time, it locally redundantly stores no less than 30 days of full feature data and original event data. It has a built-in network outage resume mechanism, which automatically retransmits the monitoring data during the network outage period after the network is restored.

[0033] Example 1: Implementation of Real-time Monitoring of Conventional Operating Conditions of Mixed-flow Hydropower Units This embodiment applies to a mixed-flow hydro-generator unit with a rated capacity of 200MW and a rated speed of [missing information]. With a rated head of 120m and a speed of r / min, the unit operates under grid base load year-round, with relatively small fluctuations in operating conditions. The core requirement is to achieve millisecond-level real-time monitoring of the unit's vibration status and early warning of anomalies.

[0034] Measurement point layout and signal acquisition: Eight triaxial vibration acceleration measurement points were set up along the vibration transmission path of the unit, namely, upper frame X-axis, upper frame Y-axis, lower frame X-axis, lower frame Y-axis, top cover vertical direction, turbine main shaft radial X-axis, turbine main shaft radial Y-axis, and tailrace pipe inlet wall. Set the multiplier. The real-time sampling frequency is calculated based on the unit's real-time speed. When the unit is at its rated speed... At r / min Hz, achieving frequency doubling synchronization locking between sampling frequency and rotational speed, synchronously acquiring triaxial vibration signals and key phase pulse signals from each measuring point, and the number of sampling points per frame signal. .

[0035] Data preprocessing is based on the key phase pulse signal, and the cross-correlation delay estimation method is used to achieve time-domain synchronization and alignment of the vibration signals from eight channels. For the signal from the third lower frame X-axis measuring point, its cross-correlation function with the key phase reference signal is calculated. The peak value of the cross-correlation function was calculated. One sampling period is used as the time delay compensation value to complete the signal alignment of the channel. After alignment, the time delay deviation of all channels is less than one sampling period. After alignment, signal detrending, 50Hz power frequency and its harmonics filtering, and 3σ criterion outlier removal are performed in sequence to output a standardized multi-measurement point synchronous vibration signal.

[0036] Spatiotemporal coupling feature extraction performs adaptive variational mode decomposition on the standardized signal at each measurement point, setting the number of decomposition layers. Eight IMF components were obtained, and the effective value, peak value, kurtosis, spectral kurtosis, and 1st-10th octave amplitude values ​​of each IMF component were extracted. Single-point feature dimensions were also calculated. Calculate the spatial weighting coefficient for each measuring point, and for steel structure media, take the vibration transmission attenuation coefficient. The spatial distance from the vertical measuring point on the top cover to the main shaft vibration source. m, the total energy of the vibration signal at the measuring point The total signal energy at all 8 measuring points Then the spatial weight coefficient of the measuring point The spatial weight coefficients of all eight measuring points were calculated in the same manner to construct a multi-measuring-point spatiotemporal feature matrix. Weighted fusion yields a spatiotemporal coupled feature set. .

[0037] The adaptive threshold update collects normal operating data of the unit over the past 12 months, correlates it with head, speed, and output parameters, and divides it into 5 operating condition clusters using a density peak clustering algorithm. For the rated output operating condition cluster, the probability density distribution of each spatiotemporal coupling feature is calculated using a kernel density estimation method, and the upper boundary of the 99.7% confidence interval is taken as the warning threshold. The number of samples within each operating condition cluster is determined based on the effective vibration value characteristics of the vertical measuring points on the top cover. Optimal bandwidth The calculated warning threshold for this feature is 3.2 mm / s. Based on this, the alarm threshold is calibrated as follows: mm / s, emergency threshold is mm / s.

[0038] Real-time calculation of weighted Mahalanobis distance anomaly for state recognition and trend prediction Real-time feature vector at a certain moment with the mean vector of the rated operating condition cluster The deviation vector is Feature weight diagonal matrix The diagonal elements are the sensitivity coefficients of each feature, and the inverse of the covariance matrix is... Calculated based on historical data, the final result is... If the value is less than the warning threshold of 3.2, the unit is determined to be in normal operating condition. Simultaneously, a gated recurrent unit network with a spatiotemporal attention mechanism is employed. Inputting the spatiotemporally coupled feature set of the past 60 time steps, it predicts the feature change trend for the next 30 time steps. The prediction results show that the feature values ​​do not show a continuous upward trend and there is no risk of degradation.

[0039] Early warning output and deployment: The system edge computing terminal is deployed in the local control cabinet of the unit. The total processing latency of a single frame signal is 3.2ms, which meets the requirements of millisecond-level real-time monitoring. After the edge terminal completes the processing, it encrypts and uploads the feature data and status results to the host computer monitoring platform. The full amount of data is stored locally redundantly. During this monitoring process, the unit status was normal, and only the normal operation status indicator was output. No early warning signal was triggered.

[0040] Example 2: Implementation of Variable Operating Condition Monitoring for High-Head Impulse Hydropower Units This embodiment is applied to a high-head impulse hydropower unit with a rated capacity of 80MW. The unit has a rated head of 620m and a rated speed of 500r / min. The unit undertakes the task of grid peak shaving. The operating conditions change frequently every day, and the output fluctuates from 10% to 100% of the rated output. The core requirement is to realize adaptive state monitoring and anomaly identification under changing operating conditions to avoid false alarms caused by operating condition fluctuations.

[0041] For the vibration transmission characteristics of impulse hydroelectric generators, 10 triaxial vibration acceleration measuring points were set up along the vibration transmission path of the unit. These points were located at: upper frame (X-axis), upper frame (Y-axis), lower frame (X-axis), lower frame (Y-axis), vertical direction of the turbine base, radial direction of the turbine main shaft (X-axis), radial direction of the turbine main shaft (Y-axis), X-axis of the nozzle mechanism, Y-axis of the nozzle mechanism, and the tailrace branch pipe wall. Set the multiplier. Rated speed of the unit At r / min, the real-time sampling frequency Hz, the sampling frequency is randomly adjusted in real time by the rotational speed of the group to achieve frequency doubling and synchronous locking, and the triaxial vibration signal and key phase pulse signal of each measuring point are collected synchronously. The number of sampling points per frame signal is... It is adapted to the signal acquisition requirements of high-speed generator units.

[0042] Data preprocessing addresses the non-stationary characteristics of signals under varying operating conditions. It achieves time-domain synchronization and alignment of multi-channel signals based on the key phase pulse signal. For the X-axis measuring point signal of the nozzle mechanism, it calculates the cross-correlation function between the signal and the key phase reference signal. The cross-correlation peak corresponding to the calculated value is obtained. Each sampling period completes time delay compensation and signal alignment, ensuring that the time delay deviation of each channel is less than one sampling period after alignment. After alignment, adaptive filtering is used to remove power frequency harmonics, and a sliding window is used to de-trend the signal and remove outliers, outputting a standardized multi-point synchronous vibration signal to meet the needs of non-stationary signal processing under varying operating conditions.

[0043] Spatiotemporal coupling feature extraction performs adaptive variational mode decomposition on the standardized signal at each measurement point, setting the number of decomposition layers. Ten IMF components were obtained, and the effective value, peak value, kurtosis, spectral kurtosis, 1st-10th octave amplitude, and envelope entropy of each IMF component were extracted. The single-point feature dimension was also calculated. Calculate the spatial weighting coefficients for each measuring point, taking into account the steel structure medium. For the measuring point of the nozzle mechanism, the spatial distance from it to the main shaft vibration source m, the total energy of the signal at this measuring point The total energy of all 10 measuring points Then the spatial weight coefficient of the measuring point The spatial weight coefficients of all measuring points are calculated in the same manner to construct a multi-measuring-point spatiotemporal feature matrix. The weighted fusion yields a spatiotemporal coupling feature set, which weakens the feature disturbances caused by chemical condition fluctuations and strengthens the feature information related to the vibration source.

[0044] The adaptive threshold update system collects six months of full-condition normal operation data from the generator unit, correlates it with parameters such as head, speed, output, and nozzle opening, and divides it into 12 operating condition clusters using a density peak clustering algorithm, covering the full operating range from 10% to 100% of rated output. For each operating condition cluster, the probability density distribution of each feature is calculated using a kernel density estimation method, and the upper boundary of the 99.7% confidence interval is taken as the warning threshold for the corresponding operating condition. For the 30% rated output operating condition cluster, the number of samples for the vibration kurtosis feature of the nozzle mechanism measuring points is... Optimal bandwidth The calculated warning threshold for this feature is 4.2, and the calibrated alarm threshold is... The emergency threshold is The system tracks the unit's operating parameters in real time and automatically switches the warning thresholds for the corresponding operating condition clusters, achieving adaptive threshold updates under changing operating conditions and avoiding false warnings caused by operating condition switching.

[0045] During the process of switching the unit from 80% rated output to 30% rated output, the system matches the warning threshold of the 30% rated output operating condition cluster in real time and calculates the weighted Mahalanobis distance anomaly. After calculating the deviation vector between the real-time feature vector and the mean vector of the operating condition cluster, we obtain... The value exceeded the warning threshold of 4.2 but did not reach the alarm threshold of 5.04. Simultaneously, a gated recurrent unit network with a spatiotemporal attention mechanism was employed. Inputting the feature sequence of the past 100 time steps, it predicted the feature change trend for the next 50 time steps. The prediction results showed that the feature value did not show a continuous upward trend, indicating that the fluctuation was due to a change in operating conditions, and there was no abnormal degradation.

[0046] Early Warning Output and Deployment: The system's edge computing terminal adopts a wide-temperature design, adapting to the low-temperature operating environment of high-head power stations. The total processing latency for a single frame signal is 4.1ms, meeting real-time monitoring requirements. After completing feature extraction and initial status judgment, the edge terminal uploads the compressed and encrypted feature data and status results to the host computer. During this operating condition switch, the system adaptively switches the threshold, without triggering false early warnings, and only outputs a normal operating status indicator, ensuring the accuracy of monitoring under changing operating conditions.

[0047] Example 3: Implementation of Two-Way Operation Status Monitoring for Pumped Storage Units This embodiment applies to a pumped storage unit with a rated capacity of 300MW. The unit can operate in both power generation and pumping modes. The rated speed is 500r / min, the rated head in power generation mode is 430m, and the rated head in pumping mode is 440m. The unit frequently switches between power generation, pumping, and phase adjustment modes every day. The core requirement is to achieve vibration monitoring and early prediction of deterioration trends under both modes to ensure the safety of the unit's bidirectional operation.

[0048] For the vibration characteristics of the pumped storage unit operating in both directions, 12 triaxial vibration acceleration measurement points were set up along the two-way vibration transmission path of the unit. These points are located at: upper frame (X-axis), upper frame (Y-axis), lower frame (X-axis), lower frame (Y-axis), top cover (vertical direction), turbine main shaft (radial X-axis), turbine main shaft (radial Y-axis), movable guide vane mechanism, tailrace inlet wall, pressure steel pipe inlet wall, pump-turbine spiral casing wall, and thrust bearing housing. Set the multiplier. Rated speed of the unit At r / min, the real-time sampling frequency The sampling frequency (Hz) is randomly adjusted in real time according to the rotational speed and operating conditions to achieve frequency doubling synchronization locking under bidirectional operating conditions. It synchronously acquires triaxial vibration signals and bidirectional key phase pulse signals from each measuring point, with a single frame signal sampling point count of [number missing]. It is suitable for high-frequency signal acquisition requirements under bidirectional operating conditions.

[0049] Data preprocessing addresses the signal characteristics of bidirectional operation of pumped storage units. Time-domain synchronization and alignment of multi-channel signals are achieved based on bidirectional key-phase pulse signals. Reference key-phase channels are set up for both power generation and pumping operations. For the volute wall measuring point signals, the cross-correlation function between the signals and the key-phase reference signals under power generation conditions is calculated. The cross-correlation peak corresponding to the calculated value is obtained. Each sampling period is used to complete time delay compensation and signal alignment. After alignment, the time delay deviation of each channel is less than one sampling period. After alignment, adaptive notch filtering is used to filter out power frequency and harmonic frequencies according to the different spectral characteristics of bidirectional operating conditions. Signal detrending and outlier removal are completed based on a sliding window, and standardized multi-point synchronous vibration signals are output for power generation and pumping conditions, respectively.

[0050] Spatiotemporal coupling feature extraction performs adaptive variational mode decomposition on the standardized signal at each measurement point, setting the number of decomposition layers. Eight IMF components were obtained, and the effective value, peak value, kurtosis, spectral kurtosis, 1st-10th octave amplitude, envelope entropy, and correlation coefficient of each IMF component were extracted. Single-point feature dimensions were also calculated. For both power generation and pumping operations, spatial weighting coefficients for each measuring point were calculated. Under power generation conditions, the vibration source was primarily the turbine runner, with the steel structure serving as the medium. water medium Spatial distance from the measuring point on the volute wall to the vibration source of the rotor m, the total energy of the signal at this measuring point The total energy of all 12 measuring points Then, under the power generation condition, the spatial weighting coefficient of this measuring point is... Under pumping conditions, the vibration source is mainly the pump impeller. The spatial weighting coefficients of each measuring point are recalculated, and a multi-measuring-point spatiotemporal feature matrix corresponding to the operating condition is constructed. The spatiotemporal coupling feature set of the corresponding working condition is obtained by weighted fusion, which is adapted to the vibration transmission characteristics of the bidirectional working condition.

[0051] The adaptive threshold update system collects normal operating data from the past 12 months for the generating, pumping, and phasing operation modes of the unit. This data is correlated with operating mode, head / lift, speed, and output / power parameters, and divided into 20 operating mode clusters using a density peak clustering algorithm, covering the entire bidirectional operating mode range. For each operating mode cluster, the probability density distribution of each feature is calculated using kernel density estimation, and the upper boundary of the 99.7% confidence interval is taken as the warning threshold for the corresponding operating mode. For the rated output operating mode cluster, the number of samples for the effective value features of the vertical vibration at the top cover measuring point is calculated. Optimal bandwidth The calculated warning threshold for this feature is 3.5 mm / s, and the calibrated alarm threshold is... mm / s, emergency threshold is mm / s. For the rated head operating condition cluster of pumping conditions, the warning threshold for the same characteristic is calibrated to 3.8 mm / s. The system identifies the unit's operating condition in real time and automatically switches the warning threshold of the corresponding operating condition cluster, realizing adaptive threshold update under bidirectional operating conditions.

[0052] Status identification and trend prediction: When the unit is operating at its rated head under pumping conditions, the system matches the early warning threshold of the pumping conditions in real time and calculates the weighted Mahalanobis distance anomaly. After calculating the deviation vector between the real-time feature vector and the mean vector of the rated operating condition cluster of pumping conditions, the following is obtained: The value exceeded the warning threshold of 3.8 but did not reach the alarm threshold of 4.56. Simultaneously, a gated recurrent unit network with a spatiotemporal attention mechanism was employed. The spatiotemporal coupled feature sequence of the past 80 time steps was input to predict the feature change trend for the next 60 time steps. The prediction results showed that within the next 40 time steps, the feature value would exceed the alarm threshold of 4.56, indicating a vibration deterioration trend in the unit.

[0053] Early Warning Output and Deployment: The system's edge computing terminal adopts a design resistant to strong electromagnetic interference, adapting to the complex electromagnetic environment of pumped storage power stations. The total processing latency for a single frame signal is 3.8ms, meeting real-time monitoring requirements. Based on the predicted degradation trend, the system outputs alarm signals, triggers local audible and visual alerts, pushes the location of abnormal measuring points and detailed characteristic anomalies to the host computer, and simultaneously triggers high-density raw data acquisition and storage. This provides data support for subsequent fault investigation, achieving advanced early warning of unit degradation trends and preventing the fault from escalating.

[0054] This invention addresses the shortcomings of existing hydraulic turbine vibration monitoring technologies, such as independent processing of multi-measurement point signals, lack of spatial correlation, and insufficient sensitivity in early anomaly identification. Through a core design of multi-measurement point spatiotemporal coupling feature extraction, it achieves high-precision feature extraction and anomaly characterization of hydraulic turbine vibration signals. The invention deploys a multi-measurement point acquisition architecture along the vibration transmission path, combining cross-correlation delay estimation of key phase pulses to achieve precise synchronization and alignment of multi-channel signals. Coupled with a speed-frequency-doubled synchronization locking sampling mechanism, it ensures the temporal synchronization and consistency of multi-measurement point data from the signal acquisition source. Simultaneously, it constructs spatial weighting coefficients based on the attenuation characteristics of the vibration transmission path, fusing the time-domain and frequency-domain features of multiple measurement points to form a spatiotemporal coupling feature set. This breaks through the limitations of isolated analysis of signals from each measurement point in existing technologies, effectively weakening the interference of operating condition fluctuations and environmental noise on monitoring results, accurately enhancing the characteristic information related to the vibration source, and significantly improving the identification accuracy of early weak vibration anomalies in the unit. It also solves the technical problems of inaccurate capture of early fault characterization and weak anti-interference capability in comparison files.

[0055] This invention addresses the technical problems of existing monitoring schemes, such as fixed thresholds being unable to adapt to the wide range of variable operating conditions of the unit, poor early warning accuracy, and lack of ability to predict deterioration trends. Through the collaborative design of dynamic threshold updates adapted to operating conditions and real-time prediction of vibration deterioration trends, this invention achieves accurate early warning and forward-looking control of the unit under all operating conditions. This invention uses a density peak clustering algorithm to divide historical normal operating data into operating condition clusters, and combines this with a kernel density estimation method to achieve adaptive dynamic updates of warning thresholds for each feature. This breaks through the technical constraints of fixed thresholds in existing technologies, effectively avoiding false and missed warnings caused by frequent switching of unit operating conditions. It can fully adapt to the full-condition operation monitoring needs of various types of hydraulic units, such as mixed-flow, impulse, and pumped storage units. At the same time, it uses a gated cyclic unit network with a spatiotemporal attention mechanism to construct a time-series prediction model, realizing the advanced prediction of unit vibration deterioration trends. This breaks through the technical barrier of existing technologies that can only achieve passive monitoring of the current operating state, providing forward-looking technical support for unit operation and maintenance, fundamentally improving the safety redundancy of unit operation, and filling the technical gaps in the comparative documents that lack adaptive capabilities for changing operating conditions and advanced warning functions.

[0056] This invention addresses the technical shortcomings of existing monitoring systems, such as large transmission latency, insufficient real-time performance, and poor field adaptability due to their centralized processing architecture. By designing a deployment architecture that coordinates edge terminals with host computers, it achieves low-latency, high-reliability real-time monitoring of the vibration status of hydraulic turbine units. This invention deploys signal acquisition, data preprocessing, spatiotemporal coupling feature extraction, and initial state judgment units on the edge computing terminal at the unit's site. Through refined computing power allocation and latency constraints, it achieves millisecond-level end-to-end processing of single-frame signals, meeting the low-latency application requirements for real-time monitoring and emergency protection of hydraulic units. Simultaneously, it uses a sparse coding algorithm to compress feature data at a high rate, combined with encrypted transmission using the national cryptographic SM4 algorithm, significantly reducing data transmission bandwidth consumption and ensuring the security of monitoring data transmission. The edge terminal adopts a high-protection, strong anti-interference design, adapting to the complex operating environment of strong electromagnetic fields and high humidity in hydraulic power plants. Combined with local redundant storage and a network interruption resumption mechanism, it comprehensively ensures the integrity of monitoring data and the stability of system operation, giving the system strong engineering practicality and field adaptability. This solves the technical problems of insufficient real-time performance and poor adaptability to the field environment in the centralized analysis architecture described in the comparison documents.

[0057] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A real-time monitoring system for the operating status of a hydraulic turbine unit based on multi-point vibration signals, comprising a multi-point vibration signal acquisition module, a data preprocessing module, a status identification module, and a real-time early warning module, characterized in that, It also includes a multi-point spatiotemporal coupling feature extraction module, an adaptive threshold dynamic update module for working conditions module, and a real-time prediction module for vibration degradation trend; The multi-point vibration signal acquisition module deploys M triaxial vibration acceleration measurement points along the vibration transmission path of the hydraulic turbine unit, and synchronously acquires the triaxial vibration time-domain signal and the key phase pulse signal of the unit at each measurement point during unit operation. The sampling frequency is synchronously locked with the real-time rotation speed of the unit. The data preprocessing module completes the time-domain synchronous alignment of multi-channel vibration signals based on the key phase pulse signal, and sequentially completes signal detrending, power frequency harmonic filtering and outlier removal, outputting a standardized multi-measurement point synchronous vibration signal. The multi-measurement point spatiotemporal coupling feature extraction module constructs a multi-measurement point spatiotemporal feature matrix based on standardized signals, and extracts a set of spatiotemporal coupling features from multiple measurement points by combining the spatial weight coefficients of the vibration transmission path. The adaptive threshold dynamic update module for operating conditions associates the real-time head, speed, and output operating condition parameters of the unit, and dynamically updates the warning thresholds of each spatiotemporal coupling feature based on the historical normal operating condition dataset. The status recognition module completes the real-time discrimination of the unit's operating status based on the spatiotemporal coupling feature set and the real-time early warning threshold. The vibration deterioration trend real-time prediction module completes the advanced prediction of the unit's vibration deterioration trend based on the time-series feature sequence. The real-time early warning module outputs graded early warning signals according to the status discrimination results and trend prediction results.

2. The real-time monitoring system for the operating status of a hydraulic turbine unit based on multi-point vibration signals according to claim 1, characterized in that, The M triaxial vibration acceleration measurement points of the multi-point vibration signal acquisition module are respectively deployed at vibration-sensitive locations on the upper frame (X / Y direction), lower frame (X / Y direction), top cover (vertical direction), turbine main shaft (radial direction), and tailrace inlet wall of the hydraulic turbine unit, where M ≥ 8; the sampling frequency is synchronously locked with the real-time rotational speed of the unit to satisfy the following: In the formula, For real-time sampling frequency, This is the harmonic factor, which can be 128, 256, or 512. This represents the real-time rotational speed of the unit, expressed in r / min.

3. The real-time monitoring system for the operating status of a hydraulic turbine unit based on multi-point vibration signals according to claim 1, characterized in that, The multi-channel vibration signals in the data preprocessing module are synchronized in the time domain using a cross-correlation delay estimation method based on key phase pulses. After alignment, the delay deviation of each channel signal is no greater than one sampling period. The cross-correlation delay estimation satisfies the following: In the formula, Let be the cross-correlation function between the signal at the i-th measurement point and the key phase signal of the reference channel. The original vibration signal at the i-th measuring point. This is the key phase reference signal. The time delay offset is N, where N is the number of signal sampling points, and the value corresponding to the peak value of the cross-correlation function is taken. This is the delay compensation value for this channel.

4. The real-time monitoring system for the operating status of a hydraulic turbine unit based on multi-point vibration signals according to claim 1, characterized in that, The spatial weighting coefficients of the multi-point spatiotemporal coupling feature extraction module are constructed based on the vibration transmission path attenuation characteristics between the measuring points and the vibration source. satisfy: In the formula, The vibration transmission attenuation coefficient is determined by taking specific values ​​for both the steel structure of the hydraulic turbine unit and the water medium. Let be the linear distance in space from the i-th measuring point to the vibration source of the unit's main shaft. Let be the total energy of the vibration signal at the i-th measuring point. This represents the total number of measurement points.

5. The real-time monitoring system for the operating status of a hydraulic turbine unit based on multi-point vibration signals according to claim 1 or 4, characterized in that, The multi-point spatiotemporal coupling feature extraction module first performs adaptive variational mode decomposition on the standardized vibration signal of each measuring point to obtain K intrinsic mode function (IMF) components. It then extracts the effective value, peak value, kurtosis, spectral kurtosis, and 1-10 octave amplitude values ​​of each IMF component as a single-point time-domain and frequency-domain feature set. Finally, it combines spatial weighting coefficients... Construct a multi-point spatiotemporal feature matrix Where F is the feature dimension of a single measurement point, and the spatiotemporal coupled feature set is obtained through weighted fusion. ,in Let be the single-point feature vector of the i-th measurement point.

6. The real-time monitoring system for the operating status of a hydraulic turbine unit based on multi-point vibration signals according to claim 1, characterized in that, The adaptive threshold dynamic update module first divides the historical normal operating condition dataset of the unit into multiple operating condition clusters using a density peak clustering algorithm. For each operating condition cluster, a kernel density estimation method is used to calculate the probability density distribution of each spatiotemporal coupling feature. The upper boundary corresponding to the 99.7% confidence interval is taken as the warning threshold for the feature under that operating condition. The kernel density estimation satisfies: In the formula, The kernel density estimation function for the eigenvalues. The number of samples within the operating condition cluster. To determine the optimal bandwidth, cross-validation is used. For Gaussian kernel function, Let be the eigenvalues ​​to be estimated. It is the i-th feature sample within the operating condition cluster.

7. The real-time monitoring system for the operating status of a hydraulic turbine unit based on multi-point vibration signals according to claim 1, characterized in that, The vibration degradation trend real-time prediction module uses a gated recurrent unit network with a spatiotemporal attention mechanism to perform temporal modeling on the spatiotemporal coupled feature set within a continuous time window, and predicts the feature change trend in the next T time steps in advance. The core calculation of the gated loop unit satisfies: In the formula, To update the gate output, To reset the gate output, This represents the hidden layer state from the previous time step. The spatiotemporal coupling features input at the current moment, It is the sigmoid activation function. This is the weight matrix. For bias terms, This represents the candidate hidden layer state. This represents the current hidden layer state.

8. The real-time monitoring system for the operating status of a hydraulic turbine unit based on multi-point vibration signals according to claim 1, characterized in that, The status recognition module, based on a real-time spatiotemporal coupled feature set and a warning threshold corresponding to the operating condition, determines the anomaly index of the computer group's operating status. This anomaly index is calculated using weighted Mahalanobis distance and satisfies the following: In the formula, To weighted Mahalanobis distance outlier, This is a real-time spatiotemporal coupled feature vector. This represents the mean vector of features within the corresponding working condition cluster. This is a feature weight diagonal matrix, where the diagonal elements are the sensitivity coefficients of each feature to the vibration degradation of the unit. is the inverse of the characteristic covariance matrix.

9. The real-time monitoring system for the operating status of a hydraulic turbine unit based on multi-point vibration signals according to claim 8, characterized in that, The real-time early warning module has a built-in four-level hierarchical early warning logic, which corresponds to four levels: normal, early warning, alarm, and emergency shutdown. Based on the single-feature early warning threshold output by the working condition adaptive threshold dynamic update module, the abnormality boundary thresholds of the corresponding levels are calibrated in sequence: the early warning threshold is the benchmark threshold, the alarm threshold is 1.2 times the benchmark threshold, and the emergency threshold is 1.5 times the benchmark threshold. The early warning trigger adopts a dual-factor coupled criterion of real-time anomaly degree and deterioration trend: the normal level corresponds to an anomaly degree lower than the early warning threshold and no deterioration trend, and only outputs the unit's normal operating status indicator; If the anomaly level corresponds to an abnormality exceeding the warning threshold but not reaching the alarm threshold, and the predicted deterioration trend does not show a continuous upward trend, a warning signal will be output, and high-density data collection and storage and event timestamp marking will be triggered. The alarm level corresponds to an anomaly exceeding the alarm threshold but not reaching the emergency threshold, or a degradation trend prediction indicating that the characteristic value will exceed the alarm threshold within the next T time steps. An alarm signal is output and linked with local audible and visual prompts, and the location of the abnormal measurement point and details of the characteristic anomaly are pushed to the host computer simultaneously. The emergency shutdown warning level corresponds to an anomaly exceeding the emergency threshold, or a degradation trend prediction indicating that the characteristic value will exceed the emergency threshold within the next T time steps. An emergency shutdown warning signal is output and a hard-wired linkage interface with the unit speed control system and emergency shutdown circuit is reserved.

10. The real-time monitoring system for the operating status of a hydraulic turbine unit based on multi-point vibration signals according to claim 1, characterized in that, The multi-point vibration signal acquisition module, data preprocessing module, multi-point spatiotemporal coupling feature extraction module, and state identification module's initial state judgment unit are all deployed on the edge computing terminal at the unit's site. The edge computing terminal adopts an IP65 protection level design, is suitable for the strong electromagnetic and high humidity operating environment of the hydraulic power plant, and has a built-in IEEE1588 precision clock synchronization unit to achieve co-source clock synchronization with the key phase pulse signal of the multi-point vibration signal acquisition module, with a synchronization error of no more than 1μs. The edge computing terminal allocates computing power and imposes latency constraints on each module. The preprocessing latency of a single frame vibration signal is ≤1ms, the latency of spatiotemporal coupling feature extraction is ≤2ms, the latency of initial state judgment is ≤1ms, and the total end-to-end processing latency is ≤5ms. After the edge terminal completes the processing, it uses a sparse coding algorithm to compress the spatiotemporal coupling feature set with a compression ratio ≥5:

1. After encryption using the national cryptographic SM4 algorithm, only the compressed feature set, state judgment results, trend prediction results, and abnormal event data are uploaded to the host computer monitoring platform. At the same time, it locally redundantly stores no less than 30 days of full feature data and original event data. It has a built-in network outage resume mechanism, which automatically retransmits the monitoring data during the network outage period after the network is restored.

Images