A Diagnostic and Early Warning Method and System for Hydropower Equipment that Integrates Voiceprint Perception and Mechanism

Abstract

This invention discloses a diagnostic and early warning method and system for hydropower equipment that integrates acoustic signature perception and mechanism fusion, belonging to the field of equipment fault diagnosis technology. The method includes: partitioning based on a sound field coupling mechanism to form K monitoring zones; deploying K acoustic signature acquisition terminal arrays according to sound field feature topology; a diagnostic server receiving K channel signals, obtaining P target acoustic signatures through sound source decoupling, and obtaining cross-operating condition feature vectors through spatiotemporal feature mining; inputting a fault propagation graph network, outputting diagnostic results through mechanism constraint modeling; and linking SCADA data for collaborative fault early warning. This invention solves the technical problems of severe sound source aliasing, difficulty in accurately locating fault sound sources, and lack of fault propagation mechanism constraints in existing hydropower equipment fault diagnosis methods, leading to insufficient diagnostic accuracy and early warning reliability. It achieves the technical effect of high-precision fault spatial positioning, type identification, and risk level assessment through acoustic signature perception and mechanism fusion, thereby improving the operational safety and reliability of equipment.

Description

Technical Field

[0001] This invention relates to the field of equipment fault diagnosis technology, specifically to a diagnostic and early warning method and system for hydropower equipment that integrates voiceprint perception and mechanism fusion. Background Technology

[0002] As the core equipment of a hydropower energy production system, the operating status of hydropower station units directly affects the safety and power generation efficiency of the power station. During long-term operation, turbines, generators, and their auxiliary machinery can experience various faults such as bearing wear, blade cavitation, rotor imbalance, and structural loosening. To ensure stable operation of the units, the hydropower industry typically monitors equipment status online through vibration monitoring, temperature monitoring, and electrical parameter monitoring, and integrates these methods with SCADA systems for operational status management.

[0003] However, in the operating environment of large hydropower units, the complex spatial structure of the equipment, the large number of mechanical devices, and the intersecting acoustic propagation paths lead to significant acoustic field coupling and superposition phenomena within the unit due to the acoustic signals generated by various devices during operation. Simultaneously, the unit's operating conditions continuously change with load variations, head variations, and speed regulation processes, resulting in significant cross-condition fluctuations in the equipment's acoustic signature. In hydropower unit operation monitoring scenarios, the multi-source coupling of sound sources within the unit, the complex sound propagation paths, and the dynamic changes in operating conditions cause existing fault diagnosis methods based on single-point acoustic or vibration monitoring to suffer from significant deficiencies in fault location accuracy and cross-condition identification stability. The root cause lies in the fact that existing methods typically lack modeling of complex sound field coupling mechanisms and fail to effectively integrate equipment fault propagation mechanisms with multi-source acoustic features for joint analysis. Furthermore, traditional equipment diagnosis methods often rely on single sensor signals or simple feature extraction methods, lacking effective decoupling means for aliased sound sources generated by multiple devices operating simultaneously, making it difficult to accurately identify specific faulty equipment and its propagation path. Meanwhile, under complex and changing operating conditions, traditional feature extraction methods struggle to obtain stable fault characterization features, resulting in insufficient generalization ability of diagnostic models and thus affecting the reliability and timeliness of fault early warning. Summary of the Invention

[0004] This application provides a diagnostic and early warning method and system for hydropower equipment that integrates acoustic signature perception and mechanism. The key aspect lies in addressing the technical obstacles in the complex operation monitoring scenarios of hydropower units, such as the difficulty in accurately decoupling fault sound sources, the difficulty in fault location, and the insufficient stability of diagnostic models caused by the coupling of multiple sound sources, acoustic signature signal aliasing, and significant cross-condition fluctuations in acoustic characteristics due to dynamic changes in operating conditions. This is achieved through sound source decoupling combining sound field coupling mechanism partitioning, beamforming, and blind source separation, as well as a fusion modeling method combining a CNN-LSTM hybrid model and a fault propagation graph network. This is further supported by a data processing workflow that includes topological deployment of the acoustic signature perception network, cross-condition feature normalization, and collaborative analysis of SCADA operating condition data. Together, these methods enable accurate separation of multiple sound sources, extraction of stable features across operating conditions, and accurate diagnosis and collaborative early warning of equipment faults based on fault propagation mechanisms. This improves the accuracy of fault identification, spatial positioning capability, and reliability of early warning for hydropower equipment.

[0005] The first aspect of this application provides a method for diagnosing and providing early warning for hydroelectric equipment by fusing voiceprint perception and mechanism, the method comprising: Based on the sound field coupling mechanism, the main unit of the hydropower equipment is partitioned into K main unit sound field monitoring areas. According to the sound field distribution characteristics of the K main unit sound field monitoring areas, a topological deployment of the acoustic signature sensing network is performed to obtain K acoustic signature acquisition terminal arrays. The diagnostic server receives K channel acoustic signature signals uploaded by the K acoustic signature acquisition terminal arrays, performs sound source decoupling to obtain P target equipment acoustic signature signals, and then obtains P cross-operating condition normalized feature vectors through spatiotemporal sensitive feature mining. The P cross-operating condition normalized feature vectors are input into a pre-constructed equipment fault propagation graph network to model the equipment fault propagation path constrained by the fault propagation mechanism, and output equipment fault diagnosis results. The equipment fault diagnosis results include fault spatial location, fault type, and fault risk level. Based on the equipment fault diagnosis results, SCADA operating condition data is linked to perform collaborative early warning of hydropower equipment faults.

[0006] A second aspect of this application provides a diagnostic and early warning system for hydroelectric equipment that integrates voiceprint perception and mechanism fusion, the system comprising: The system comprises the following modules: Host Partitioning Module: Partitioning the hydropower equipment host based on a sound field coupling mechanism, creating K host sound field monitoring zones; Acquisition Terminal Deployment Module: Deploying a topologically-based acoustic signature sensing network based on the sound field distribution characteristics of the K host sound field monitoring zones, resulting in a K acoustic signature acquisition terminal array; Feature Mining Module: Receiving K-channel acoustic signature signals uploaded by the K acoustic signature acquisition terminal arrays from the diagnostic server, decoupling the sound sources to obtain P target equipment acoustic signature signals, and then mining these signals using spatiotemporally sensitive features to obtain P cross-condition normalized feature vectors; Fault Diagnosis Module: Inputting the P cross-condition normalized feature vectors into a pre-constructed equipment fault propagation graph network, modeling the equipment fault propagation path constrained by the fault propagation mechanism, and outputting equipment fault diagnosis results, including fault spatial location, fault type, and fault risk level; Fault Early Warning Module: Based on the equipment fault diagnosis results, linking SCADA operating condition data to provide collaborative early warning for hydropower equipment faults.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, based on the acoustic field coupling characteristics within the hydroelectric equipment, the space of the main equipment is rationally divided into multiple acoustic field monitoring areas. Corresponding acoustic signature acquisition terminal arrays are deployed according to the acoustic field distribution characteristics of each area to achieve multi-area acoustic signal acquisition. Subsequently, the diagnostic server receives multi-channel acoustic signature signals uploaded from each acquisition terminal, separates the target acoustic signature signals corresponding to different devices using sound source decoupling technology, and further extracts cross-condition normalized feature vectors with spatiotemporal characteristics. Then, these features are input into a pre-constructed equipment fault propagation graph network. By combining the equipment fault propagation mechanism with modeling and analysis of the fault propagation path, the fault diagnosis results are obtained, including fault location, fault type, and risk level. Finally, the diagnostic results are linked with the operating condition data in the SCADA system for collaborative early warning of the hydroelectric equipment's operating status. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a schematic diagram of the process for a hydroelectric equipment diagnosis and early warning method that combines voiceprint perception and mechanism fusion, as provided in an embodiment of this application.

[0010] Figure 2 A schematic diagram of the structure of a hydroelectric equipment diagnostic and early warning system that integrates voiceprint perception and mechanism, provided in an embodiment of this application.

[0011] Figure labeling: Host partitioning module 11, data acquisition terminal deployment module 12, feature mining module 13, fault diagnosis module 14, fault early warning module 15. Detailed Implementation

[0012] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0013] Example 1, as Figure 1 As shown, this application provides a diagnostic and early warning method for hydroelectric equipment that integrates voiceprint perception and mechanism fusion, wherein the method includes: Based on the acoustic field coupling mechanism, the main unit of the hydroelectric equipment is divided into K main unit acoustic field monitoring zones.

[0014] In this embodiment, considering the dense internal equipment, numerous sound sources, and complex propagation paths of the hydropower equipment main unit, the system utilizes a sound field coupling mechanism to acoustically partition the internal space of the main unit. Specifically, firstly, the structural layout information of the hydropower equipment main unit is acquired, including the spatial positions and structural connections of the turbine, generator, bearing housing, guide vane mechanism, and related auxiliary equipment, and a three-dimensional physical space model of the main unit's interior is established accordingly. Based on this three-dimensional model, the internal space of the main unit is discretized into a grid according to a preset grid scale, dividing the overall space into multiple acoustic grids, so that each grid can reflect local acoustic propagation characteristics. Subsequently, virtual sound pressure sampling points are set in each acoustic grid, and combined with the main sound source positions under the equipment's operating state, acoustic simulation is used to obtain acoustic characteristics such as sound pressure distribution, sound pressure gradient, and energy distribution of the main frequency bands at each sampling point, thereby analyzing the degree of sound field coupling between different grids. Afterwards, based on the acoustic similarity and coupling strength between grid units, adjacent grids with similar acoustic characteristics and close sound propagation connections are clustered and merged, gradually forming multiple regions with relatively independent sound field characteristics. Finally, considering the actual internal structural boundaries of the host and the equipment layout, the structural boundaries of the aforementioned clustered regions are corrected to ensure that each region has a clearly defined structural range in physical space, thus obtaining K host acoustic field monitoring zones. This approach ensures that the acoustic characteristics within each acoustic field monitoring zone are relatively consistent, while minimizing acoustic field interference between different monitoring zones. This lays the foundation for the rational deployment of subsequent acoustic signature acquisition terminals and the accurate identification of equipment sound sources.

[0015] Furthermore, based on the acoustic field coupling mechanism, the main unit of the hydroelectric equipment is divided into K main unit acoustic field monitoring zones. The method includes: The physical space of the hydroelectric equipment host is discretized into an initial acoustic grid based on a preset grid scale. After deploying virtual sound pressure sampling points on the initial acoustic grid, CFD simulation is used to calculate multiple dominant sound source features of multiple initial grids in the initial acoustic grid. The dominant sound source features include frequency band energy proportion and sound pressure gradient. Based on the multiple dominant sound source features, a watershed algorithm is introduced to merge and iterate adjacent grids of the multiple initial grids to obtain K preliminary acoustic partitions. The physical space of the host is structurally corrected and cut according to the K preliminary acoustic partitions to output the K host sound field monitoring areas.

[0016] Preferably, to achieve acoustic field zoning of the hydropower equipment main unit based on the acoustic field coupling mechanism, the structural parameters and spatial layout information of the hydropower unit's main unit are first obtained, including the position, size, and spatial relationship of key components such as the turbine runner, main shaft, guide bearing, thrust bearing, generator stator, rotor, and casing. Based on this, a three-dimensional physical space model of the hydropower equipment main unit is constructed. Subsequently, a preset grid scale is set according to the sound wave propagation characteristics and the equipment size range. For example, the internal space of the main unit is uniformly gridded and discretized according to a spatial scale of 0.5m to 1m, thereby dividing the entire physical space of the main unit into multiple regular initial acoustic grid units, each grid unit representing a local acoustic field calculation region.

[0017] After completing the initial acoustic mesh division, at least one virtual sound pressure sampling point is set within each initial mesh cell to simulate the sound pressure response at that location during equipment operation. Based on typical operating parameters of the hydropower unit, such as unit speed, load, head, and flow rate, and combined with the locations of major sound sources, such as turbine cavitation noise, bearing friction noise, and electromagnetic noise, CFD simulation is used to jointly simulate and calculate the internal fluid disturbance and acoustic propagation process of the main unit. The sound pressure values ​​of each virtual sound pressure sampling point in different frequency ranges are obtained through simulation calculations, and the acoustic characteristic parameters of each initial grid are further calculated as the dominant sound source characteristics of the corresponding initial grid. These dominant sound source characteristics include the frequency band energy ratio and the sound pressure gradient. The frequency band energy ratio is the proportion of sound energy to total sound energy within several preset frequency ranges (such as 0–1kHz, 1–3kHz, 3–5kHz, etc.), used to characterize the spectral structure of the sound source. The sound pressure gradient is the ratio of the sound pressure difference between adjacent grids to the spatial distance, used to reflect the direction of sound wave propagation and attenuation trend.

[0018] After obtaining the dominant sound source features of all initial grids, a sound field feature matrix is ​​constructed using the frequency band energy proportion and sound pressure gradient as feature parameters, and this feature matrix is ​​mapped to an acoustic feature distribution map. Based on this, a watershed algorithm is introduced to divide the initial grids into regions. Specifically, the acoustic feature distribution map is treated as a topographic elevation map, where acoustic feature values ​​correspond to altitude. Local extreme regions are identified as acoustic feature centers, and these centers are used as starting points for region expansion. When the acoustic feature similarity between adjacent grids is higher than a preset similarity threshold, they are merged into the same acoustic region; when the feature difference exceeds the threshold, the partition boundary is preserved. Multiple rounds of adjacent grid merging iterations are performed in this manner until the region merging stabilizes, resulting in K preliminary acoustic partitions with relatively consistent acoustic features.

[0019] After obtaining K preliminary acoustic zones, to ensure the acoustic zoning results match the actual equipment structure layout, a structural correction and cutting process is performed on the preliminary acoustic zones. This involves spatially matching the boundaries of the preliminary acoustic zones with the boundaries of the internal equipment structures of the host unit. Zone boundaries that cross critical equipment structures are adjusted to align as closely as possible with the boundaries of physical structures such as housing walls, bearing housing boundaries, vibration isolation structures, and equipment mounting support structures. This prevents a single acoustic zone from spanning multiple physical equipment areas. Simultaneously, for areas with significant structural barriers, such as housing partitions or structural walls, structural constraints are added to re-correct the zone boundaries, creating relatively independent monitoring areas for each acoustic zone in physical space. Through this structural correction and cutting process, the final sound field monitoring area maintains consistency in acoustic feature distribution and aligns with the actual equipment structure layout, resulting in K host unit sound field monitoring areas. These host unit sound field monitoring areas effectively reflect the acoustic characteristics of different equipment areas and reduce sound source interference between different devices, providing a reliable foundation for the subsequent topological deployment of the acoustic signature acquisition terminal array and the accurate identification of equipment fault sound sources.

[0020] Based on the sound field distribution characteristics of the K host sound field monitoring areas, the topology of the voiceprint sensing network is deployed to obtain an array of K voiceprint acquisition terminals.

[0021] In one embodiment, after dividing the K host acoustic field monitoring areas, acoustic field mapping is carried out for each host acoustic field monitoring area to obtain the spatial distribution of sound pressure level, dominant frequency band distribution, sound attenuation direction, noise interference sources, and possible sound reflection paths under typical operating conditions. Subsequently, based on the acoustic field distribution characteristics, a corresponding acquisition terminal deployment strategy is selected for each host acoustic field monitoring area. When the sound source distribution is relatively concentrated and the sound pressure peak is obvious in a certain monitoring area, multiple acoustic fingerprint acquisition terminals are arranged in a surround or fan-shaped coverage manner around the target area, so that the acquisition direction is pointed towards the sound energy focusing area; when the sound propagation direction in a certain monitoring area has obvious main channel characteristics, multiple acoustic fingerprint acquisition terminals are deployed linearly or at gradient intervals along the main sound propagation direction to enhance the ability to capture signal changes on the sound propagation path; when there are multiple reflective surfaces or equipment obstructions inside a certain monitoring area, a layered and anisotropic combination deployment method is adopted, so that different acquisition terminals face different reflection paths and the rear of the obstruction area, thereby reducing the acquisition blind zone. When determining the deployment locations, constraints are set for the installation height, orientation angle, terminal spacing, and number of each acoustic signature acquisition terminal. The installation height is preferably corresponding to the height of the main radiating sound source of the target device; the orientation angle is preferably pointing towards the peak sound pressure point or the dominant sound propagation direction; and the terminal spacing is determined based on the size of the monitoring area and the wavelength of the dominant frequency band to avoid spatial aliasing and sampling redundancy. After completing the terminal position and orientation planning for each monitoring area, multiple acoustic signature acquisition terminals within the same monitoring area are grouped into an acoustic signature acquisition terminal array according to preset communication and acquisition coordination relationships. This allows the terminals within the array to sample synchronously in time and form complementary coverage in space. The arrays corresponding to different monitoring areas are numbered respectively, ultimately resulting in K acoustic signature acquisition terminal arrays. Through this topological deployment method, each host sound field monitoring area possesses acoustic signature acquisition capabilities adapted to its sound field distribution characteristics, thereby improving the acquisition integrity, spatial resolution, and anti-interference capability of the target device's acoustic signature signal, providing high-quality input data for subsequent multi-channel acoustic signature signal source decoupling and fault diagnosis.

[0022] Furthermore, based on the sound field distribution characteristics of the K host sound field monitoring areas, a topological deployment of the voiceprint sensing network is performed to obtain an array of K voiceprint acquisition terminals. The method includes: Multi-scale operating condition tests are performed on the first host sound field monitoring area to obtain the multi-scale sound pressure level spatial distribution; a three-dimensional sound pressure cloud map is constructed based on the multi-scale sound pressure level spatial distribution to identify the sound pressure peak point and sound attenuation gradient; the dominant frequency band energy focusing of the first host sound field monitoring area is extracted; according to the sound pressure peak point, sound attenuation gradient and dominant frequency band energy focusing, the array configuration strategy is matched in the terminal deployment strategy library to perform the topological deployment of the voiceprint sensing network in the first host sound field monitoring area to obtain the first voiceprint acquisition terminal array.

[0023] Preferably, to achieve precise deployment of the acoustic signature acquisition terminal within the first main unit's acoustic field monitoring area, multi-scale operating condition testing is first conducted on the first main unit's acoustic field monitoring area. During this process, acoustic tests are performed on the hydropower unit under different operating conditions, including the unit startup phase, stable low-load operation phase, rated load operation phase, and load regulation phase. For each operating condition, temporary sound pressure test points are arranged within the first main unit's acoustic field monitoring area at preset spatial intervals. For example, two-dimensional grid measurement points are formed on the horizontal plane at intervals of 0.5m to 1m, and layered sampling points are set at different height levels, such as 0.5m, 1.5m, and 2.5m, thus forming a three-dimensional sound pressure sampling grid. Each sampling point records sound pressure level data within a certain time window using a mobile acoustic sensor, and simultaneously records the unit's operating condition parameters. By summarizing and processing the sound pressure level data collected under different operating conditions and spatial locations, a multi-scale spatial distribution dataset of sound pressure levels in the first main unit's acoustic field monitoring area under multiple operating conditions is formed.

[0024] After obtaining multi-scale sound pressure level spatial distribution data, the sound pressure level at each sampling point is treated as discrete spatial node data. Inverse distance weighted interpolation or three-dimensional spline interpolation methods are used to perform three-dimensional continuous processing of the sound pressure data, thereby generating a three-dimensional sound pressure cloud map of the first host sound field monitoring area. This three-dimensional sound pressure cloud map represents the sound pressure level distribution at different spatial locations in the form of color gradients or isosurfaces. Subsequently, local extremum detection is performed using the three-dimensional sound pressure cloud map to identify sound pressure peak points where the sound pressure level is significantly higher than the surrounding area, and the corresponding spatial coordinates are recorded. Simultaneously, by calculating the ratio of the sound pressure level difference between adjacent spatial points to the spatial distance, the sound pressure gradient vector is obtained, thereby determining the main propagation direction of sound energy and the sound pressure attenuation trend.

[0025] After determining the spatial distribution characteristics of sound pressure, frequency domain analysis is performed on the acoustic signals collected by the first host sound field monitoring area. Specifically, a fast Fourier transform is performed on the acoustic signal at each sampling point to obtain the energy spectrum of the sound pressure signal in different frequency ranges. The proportion of sound energy in each frequency band is then statistically analyzed according to preset frequency ranges, such as 0–500Hz, 500Hz–2kHz, and 2kHz–5kHz. Through statistical analysis of all sampling points and multi-condition data, frequency ranges with a high energy proportion under most conditions are selected and defined as the dominant frequency band of the first host sound field monitoring area. Spatial energy distribution analysis is then performed on the dominant frequency band signal to identify the focusing area of ​​the dominant frequency band sound energy in space, thereby obtaining the energy focusing position of the dominant frequency band.

[0026] After obtaining the sound pressure peak point, sound pressure attenuation gradient, and dominant frequency band energy focusing area, these parameters are used as deployment features input into a pre-built terminal deployment strategy library. This library stores various acoustic signature acquisition array configuration strategies and their applicable conditions. For example, when the sound pressure peak point is obvious and the energy concentration area is small, a surround array strategy is matched, so that multiple acquisition terminals are evenly distributed around the sound pressure peak point; when the sound pressure gradient direction has a clear main propagation channel, a linear array strategy is matched, so that the acquisition terminals are arranged at a certain interval along the sound propagation direction; when the sound energy is distributed in a fan shape, a fan-shaped array strategy is matched, so that the acquisition terminals form angular coverage towards the energy diffusion area; when there are sound sources at different heights within the monitoring area, a layered array strategy is matched, setting acquisition terminals at different heights to enhance spatial coverage.

[0027] After determining the array configuration strategy, the specific deployment parameters of the acquisition terminals are further calculated based on the spatial dimensions of the first host acoustic field monitoring area, the dominant frequency band wavelength, and the structural position of the equipment. The terminal spacing is determined according to the dominant frequency band wavelength, generally set to 0.5 to 1 times the dominant wavelength to avoid spatial aliasing. The terminal installation height is usually consistent with the radiation height of the main sound source. The orientation angle of the acquisition terminals is preferentially pointed towards the sound pressure peak point or the direction of sound energy propagation. Simultaneously, the terminal installation position is corrected during deployment based on the equipment structural position to avoid the acquisition terminals being obstructed by the casing, support structure, or other equipment. Finally, the topological deployment of the acoustic signature acquisition terminals is completed within the first host acoustic field monitoring area according to the matched array configuration strategy. Multiple acoustic signature acquisition terminals within the same monitoring area are logically associated through an industrial communication network, ensuring synchronized sampling in time and complementary spatial coverage, thus forming the first acoustic signature acquisition terminal array for the coordinated acquisition and transmission of acoustic signature signals generated by the equipment operation within the monitoring area.

[0028] Furthermore, the array of K voiceprint acquisition terminals is connected to the diagnostic server via an industrial Ethernet network.

[0029] Optionally, K acoustic signature acquisition terminal arrays are respectively set up within K host acoustic field monitoring areas. Each acoustic signature acquisition terminal array includes multiple acoustic acquisition nodes, an array control unit, and a network communication interface. Each acoustic acquisition node is used to complete on-site acoustic signal sampling; the array control unit is used to coordinate and control the sampling time, data buffering, and upload rhythm of each acquisition node within the same array; the network communication interface is used to connect the data acquired by the array to an industrial Ethernet network. The industrial Ethernet network adopts an industrial switching network structure that supports real-time communication, such as a star topology, tree topology, or redundant ring network topology, enabling each acoustic signature acquisition terminal array to access the diagnostic server through an industrial switch. To ensure unified joint analysis of acoustic signature signals from different monitoring areas, unified clock synchronization control is performed on each acoustic signature acquisition terminal array during industrial Ethernet communication. This clock synchronization control can use network time synchronization or master-slave clock synchronization to ensure that the acoustic signature data uploaded by each array has a unified time reference and timestamp, thereby ensuring that the K channels of acoustic signature signals received by the diagnostic server are time-aligned. Each acoustic signature acquisition terminal array completes analog-to-digital conversion, frame buffering, and data encapsulation of the raw acoustic signal locally. Then, it uploads the data to the diagnostic server via industrial Ethernet according to a preset communication protocol. The encapsulated data includes at least the array number, acquisition node number, sampling time, sampling frequency, acoustic signature data payload, and device status identification information. The diagnostic server receives, verifies, sorts, and buffers the data from the K acoustic signature acquisition terminal arrays. Once it detects that all K channels of data within the corresponding time window have arrived, it triggers subsequent beamforming, sound source decoupling, and feature mining processing. The use of industrial Ethernet connectivity meets the requirements for data bandwidth, communication reliability, and anti-interference performance under long-term continuous monitoring conditions of hydropower equipment. Furthermore, it facilitates the unified access of multiple acoustic signature acquisition terminal arrays to the same diagnostic server, enabling collaborative acquisition, centralized management, and real-time fault diagnosis of acoustic signature signals from multiple monitoring areas.

[0030] The diagnostic server receives the K-channel voiceprint signals uploaded by the K voiceprint acquisition terminal array, performs sound source decoupling to obtain P target device voiceprint signals, and then obtains P cross-operating condition normalized feature vectors through spatiotemporal sensitive feature mining.

[0031] In one embodiment, the diagnostic server establishes a communication connection with K acoustic signature acquisition terminal arrays corresponding to each host acoustic field monitoring area. After each acoustic signature acquisition terminal array completes synchronous sampling, it receives K-channel acoustic signature signals uploaded by each array. The K-channel acoustic signature signals correspond to the raw acoustic data acquired in different host acoustic field monitoring areas, and each channel signal has a unified timestamp to ensure timing consistency during subsequent multi-channel joint processing. Subsequently, the diagnostic server preprocesses the received K-channel acoustic signature signals. This preprocessing includes removing DC components, bandpass filtering, amplitude normalization, and frame processing according to fixed time windows to eliminate the influence of low-frequency drift, high-frequency random noise, and differences in sampling amplitude between different channels, resulting in K preprocessed acoustic signature signals. Then, considering the possibility of multiple devices emitting sound simultaneously in each monitoring area and the superposition of sound sources, the diagnostic server performs coupled sound source decomposition on the K preprocessed acoustic signature signals, and then matches the decomposed independent sound source components with a pre-established acoustic signature feature rule library to filter out P target device acoustic signature signals corresponding to P target devices. To further enhance the adaptability of subsequent diagnostics to changes in operating conditions, the diagnostic server performs spatiotemporal sensitive feature mining on the acoustic signature signals of P target devices. Specifically, it extracts time-domain, frequency-domain, time-frequency-domain, and spatial correlation features from the acoustic signature signals of each target device. Time-domain features include envelope variation, root mean square (RMS), peak factor, impulse factor, and short-time energy variation. Frequency-domain features include band energy proportion, spectral peak position, octave distribution, and spectral centroid. Time-frequency-domain features include short-time Fourier spectrum features, Mel-frequency cepstral features, or continuous wavelet energy ridge features. Spatial correlation features include time delay relationships, coherence, and sound pressure attenuation relationships between channels in different monitoring areas. Then, the acoustic signature signal of each target device is input into a CNN-LSTM hybrid model. The CNN branch extracts local spatial texture features from the spectrum, while the LSTM branch extracts the dynamic evolution features of the acoustic signature within a continuous time window. The outputs of the two branches are then fused to obtain a joint feature representation of the corresponding target device, which serves as a cross-operating-condition normalized feature vector representing its current state, providing input for subsequent equipment fault propagation modeling and fault diagnosis.

[0032] Furthermore, the diagnostic server receives the K channels of voiceprint signals uploaded by the K voiceprint acquisition terminal array, performs sound source decoupling to obtain P target device voiceprint signals, and then obtains P cross-condition normalized feature vectors through spatiotemporal sensitive feature mining. The method includes: A beamforming algorithm is applied to the K-channel acoustic signature signals to generate directional acoustic beams to suppress noise in non-target areas, resulting in K directional acoustic beam signals. Coupled source decomposition based on blind source separation is then performed on the K directional acoustic beam signals to filter out the P target device acoustic signature signals corresponding to the P device fault mechanisms. The P target device acoustic signature signals are input into a CNN-LSTM hybrid model, and spatial and temporal feature extraction is performed in parallel via parallel CNN and LSTM channels to obtain P spatial feature vectors and P temporal feature vectors. The P spatial feature vectors and P temporal feature vectors are dynamically normalized in conjunction with the SCADA operating condition data to output the P cross-operating condition normalized feature vectors.

[0033] Preferably, after receiving the K-channel acoustic signature signals uploaded by each acoustic signature acquisition terminal array, the diagnostic server performs beamforming processing on the K-channel acoustic signature signals to enhance the acoustic signal in the direction of the target device and suppress background noise and device crosstalk from other areas. Specifically, firstly, a geometric model of the array is established based on the spatial deployment coordinates of the acoustic signature acquisition terminal array, and the propagation distance difference from each acquisition terminal in the array to the target device is calculated. Then, the corresponding signal delay compensation is calculated based on the propagation distance difference to perform time alignment processing on each channel acoustic signature signal. After completing the delay compensation, a preset weighting coefficient is applied to each channel signal, and weighted superposition is performed to construct a spatially directional beam signal. This weighting coefficient can be calculated according to the minimum variance distortionless response criterion or the delay summation method, so that the acoustic signal from the target direction is phase consistent and enhanced during superposition, while noise signals from non-target directions are canceled out due to phase differences. Through the above processing, K directional acoustic beam signals are obtained, each directional acoustic beam signal mainly corresponding to the target sound source area within the host sound field monitoring area.

[0034] After obtaining K directional acoustic beam signals, they are coupled and decomposed. During this process, a short-time Fourier transform is performed on each directional acoustic beam signal to convert the original time-domain signal into a time-spectrum representation. The signal is then divided into sub-bands according to a preset frequency range, for example, into multiple sub-bands such as 0–500Hz, 500Hz–2kHz, 2kHz–5kHz, and above 5kHz, thus obtaining multi-source acoustic sub-band signals. Next, the sub-band signals from multiple channels are processed using Independent Component Analysis (ICA) algorithm for blind source separation, thereby decomposing the original mixed signal into multiple independent acoustic source components. Each independent acoustic source component corresponds to a potential device acoustic source signal. These multiple independent acoustic source components are then matched with a pre-established device fault mechanism acoustic signature feature rule library to select the acoustic source component that best matches P device fault mechanism features, and these are identified as P target device acoustic signature signals.

[0035] After obtaining P target device voiceprint signals, they are input into a CNN-LSTM hybrid model for spatiotemporal feature extraction. In the CNN channel, a short-time Fourier transform or Mel-frequency transform is first performed on each target device voiceprint signal to generate a two-dimensional time-spectrum map. Then, a multi-layer convolutional neural network is used to perform convolution operations on the time-spectrum map. The convolution kernel is used to extract local spectral texture features, and the pooling layer is used to reduce the feature dimensionality and enhance the robustness of the model. After multi-layer convolution and pooling processing, a spatial feature representation reflecting the spectral structure, harmonic characteristics, and energy distribution pattern is obtained. P spatial feature vectors are output through a fully connected layer. These spatial feature vectors are used to characterize the features of the target device voiceprint signal in the spatial and spectral structure dimensions, including spatial sequence features such as low-frequency energy proportion, mid-frequency energy proportion, high-frequency energy proportion, dominant frequency amplitude, sound source spatial weight, and channel correlation coefficient. Meanwhile, in the LSTM channel, the target device's acoustic signature signal is divided into sequences according to fixed time windows. For example, time-series input data is constructed with a time step of 0.5 seconds or 1 second and then input into the LSTM model. The LSTM model uses input gates, forget gates, and output gates to remember and update the time-series data, thereby capturing the dynamic changes of the acoustic signature signal in the time dimension, such as periodic vibration modes, changes in impact signal intervals, and trends in acoustic energy fluctuations. After processing by a multi-layer LSTM network, P time-series feature vectors reflecting the dynamic characteristics of the device's operation are obtained. These time-series feature vectors are used to characterize the dynamic changes of the target device's acoustic signature signal in the time dimension, and all include time-series features such as periodic stability, periodic energy fluctuation rate, impact density, impact peak intensity, dominant frequency drift rate, and short-term energy growth rate.

[0036] After obtaining P spatial feature vectors and P temporal feature vectors, they are linked with SCADA operating condition data at the corresponding timestamps for processing. Specifically, the current operating status parameters of the unit are obtained from the SCADA system, including key operating parameters such as unit speed, load level, head height, flow rate, and guide vane opening, and the current equipment operating condition category is determined based on these parameters. Then, a pre-established operating condition normalization model is used to dynamically normalize the spatial and temporal feature vectors. For example, cosine similarity is used to match the operating condition parameters and operating condition templates, and the operating condition template with the highest similarity is selected as the current equipment operating condition category. The normalization function corresponding to the equipment operating condition category is then used to perform amplitude scaling and frequency band offset correction on the P spatial feature vectors and P temporal feature vectors, so that the acoustic signature features under different operating conditions are mapped to a unified feature space, thereby reducing the impact of operating condition changes on fault feature expression. Finally, the normalized spatial feature vectors and temporal feature vectors are fused. For example, a unified feature expression is generated by feature splicing or weighted fusion to obtain P cross-condition normalized feature vectors that can reflect the actual operating status of the equipment and have cross-condition consistency. These cross-condition normalized feature vectors will be used as inputs to the subsequent equipment fault propagation graph network to realize fault propagation path modeling and fault diagnosis analysis of hydropower equipment.

[0037] Furthermore, the method involves performing coupled sound source decomposition based on blind source separation on the K directional sound beam signals to filter out the P target device acoustic signature signals corresponding to P device fault mechanisms. The K directional acoustic beam signals are divided into frequency bands based on time-frequency alignment to obtain multi-source acoustic sub-band signals; the multi-source acoustic sub-band signals are separated into blind sources based on ICA to obtain multiple independent sound source components; the P soundprint feature rules of the P device fault mechanisms are used to match the soundprint features of the multiple independent sound source components, and the P target device soundprint signals are filtered and output.

[0038] Optionally, to separate effective acoustic signature signals that characterize different device fault mechanisms from the K directional acoustic beam signals, a time-frequency aligned frequency band segmentation process is first performed on the K directional acoustic beam signals. In this process, the diagnostic server first performs frame segmentation on each directional acoustic beam signal, dividing the continuous acoustic signal according to a preset time window length. For example, a Hamming window with a length of 25ms to 50ms is used for windowing, and a window shift step size of 10ms to 20ms is set to obtain a continuous time frame sequence. Subsequently, a short-time Fourier transform is performed on each frame signal to convert the time-domain signal into a time-spectrum matrix, where the horizontal axis represents the time frame sequence, the vertical axis represents the frequency components, and the matrix elements represent the acoustic energy intensity at the corresponding time and frequency. To eliminate time delay differences between different array channels caused by variations in propagation paths and array spatial distribution, the cross-correlation function between the signals of each channel is calculated, and the relative time delay is estimated based on the position of the cross-correlation peak. Then, the time-frequency matrix of each channel is time-shifted and corrected to ensure that the acoustic energy from the same physical sound source has a consistent time alignment relationship in different channels. At the same time, the spectrum of different channels is fine-tuned in the frequency dimension based on the position of the spectral peak and the energy concentration region, so that the main frequency components of the same sound source fall into the same frequency range in different channels, thereby completing the time-frequency alignment process.

[0039] After completing time-frequency alignment, the entire effective frequency range is divided into multiple sub-bands according to preset frequency division rules. For example, the spectrum range of 0Hz to 8kHz is divided into low frequency band (0Hz to 500Hz), mid-low frequency band (500Hz to 2kHz), mid-high frequency band (2kHz to 5kHz), and high frequency band (5kHz to 8kHz). Then, the time-frequency energy corresponding to each frequency band is extracted, and the signals of the same frequency band from K channels are combined to form a sub-band observation signal matrix, thereby obtaining multiple multi-source acoustic sub-band signals. Each sub-band signal matrix contains mixed sound source information from different channels.

[0040] After obtaining the multi-source acoustic sub-band signals, blind source separation processing based on ICA is performed on the signal matrix of each sub-band. Specifically, the signal matrix of each sub-band is first mean-reduced, that is, the average value of each channel signal is subtracted to eliminate the DC component. Then, the mean-reduced signal is whitened by eigenvalue decomposition of the covariance matrix to reduce the correlation between the signal components and transform the covariance matrix into an identity matrix. After preprocessing, the whitened signal matrix is ​​input into the ICA algorithm for demixing. In one feasible approach, the FastICA algorithm can be used, which maximizes the non-Gaussianity of the signal as the optimization objective and iteratively solves the demixing matrix to decompose the original mixed signal into multiple statistically independent sound source components. Several independent sound source components can be obtained for each sub-band. Based on the continuity of component energy distribution in adjacent time frames and adjacent frequency bands, components belonging to the same physical sound source are aggregated and reconstructed, ultimately yielding multiple independent sound source components.

[0041] Subsequently, a pre-established audioprint feature rule library for equipment fault mechanisms is used to perform audioprint feature matching on multiple independent sound source components. This rule library establishes corresponding audioprint feature rules for P types of equipment fault mechanisms. Each rule includes typical audioprint feature parameters related to that fault mechanism, such as dominant frequency range, harmonic overtone structure, peak amplitude ratio, envelope spectrum characteristic frequency, impulse pulse period, and energy proportion of different frequency bands. For each independent sound source component, its corresponding audioprint feature parameters are extracted, including root mean square value, peak position, frequency band energy distribution, harmonic structure, and envelope spectrum characteristics. The similarity between the feature parameters of the independent sound source component and each fault mechanism template in the rule library is then calculated, for example, by using cosine similarity, Euclidean distance, or correlation coefficient for matching scoring. If the matching score of an independent sound source component with a fault mechanism rule exceeds a preset threshold, and this score is the maximum value among all mechanism rules, then the independent sound source component is determined to correspond to that equipment fault mechanism. Following the above method, all independent sound source components are matched and filtered, and the P sound source components that best match the fault mechanisms of P devices are selected as target outputs, thus obtaining P target device acoustic signature signals. Through the above processing, the sound source signals of multiple devices that were originally superimposed in a complex sound field environment can be effectively separated and accurately mapped to the corresponding devices and their fault mechanisms, providing a reliable data foundation for subsequent acoustic signature feature extraction and fault diagnosis modeling.

[0042] The P cross-condition normalized feature vectors are input into a pre-constructed equipment fault propagation graph network to model the equipment fault propagation path under fault propagation mechanism constraints, and output the equipment fault diagnosis results, wherein the equipment fault diagnosis results include fault spatial location, fault type and fault risk level.

[0043] In one embodiment, after obtaining P cross-condition normalized feature vectors, these P feature vectors are used as feature inputs for the current operating state of the equipment and fed into a pre-constructed equipment fault propagation graph network for inference calculation. This network uses various typical fault mechanisms within hydropower equipment as graph nodes, with each node corresponding to a fault state or fault evolution mechanism. The connections between nodes represent the propagation correlations between different faults in equipment physical connections, energy transfer paths, structural coupling relationships, or historical cases. The network performs cross-node information propagation calculations according to a preset directed topology and edge weights. In each round of propagation, each node not only retains its own input features but also receives state information transmitted from its upstream associated nodes. The received information is weighted and fused according to the edge weights between nodes to simulate the diffusion process of fault influence in the equipment structure. For example, when the fault mechanism corresponding to a node has a strong causal propagation relationship with another node, the state change of the former node will affect the state update of the latter node through a higher weight; conversely, if the propagation correlation is weak, the corresponding influence is lower. After multiple rounds of iterative propagation, the equipment fault propagation graph network gradually forms a stable node state distribution reflecting the propagation intensity and evolution path between various fault mechanisms. When the change in node state is less than a preset threshold in two consecutive iterations, the propagation calculation ends, resulting in P node state confidence vectors. These vectors characterize the probability of various fault mechanisms occurring under the current equipment operating state and their propagation activity. Finally, the P node state confidences are fused for decision-making, forming equipment fault diagnosis results including fault spatial location, fault type, and fault risk level. This approach not only allows for determining whether equipment is abnormal based on current acoustic signature characteristics but also identifies possible propagation paths and impact ranges of faults under the constraints of fault propagation mechanisms, thereby improving the accuracy, interpretability, and engineering applicability of fault diagnosis for complex hydropower equipment.

[0044] Furthermore, the method also includes: P fault mechanism nodes are predefined, and a directed edge topology of the P fault mechanism nodes is constructed based on the physical connection relationship of the equipment. The edge weights of the directed edge topology are corrected according to the co-occurrence frequency of the P fault mechanism nodes in historical fault case data to obtain a fault propagation directed topology. Based on the timestamps of multiple historical fault diagnosis results, the working condition association backtracking slice of the historical fault case data is performed to obtain multiple cross-working condition fault sample sets, multiple sample propagation path labels, and multiple sample fault diagnosis results. Vectorized feature encoding based on the P fault mechanism nodes is performed on the multiple cross-working condition fault sample sets to obtain multiple sets of sample fault feature vectors. The multiple sets of sample fault feature vectors, multiple sample propagation path labels, and multiple sample fault diagnosis results are used as training data to perform GCN supervised learning training of the fault propagation directed topology to obtain the equipment fault propagation graph network.

[0045] Optionally, to construct a fault propagation graph network that reflects the occurrence, expansion, and mutual influence of faults in hydropower equipment, the typical fault mechanisms that may occur in different equipment parts are first enumerated and classified based on the structural composition, operating principle, and historical common fault types of the main hydropower equipment. Each fault mechanism is defined as a graph node. These fault mechanism nodes include at least a node number, corresponding equipment part, fault category, mechanism description, and basic characteristic items corresponding to the fault. For example, turbine blade cavitation, main shaft eccentricity, thrust bearing wear, guide bearing loosening, rotor imbalance, stator-rotor friction, electromagnetic anomaly, and structural resonance can be defined as different fault mechanism nodes.

[0046] After defining P fault mechanism nodes, a physical connection table for the main hydropower equipment is established based on the physical connection relationships of the equipment. This table includes at least mechanical connection relationships, support connection relationships, energy transfer relationships, and fluid coupling relationships. Mechanical connection relationships indicate whether there are rigid or flexible connections between components, such as the connection between the main shaft and the impeller, the main shaft and the rotor, or the bearing and the support structure. Support connection relationships indicate whether vibration or load can be transmitted along the structural path. Energy transfer relationships indicate whether the impact of a fault can propagate along the power chain or electromagnetic chain. Fluid coupling relationships indicate whether water flow disturbances will have a continuous impact on adjacent components. Subsequently, for any two fault mechanism nodes, it is determined whether they meet the preset propagation edge establishment conditions. If the first fault mechanism may induce a second fault mechanism structurally, energetically, or technologically, then directed edges are established between the nodes of the first fault mechanism and the nodes of the second fault mechanism. These edge establishment conditions can be implemented using a regularized approach. For example, if the components to which the two nodes belong have a direct physical connection, and a fault in the former node can cause changes in force, friction, vibration, or flow field in the latter node, then a first-order propagation edge is established. If there is no direct connection between the two nodes, but a clear indirect propagation path exists through an intermediate component, then a second-order propagation edge is established and assigned a lower initial weight. The initial edge weights can be set according to the strength of the propagation relationship; for example, a direct coupling edge is assigned a value of 1.0, an indirect coupling edge is assigned a value of 0.5, and a weakly coupled edge is assigned a value of 0.2, thus forming an initial directed topology graph.

[0047] After constructing the initial directed topology, the edge weights of the directed topology are corrected based on the co-occurrence frequency of P fault mechanism nodes in historical fault case data to obtain the fault propagation directed topology. Specifically, a historical fault case database is first established. Each case in this database includes at least the faulty equipment, fault type, fault discovery time, fault handling time, monitoring data before and after the fault, operating condition data, and maintenance confirmation results. Next, the historical fault cases are analyzed in chronological order to identify the chronological relationship of the fault mechanisms in each case. For example, in a certain case, rotor imbalance occurs first, and then thrust bearing abnormality occurs after a period of time, which is recorded as a directed co-occurrence of "rotor imbalance → thrust bearing abnormality". For any directed edge e(i,j), count the number of cases where node i appears first and node j appears later, with the time interval between them falling within a preset propagation time window. This number is denoted as the co-occurrence frequency N(i,j). Simultaneously, count the total number of times node i appears alone, N(i). Calculate the conditional propagation probability Pr(i,j) = N(i,j) / N(i). Then, adjust this probability based on the propagation time interval. For example, if node j appears very shortly after node i, it indicates a strong propagation relationship, and a time decay coefficient e can be introduced. -Δt / TWeighting is applied, where Δt is the propagation time interval and T is the preset time scale.

[0048] After obtaining the directed topology of fault propagation, for each historical fault diagnosis result, using its final confirmation timestamp t0 as the baseline time point, a preset time window ΔT is traced backward to extract the acoustic fingerprint monitoring data, SCADA operating condition data, maintenance logs, and alarm records within the time interval [t0-ΔT,t0]. This ΔT can be set according to the fault type; for example, it can be set to 24 to 72 hours for progressive mechanical faults and 10 minutes to 2 hours for sudden faults. To preserve the dynamic information during the fault evolution process, sliding slices can be performed at fixed step sizes within this backtracking interval, splitting the original time series into multiple continuous sample segments. Each sample segment contains monitoring features and SCADA operating condition features related to P fault mechanism nodes within the corresponding time period. Furthermore, to reflect cross-condition characteristics, it is necessary to retain the condition labels for samples formed under different loads, heads, flow rates, speeds, and guide vane openings. For example, the load can be divided into three intervals: low load, medium load, and rated load; the head can be divided into two intervals: high head and low head. These conditions can then be combined to form multiple condition categories. In this way, the same fault mechanism can correspond to multiple different samples under different conditions, thus forming multiple cross-condition fault sample sets. For each sample set, a sample propagation path label needs to be given based on maintenance records, expert annotation results, or fault event tracking results. For example, if the final confirmed propagation path of a case is "cavitation → enhanced structural vibration → bearing loosening," then the nodes and edges on this path are marked as positive sample propagation paths, and the remaining nodes and edges that do not participate in the propagation are marked as negative samples. At the same time, the final fault diagnosis result corresponding to the sample is used as a supervision label. This fault diagnosis result includes the final fault type, fault spatial location, and fault risk level. In this way, each training sample has three types of information: input features, propagation path label, and diagnosis result label.

[0049] After obtaining multiple cross-condition fault sample sets, for any training sample, node feature vectors are constructed one by one according to the fixed order of P fault mechanism nodes in the fault propagation directed topology. Each node feature vector includes at least the cross-condition normalized acoustic signature features related to the corresponding equipment, the abnormal score of the node in the current time slice, the SCADA operating condition parameters directly associated with the node, the alarm statistics of the area of ​​the corresponding equipment, and the state memory of the node in the previous time slice. For example, each node feature can be encoded as a d-dimensional vector, where the first m dimensions represent acoustic signature statistical features, the middle n dimensions represent operating condition features, and the last q dimensions represent state auxiliary features. In this way, a sample can form a node feature matrix X of dimension P×d. At the same time, a corresponding adjacency matrix A is generated according to the fault propagation directed topology, and the element A(i,j) in the adjacency matrix is ​​the corrected weight of edge e(i,j). For each training sample, a path supervision matrix Ypath and a result supervision label Ydiag can also be added, where Ypath is used to represent the real propagation edges in the sample, and Ydiag is used to represent the final diagnostic category and risk level of the sample. Therefore, each training sample can be represented as graph-structured data G={X,A,Ypath,Ydiag}.

[0050] After encoding the training samples, multiple sets of sample fault feature vectors, multiple sample propagation path labels, and multiple sample fault diagnosis results are used as training data to perform GCN supervised learning training on the directed topology of fault propagation. Specifically, the node feature matrix X and adjacency matrix A are first input into the graph convolutional network. The graph convolutional layer performs neighborhood feature aggregation calculation on each node, so that when a node is updated, it not only retains its own information but also integrates the state information of its neighboring nodes after edge weighting. After two or more layers of graph convolution operations, a high-level representation of each node under the constraints of global topological relationships can be obtained. In order to learn the propagation path and the final diagnosis result simultaneously, a dual output branch can be set after the graph convolutional network. The first output branch is used for propagation path prediction, outputting the propagation probability for each directed edge to form the path prediction result; the second output branch is used for fault diagnosis prediction, classifying and risk-assessing the aggregated representation of the entire graph or key nodes, and outputting the fault type, spatial location, and risk level.

[0051] During training, a propagation path loss function and a fault diagnosis loss function are constructed. The propagation path loss function measures the difference between the predicted propagation edge and the actual propagation path label, and can use binary cross-entropy loss. The fault diagnosis loss function measures the difference between the predicted fault type, location, and risk level and the actual label, and can use a combination of multi-class cross-entropy loss and mean squared error loss. During training, gradient descent or Adam optimization algorithms are used to iteratively update the network parameters. After each training round, the model performance is evaluated using a validation set. Training stops when the validation loss no longer decreases for several consecutive rounds or reaches the preset number of training rounds, and the trained equipment fault propagation graph network model is output. After training, the equipment fault propagation graph network can identify the activity level of each fault mechanism node based on the input cross-condition fault features, infer the propagation path of the fault within the equipment based on the graph topology and edge weights, and output the final fault type, fault spatial location, and fault risk level by combining node states and propagation results. Since this graph network is obtained under the joint constraints of prior physical connections of hydropower equipment, historical fault propagation patterns, and cross-operating condition training samples, it not only has high diagnostic accuracy, but also strong interpretability and engineering applicability, and can be used for subsequent online fault diagnosis reasoning.

[0052] Furthermore, the method involves inputting the P cross-condition normalized feature vectors into a pre-constructed equipment fault propagation graph network to model the equipment fault propagation path constrained by the fault propagation mechanism, and outputting the equipment fault diagnosis result. The method includes: After inputting the P cross-condition normalized feature vectors into the P fault mechanism nodes of the equipment fault propagation graph network, the fault state is iteratively propagated across nodes based on the fault propagation directed topology until the change in node state is less than a preset threshold, and P node state confidence vectors are output; the P node state confidence vectors are fused and decided to output the equipment fault diagnosis result.

[0053] Optionally, after obtaining P cross-condition normalized feature vectors, they are input into a pre-constructed equipment fault propagation graph network. Based on the order of the P predefined fault mechanism nodes in the fault propagation graph network, the P cross-condition normalized feature vectors are mapped to the corresponding P fault mechanism nodes, serving as the initial state features of each node. The initial state vector of each node represents the potential activity or abnormality of the fault mechanism in the equipment at the current moment. After completing the node feature input, cross-node iterative propagation calculations of the fault state are performed based on the directed fault propagation topology. These directed fault propagation topologies are jointly represented by the node adjacency matrix and the corresponding edge weight matrix, where the edge weights describe the propagation strength between different fault mechanisms. In each round of iterative calculation, each node not only retains its current state information but also receives state information transmitted from its upstream neighboring nodes, and weights and fuses the received information according to the corresponding edge weights. In one iteration, the state update of the i-th fault mechanism node can be calculated by weighting its current state vector and the state vectors of all its neighboring nodes with directed connections. This allows the node state to gradually reflect the propagation impact of the fault in the equipment structure. For example, the current state of the node and the weighted states of its neighboring nodes can be linearly combined and updated using a nonlinear activation function to obtain a new node state representation. The above node state update process is performed synchronously throughout the graph network and continues to execute according to a preset number of iterations or convergence conditions. When the change between all node state vectors in two adjacent iterations is less than a preset threshold, the fault state propagation process in the graph network is considered to have reached a stable state, thus ending the iteration calculation. At this point, each fault mechanism node in the graph network corresponds to a stable node state value. This state value can be normalized to a node state confidence value, resulting in P node state confidence vectors. Each element in these node state confidence vectors represents the probability or activity level of the corresponding fault mechanism in the current equipment operating state.

[0054] After obtaining P node state confidence vectors, the P node state confidence scores are sorted by value, and the node with the highest confidence score is selected as the primary fault mechanism. The category corresponding to this primary fault mechanism is then designated as the equipment fault type. Simultaneously, based on the confidence score of the primary fault mechanism node, the fault risk is classified into different levels. For example, a confidence score below a first threshold is considered low risk, a confidence score between the first and second thresholds is considered medium risk, and a confidence score above the second threshold is considered high risk. Then, based on the equipment location information corresponding to the primary fault mechanism node and its associated host acoustic field monitoring area, the node is mapped to a specific region in the equipment physical space model, thereby determining the spatial location of the fault. Finally, a comprehensive equipment fault diagnosis result, including the fault spatial location, fault type, and fault risk level, is obtained and used for subsequent equipment fault early warning and maintenance decisions.

[0055] Furthermore, the method involves fusing the state confidence vectors of the P nodes to make a decision and outputting the equipment fault diagnosis result. The P node state confidence vectors are sorted in descending order, the main fault mechanism is extracted as the fault type, and the fault risk level is mapped according to the corresponding node state confidence vector; the host acoustic field monitoring area is backtracked according to the main fault mechanism, and the mapped location physical space coordinates are used as the fault spatial location.

[0056] Optionally, after obtaining P node state confidence vectors, the confidence values ​​of each node in the P node state confidence vectors are sorted from largest to smallest to form a fault mechanism priority sequence. Subsequently, the fault mechanism corresponding to the first node in the sorted sequence is extracted as the primary fault mechanism. In addition, to avoid misjudgment caused by occasional noise or local anomalies, a primary fault extraction judgment condition can be set. For example, when the state confidence of the first-ranked node is higher than a preset primary fault threshold, and the difference between its state confidence and that of the second-ranked node is greater than a preset distinction threshold, the fault mechanism corresponding to the first-ranked node is directly determined as the primary fault mechanism. When the confidence of the first and second-ranked nodes is relatively close, the upstream and downstream relationships, equipment parts, and historical co-occurrence relationships of the two nodes in the directed fault propagation topology are combined for auxiliary judgment. If there is a clear propagation link between the two nodes, the fault mechanism corresponding to the node at the beginning of the propagation link or the node with stronger propagation influence is taken as the primary fault mechanism, and the fault category corresponding to the determined primary fault mechanism is output as the fault type.

[0057] After determining the fault type, the fault risk level is mapped based on the state confidence value of the corresponding main fault mechanism node. For example, a node state confidence value between 0 and 0.4 is defined as low risk, between 0.4 and 0.7 as medium risk, and above 0.7 as high risk. When the state confidence value of the main fault mechanism node falls into the corresponding range, the corresponding fault risk level is determined. After determining the main fault mechanism and fault risk level, a correspondence table is established for each fault mechanism node in the system's preset fault mechanism node attributes, relating it to the target device, the host acoustic field monitoring area, and the physical space model. This correspondence table includes at least the fault mechanism node number, the corresponding device name, the host acoustic field monitoring area number, the boundary range of the monitoring area in the host's three-dimensional physical space model, and the coordinates of key equipment parts within the monitoring area. After determining the main fault mechanism, the corresponding target device and its host acoustic field monitoring area number are queried based on the main fault mechanism node, thus tracing back to the target device acoustic signature signal source area obtained in the previous sound source decoupling stage, and the coverage information of the acoustic signature acquisition terminal array for that device area. Next, the host sound field monitoring area number is mapped to the host three-dimensional physical space model to determine the spatial boundary range of the corresponding monitoring area.

[0058] Subsequently, to improve the accuracy of fault spatial location, the most likely radiation point of the fault sound source within the monitoring area can be calculated based on the arrival time difference, sound pressure intensity difference, and energy focusing position of the dominant frequency band of the target acoustic signature signals received from multiple acquisition terminals. If the spatial coordinates of key parts of the equipment are preset in the system, such as the center point of the bearing housing, the edge area of ​​the impeller, or the position of the inner ring of the stator, the fault sound source location is preferentially mapped to the coordinates of the key parts of the equipment that best match the main fault mechanism. If the coordinates of a single point cannot be located, a spatial region coordinate range can be output, for example, in the form of a cuboid bounding box, a spherical neighborhood, or a sub-region of the monitoring area to represent the fault spatial location, and the determined physical spatial coordinates or spatial region is output as the fault spatial location. Through the above processing, a complete decision transformation from node state confidence vector to fault type, fault risk level, and fault spatial location is achieved. The fault type is directly determined by the main fault mechanism, the fault risk level is obtained by mapping the node state confidence corresponding to the main fault mechanism, and the fault spatial location is obtained by backtracking the main fault mechanism with the host acoustic field monitoring area and the three-dimensional physical space model. This process can combine the graph network reasoning results with the actual spatial structure of the equipment to accurately identify the fault mechanism of hydropower equipment, accurately locate the spatial location of the fault, and reasonably assess the fault risk level, thereby improving the accuracy of equipment fault diagnosis and the reliability of early warning.

[0059] Based on the equipment fault diagnosis results, the SCADA operating data is linked to provide collaborative early warning for water and electricity equipment faults.

[0060] In one embodiment, after obtaining the equipment fault diagnosis results, the equipment fault diagnosis results are further linked with SCADA operating data for analysis to achieve coordinated early warning of abnormal operation of hydropower equipment. Specifically, after outputting the fault spatial location, fault type, and fault risk level, the diagnostic server synchronously retrieves the SCADA operating data corresponding to the current diagnosis time. This SCADA operating data includes at least the unit's active load, speed, head, flow rate, guide vane opening, pressure, temperature, vibration, current, voltage, and start-up / shutdown status, among other operating parameters. Subsequently, the fault diagnosis results and SCADA operating data are time-series aligned and correlated, that is, data with the same timestamp or within a preset time tolerance range are used as the same early warning analysis unit to determine whether the current fault corresponds to specific operating condition fluctuations, load switching, start-up / shutdown processes, or local operational anomalies. Subsequently, based on a pre-established fault-condition collaborative early warning rule base, the current fault diagnosis results are used to determine the early warning level. This rule base stores the risk triggering conditions, alarm thresholds, and handling levels for different fault types under different operating conditions. For example, if the diagnosis results indicate an increased risk of bearing wear, and SCADA data shows that the corresponding bearing temperature continues to rise and the vibration amplitude exceeds the set threshold, it is determined to be an enhanced early warning. If the diagnosis results indicate an increased probability of cavitation faults, and SCADA data shows that the unit is in a low-head, high-flow condition or a state of frequent fluctuations in guide vane opening, it is determined that the fault has a continuous deterioration trend, and the early warning level is upgraded. In addition, the fault risk level and SCADA anomaly degree can be weighted and fused to generate a comprehensive early warning index. The fault risk level reflects the fault activity level on the acoustic signature side, and the SCADA anomaly degree reflects the degree of deviation of the operating condition on the operating parameter side. Both are used together to measure the overall operating risk of the equipment. When the comprehensive early warning index reaches the preset early warning threshold, the corresponding level of early warning information is output. This early warning information includes the faulty equipment name, fault type, fault spatial location, current risk level, associated abnormal operating condition parameters, and suggested handling methods. When the comprehensive early warning index continues to rise or exceeds the threshold for multiple consecutive time windows, a higher level alarm or coordinated operation and maintenance strategy can be triggered. Through this method, collaborative analysis and joint judgment of fault diagnosis results and SCADA operating data are achieved. This ensures that the early warning results not only reflect the internal fault mechanism of the equipment but also assess the triggering background and actual impact of the fault in conjunction with real-time operating conditions, thereby improving the accuracy, timeliness, and engineering applicability of hydropower equipment fault early warning.

[0061] In summary, the embodiments of this application have at least the following technical effects: First, based on the sound field coupling mechanism, the main unit of the hydropower equipment is partitioned into K main unit sound field monitoring areas. Then, according to the sound field distribution characteristics of the K main unit sound field monitoring areas, a topological deployment of the acoustic signature sensing network is performed, resulting in a K acoustic signature acquisition terminal array. Next, the diagnostic server receives the K channels of acoustic signature signals uploaded by the K acoustic signature acquisition terminal arrays, performs sound source decoupling to obtain P target equipment acoustic signature signals, and then obtains P cross-operating condition normalized feature vectors through spatiotemporal sensitive feature mining. Then, the P cross-operating condition normalized feature vectors are input into a pre-constructed equipment fault propagation graph network to model the equipment fault propagation path constrained by the fault propagation mechanism, outputting equipment fault diagnosis results, whereby the equipment fault diagnosis results include fault spatial location, fault type, and fault risk level. Finally, based on the equipment fault diagnosis results, coordinated early warning of hydropower equipment faults is performed in conjunction with SCADA operating condition data. This invention addresses the technical problems in existing hydropower equipment fault diagnosis methods, such as severe sound source aliasing, difficulty in accurately locating fault sound sources, and lack of fault propagation mechanism constraints, which lead to insufficient diagnostic accuracy and early warning reliability. It achieves high-precision fault spatial positioning, type identification, and risk level assessment through the fusion of acoustic signature perception and mechanism, and combines SCADA operating condition data for collaborative early warning, thereby improving the safety and reliability of equipment operation.

[0062] Example 2, based on the same inventive concept as the hydroelectric equipment diagnosis and early warning method that fuses voiceprint perception and mechanism in the aforementioned examples, such as... Figure 2 As shown, this application provides a diagnostic and early warning system for hydroelectric equipment that integrates voiceprint perception and mechanism fusion, wherein the system includes: Host partitioning module 11: Based on the sound field coupling mechanism, the host of hydropower equipment is partitioned into K host sound field monitoring areas; Acquisition terminal deployment module 12: According to the sound field distribution characteristics of the K host sound field monitoring areas, the topology deployment of the acoustic fingerprint sensing network is carried out to obtain K acoustic fingerprint acquisition terminal arrays; Feature mining module 13: The diagnostic server receives the K channel acoustic fingerprint signals uploaded by the K acoustic fingerprint acquisition terminal arrays, performs sound source decoupling to obtain P target equipment acoustic fingerprint signals, and obtains P cross-operating condition normalized feature vectors through spatiotemporal sensitive feature mining; Fault diagnosis module 14: The P cross-operating condition normalized feature vectors are input into a pre-constructed equipment fault propagation graph network to perform equipment fault propagation path modeling constrained by fault propagation mechanism, and output equipment fault diagnosis results, wherein the equipment fault diagnosis results include fault spatial location, fault type and fault risk level; Fault early warning module 15: Based on the equipment fault diagnosis results, the SCADA operating condition data is linked to carry out collaborative early warning of hydropower equipment faults.

[0063] Furthermore, the host partitioning module 11 is used to perform the following method: The physical space of the hydroelectric equipment host is discretized into an initial acoustic grid based on a preset grid scale. After deploying virtual sound pressure sampling points on the initial acoustic grid, CFD simulation is used to calculate multiple dominant sound source features of multiple initial grids in the initial acoustic grid. The dominant sound source features include frequency band energy proportion and sound pressure gradient. Based on the multiple dominant sound source features, a watershed algorithm is introduced to merge and iterate adjacent grids of the multiple initial grids to obtain K preliminary acoustic partitions. The physical space of the host is structurally corrected and cut according to the K preliminary acoustic partitions to output the K host sound field monitoring areas.

[0064] Furthermore, the data acquisition terminal deployment module 12 is used to perform the following method: Multi-scale operating condition tests are performed on the first host sound field monitoring area to obtain the multi-scale sound pressure level spatial distribution; a three-dimensional sound pressure cloud map is constructed based on the multi-scale sound pressure level spatial distribution to identify the sound pressure peak point and sound attenuation gradient; the dominant frequency band energy focusing of the first host sound field monitoring area is extracted; according to the sound pressure peak point, sound attenuation gradient and dominant frequency band energy focusing, the array configuration strategy is matched in the terminal deployment strategy library to perform the topological deployment of the voiceprint sensing network in the first host sound field monitoring area to obtain the first voiceprint acquisition terminal array.

[0065] Furthermore, the data acquisition terminal deployment module 12 is used to perform the following method: The array of K voiceprint acquisition terminals is connected to the diagnostic server via an industrial Ethernet network.

[0066] Furthermore, the feature mining module 13 is used to perform the following methods: A beamforming algorithm is applied to the K-channel acoustic signature signals to generate directional acoustic beams to suppress noise in non-target areas, resulting in K directional acoustic beam signals. Coupled source decomposition based on blind source separation is then performed on the K directional acoustic beam signals to filter out the P target device acoustic signature signals corresponding to the P device fault mechanisms. The P target device acoustic signature signals are input into a CNN-LSTM hybrid model, and spatial and temporal feature extraction is performed in parallel via parallel CNN and LSTM channels to obtain P spatial feature vectors and P temporal feature vectors. The P spatial feature vectors and P temporal feature vectors are dynamically normalized in conjunction with the SCADA operating condition data to output the P cross-operating condition normalized feature vectors.

[0067] Furthermore, the feature mining module 13 is used to perform the following methods: The K directional acoustic beam signals are divided into frequency bands based on time-frequency alignment to obtain multi-source acoustic sub-band signals; the multi-source acoustic sub-band signals are separated into blind sources based on ICA to obtain multiple independent sound source components; the P soundprint feature rules of the P device fault mechanisms are used to match the soundprint features of the multiple independent sound source components, and the P target device soundprint signals are filtered and output.

[0068] Furthermore, the fault diagnosis module 14 is used to perform the following methods: P fault mechanism nodes are predefined, and a directed edge topology of the P fault mechanism nodes is constructed based on the physical connection relationship of the equipment. The edge weights of the directed edge topology are corrected according to the co-occurrence frequency of the P fault mechanism nodes in historical fault case data to obtain a fault propagation directed topology. Based on the timestamps of multiple historical fault diagnosis results, the working condition association backtracking slice of the historical fault case data is performed to obtain multiple cross-working condition fault sample sets, multiple sample propagation path labels, and multiple sample fault diagnosis results. Vectorized feature encoding based on the P fault mechanism nodes is performed on the multiple cross-working condition fault sample sets to obtain multiple sets of sample fault feature vectors. The multiple sets of sample fault feature vectors, multiple sample propagation path labels, and multiple sample fault diagnosis results are used as training data to perform GCN supervised learning training of the fault propagation directed topology to obtain the equipment fault propagation graph network.

[0069] Furthermore, the fault diagnosis module 14 is used to perform the following methods: After inputting the P cross-condition normalized feature vectors into the P fault mechanism nodes of the equipment fault propagation graph network, the fault state is iteratively propagated across nodes based on the fault propagation directed topology until the change in node state is less than a preset threshold, and P node state confidence vectors are output; the P node state confidence vectors are fused and decided to output the equipment fault diagnosis result.

[0070] Furthermore, the fault diagnosis module 14 is used to perform the following methods: The P node state confidence vectors are sorted in descending order, the main fault mechanism is extracted as the fault type, and the fault risk level is mapped according to the corresponding node state confidence vector; the host acoustic field monitoring area is backtracked according to the main fault mechanism, and the mapped location physical space coordinates are used as the fault spatial location.

[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A diagnostic and early warning method for hydroelectric equipment that integrates voiceprint perception and mechanism, characterized in that, The method includes: Based on the acoustic field coupling mechanism, the main unit of the hydropower equipment is divided into K main unit acoustic field monitoring areas. Based on the sound field distribution characteristics of the K host sound field monitoring areas, the topology of the voiceprint sensing network is deployed to obtain an array of K voiceprint acquisition terminals. The diagnostic server receives the K-channel voiceprint signals uploaded by the K voiceprint acquisition terminal array, performs sound source decoupling to obtain P target device voiceprint signals, and then obtains P cross-operating condition normalized feature vectors through spatiotemporal sensitive feature mining. The P cross-operating condition normalized feature vectors are input into a pre-constructed equipment fault propagation graph network to model the equipment fault propagation path under fault propagation mechanism constraints, and output the equipment fault diagnosis results, wherein the equipment fault diagnosis results include fault spatial location, fault type and fault risk level; Based on the equipment fault diagnosis results, coordinated early warning of water and electricity equipment faults is carried out by linking SCADA operating data; The device fault propagation graph network includes: P fault mechanism nodes are predefined, and the directed edge topology of the P fault mechanism nodes is constructed based on the physical connection relationship of the devices; The edge weights of the directed edge topology are corrected based on the co-occurrence frequency of the P fault mechanism nodes in historical fault case data to obtain the fault propagation directed topology. Based on the timestamps of multiple historical fault diagnosis results, the working condition association backtracking slice of the historical fault case data is performed to obtain multiple cross-working condition fault sample sets, multiple sample propagation path labels and multiple sample fault diagnosis results. The multiple cross-operating condition fault sample sets are subjected to vectorized feature encoding based on P fault mechanism nodes to obtain multiple sets of sample fault feature vectors. Using the multiple sets of sample fault feature vectors, multiple sample propagation path labels, and multiple sample fault diagnosis results as training data, the GCN supervised learning training of the directed topology of the fault propagation is performed to obtain the equipment fault propagation graph network.

2. The method for diagnosing and warning hydroelectric equipment by fusing voiceprint perception and mechanism as described in claim 1, characterized in that, Based on the acoustic field coupling mechanism, the main unit of the hydroelectric equipment is divided into K main unit acoustic field monitoring zones. The method includes: The physical space of the main unit of the hydropower equipment is discretized into an initial acoustic grid based on a preset grid scale. After deploying virtual sound pressure sampling points on the initial acoustic grid, CFD simulation is used to calculate the characteristics of multiple dominant sound sources in multiple initial grids of the initial acoustic grid. The dominant sound source characteristics include the frequency band energy proportion and sound pressure gradient. Based on the characteristics of the multiple dominant sound sources, a watershed algorithm is introduced to iterate the merging of adjacent grids of the multiple initial grids to obtain K preliminary acoustic partitions; Based on the K preliminary acoustic partitions, the physical space of the host is structurally modified and cut, and the K host sound field monitoring zones are output.

3. The method for diagnosing and warning hydroelectric equipment by fusing voiceprint perception and mechanism as described in claim 1, characterized in that, The diagnostic server receives K channels of voiceprint signals uploaded by the K voiceprint acquisition terminal arrays, performs sound source decoupling to obtain P target device voiceprint signals, and then obtains P cross-condition normalized feature vectors through spatiotemporal sensitive feature mining. The method includes: A beamforming algorithm is applied to the K-channel acoustic signature signal to generate directional acoustic beams to suppress noise in non-target areas, resulting in K directional acoustic beam signals. The K directional acoustic beam signals are subjected to coupled acoustic source decomposition based on blind source separation in order to filter the P target device acoustic signature signals corresponding to the P device fault mechanisms; The P target device voiceprint signals are input into the CNN-LSTM hybrid model. Spatial feature extraction and temporal feature extraction are performed in parallel through the parallel CNN and LSTM channels to obtain P spatial feature vectors and P temporal feature vectors. The P spatial feature vectors and P temporal feature vectors are dynamically normalized by linking the SCADA operating condition data, and the P cross-operating condition normalized feature vectors are output.

4. The method for diagnosing and warning hydropower equipment by fusing voiceprint perception and mechanism as described in claim 3, characterized in that, The method involves performing coupled acoustic source decomposition based on blind source separation on the K directional acoustic beam signals to filter out the P target device acoustic signature signals corresponding to P device fault mechanisms. The K directional acoustic beam signals are divided into frequency bands based on time-frequency alignment to obtain multi-source acoustic sub-band signals; Blind source separation based on ICA is performed on the multi-source acoustic sub-band signal to obtain multiple independent acoustic source components; The voiceprint features of the multiple independent sound source components are matched using the P voiceprint feature rules of the P device fault mechanisms, and the P target device voiceprint signals are filtered and output.

5. The method for diagnosing and warning hydroelectric equipment by fusing voiceprint perception and mechanism as described in claim 1, characterized in that, The method involves inputting the P cross-operating-condition normalized feature vectors into a pre-constructed equipment fault propagation graph network to model the equipment fault propagation path under fault propagation mechanism constraints, and outputting the equipment fault diagnosis results. After inputting the P cross-condition normalized feature vectors into the P fault mechanism nodes of the equipment fault propagation graph network, the fault state is iteratively propagated across nodes based on the fault propagation directed topology until the change in node state is less than a preset threshold, and P node state confidence vectors are output. The system performs a fusion decision on the state confidence vectors of the P nodes and outputs the equipment fault diagnosis result.

6. The method for diagnosing and warning hydroelectric equipment by fusing voiceprint perception and mechanism as described in claim 5, characterized in that, The method involves fusing the state confidence vectors of the P nodes to make a decision and outputting the equipment fault diagnosis result. Arrange the P node state confidence vectors in descending order, extract the main fault mechanism as the fault type, and map the fault risk level according to the corresponding node state confidence vector; Based on the main fault mechanism, the host acoustic field monitoring area is traced back to map the physical space coordinates as the fault spatial location.

7. The method for diagnosing and warning hydroelectric equipment by fusing voiceprint perception and mechanism as described in claim 1, characterized in that, Based on the sound field distribution characteristics of the K host sound field monitoring areas, a topological deployment of the voiceprint sensing network is performed to obtain a K voiceprint acquisition terminal array. The method includes: Multi-scale operating condition tests were conducted on the sound field monitoring area of ​​the first host to obtain the spatial distribution of multi-scale sound pressure level. A three-dimensional sound pressure cloud map is constructed based on the multi-scale sound pressure level spatial distribution to identify the sound pressure peak point and sound attenuation gradient. Extract and focus the dominant frequency band energy of the first host acoustic field monitoring area; Based on the sound pressure peak point, sound attenuation gradient, and dominant frequency band energy focusing, the array configuration strategy is matched in the terminal deployment strategy library, and the topology deployment of the acoustic signature sensing network in the first host sound field monitoring area is carried out to obtain the first acoustic signature acquisition terminal array.

8. The method for diagnosing and warning hydroelectric equipment by fusing voiceprint perception and mechanism as described in claim 7, characterized in that, The array of K voiceprint acquisition terminals is connected to the diagnostic server via an industrial Ethernet network.

9. A diagnostic and early warning system for hydroelectric equipment that integrates voiceprint perception and mechanism, characterized in that, The system is used to implement the hydroelectric equipment diagnostic and early warning method based on voiceprint perception and mechanism fusion as described in any one of claims 1-8, the system comprising: Host partitioning module: Based on the sound field coupling mechanism, the host of hydroelectric equipment is partitioned into K host sound field monitoring areas; Acquisition terminal deployment module: Based on the sound field distribution characteristics of the K host sound field monitoring areas, the topology deployment of the voiceprint perception network is carried out to obtain an array of K voiceprint acquisition terminals; Feature mining module: The diagnostic server receives the K channels of voiceprint signals uploaded by the K voiceprint acquisition terminal array, performs sound source decoupling to obtain P target device voiceprint signals, and then obtains P cross-operating condition normalized feature vectors through spatiotemporal sensitive feature mining. Fault diagnosis module: Input the P cross-operating condition normalized feature vectors into the pre-constructed equipment fault propagation graph network, perform equipment fault propagation path modeling with fault propagation mechanism constraints, and output equipment fault diagnosis results, wherein the equipment fault diagnosis results include fault spatial location, fault type and fault risk level; Fault early warning module: Based on the equipment fault diagnosis results, it coordinates with SCADA operating data to provide early warning of faults in hydropower equipment.

Images