A multi-source sensing fusion-based intelligent diagnosis and full-condition monitoring system for hydraulic units

Abstract

The present application relates to the technical field of intelligent data processing, in particular to a water turbine unit intelligent diagnosis and full working condition monitoring system based on multi-source sensor fusion, comprising multi-source sensor acquisition units, space-time synchronization preprocessing units, adaptive multi-source heterogeneous data fusion units, full working condition boundary adaptive matching units, intelligent fault diagnosis units and edge-cloud collaborative management and control units which are sequentially connected in communication. The present application realizes multi-source feature fusion through confidence optimization adaptive weighted fusion algorithm, realizes accurate fault identification and positioning through improved D-S evidence theory, realizes deterioration trend prediction through bidirectional LSTM network, can realize accurate monitoring, intelligent diagnosis and early warning of water turbine units under full working conditions, improve the operation safety and control efficiency of the units, and adapt to various water turbine unit engineering applications.

Description

Technical Field

[0001] This invention relates to the field of intelligent data processing technology, specifically to an intelligent diagnostic and full-condition monitoring system for hydraulic units based on multi-source sensor fusion. Background Technology

[0002] Hydropower generating units are the core equipment for hydropower energy conversion, and their operational stability and safety directly affect the safe and stable operation of the power system and the efficient utilization of water resources in the basin. Hydropower generating units operate under complex conditions, encompassing various typical scenarios such as shutdown, turning gear operation, no-load operation, rated load operation, load shedding, and phase-shifting operation. Furthermore, the equipment is highly coupled, and the causes of failures are complex. Single-sensor monitoring data cannot fully characterize the true operating status of the unit. Therefore, unit condition monitoring, fault diagnosis, and health assessment technologies based on multi-source information fusion have become the core technological direction for condition-based maintenance and safety management of hydropower generating units.

[0003] A Chinese invention patent with publication number CN115577808A discloses a method and system for evaluating the health status of hydropower units based on multi-source information fusion. This scheme first constructs a four-level hierarchical analysis structure for evaluating the health status of hydropower units: "target layer - system layer - project layer - indicator layer." A judgment matrix is ​​constructed based on pairwise comparison and a 1-9 comparison scale table. The weights of indicators at each level are calculated using the analytic hierarchy process (AHP) and consistency checks are performed. The system then divides the quantitative evaluation status score range and indicator threshold range using industry standards and the 3σ criterion, distinguishing between indicators trending towards large and small values ​​to calculate deterioration degree and membership degree. A fuzzy comprehensive evaluation method is used to construct indicator scoring standards. Through a multi-source data acquisition module, online monitoring data, offline test data, and manual inspection data of the unit are acquired. From bottom to top, the scoring calculation and anomaly marking at the indicator layer, project layer, and system layer are completed, ultimately achieving an overall evaluation of the health status of the hydropower unit and outputting operation and maintenance guidance suggestions. This scheme combines hierarchical analysis and fuzzy evaluation to achieve fusion evaluation of multi-source data, improves the comprehensiveness of unit health status evaluation, and provides a feasible technical path for the health status evaluation of hydropower units.

[0004] However, the existing technical solutions represented by the aforementioned comparative documents still have significant technical defects in practical engineering applications. First, the multi-source information fusion of this solution only reaches the decision-level weight allocation and scoring fusion level, without performing spatiotemporal synchronization preprocessing and feature-level adaptive fusion of multi-source heterogeneous sensor data. It lacks microsecond-level time synchronization and spatial reference registration processing for multi-source sensor data, and cannot solve the problems of time sequence misalignment and inconsistent spatial references of different sensor data. Furthermore, it does not consider the impact of the real-time confidence level of sensor data on the fusion results. When single-source sensor data is abnormal or fails, it is easy to lead to distortion of the final evaluation results, and its anti-interference ability and data representation ability are insufficient. Second, this solution adopts a fixed index threshold range and state classification standard, without considering the differences in the operating characteristics of hydropower units under all operating conditions. It cannot achieve dynamic correction of feature thresholds under transient and transitional operating conditions such as load change, load shedding, and phase adjustment. During the switching of unit operating conditions, misjudgment and omission of state evaluation are prone to occur, and it cannot adapt to the monitoring and evaluation needs of the unit under all operating conditions. Third, this solution can only achieve post-event graded evaluation of unit health status, and cannot achieve accurate identification of unit fault types or precise fault location. It also lacks a module for predicting unit performance degradation trends, thus failing to provide early warnings of early unit faults. Furthermore, its feature extraction does not distinguish between steady-state and transient features, resulting in insufficient ability to capture fault features under transient operating conditions and difficulty in identifying early, minor unit faults. Fourth, this solution does not adopt an edge-cloud collaborative system architecture, failing to balance the real-time requirements of on-site unit monitoring with the continuous iterative optimization needs of the diagnostic model. It struggles to achieve millisecond-level local alarms for abnormal unit operating conditions, and also cannot achieve online updates of the diagnostic model or unified management of multiple unit clusters, resulting in insufficient system scalability and engineering adaptability. Fifth, this solution relies on qualitative data entry from manual inspections for multi-source data, failing to achieve fully automated real-time collection and diagnosis of all dimensions of unit operating data. Moreover, it lacks sufficient collection and analysis of characteristic quantities such as high-frequency acoustic emissions and high-frequency vibrations corresponding to early unit faults, leading to low sensitivity in identifying early unit faults. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent diagnostic and full-condition monitoring system for hydraulic units based on multi-source sensor fusion, in order to solve the problems mentioned in the background art, such as inconsistent spatiotemporal references of existing multi-source data, poor adaptability to full-condition operation, insufficient fault diagnosis accuracy, and inability to balance real-time performance and iterative capabilities.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A hydraulic unit intelligent diagnosis and full-condition monitoring system based on multi-source sensor fusion includes a multi-source sensor acquisition unit, a spatiotemporal synchronous preprocessing unit, an adaptive multi-source heterogeneous data fusion unit, a full-condition boundary adaptive matching unit, an intelligent fault diagnosis unit, and an edge-cloud collaborative management and control unit that are connected in sequence via communication. The multi-source sensing and acquisition unit is deployed on the main shaft, guide bearing, turbine runner, stator, excitation system and water pipeline of the hydraulic unit, and is used to synchronously collect multi-source heterogeneous operating data of vibration, acoustic emission, temperature, pressure, flow rate, stator current and excitation voltage under all operating conditions of the unit. The spatiotemporal synchronization preprocessing unit is used to remove outliers, filter and reduce noise in the collected multi-source data, and complete the spatiotemporal reference registration and sequence alignment of the multi-source data. The adaptive multi-source heterogeneous data fusion unit is used to perform feature-level adaptive weighted fusion on the registered multi-source data to generate a global feature vector for unit operation. The full-condition boundary adaptive matching unit is used to match the corresponding operating condition boundary based on the real-time operating parameters of the unit, and to complete the dynamic correction of the feature threshold under the full operating condition. The intelligent fault diagnosis unit is used to identify the type of unit fault, locate the fault, and predict the deterioration trend based on the corrected global feature vector. The edge-cloud collaborative management and control unit is used to realize real-time monitoring and alarm at the edge, iterative updates of cloud models, data traceability, and full lifecycle management.

[0007] Preferably, the spatiotemporal synchronization preprocessing unit uses the rising edge of the key phase signal of the unit's main shaft as the time reference trigger point to mark a synchronization timestamp for each set of multi-source sensor data frames, completing the microsecond-level time synchronization alignment of the multi-source sensor data; it constructs a three-dimensional spatial coordinate system of the unit with the rotation center of the unit's main shaft as the origin of the spatial coordinate system, and uses a spatial coordinate mapping matrix to complete the spatial reference registration of sensor data at different installation positions, mapping each sensor data to a unified three-dimensional spatial coordinate system of the unit; at the same time, it uses an improved 3σ criterion, combined with the rate of change of adjacent data points of the sensor data, to complete the identification and removal of outliers in the data; it uses a wavelet soft threshold noise reduction algorithm, selects the db6 wavelet basis function to perform 5-level wavelet decomposition, and completes the suppression of environmental noise and electromagnetic interference in the collected data; and it uses cubic spline interpolation to unify the sampling frequency and align the data sequence for the processed multi-source data.

[0008] Preferably, the adaptive multi-source heterogeneous data fusion unit employs a confidence-optimized adaptive weighted fusion algorithm based on the minimum mean square error criterion to perform feature-level fusion on the registered multi-source heterogeneous data. Its basic weight calculation formula is as follows: ; In the formula, The total number of feature dimensions of multi-source sensor data. For the first Initial fusion weights for 3D sensing features, For the first Variance of dimensional sensing features; Simultaneously, a sensor data confidence correction factor is introduced. The initial weights are then optimized a second time, and the optimized fusion weights are calculated as follows: ; In the formula, For the first The confidence coefficient of the dimensional sensor data is calculated by comprehensively considering the real-time signal-to-noise ratio, online calibration deviation, and historical operational stability coefficient of the sensor data. Finally, based on the optimized weights, multi-source feature fusion is completed to generate a unified-dimensional global feature vector for unit operation.

[0009] Preferably, the adaptive multi-source heterogeneous data fusion unit first extracts single-dimensional features from the sensor data of each dimension. For vibration and acoustic emission signals, it extracts time-domain statistical features, frequency-domain spectral features, and time-frequency-domain wavelet packet energy features. For temperature, pressure, flow, and electrical quantity signals, it extracts time-domain trend features and abrupt change features. The time-domain statistical features include mean, root mean square, peak value, kurtosis, margin factor, and impulse factor. The frequency-domain spectral features include fundamental frequency amplitude, harmonic amplitude, harmonic energy ratio, spectral centroid, and spectral kurtosis. The time-frequency-domain wavelet packet energy features include the energy ratio of each frequency band after three-layer wavelet packet decomposition and the wavelet packet energy entropy. The extracted single-dimensional feature sets are then subjected to min-max normalization to map all feature values ​​to the [0,1] interval. The normalized feature sets are then subjected to dimension alignment to generate a single-dimensional feature matrix with unified dimensions. Finally, an adaptive weighted fusion with confidence optimization is performed.

[0010] Preferably, the full-condition boundary adaptive matching unit has a preset full-condition type library for hydraulic turbine units. This library includes seven typical operating conditions: shutdown, turning gear, no-load, rated load, load change, load shedding, and phase-shifting operation. Each typical operating condition has a preset multi-parameter boundary threshold range, which includes upper and lower limits for active power, unit speed, operating head, guide vane opening, and excitation current. Based on the real-time collected operating parameters of active power, speed, head, guide vane opening, and excitation current, multi-parameter joint matching is performed with the multi-parameter boundary threshold ranges of each typical operating condition to achieve accurate identification of the real-time operating condition. For transitional operating conditions during the switching process, linear interpolation is performed based on the boundary parameters of two adjacent typical operating conditions to generate a condition correction coefficient corresponding to the transitional operating condition. Based on the condition correction coefficients corresponding to the matched and identified operating conditions, dynamic correction of the feature thresholds under all operating conditions is achieved. The threshold correction calculation formula is as follows: ; In the formula, For the first Under the first type of working condition Dynamic correction threshold for feature item For the first Under the first type of working condition The baseline threshold for the feature. For the first The operating condition correction factor corresponding to a certain type of operating condition is obtained by fitting the unit design parameters with historical operating data under the same operating conditions.

[0011] Preferably, the intelligent fault diagnosis unit incorporates a dual-branch feature fusion fault diagnosis model and an improved DS evidence theory fault decision-making module. The dual-branch feature fusion fault diagnosis model extracts steady-state and transient features from the global feature vector in parallel, performs weighted feature fusion through an attention mechanism, and outputs preliminary identification results of the fault type and location. The improved DS evidence theory fault decision-making module performs evidence fusion and conflict resolution on the preliminary identification results. The conflict-corrected evidence fusion calculation formula is as follows: ; In the formula, Problem proposition after fusion The basic probability allocation value, The basic probability assignment values ​​for the two sets of independent evidence. The coefficient of conflict of evidence. Problem proposition The credibility coefficient of the evidence Problem proposition The average probability distribution value, This is a fault identification framework.

[0012] Preferably, the intelligent fault diagnosis unit also incorporates a unit performance degradation trend prediction module based on a long short-term memory network. The trend prediction module uses a fixed-step sliding time window to extract global feature vectors from continuous time periods to construct a time-series feature sequence. After Z-score standardization, the extracted time-series feature sequence is input into the long short-term memory network. The long short-term memory network is configured with an input layer, a bidirectional LSTM hidden layer, a dropout layer, a fully connected layer, and an output layer. The bidirectional LSTM hidden layer extracts the forward and reverse temporal dependencies of the time-series feature sequence in parallel. The dropout layer randomly deactivates a preset proportion of neurons. The fully connected layer completes the dimensional mapping of the time-series features. The output layer outputs the predicted sequence of key operating parameters of the unit within a set future time period, as well as the quantitative values ​​of feature degradation rate and fault degradation degree.

[0013] Preferably, the edge-cloud collaborative management and control unit includes an edge terminal module and a cloud sub-module. The edge terminal module uses an industrial-grade edge computing gateway as its hardware carrier, and is configured with multiple analog signal acquisition interfaces, digital signal acquisition interfaces, gigabit Ethernet interfaces, and fiber optic communication interfaces. It establishes bidirectional communication with the multi-source sensor acquisition unit, the unit's local control cabinet, and the unit's monitoring system, and is deployed within the unit's local control cabinet. The edge terminal module has a built-in hard real-time operating system, which performs synchronous acquisition, preprocessing, local feature fusion, and fault diagnosis of multi-source sensor data according to preset acquisition and processing cycles, and handles abnormal data exceeding the dynamic correction threshold. The system triggers local audible and visual alarms and relay dry contact signal outputs. The cloud submodule is built on a flexibly expandable cloud server cluster and is configured with a time-series database, a relational database, and a distributed file storage database, which respectively store the unit's full-cycle time-series operation data, equipment basic parameter data, fault sample data, and model file data. The cloud submodule has built-in distributed model training cluster, fault case library management module, multi-unit cluster management module, and equipment full life cycle management module, which respectively complete the training and optimization of diagnostic models, the archiving and management of fault cases, the unified management and control of multi-unit clusters within the basin, and the traceability of equipment operation data throughout its entire life cycle.

[0014] Preferably, the edge-cloud collaborative management and control unit adopts a lightweight model migration mechanism. First, it performs a combination of unstructured and structured pruning on the cloud-based iteratively optimized fault diagnosis model and trend prediction model, removing redundant neurons and ineffective convolutional and fully connected layers with weights below a preset threshold. Then, it performs INT8 linear quantization compression on the pruned model to generate a lightweight model adapted to the edge hardware's computing power and memory space. The cloud, through a communication channel encrypted with national cryptographic algorithms, sends the lightweight model and verification files to the edge terminal module. During unit downtime or low-load stable operation, the edge terminal module performs legality verification and online replacement of the lightweight model, achieving uninterrupted online updates of the edge diagnosis and prediction models. The edge terminal module, according to a preset upload cycle, uploads manually labeled abnormal operating condition data, fault sample data, and normal operation comparison sample data under the same conditions to the cloud sub-module through an encrypted communication channel. The cloud sub-module supplements the uploaded data to the corresponding fault case library and sample library, and completes incremental training, parameter tuning, and iterative optimization of the fault diagnosis model and trend prediction model based on an incremental learning algorithm.

[0015] Preferably, the multi-source sensing acquisition unit includes a vibration acceleration sensor, an acoustic emission sensor, a platinum resistance temperature sensor, a pressure transmitter, an electromagnetic flowmeter, a current transformer, a voltage transformer, and a key phase sensor; all sensors support the IEEE 1588PTP precision time synchronization protocol to achieve microsecond-level synchronous acquisition of multi-source operating data.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention, through the coordinated configuration of a multi-source sensor acquisition unit and a spatiotemporal synchronous preprocessing unit, achieves microsecond-level spatiotemporal synchronous acquisition and standardized preprocessing of heterogeneous operational data across all dimensions, including vibration, acoustic emission, temperature, pressure, and electrical quantities of hydraulic turbine units. This effectively solves the technical problems of inconsistent spatiotemporal benchmarks for multi-source data and low signal-to-noise ratios under environmental interference in existing monitoring systems. The confidence-optimized adaptive weighted fusion algorithm employed in this invention calculates initial weights based on the minimum mean square error criterion and performs secondary weight optimization using confidence coefficients constructed from real-time signal-to-noise ratios of sensor data, calibration deviations, and operational stability. This achieves effective feature-level fusion of multi-source heterogeneous data, avoiding monitoring distortion caused by single-source sensor failure and improving the accuracy and anti-interference capability of the overall characteristic representation of unit operation. Meanwhile, this invention completes the joint identification of multiple parameters for seven typical operating conditions and transitional operating conditions of the unit through the full-condition boundary adaptive matching unit, and realizes the dynamic adaptation of feature thresholds based on the operating condition correction coefficient. It effectively solves the defect of high false alarm and false alarm rates of fixed thresholds under changing and transient operating conditions, and ensures the effectiveness and stability of the monitoring system in the full operating scenarios of the unit.

[0017] 2. This invention achieves accurate identification and location of hydraulic turbine unit faults through the combination of a dual-branch feature fusion fault diagnosis model and an improved DS evidence theory fault decision-making module. It effectively solves the technical problems of existing diagnostic technologies' insufficient consideration of both steady-state and transient features, and low reliability of fault decision-making under conflicting multi-source evidence. The dual-branch fault diagnosis model extracts steady-state and transient features from the global feature vector in parallel, and completes feature weighted fusion through an attention mechanism, achieving comprehensive capture of fault features under different operating conditions. The improved DS evidence fusion algorithm introduces an evidence credibility coefficient to correct conflicting evidence, effectively resolving evidence conflicts in multi-source diagnostic results and improving the accuracy of fault type identification and location. Simultaneously, the invention's built-in degradation trend prediction module based on a bidirectional LSTM network can quantitatively predict the changing trends of key unit parameters and the degree of fault degradation based on the time-series global feature vector, achieving a technological breakthrough from post-event diagnosis to pre-event early warning of unit faults, providing accurate data support for condition-based maintenance of units.

[0018] 3. This invention adopts an edge-cloud collaborative management system architecture, realizing the collaborative operation of real-time monitoring and alarm at the edge and model iteration and full lifecycle management at the cloud. This effectively solves the technical shortcomings of existing monitoring systems, such as the inability to simultaneously achieve real-time performance and model iteration capabilities, and insufficient system scalability. The edge sub-module, based on an industrial-grade edge computing gateway, completes local data processing, real-time diagnosis, and millisecond-level alarms, avoiding processing delays caused by cloud transmission and ensuring rapid response under abnormal unit operating conditions. The cloud sub-module completes full-cycle data storage for the unit through multiple types of databases, continuously optimizes the diagnostic model based on incremental learning algorithms, and achieves non-intrusive online updates of the edge model through a lightweight model migration mechanism, ensuring continuous adaptation of the diagnostic model to the long-term operating characteristics of the unit. Simultaneously, the system architecture of this invention supports cluster management and control of multiple hydraulic units within a river basin and traceability of equipment operation data throughout its entire lifecycle. The sensor deployment scheme and communication protocol have standardized adaptability, are compatible with different types and capacities of hydraulic units, and possess good engineering feasibility and scalability. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are explained in detail together with the embodiments of the invention, but do not constitute a limitation thereof.

[0020] Figure 1 To illustrate the six core units of this invention—multi-source data acquisition, preprocessing, fusion, operating condition matching, diagnosis, and collaborative management—and the data flow diagram; Figure 2 This is a block diagram of the multi-source sensing acquisition and spatiotemporal synchronization preprocessing unit of the present invention; Figure 3 This is a block diagram of the adaptive multi-source heterogeneous data fusion unit of the present invention; Figure 4 This is a block diagram of the full-condition boundary matching and intelligent fault diagnosis unit of the present invention; Figure 5 This is a block diagram of the edge-cloud collaborative management and control unit of the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figures 1-5As shown, the intelligent diagnosis and full-condition monitoring system for hydraulic units based on multi-source sensor fusion of the present invention achieves accurate perception of the full-condition operation status of hydraulic units through synchronous acquisition and spatiotemporal registration of multi-source heterogeneous data, achieves effective fusion of multi-source features through adaptive weighted fusion algorithm, achieves dynamic correction of feature thresholds through full-condition boundary adaptive matching, achieves accurate identification and location of faults through dual-branch fault diagnosis model and improved evidence fusion algorithm, and achieves collaborative operation of real-time monitoring and model iterative optimization through edge-cloud collaborative architecture.

[0023] This system comprises, in sequence, a multi-source sensor acquisition unit, a spatiotemporal synchronization preprocessing unit, an adaptive multi-source heterogeneous data fusion unit, a full-condition boundary adaptive matching unit, an intelligent fault diagnosis unit, and an edge-cloud collaborative management and control unit. The specific implementation methods of each unit are as follows: Multi-source sensor acquisition units are deployed on the main shaft, guide bearings, turbine runner, stator, excitation system, and water pipelines of the hydraulic turbine unit. These units include vibration acceleration sensors, acoustic emission sensors, platinum resistance temperature sensors, pressure transmitters, electromagnetic flowmeters, current transformers, voltage transformers, and key phase sensors.

[0024] Vibration acceleration sensors are deployed on the horizontal radial and vertical axial mounting surfaces of the upper and lower guide bearings, turbine guide bearings, and thrust bearings of the unit to collect radial and axial vibration acceleration signals under all operating conditions. Acoustic emission sensors are attached to the turbine runner chamber, guide bearing housing, and stator core using couplant to collect high-frequency acoustic emission signals generated by local structural damage, fluid excitation, and partial discharge. Platinum resistance temperature sensors are pre-embedded in each bearing pad, stator winding, stator core, excitation winding, and at the inlet and outlet of the water supply pipeline to collect real-time temperature signals at the corresponding locations. Pressure transmitters are deployed at the turbine spiral casing inlet, tailrace outlet, before and after the movable guide vanes, and pressure steel pipe section to collect pressure and pressure pulsation signals throughout the hydraulic system. Electromagnetic flowmeters are deployed in the straight section of the unit's water supply pipeline inlet to collect real-time flow signals through the unit. Current transformers and voltage transformers are deployed on the stator output side and the excitation system power output side of the unit, respectively, to collect the three-phase stator current, three-phase stator voltage, excitation voltage, and excitation current signals. Key phase sensors are deployed non-contactly on the end face of the unit's main shaft to collect the real-time speed of the main shaft and the synchronous key phase pulse signal.

[0025] All sensors have a built-in IEEE 1588PTP precision time synchronization protocol slave module, which communicates bidirectionally with the system master clock unit via industrial Ethernet to achieve real-time synchronization calibration of the acquisition clock and microsecond-level synchronous acquisition of multi-source operating data.

[0026] The spatiotemporal synchronization preprocessing unit uses the rising edge of the key phase signal on the unit's main shaft as the time reference trigger point to mark a synchronization timestamp for each set of multi-source sensor data frames, completing the microsecond-level time synchronization alignment of the multi-source sensor data. The spatiotemporal synchronization preprocessing unit constructs a three-dimensional spatial coordinate system for the unit using the rotation center of the unit's main shaft as the spatial coordinate origin. It then uses a spatial coordinate mapping matrix to complete the spatial reference registration of sensor data from different installation locations, mapping each sensor data to a unified three-dimensional spatial coordinate system of the unit.

[0027] The spatiotemporal synchronization preprocessing unit employs an improved 3σ criterion, combining the rate of change of adjacent data points in the sensor data to identify and remove outliers. It also uses a wavelet soft thresholding denoising algorithm, selecting the db6 wavelet basis function for 5-level wavelet decomposition to suppress environmental noise and electromagnetic interference in the acquired data. Finally, the unit uses cubic spline interpolation to unify the sampling frequency and align the data sequence for the processed multi-source data.

[0028] The adaptive multi-source heterogeneous data fusion unit first extracts single-dimensional features from the sensor data of each dimension. For vibration and acoustic emission signals, it extracts time-domain statistical features, frequency-domain spectral features, and time-frequency domain wavelet packet energy features. For temperature, pressure, flow, and electrical quantity signals, it extracts time-domain trend features and abrupt change features. Time-domain statistical features include mean, root mean square, peak value, kurtosis, margin factor, and impulse factor. Frequency-domain spectral features include fundamental frequency amplitude, harmonic amplitude, harmonic energy proportion, spectral centroid, and spectral kurtosis. Time-frequency domain wavelet packet energy features include the energy proportion of each frequency band after three-level wavelet packet decomposition and the wavelet packet energy entropy.

[0029] The adaptive multi-source heterogeneous data fusion unit performs min-max normalization on each extracted single-dimensional feature set, mapping all feature values ​​to the 0-1 range. After dimensional alignment of the normalized feature sets to generate a single-dimensional feature matrix with uniform dimensions, it performs confidence-optimized adaptive weighted fusion.

[0030] The adaptive multi-source heterogeneous data fusion unit employs a confidence-optimized adaptive weighted fusion algorithm based on the minimum mean square error criterion to perform feature-level fusion on the registered multi-source heterogeneous data. The basic weight calculation formula is: In the formula, This represents the total number of feature dimensions of the multi-source sensor data. For the first Initial fusion weights for 3D sensing features. For the first Variance of the sensing features.

[0031] The adaptive multi-source heterogeneous data fusion unit introduces a sensor data confidence correction factor. The initial weights are then optimized a second time, and the optimized fusion weights are calculated as follows: In the formula, For the first Confidence coefficient of dimensional sensor data. It is calculated by combining the real-time signal-to-noise ratio of the sensor data, the online calibration deviation, and the historical operational stability coefficient.

[0032] The adaptive multi-source heterogeneous data fusion unit ultimately completes the fusion of multi-source features based on the optimized weights, generating a unified-dimensional global feature vector for unit operation.

[0033] The full-condition boundary adaptive matching unit has a pre-set library of hydraulic turbine operating condition types. This library includes seven typical operating conditions: shutdown, turning gear, no-load, rated load, variable load, load shedding, and phase-shifting operation. Each typical operating condition has a pre-set multi-parameter boundary threshold range. This range includes upper and lower limits for active power, turbine speed, operating head, guide vane opening, and excitation current.

[0034] The full-condition boundary adaptive matching unit performs multi-parameter joint matching based on the real-time collected operating parameters of active power, speed, head, guide vane opening, and excitation current of the unit, and the multi-parameter boundary threshold ranges of various typical operating conditions to achieve accurate identification of real-time operating conditions. For the transitional operating conditions during the switching process, the full-condition boundary adaptive matching unit performs linear interpolation based on the boundary parameters of two adjacent typical operating conditions to generate the corresponding operating condition correction coefficient.

[0035] The full-condition boundary adaptive matching unit dynamically corrects the feature threshold under all working conditions based on the working condition correction coefficients corresponding to the matched and identified working conditions. The threshold correction calculation formula is as follows: In the formula, For the first Under the first type of working condition Dynamically corrected threshold for a feature. For the first Under the first type of working condition The baseline threshold for each feature. For the first The working condition correction factor corresponding to the type of working condition. It is obtained by fitting the unit design parameters with historical operating data under the same conditions.

[0036] The intelligent fault diagnosis unit incorporates a dual-branch feature fusion fault diagnosis model and an improved DS evidence theory fault decision-making module. The dual-branch feature fusion fault diagnosis model extracts steady-state and transient features from the global feature vector in parallel, and after weighted fusion of features through an attention mechanism, outputs preliminary identification results of fault type and fault location.

[0037] The improved DS evidence theory fault decision module performs evidence fusion and conflict resolution on the preliminary identification results. The evidence fusion calculation formula after conflict correction is as follows: In the formula, Problem proposition after fusion The basic probability allocation value. , The basic probability assignment values ​​for the two sets of independent evidence. This represents the conflict coefficient of evidence. Problem proposition The credibility coefficient of the evidence. Problem proposition The average probability distribution value. This is a fault identification framework.

[0038] The intelligent fault diagnosis unit also incorporates a module for predicting unit performance degradation trends based on a Long Short-Term Memory (LSTM) network. This module uses a fixed-step sliding time window to extract global feature vectors from continuous time periods to construct a temporal feature sequence. The module then performs Z-score standardization on the extracted temporal feature sequence before inputting it into the LSTM network. The LSTM network consists of an input layer, a bidirectional LSTM hidden layer, a dropout layer, a fully connected layer, and an output layer. The bidirectional LSTM hidden layer extracts the forward and reverse temporal dependencies of the temporal feature sequence in parallel. The dropout layer randomly deactivates a predetermined proportion of neurons. The fully connected layer performs dimensionality mapping of the temporal features. The output layer outputs the predicted sequence of key operating parameters for the unit within a set future timeframe, along with quantified values ​​of feature degradation rate and fault degradation degree.

[0039] The edge-cloud collaborative management and control unit comprises an edge terminal module and a cloud sub-module. The edge terminal module uses an industrial-grade edge computing gateway as its hardware carrier, configured with multiple analog acquisition interfaces, digital acquisition interfaces, gigabit Ethernet interfaces, and fiber optic communication interfaces. The edge terminal module establishes bidirectional communication with the multi-source sensor acquisition unit, the unit's local control cabinet, and the unit's monitoring system, and is deployed within the unit's local control cabinet. The edge terminal module has a built-in hardware real-time operating system, which performs synchronous acquisition, preprocessing, local feature fusion, and fault diagnosis of multi-source sensor data according to preset acquisition and processing cycles. For abnormal data exceeding the dynamic correction threshold, the edge terminal module triggers local audible and visual alarms and relay dry contact signal outputs.

[0040] The cloud-based submodule is built on a dynamically scalable cloud server cluster, configured with a time-series database, a relational database, and a distributed file storage database. The time-series database stores the entire lifecycle of the generator set's time-series operational data. The relational database stores basic device parameter data. The distributed file storage database stores fault sample data and model file data. The cloud-based submodule includes a distributed model training cluster, a fault case library management module, a multi-generator cluster management module, and a device lifecycle management module. The distributed model training cluster performs training and optimization of diagnostic models. The fault case library management module manages the archiving of fault cases. The multi-generator cluster management module provides unified management of multiple generator clusters within the watershed. The device lifecycle management module enables the tracing of operational data throughout the entire lifecycle of the equipment.

[0041] The edge-cloud collaborative management and control unit employs a lightweight model migration mechanism. First, it performs a combination of unstructured and structured pruning on the cloud-based iteratively optimized fault diagnosis and trend prediction models, removing redundant neurons and ineffective convolutional and fully connected layers with weights below a preset threshold. Then, the pruned models are compressed using INT8 linear quantization to generate a lightweight model adapted to the edge hardware's computing power and memory space. The cloud then distributes the lightweight model and verification files to the edge terminal modules via a communication channel encrypted with national cryptographic algorithms. During unit downtime or low-load stable operation, the edge terminal modules perform legality verification and online replacement of the lightweight model, achieving uninterrupted online updates of the edge diagnostic and prediction models.

[0042] The edge terminal module uploads manually tagged abnormal operating condition data, fault sample data, and normal operation comparison sample data under the same conditions to the cloud sub-module via an encrypted communication channel according to a preset upload cycle. The cloud sub-module supplements the uploaded data into the corresponding fault case library and sample library, and completes incremental training, parameter tuning, and iterative optimization of the fault diagnosis model and trend prediction model based on incremental learning algorithms.

[0043] Example 1: Implementation of Full-Condition Monitoring and Multi-Source Data Fusion under Rated Load Conditions This embodiment is for a mixed-flow hydraulic turbine unit with a rated capacity of 200MW, a rated head of 120m, a rated speed of 150r / min, and the unit is operating under rated load conditions.

[0044] The first step involves the multi-source sensor acquisition unit synchronously acquiring unit operating data. The acquired multi-source sensor data includes 8 vibration acceleration signals, 4 acoustic emission signals, 12 temperature signals, 6 pressure signals, 1 flow signal, 6 electrical quantity signals, and 1 key phase signal. All sensors are synchronized via the IEEE 1588 PTP protocol, achieving a synchronization accuracy of ±1 μs.

[0045] The second step involves a spatiotemporal synchronization preprocessing unit that processes the acquired data. Using the rising edge of the key phase signal as the time reference, timestamps and sequence alignment are completed for all data frames. Spatial coordinate mapping and registration of all sensor data are completed using the spindle rotation center as the origin. An improved 3σ criterion is used to remove outliers, 5-level decomposition using the db6 wavelet basis is employed for noise reduction, and cubic spline interpolation is used to unify the sampling frequency of all data to 10kHz.

[0046] The third step involves the adaptive multi-source heterogeneous data fusion unit completing feature extraction and fusion. Feature extraction is performed on the preprocessed data across various dimensions, yielding a total of [number missing]. Effective features in each dimension. Calculate the variance of each feature dimension. Taking the root mean square characteristic of the first-dimensional vibration acceleration as an example, the following is calculated: Corresponding to the initial weight The calculation process is as follows: Substitute into the calculation to obtain ,therefore .

[0047] Introducing confidence correction factor The initial weights are optimized, and the optimized weights are... The calculation process is as follows: Substitute into the calculation to obtain ,therefore .

[0048] The weight optimization and fusion of all 32-dimensional features are completed according to the above process to generate a global feature vector for unit operation.

[0049] The fourth step involves the full-condition boundary adaptive matching unit completing condition matching and threshold correction. Real-time data is collected on the unit's active power (200MW), speed (150r / min), head (120m), guide vane opening (95%), and excitation current (1200A). This data is then matched against the condition type library and identified as the rated load condition. The corresponding condition correction coefficient for the rated load condition is then determined. Taking the root mean square feature of vibration as an example, the benchmark threshold Dynamically correct threshold .

[0050] Fifth, the intelligent fault diagnosis unit analyzes the global feature vector and outputs a result indicating that the unit is operating normally and there are no abnormal alarms. The edge terminal module completes local real-time monitoring and uploads normal operation data to the cloud at preset intervals for data archiving and storage.

[0051] Example 2: Implementation of bearing fault diagnosis under variable load transition conditions This embodiment refers to the 200MW mixed-flow hydraulic turbine unit in Embodiment 1, where the unit is in a variable load transition condition from 50% rated load to 100% rated load.

[0052] The first step is to use a multi-source sensor acquisition unit to synchronously acquire multi-source operating data of the unit during the load change process. The acquisition parameters are the same as in Example 1, and the synchronization accuracy is maintained at ±1μs.

[0053] The second step involves the spatiotemporal synchronization preprocessing unit completing outlier removal, noise reduction, spatiotemporal registration, and sequence alignment of the data. The processing flow is the same as in Example 1.

[0054] The third step involves the adaptive multi-source heterogeneous data fusion unit completing feature extraction and weight optimization fusion, increasing the feature dimension. The fusion process is the same as in Example 1, generating a global feature vector sequence during the variable load process.

[0055] The fourth step involves the full-condition boundary adaptive matching unit completing condition matching and threshold correction. Real-time data is collected showing the unit's active power increasing from 100MW to 200MW, speed fluctuations ranging from 148r / min to 152r / min, head from 118m to 122m, and guide vane opening increasing from 48% to 95%, identifying this as a variable load transition condition. Linear interpolation is then performed based on the boundary parameters of the no-load and rated load conditions to calculate the corresponding condition correction coefficient for the variable load condition. Taking the root mean square characteristic of horizontal vibration of a turbine guide bearing as an example, the benchmark threshold... Dynamically correct threshold .

[0056] The fifth step involves the intelligent fault diagnosis unit analyzing the global feature vector. The real-time collected root mean square value of the horizontal vibration of the turbine guide bearing is 10.5 mm / s, exceeding the dynamic correction threshold. The dual-branch feature fusion fault diagnosis model extracts steady-state and transient features in parallel, outputting preliminary identification results, the first set of evidence. The basic probability distribution value for the corresponding bearing misalignment fault is 0.82, the second set of evidence. The basic probability distribution value for the corresponding bearing misalignment fault is 0.78.

[0057] Calculate the coefficient of evidence conflict The credibility coefficient of evidence corresponding to the proposition of bearing misalignment failure Average probability distribution value Substituting into the improved DS evidence fusion formula: Step-by-step calculations yielded the molecule as follows: The denominator is ,final .

[0058] Based on the fusion results, the fault decision module outputs the identification result of the turbine guide bearing misalignment fault, locates the fault to the position of the turbine guide bearing, triggers the ground audible and visual alarm of the edge terminal module, and uploads the fault data to the cloud to supplement the fault case library.

[0059] Example 3: Implementation of Unit Deterioration Trend Prediction under Load Shedding Transient Conditions This embodiment focuses on the 200MW mixed-flow hydraulic turbine unit in Embodiment 1, where the unit experiences a transient load shedding condition at 100% rated load.

[0060] The first step involves the multi-source sensor acquisition unit synchronously acquiring multi-source operating data during the load shedding process at a sampling frequency of 20kHz. The acquisition parameters are consistent with those in Example 1, and the time synchronization accuracy is maintained at ±1μs.

[0061] The second step involves the spatiotemporal synchronization preprocessing unit completing outlier removal, noise reduction, spatiotemporal registration, and sequence alignment of the data. The processing flow is the same as in Example 1.

[0062] The third step involves the adaptive multi-source heterogeneous data fusion unit completing feature extraction and weight optimization fusion, increasing the feature dimension. The fusion process is the same as in Example 1, generating a global feature vector time sequence during the load shedding process.

[0063] The fourth step involves the full-condition boundary adaptive matching unit completing condition matching and threshold correction. Real-time data acquisition shows the unit's active power dropping from 200MW to 0MW, the maximum speed rising to 165r / min, and the guide vane opening rapidly closing from 95% to 0%, identifying this as a load shedding condition. The corresponding condition correction coefficient for the load shedding condition is then determined. Taking the root mean square characteristic of vertical vibration of a thrust bearing as an example, the benchmark threshold Dynamically correct threshold .

[0064] The fifth step involves the degradation trend prediction module of the intelligent fault diagnosis unit. Using a sliding time window with a step size of 100, it extracts a 10-second global feature vector time series during the load shedding process. After Z-score standardization, this sequence is input into a bidirectional LSTM network. The bidirectional LSTM network outputs a predicted sequence for the thrust bearing temperature over the next 30 seconds. The predicted temperature of the thrust bearing pads will rise from 65℃ to 78℃ within the next 30 seconds, with a degradation rate of 0.43℃ / s and a fault degradation quantification value of 0.72.

[0065] Based on the prediction results, the edge terminal module triggers an early warning for the unit and uploads transient operating condition data and prediction results to the cloud. The cloud uses the uploaded load shedding data to complete incremental training of the diagnostic model. The optimized model is then processed to be lightweight and distributed to the edge terminal, completing the online update of the edge terminal model.

[0066] This invention achieves microsecond-level spatiotemporal synchronous acquisition and standardized preprocessing of heterogeneous operational data across all dimensions, including vibration, acoustic emission, temperature, pressure, and electrical quantities, from hydraulic turbine units through the coordinated configuration of a multi-source sensor acquisition unit and a spatiotemporal synchronous preprocessing unit. This effectively solves the technical problems of inconsistent spatiotemporal benchmarks for multi-source data and low signal-to-noise ratios under environmental interference in existing monitoring systems. The confidence-optimized adaptive weighted fusion algorithm employed in this invention calculates initial weights based on the minimum mean square error criterion and performs secondary weight optimization using confidence coefficients constructed from real-time signal-to-noise ratios, calibration deviations, and operational stability of the sensor data. This achieves effective feature-level fusion of multi-source heterogeneous data, avoiding monitoring distortion caused by single-source sensor failure and improving the accuracy and anti-interference capability of the overall operational characteristics of the unit. Meanwhile, this invention completes the joint identification of multiple parameters for seven typical operating conditions and transitional operating conditions of the unit through the full-condition boundary adaptive matching unit, and realizes the dynamic adaptation of feature thresholds based on the operating condition correction coefficient. It effectively solves the defect of high false alarm and false alarm rates of fixed thresholds under changing and transient operating conditions, and ensures the effectiveness and stability of the monitoring system in the full operating scenarios of the unit.

[0067] This invention achieves accurate identification and location of faults in hydraulic turbine units through the combination of a dual-branch feature fusion fault diagnosis model and an improved DS evidence theory fault decision-making module. It effectively solves the technical problems of existing diagnostic technologies, such as insufficient consideration of both steady-state and transient features and low reliability of fault decision-making under conflicting multi-source evidence. The dual-branch fault diagnosis model extracts steady-state and transient features from the global feature vector in parallel, and completes feature weighted fusion through an attention mechanism, achieving comprehensive capture of fault features under different operating conditions. The improved DS evidence fusion algorithm introduces an evidence credibility coefficient to correct conflicting evidence, effectively resolving evidence conflicts in multi-source diagnostic results and improving the accuracy of fault type identification and location. Simultaneously, the invention's built-in degradation trend prediction module based on a bidirectional LSTM network can quantitatively predict the changing trends of key unit parameters and the degree of fault degradation based on the time-series global feature vector, achieving a technological breakthrough from post-fault diagnosis to pre-fault warning for unit faults, and providing accurate data support for condition-based maintenance of units.

[0068] This invention adopts an edge-cloud collaborative management system architecture, realizing the coordinated operation of real-time monitoring and alarm at the edge and model iteration and full lifecycle management in the cloud. This effectively solves the technical shortcomings of existing monitoring systems, such as the inability to simultaneously achieve real-time performance and model iteration capabilities, and insufficient system scalability. The edge sub-module, based on an industrial-grade edge computing gateway, completes local data processing, real-time diagnosis, and millisecond-level alarms, avoiding processing delays caused by cloud transmission and ensuring rapid response under abnormal unit operating conditions. The cloud sub-module completes full-cycle data storage for the unit through multiple types of databases, continuously optimizes the diagnostic model based on incremental learning algorithms, and achieves non-intrusive online updates of the edge model through a lightweight model migration mechanism, ensuring continuous adaptation of the diagnostic model to the long-term operating characteristics of the unit. Simultaneously, the system architecture of this invention supports cluster management and control of multiple hydraulic units within a river basin and traceability of equipment operation data throughout its entire lifecycle. The sensor deployment scheme and communication protocol have standardized adaptability, are compatible with different types and capacities of hydraulic units, and possess good engineering feasibility and scalability.

[0069] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A smart diagnostic and full-condition monitoring system for hydraulic turbine units based on multi-source sensor fusion, characterized in that: It includes a multi-source sensor acquisition unit, a spatiotemporal synchronous preprocessing unit, an adaptive multi-source heterogeneous data fusion unit, a full-condition boundary adaptive matching unit, an intelligent fault diagnosis unit, and an edge-cloud collaborative management and control unit that are connected in sequence. The multi-source sensing and acquisition unit is deployed on the main shaft, guide bearing, turbine runner, stator, excitation system and water pipeline of the hydraulic unit, and is used to synchronously collect multi-source heterogeneous operating data of vibration, acoustic emission, temperature, pressure, flow rate, stator current and excitation voltage under all operating conditions of the unit. The spatiotemporal synchronization preprocessing unit is used to remove outliers, filter and reduce noise in the collected multi-source data, and complete the spatiotemporal reference registration and sequence alignment of the multi-source data. The adaptive multi-source heterogeneous data fusion unit is used to perform feature-level adaptive weighted fusion on the registered multi-source data to generate a global feature vector for unit operation. The full-condition boundary adaptive matching unit is used to match the corresponding operating condition boundary based on the real-time operating parameters of the unit, and to complete the dynamic correction of the feature threshold under the full operating condition. The intelligent fault diagnosis unit is used to identify the type of unit fault, locate the fault, and predict the deterioration trend based on the corrected global feature vector. The edge-cloud collaborative management and control unit is used to realize real-time monitoring and alarm at the edge, iterative updates of cloud models, data traceability, and full lifecycle management.

2. The intelligent diagnostic and full-condition monitoring system for hydraulic units based on multi-source sensor fusion as described in claim 1, characterized in that, The spatiotemporal synchronization preprocessing unit uses the rising edge of the key phase signal of the unit's main shaft as the time reference trigger point to mark a synchronization timestamp for each set of multi-source sensor data frames, thereby completing the microsecond-level time synchronization and alignment of multi-source sensor data. Using the rotation center of the unit's main shaft as the origin of the spatial coordinate system, a three-dimensional spatial coordinate system for the unit is constructed. A spatial coordinate mapping matrix is ​​used to complete the spatial reference registration of sensor data from different installation locations, mapping each sensor data to a unified three-dimensional spatial coordinate system of the unit. At the same time, an improved 3σ criterion is adopted, combined with the rate of change of adjacent data points of the sensor data, to complete the identification and removal of outliers. A wavelet soft thresholding denoising algorithm is used, and the db6 wavelet basis function is selected for 5-level wavelet decomposition to suppress environmental noise and electromagnetic interference in the collected data. Cubic spline interpolation is used to unify the sampling frequency and align the data sequence for the processed multi-source data.

3. The intelligent diagnostic and full-condition monitoring system for hydraulic units based on multi-source sensor fusion as described in claim 1, characterized in that, The adaptive multi-source heterogeneous data fusion unit employs a confidence-optimized adaptive weighted fusion algorithm based on the minimum mean square error criterion to perform feature-level fusion on the registered multi-source heterogeneous data. Its basic weight calculation formula is as follows: ； In the formula, The total number of feature dimensions of multi-source sensor data. For the first Initial fusion weights for 3D sensing features, For the first Variance of dimensional sensing features; Simultaneously, a sensor data confidence correction factor is introduced. The initial weights are then optimized a second time, and the optimized fusion weights are calculated as follows: ； In the formula, For the first The confidence coefficient of the dimensional sensor data is calculated by comprehensively considering the real-time signal-to-noise ratio, online calibration deviation, and historical operational stability coefficient of the sensor data. Finally, based on the optimized weights, multi-source feature fusion is completed to generate a unified-dimensional global feature vector for unit operation.

4. The intelligent diagnostic and full-condition monitoring system for hydraulic units based on multi-source sensor fusion according to claim 3, characterized in that, The adaptive multi-source heterogeneous data fusion unit first performs single-dimensional feature extraction on the sensor data of each dimension. For vibration and acoustic emission signals, it extracts time-domain statistical features, frequency-domain spectral features and time-frequency-domain wavelet packet energy features. For temperature, pressure, flow rate and electrical quantity signals, it extracts time-domain trend features and abrupt change features. The time-domain statistical features include mean, root mean square, peak value, kurtosis, margin factor, and impulse factor. The frequency-domain spectral features include fundamental frequency amplitude, harmonic amplitude, harmonic energy ratio, spectral centroid, and spectral kurtosis. The time-frequency domain wavelet packet energy features include the energy ratio of each frequency band after three-layer wavelet packet decomposition and the wavelet packet energy entropy. The extracted single-dimensional feature sets are subjected to min-max normalization to map all feature values ​​to the [0,1] interval. The normalized feature sets are then subjected to dimension alignment to generate a single-dimensional feature matrix with unified dimensions. Finally, confidence-optimized adaptive weighted fusion is performed.

5. The intelligent diagnostic and full-condition monitoring system for hydraulic units based on multi-source sensor fusion as described in claim 1, characterized in that, The full-condition boundary adaptive matching unit has a pre-set library of hydraulic turbine unit operating condition types. This library includes seven typical operating conditions: shutdown, turning gear, no-load, rated load, load change, load shedding, and phase-shifting operation. Each typical operating condition has a corresponding multi-parameter boundary threshold range, which includes upper and lower limits for active power, turbine speed, operating head, guide vane opening, and excitation current. Based on the real-time collected operating parameters of active power, speed, head, guide vane opening, and excitation current, multi-parameter joint matching is performed with the multi-parameter boundary threshold ranges of each typical operating condition to achieve accurate identification of the real-time operating condition. For transitional operating conditions during the switching process, linear interpolation is performed based on the boundary parameters of two adjacent typical operating conditions to generate a corresponding operating condition correction coefficient. Based on the matching and identified operating condition correction coefficients, dynamic correction of the feature thresholds under all operating conditions is achieved. The threshold correction calculation formula is as follows: ； In the formula, For the first Under the first type of working condition Dynamic correction threshold for feature item For the first Under the first type of working condition The baseline threshold for the feature. For the first The operating condition correction coefficient corresponding to a certain operating condition is obtained by fitting the unit design parameters with historical operating data under the same operating condition.

6. The intelligent diagnostic and full-condition monitoring system for hydraulic units based on multi-source sensor fusion as described in claim 1, characterized in that, The intelligent fault diagnosis unit incorporates a dual-branch feature fusion fault diagnosis model and an improved DS evidence theory fault decision-making module. The dual-branch feature fusion fault diagnosis model extracts steady-state and transient features from the global feature vector in parallel, performs weighted feature fusion through an attention mechanism, and outputs preliminary identification results of the fault type and location. The improved DS evidence theory fault decision-making module performs evidence fusion and conflict resolution on the preliminary identification results. The conflict-corrected evidence fusion calculation formula is as follows: ； In the formula, Problem proposition after fusion The basic probability allocation value, The basic probability assignment values ​​for the two sets of independent evidence. The coefficient of conflict of evidence. Problem proposition The credibility coefficient of the evidence Problem proposition The average probability distribution value, This is a fault identification framework.

7. The intelligent diagnostic and full-condition monitoring system for hydraulic units based on multi-source sensor fusion as described in claim 6, characterized in that, The intelligent fault diagnosis unit also incorporates a unit performance degradation trend prediction module based on a long short-term memory network. The trend prediction module uses a fixed-step sliding time window to extract global feature vectors from continuous time periods to construct a temporal feature sequence. After Z-score standardization, the extracted temporal feature sequence is input into the long short-term memory network. The long short-term memory network is configured with an input layer, a bidirectional LSTM hidden layer, a dropout layer, a fully connected layer, and an output layer. The bidirectional LSTM hidden layer extracts the forward and reverse temporal dependencies of the temporal feature sequence in parallel. The dropout layer randomly deactivates a preset proportion of neurons. The fully connected layer completes the dimensional mapping of the temporal features. The output layer outputs the predicted sequence of key operating parameters of the unit within a set future time period, as well as the quantitative values ​​of feature degradation rate and fault degradation degree.

8. The intelligent diagnostic and full-condition monitoring system for hydraulic units based on multi-source sensor fusion according to claim 1, characterized in that, The edge-cloud collaborative management and control unit includes an edge terminal module and a cloud sub-module. The edge terminal module uses an industrial-grade edge computing gateway as its hardware carrier, configured with multiple analog signal acquisition interfaces, digital signal acquisition interfaces, gigabit Ethernet interfaces, and fiber optic communication interfaces. It establishes bidirectional communication with the multi-source sensor acquisition unit, the unit's local control cabinet, and the unit's monitoring system, and is deployed within the unit's local control cabinet. The edge terminal module has a built-in hard real-time operating system that, according to preset acquisition and processing cycles, performs synchronous acquisition, preprocessing, local feature fusion, and fault diagnosis of multi-source sensor data. For abnormal data exceeding the dynamic correction threshold, it triggers... The system includes local audible and visual alarms and relay dry contact signal outputs. The cloud submodule is built on a flexibly expandable cloud server cluster and is configured with a time-series database, a relational database, and a distributed file storage database, which respectively store the unit's full-cycle time-series operating data, equipment basic parameter data, fault sample data, and model file data. The cloud submodule also includes a built-in distributed model training cluster, a fault case library management module, a multi-unit cluster management module, and an equipment full lifecycle management module, which respectively complete the training and optimization of diagnostic models, the archiving and management of fault cases, the unified management and control of multi-unit clusters within the basin, and the traceability of operating data throughout the equipment's lifecycle.

9. The intelligent diagnostic and full-condition monitoring system for hydraulic units based on multi-source sensor fusion as described in claim 8, characterized in that, The edge-cloud collaborative management and control unit adopts a model lightweight migration mechanism. First, it performs a combination of unstructured pruning and structured pruning on the fault diagnosis model and trend prediction model after iterative optimization in the cloud. Redundant neurons and invalid convolutional layers and fully connected layers with weights below a preset threshold are removed from the model. Then, the pruned model is compressed using INT8 linear quantization to generate a lightweight model that is adapted to the computing power and memory space of the edge hardware. The cloud-based terminal module distributes the lightweight model and verification files to the edge terminal module via a communication channel encrypted with national cryptographic algorithms. During unit shutdown periods or low-load stable operation periods, the edge terminal module performs legality verification and online replacement of the lightweight model, achieving uninterrupted online updates of the edge-end diagnostic and prediction models. According to a preset upload cycle, the edge terminal module uploads manually marked abnormal operating condition data, fault sample data, and normal operation comparison sample data under the same operating conditions to the cloud sub-module via an encrypted communication channel. The cloud sub-module supplements the uploaded data into the corresponding fault case library and sample library, and completes incremental training, parameter tuning, and iterative optimization of the fault diagnosis model and trend prediction model based on incremental learning algorithms.

10. The intelligent diagnostic and full-condition monitoring system for hydraulic units based on multi-source sensor fusion according to claim 1, characterized in that, The multi-source sensing and acquisition unit includes a vibration acceleration sensor, an acoustic emission sensor, a platinum resistance temperature sensor, a pressure transmitter, an electromagnetic flowmeter, a current transformer, a voltage transformer, and a key phase sensor; all sensors support the IEEE 1588PTP precision time synchronization protocol to achieve microsecond-level synchronous acquisition of multi-source operating data.

Images