Intelligent early warning method, device and server for full state of water turbine

Abstract

The application provides a water turbine full-state intelligent early warning method and device and a server, relates to the technical field of intelligent operation and maintenance of hydroelectric generating equipment, and comprises the following steps: performing data cleaning processing and multi-scale time alignment processing on multi-modal data, performing feature fusion processing on target multi-modal data based on a cross-modal attention model to obtain target fusion features; adopting a multi-modal space-time memory network model, performing history state memory processing on the target fusion features through a four-branch parallel memory structure to obtain a memory state set, and performing cross-modal memory fusion processing on each memory state in the memory state set through a covariance adaptive fusion model to obtain a unified memory state; inputting the unified memory state into a full connection network model to perform mapping processing on a future state to obtain a prediction value and an uncertainty degree of the future state of the water turbine, and performing hierarchical early warning based on the prediction value and the uncertainty degree. The application can significantly improve the accuracy of intelligent early warning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent operation and maintenance of hydroelectric power generation equipment, and in particular to a water turbine full-state intelligent early warning method and device and a server. BACKGROUND

[0002] The water turbine is the core equipment of the hydropower station, and its running state is directly related to the power generation efficiency and safety. At present, the related technology proposes that the existing water turbine early warning scheme has problems such as insufficient model self-adaptive ability, insufficient early warning foresight, and missing root cause analysis. In addition, since the water turbine state monitoring system usually collects multi-modal data, and the existing water turbine early warning scheme usually analyzes the multi-modal data in isolation, the scheme also causes the sensor data to be unable to be effectively fused, thereby affecting the accuracy of the early warning. SUMMARY

[0003] Therefore, the purpose of the present application is to provide a water turbine full-state intelligent early warning method, device and server, which can significantly improve the accuracy of intelligent early warning.

[0004] In a first aspect, the present application provides a water turbine full-state intelligent early warning method, which comprises: performing data cleaning processing and multi-scale time alignment processing on multi-modal data collected by a sensor set of the water turbine to obtain target multi-modal data, and performing feature fusion processing on the target multi-modal data based on a cross-modal attention model to obtain target fusion features; using a multi-modal spatiotemporal memory network model to perform historical state memory processing on the target fusion features through a four-branch parallel memory structure to obtain a memory state set, and performing cross-modal memory fusion processing on each memory state in the memory state set through a covariance adaptive fusion model to obtain a unified memory state; inputting the unified memory state into a fully connected network model to perform mapping processing on the future state to obtain a prediction value and an uncertainty of the future state of the water turbine, and performing hierarchical early warning based on the prediction value and the uncertainty.

[0005] In one embodiment, the step of performing data cleaning processing and multi-scale time alignment processing on the multi-modal data collected by the sensor set of the water turbine to obtain the target multi-modal data comprises: after performing temperature data cleaning processing and vibration data cleaning processing on the multi-modal data, sending the multi-modal data to a dynamic time warping model, and calculating the optimal matching path between different modal time series to determine a time lag estimation value; performing spline interpolation processing based on the time lag estimation value to align the multi-scale time of the multi-modal data to obtain the target multi-modal data.

[0006] In an implementation, based on the cross-modal attention model, the step of performing feature fusion processing on the target multi-modal data to obtain target fusion features includes: performing feature extraction processing on the target multi-modal data to obtain physical prior features corresponding to each modal data; and performing feature fusion processing on each physical prior feature through the cross-modal attention model to obtain the target fusion features.

[0007] In an implementation, the step of performing history state memory processing on the target fusion features through a four-branch parallel memory structure to obtain a memory state set using a multi-modal spatio-temporal memory network model includes: using a temperature memory branch, a vibration memory branch, a pressure memory branch, and a working condition memory branch to perform history state memory processing on temperature feature components, vibration feature components, pressure feature components, and working condition feature components in the target fusion features, respectively, to obtain the memory state set using a gated memory update mechanism to exponentially decay and accumulate the history evolution state of each modal within a corresponding preset time window.

[0008] In an implementation, after the step of obtaining the predicted value and the uncertainty of the future state of the hydraulic turbine, the step includes: when the unit of the hydraulic turbine is replaced, collecting small sample data of the start-stop process of the replaced hydraulic turbine, and performing model fine-tuning processing on the initial model trained by the original hydraulic turbine based on the model-agnostic meta-learning framework and the small sample data to obtain a target model corresponding to the replaced hydraulic turbine.

[0009] In an implementation, the step of performing hierarchical early warning based on the predicted value and the uncertainty includes: performing anomaly scoring processing on the predicted value and the uncertainty to obtain an anomaly score, and performing cumulative calculation on the anomaly score through sequential testing to obtain an anomaly score cumulative value; when the anomaly score cumulative value reaches the trigger threshold of any level of the preset three-level hierarchical trigger condition, performing corresponding early warning processing, and locating the root cause of the three-level early warning when the three-level early warning is reached.

[0010] In an implementation, the step of locating the root cause of the three-level early warning includes: obtaining a causal graph and a transfer entropy of the components of the hydraulic turbine, and calculating the causal strength between the bearings, the frame, and the shell based on the transfer entropy and the causal graph to determine a target causal strength; based on the target causal strength, performing similarity retrieval processing in a historical case library to match a similar historical fault set corresponding to the target causal strength, and determining a target root cause with the highest probability and a maintenance recommendation list corresponding to the target root cause according to the root cause posterior probability of each similar historical fault in the similar historical fault set.

[0011] In a second aspect, the embodiments of the present application also provide a water turbine full-state intelligent early warning device, which comprises: a feature fusion module, which performs data cleaning processing and multi-scale time alignment processing on multi-modal data collected by a sensor set of the water turbine, obtains target multi-modal data, and performs feature fusion processing on the target multi-modal data based on a cross-modal attention model to obtain target fusion features; a state memory module, which adopts a multi-modal spatio-temporal memory network model, performs historical state memory processing on the target fusion features through a four-branch parallel memory structure to obtain a memory state set, and performs cross-modal memory fusion processing on each memory state in the memory state set through a covariance adaptive fusion model to obtain a unified memory state; and a state early warning module, which inputs the unified memory state into a full connection network model to perform mapping processing on a future state, obtains a prediction value and an uncertainty of the future state of the water turbine, and performs hierarchical early warning based on the prediction value and the uncertainty.

[0012] In a third aspect, the embodiments of the present application also provide a server, which comprises a processor and a memory, the memory stores computer executable instructions capable of being executed by the processor, and the processor executes the computer executable instructions to implement the method of any one of the first aspect.

[0013] In a fourth aspect, the embodiments of the present application also provide a computer readable storage medium, which stores computer executable instructions, and the computer executable instructions, when invoked and executed by a processor, cause the processor to implement the method of any one of the first aspect.

[0014] The embodiments of the present application have the following beneficial effects:

[0015] The water turbine full-state intelligent early warning method, device and server provided by the embodiments of the present application perform data cleaning processing and multi-scale time alignment processing on multi-modal data collected by a sensor set of the water turbine, obtain target multi-modal data, perform feature fusion processing on the target multi-modal data based on a cross-modal attention model to obtain target fusion features, then adopt a multi-modal spatio-temporal memory network model, perform historical state memory processing on the target fusion features through a four-branch parallel memory structure to obtain a memory state set, perform cross-modal memory fusion processing on each memory state in the memory state set through a covariance adaptive fusion model to obtain a unified memory state, finally input the unified memory state into a full connection network model to perform mapping processing on a future state, obtain a prediction value and an uncertainty of the future state of the water turbine, and perform hierarchical early warning based on the prediction value and the uncertainty, so that the accuracy of intelligent early warning can be significantly improved.

[0016] Other features and advantages of the present application will be set forth in the descriptions below, and in part will be apparent from the description, or can be learned by practice of the application.

[0017] In order to make the above objectives, features and advantages of the present application more apparent, the following preferred embodiments are specifically described below with reference to the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the specific embodiments of the present application or the technical solutions in the prior art, the following will briefly introduce the drawings needed to be used in the specific embodiments or the prior art description. Obviously, the drawings described below are some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0019] Figure 1 A flowchart of a water turbine full-state intelligent early warning method provided by an embodiment of the present application is shown in the figure.

[0020] Figure 2 A specific flowchart of a water turbine full-state intelligent early warning method provided by an embodiment of the present application is shown in the figure.

[0021] Figure 3 A structure diagram of a multi-modal spatio-temporal memory network model provided by an embodiment of the present application is shown in the figure.

[0022] Figure 4 A structure diagram of a water turbine full-state intelligent early warning device provided by an embodiment of the present application is shown in the figure.

[0023] Figure 5 A structure diagram of a server provided by an embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0024] In order to make the objectives, technical solutions and advantages of the embodiments of the present application more apparent, the technical solutions of the present application will be described clearly and completely below with reference to the embodiments. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0025] As the core equipment of a hydropower station, the operation state of a water turbine is directly related to the power generation efficiency and safety. The current water turbine state monitoring system usually collects multi-modal data, including temperature, vibration, swing, pressure fluctuation and working condition parameters (such as rotation speed, power, water head, flow, etc.).

[0026] However, the existing monitoring methods have the following main problems: 1. Isolated analysis of multi-modal data: Temperature, vibration, pressure pulsation and other data are usually processed independently, lacking effective cross-modal collaborative analysis mechanism, and it is difficult to fully capture the overall operation state of the equipment, such as the failure to effectively fuse the data of 94 temperature sensors and more than 20 vibration sensors of a hydropower station. 2. Insufficient model adaptive ability: Traditional statistical or machine learning models (such as random forest, support vector machine, etc.) are slow to adapt to new units and new working conditions (start-stop, load fluctuation). 3. Insufficient early warning foresight: The early warning mechanism based on statistical error is mostly post-alarm, lacking early warning capability. 4. Lack of root cause analysis: Existing systems can only detect abnormalities, but cannot locate the fault source and propagation path, and cannot provide effective support for operation and maintenance decisions.

[0027] In the context of a Francis turbine, the number of temperature sensors can reach dozens to hundreds, while the sampling frequencies and response time lags of vibration, swing and pressure pulsation are different, and the signals are affected by heat-flow-structure coupling. Based on this, the full-state intelligent early warning method, device and server provided by the present application can significantly improve the accuracy of intelligent early warning.

[0028] Referring to Figure 1 the flowchart of a full-state intelligent early warning method for a hydraulic turbine, the method mainly includes the following steps S102 to S106:

[0029] Step S102, the multi-modal data collected by the sensor set of the hydraulic turbine is subjected to data cleaning and multi-scale time alignment processing to obtain target multi-modal data, and the target multi-modal data is subjected to feature fusion processing based on a cross-modal attention model to obtain target fusion features.

[0030] In an embodiment, a system deployed on a power station industrial server (or edge gateway) can be used to uniformly complete data acquisition, timestamp synchronization and preliminary caching, continuously acquire multi-modal data at a set period (for example, 1s) and write to a data bus, and then call the built-in data processing module to perform the following operations:

[0031] 1. For temperature data, the program calculates the theoretical residual based on the discretized heat conduction equation. If the residual of a certain measurement point (threshold ), it is automatically marked as abnormal, and the data of other sensors at the same site is used to repair it through a neighborhood interpolation algorithm.

[0032] 2. For vibration data, its power spectral density (PSD) is calculated in real time, and if the total energy in the 0.5 to 100 Hz frequency band compared to the historical baseline, it increases by 50%, a 5th order Butterworth bandpass filter is triggered to reconstruct the signal.

[0033] 3. Calculate the time lag of vibration, pressure fluctuation and relative temperature data based on 1-second time axis using dynamic time warping (DTW) algorithm , and synchronize all modal data to a unified timestamp based on cubic spline interpolation method to generate a synchronized sequence .

[0034] Then call the feature extraction module to calculate the physical prior features of each mode in parallel, including: temperature gradient , vibration main frequency (computed by FFT), pressure fluctuation vortex belt energy ratio, etc. After splicing the physical features with the original data, input them into a deep learning encoder (constituted by CNN and LSTM units) to obtain high-dimensional features . Then, use the cross-modal attention mechanism to automatically calculate the weight of each mode , and weightedly fuse to obtain a unified feature representation for subsequent prediction model.

[0035] Step S104, adopt a multi-modal spatio-temporal memory network model to perform historical state memory processing on the target fusion features through a four-branch parallel memory structure to obtain a memory state set, and perform cross-modal memory fusion processing on each memory state in the memory state set through a covariance adaptive fusion model to obtain a unified memory state.

[0036] In an embodiment, the core network model file can be called for online prediction. The fusion features are input into a pre-trained multi-modal spatio-temporal memory network (MSTM-Net). The network includes four memory branches:

[0037] 1. In the temperature memory branch, the time constant is minutes, the memory state is updated slowly , and the temperature rising trend of bearings and other components is captured.

[0038] 2. In the vibration memory branch, the time constant is seconds, the memory state is updated quickly , and a frequency band attention gate is equipped to focus on high-frequency components such as 80-120Hz related to faults.

[0039] 3. In the pressure memory branch, the time constant is seconds, the pressure fluctuation signal is processed, and the vortex frequency (such as 0.5-2Hz) and high-frequency water pressure fluctuation are focused on to capture hydraulic instability phenomena.

[0040] 4. In the working condition memory branch, based on the hidden semi-Markov model (HSMM), the running states such as unit startup, steady state, shutdown, etc. are automatically identified and encoded The network finally outputs temperature, vibration trend prediction value and prediction uncertainty of the next 60 minutes.

[0041] In addition, the rapid adaptation of the new unit can also be realized through meta-learning, and the dependence on historical data is reduced. In the meta-learning adaptation, when the system is applied to any new unit of the power station, the model-independent meta-learning (MAML) framework can be used for rapid adaptation. Only 5 complete start-stop process data (about 40 hours) of the new unit need to be collected, and the learning script is run, so that the model fine-tuning can be completed within 2 hours without retraining.

[0042] In step S106, the unified memory state is input into the full connection network model to perform mapping processing on the future state, to obtain the prediction value and uncertainty of the future state of the hydraulic turbine, and perform hierarchical early warning based on the prediction value and the uncertainty.

[0043] In an embodiment, the early warning and root cause module can be called. In the abnormal score and early warning, the abnormal score is calculated in real time . The cumulative calculation is performed through the CUSUM sequential test algorithm , and the hierarchical rules (assuming that the default threshold = 4.5, ) are used: when , it is level one (note), and only a prompt is displayed on the monitoring interface; when , and the temperature and vibration mode are abnormal, it is level two (warning); when , the system automatically pops up an alarm window and starts root cause analysis. In the root cause positioning, for the third-level alarm, the system automatically constructs a component causal diagram, calculates the causal strength between the bearings, the frame and the shell through the transfer entropy, simultaneously, searches the historical case library (a MySQL database), matches similar historical faults through the DTW algorithm, and finally outputs the report on the interface.

[0044] The above-mentioned intelligent early warning method for the hydraulic turbine provided by the embodiment of the application is based on a multimodal spatio-temporal memory network (MSTM-Net), and realizes the early, accurate and interpretable early warning of the hydraulic turbine state through physical constraint data cleaning and alignment, feature extraction and cross-modal fusion, multi-branch spatio-temporal memory network modeling, meta-learning rapid adaptation, abnormal score and hierarchical early warning, and causal root cause positioning and suggestion generation, and solves the problems of insufficient multi-modal cooperation, slow adaptation, post-warning and uninterpretable.

[0045] Referring to a specific flowchart of an intelligent early warning method for a hydraulic turbine shown in Figure 2 , the embodiment of the application further provides an implementation manner for early warning of the hydraulic turbine, and specifically refers to the following (1) to (6):

[0046] (1) Physical constraint cleaning and alignment: based on thermal equilibrium and PSD consistency repair anomalies, using DTW + physical time lag prior to complete multi-modal alignment. After the multi-modal data is sent to the dynamic time warping model after temperature data cleaning and vibration data cleaning, the optimal matching path between different modal time series is calculated to determine the time lag estimation value, and then based on the time lag estimation value, spline interpolation processing is performed to align the multi-scale time of the multi-modal data, and the target multi-modal data is obtained.

[0047] In an embodiment, the role of this step is to unify the format and sampling frequency of multi-source heterogeneous data, providing a basis for subsequent processing. Specifically, first, let the observation time sequence be Define the following multi-modal data:

[0048] Temperature mode: , the sampling period is , where is the number of temperature sensors (e.g. 94).

[0049] Vibration, swing mode: , the sampling period is , where is the number of vibration and swing sensors.

[0050] Pressure pulsation mode: , the sampling period is , where is the number of pressure pulsation sensors.

[0051] Working condition mode: , including speed, power, head, flow and other parameters, the sampling period is .

[0052] Therefore, the multi-modal input data is recorded as:

[0053]

[0054] After collecting temperature, vibration, swing, pressure pulsation and working condition data, the following processing is performed:

[0055] 1. Temperature data cleaning: based on the theoretical temperature change calculated by the heat conduction equation, compared with the actual measured value. If the residual exceeds the threshold (such as ), then interpolate and repair based on the neighborhood sensor data and the principle of conservation of energy.

[0056] Anomaly detection and repair based on the heat conduction equation:

[0057]

[0058] where, is the thermal diffusivity, is the discrete Laplacian (discrete form of Laplace operator on sensor graph), is the internal heat source term. The discretized temperature residual is defined as:

[0059]

[0060] If (predefined threshold), it is determined as an outlier and repaired based on neighborhood interpolation and energy conservation.

[0061] 2. Vibration data cleaning: Through power spectral density analysis, detect abnormal band energy. If a peak value is found that does not match the inherent frequency of the device, start band-pass filtering to reconstruct the signal.

[0062] Perform power spectral density (PSD) analysis on the vibration signal and constrain the band energy:

[0063]

[0064] where, is the PSD at time , is the frequency band of interest, is the energy of this band. If the integral value deviates from the historical statistical range, band-pass reconstruction and harmonic suppression will be triggered to correct the abnormal signal.

[0065] 3. Multi-scale time alignment: Use dynamic time warping (DTW) algorithm, combined with physical priors such as vibration propagation speed (about 500 m / s), to estimate the time lag between each mode, and achieve data synchronization through spline interpolation.

[0066] Establish a unified time axis and estimate the cross-modal time lag matrix . Use the dynamic time warping algorithm (DTW) based on Euclidean distance to calculate the optimal matching path between different modal time series, and thus estimate the time lag:

[0067]

[0068] where, is the physical lag prior (such as vibration propagation delay), is the regularization coefficient. Based on the estimated time lag, resample and interpolate the data to obtain the synchronized sequence .

[0069] ​(2) Physical guiding features and cross-modal fusion: Extract features such as temperature gradient, Laplacian, dominant frequency, bandwidth energy, pulsatile coherence, and realize feature-level fusion through cross-modal attention. Perform feature extraction processing on the target multi-modal data to obtain the physical prior features corresponding to each modal data; and through the cross-modal attention model, perform feature fusion processing on each physical prior feature to obtain the target fusion feature.

[0070] In an implementation, the role of this step is to extract features with clear physical meaning and enhance model interpretability. Specifically, physical prior features are extracted for each modal:

[0071] Temperature modal: Extract temperature gradient (Difference between adjacent measurement points according to adjacency graph), Laplacian operator (Difference between average temperature difference of each measurement point and its neighborhood), convection term estimation, etc.

[0072] Vibration, swing modal: Extract dominant frequency (PSD peak corresponding frequency), wideband energy Envelope spectrum statistics: mean, peak, kurtosis of Hilbert envelope (used for bearing defect features).

[0073] Pressure pulsation modal: Extract vortex band energy, pulsatile coherence coefficient, etc.

[0074] Working condition modal: Extract state embedding .

[0075] Then map each modal feature to high-dimensional features through an encoder :

[0076]

[0077] Wherein, is the physical prior feature.

[0078] Further, a cross-modal attention mechanism is used for feature fusion: a convolutional neural network (CNN) is used to extract local patterns, a long short-term memory network (LSTM) is used to capture temporal dependencies, and an attention mechanism is used to dynamically weight the contributions of each modal, for example, vibration features are more important in the startup phase, and temperature features are more important in the steady state phase.

[0079] Define query, key, and value vectors:

[0080]

[0081]

[0082] Use attention weights The correlation score for each modality is scored and normalized, which can be interpreted as the importance of each modality at this moment:

[0083]

[0084] Fusion characteristics:

[0085]

[0086] The output of the aforementioned feature-level fusion contains cross-modal information and provides interpretable weights for which features are more important.

[0087] (3) Multimodal spatiotemporal memory network: Construct a four-branch memory system with memory-level fusion for temperature (slow timescale), vibration (fast timescale), pressure (fast timescale), and operating condition (state switching), and output a unified state representation. Using the temperature memory branch, vibration memory branch, pressure memory branch, and operating condition memory branch, historical state memory processing is performed on the temperature feature component, vibration feature component, pressure feature component, and operating condition feature component in the target fusion feature, respectively. With a gated memory update mechanism, the historical evolution state of each mode within the corresponding preset time window is exponentially decayed and accumulated to obtain the memory state set.

[0088] In one implementation, this step aims to capture the dynamic characteristics of the multimodal data of the turbine at different time scales, achieving accurate state prediction and uncertainty quantification. A four-branch parallel memory structure processes thermodynamic, mechanical vibration, hydraulic pulsation, and operational state information separately, ultimately fusing them into a unified state representation to provide a reliable basis for early warning. For details, see [link to details]. Figure 3 The diagram shows the structure of a multimodal spatiotemporal memory network model. It utilizes a four-branch memory structure to represent the temporal accumulation of historical information for each modality, and then uses an attention mechanism to fuse these memory states into a unified state vector for subsequent prediction and early warning.

[0089] 1. Temperature memory branch (slow time scale, minutes to hours):

[0090] Input fused feature vector The temperature-related characteristic components in the data are processed through a gate to handle temperature signals related to thermodynamic processes, whose dynamic changes are slow (on the order of minutes to hours). An exponentially decaying memory kernel is used to capture the slow temperature rise process of components such as bearings and stators. The memory update formula is:

[0091]

[0092] in, Forget Gate (usually set) ), controlling the rate of decay of historical memory; To extract temperature features for temperature feature extraction functions (e.g., fully connected networks), separate and strengthen temperature trend information from the fused features. This branch innovatively introduces thermal conduction physical constraints to ensure that the memory state complies with the law of conservation of energy.

[0093] Final output slow-time memory state , representing the long-term thermal accumulation effect of the device.

[0094] 2. Vibration memory branch (fast time scale, seconds to minutes):

[0095] Input fused feature vector Vibration, swing-related feature components in the input fused feature vector are processed. The mechanical vibration and swing signals have fast dynamic changes (seconds to minutes). A high-frequency update mechanism (forget gate ) is used in combination with a frequency band attention gate to highlight fault-sensitive frequency bands (such as bearing defect frequencies or rotor imbalance frequencies in the 80-120Hz range). The memory update formula is:

[0096]

[0097] where, is the vibration feature extraction function, which integrates the frequency band attention mechanism to automatically weight the energy of important frequency bands. The innovative point of this branch is to embed the mechanical vibration spectrum characteristics into the memory update, enhancing the ability to capture sudden impact events.

[0098] Final output fast-time memory state , representing the instantaneous dynamic response of the mechanical structure.

[0099] 3. Pressure memory branch (fast time scale, seconds to minutes):

[0100] Input fused feature vector Pressure pulsation-related feature components in the input fused feature vector are processed. The pressure pulsation signals of the hydraulic system have similar dynamic characteristics to vibration but different physical sources. This branch focuses on periodic or intermittent pressure fluctuations caused by vortex shedding, cavitation, and other hydraulic phenomena (such as vortex shedding frequencies of 0.5-2Hz and high-frequency water pressure pulsations). The memory update formula is:

[0101]

[0102] where, is the forget gate; is the pressure feature extraction function, which emphasizes the coherence of vortex energy and the statistical amplitude of pulsations. The innovation lies in independently processing hydraulic excitation signals to avoid confusion with mechanical vibrations and enhancing the rationality of memory through physical guidance.

[0103] Final output fast-time memory state , representing the instability of the hydraulic system and the propagation of pressure fluctuations.

[0104] 4. Working condition memory branch (state switching):

[0105] Input fusion feature vector with working condition parameter feature components (such as speed, power, and water head). Based on the hidden semi-Markov model (HSMM), the running state switching (start-up-transition-steady state-shutdown) is encoded. This branch does not directly process sensor data, but learns the state transition probability:

[0106]

[0107] and outputs state-dependent memory:

[0108]

[0109] wherein, is the current state identifier; is the state encoder. The innovation lies in the continuous discretization of the discrete running state, providing context guidance for other branches (such as increasing vibration weight in the start-up phase).

[0110] Final output state memory representing the macroscopic operation mode of the unit.

[0111] After obtaining the above four memory states, cross-modal memory fusion is performed:

[0112] Input four-branch memory state , , , . A covariance adaptive fusion mechanism is used to dynamically weight the contributions of each branch. The fusion function is:

[0113]

[0114] wherein, can be instantiated as attention weighting, for example:

[0115]

[0116] weight is automatically calculated from the current state and branch uncertainty (such as reducing vibration branch weight at high temperature). The innovation lies in the realization of physically guided dynamic fusion, ensuring that the memory state retains both modality specificity and global consistency.

[0117] Final output unified memory state integrating multi-modal spatio-temporal information.

[0118] Further, according to the above unified memory state, prediction is performed:

[0119] Input unified memory state . The future H-step state prediction and uncertainty are mapped by the prediction head (fully connected network layer):

[0120]

[0121] where, is the future state prediction value (e.g. temperature, vibration amplitude); is the prediction covariance matrix, quantifying the uncertainty. The innovation lies in the joint output of point prediction and probability distribution, providing a basis for risk-aware early warning.

[0122] Final output prediction value and uncertainty for anomaly scoring.

[0123] (4) Meta-learning fast adaptation: Small sample adaptation is performed on units and working conditions as tasks. When the unit of the water turbine is replaced, small sample data of the start-stop process of the replaced water turbine is collected, and based on the model-independent meta-learning framework and small sample data, the initial model trained by the original water turbine is fine-tuned to obtain the target model corresponding to the replaced water turbine.

[0124] In an embodiment, the step functions to realize small sample fast adaptation of new units and working conditions. By regarding each unit and each working condition as a task , a model-independent meta-learning (MAML) framework is adopted:

[0125]

[0126] Task set: represents the meta-task set of units and working conditions; each contains a number of continuous time slice samples.

[0127] Parameter: is the shared base model parameter; is the inner loop update step (which can be a scalar or a hierarchical step).

[0128] Inner loop loss: is used to make 1-K step gradient updates on task to obtain , and the inner loop loss is the mean square error between the predicted value and the true value.

[0129] Outer loop target: ; that is, the validation loss is calculated on the updated parameter and is reversed to the initial parameter . The loss function includes the prediction error, physical constraints, and consistency terms:

[0130]

[0131] wherein the prediction error term (measuring the deviation of the output from the measurement) is the predicted head (with as input) from step S3, the measured vector for the corresponding time slice. the physical regularization term, writing the physical laws such as temperature-heat transfer, vibration-spectrum bias as soft constraints to suppress unreasonable, for example:

[0132]

[0133] wherein for the temperature branch: with the meaning that the temporal variation of the predicted temperature should be explained by the heat conduction and heat generation terms; wherein is the effective thermal diffusivity, is the effective internal heat source (obtained from the regression of the load). For the vibration branch: , guiding the spectrum energy to concentrate on the engineering related frequency band, and suppressing unreasonable high frequency noise is the weight.

[0134] the cross-modal consistency term, which constrains different modalities to have consistent abnormal patterns after aligning the estimated physical time lag, avoiding single-modal false positives. Implementation:

[0135]

[0136] wherein, denotes the calculation of the residual after aligning the modalities with the estimated time lag . DTW+physical prior time lag estimation from step S1.

[0137] (5) Early warning judgment and grading: combine residual and uncertainty to get abnormal score and realize grading warning through sequential test. Abnormal score processing is carried out for the predicted value and uncertainty, and the abnormal score is obtained. Through sequential test, the abnormal score is accumulated to obtain the abnormal score cumulative value. When the abnormal score cumulative value reaches the trigger threshold of any level of the preset three-level grading trigger condition, the corresponding early warning processing is executed, and when the three-level warning is reached, the root cause of the three-level warning is located.

[0138] In an embodiment, the function of this step is to generate graded warning to support operation and maintenance decision. Specifically, it includes the following:

[0139] 1. Abnormal score (taking into account deviation and uncertainty):

[0140]

[0141] where, and are weight coefficients, the first term is the Mahalanobis residual (the greater the deviation, the more abnormal), and the second term is the uncertainty.

[0142] 2. Sequential inspection (CUSUM algorithm) forms an advance controllable alarm:

[0143]

[0144] where, is the expected drift under normal conditions, selected by minimizing the false alarm cost of the validation set; set the upper and lower hysteresis and the minimum duration (such as ≥ 30s) to suppress jitter.

[0145] 3. Three-level hierarchical trigger conditions (thresholds can be site-specific fine-tuning):

[0146] First (attention): short-term abnormality, ; second (warning): , and a consistent abnormality of multiple modes (at least two types of modes exceed the limit in the same window); third (serious): . Among them, the first level needs to be strengthened; the second level needs to automatically generate a patrol task and shorten the aggregation window; the third level needs to trigger a repair work order, reserve a downtime window, and enter root cause analysis.

[0147] (6) Causal root cause positioning: under the constraints of component-sensor topology and causal strength, retrieve historical cases to obtain root cause posterior and maintenance recommendations. Obtain the causal graph of the constructed components of the water turbine and the transfer entropy, and based on the transfer entropy and the causal graph, calculate the causal strength between the bearing, the frame and the shell, and determine the target causal strength; based on the target causal strength, perform similarity retrieval processing in the historical case library, match the similar historical fault set corresponding to the target causal strength, and according to the root cause posterior probability corresponding to each similar historical fault in the similar historical fault set, determine the target root cause with the highest probability and the maintenance recommendation list corresponding to the target root cause.

[0148] In one embodiment, the function of this step is to locate the fault source and provide maintenance recommendations. Specifically, it includes the following:

[0149] 1. Construct component graph G=(V,E), with edge weight as time delay correlation and causal strength:

[0150]

[0151] where, To transfer entropy. Causal strength is quantified by calculating transfer entropy. For example, the transfer entropy of the bearing vibration signal to the rack vibration signal is calculated, and if the value is significantly higher than the reverse transfer, it indicates that the bearing may be the source of the rack vibration.

[0152] 2. Retrieving the historical case library :

[0153]

[0154] 3. Root cause posterior probability:

[0155]

[0156] wherein, is the physical prior (higher for critical components), can be given by the set of affected sensors and causal paths (e.g. the more paths covered, the greater the strength, the higher the probability). The final output is the most likely root cause and a list of maintenance recommendations.

[0157] In summary, the present application can realize multi-modal collaborative early warning, fuse multi-source data such as temperature, vibration, pressure pulsation, etc., improve the accuracy of early warning, and adopt small sample rapid adaptation, realize rapid adaptation of new units through meta-learning, reduce the dependence on historical data, in addition, the present application also has early warning capability, and supports root cause diagnosis: based on sequential inspection and hierarchical rules, early warning can be realized 30-60 minutes in advance; combined with causal reasoning and case retrieval, the fault source can be located and maintenance suggestions can be given, finally, the engineering applicability of the present application is strong, when applied to actual hydropower stations, the early warning accuracy can be improved to more than 95%, and the false positive rate can be reduced to less than 5%.

[0158] For the water turbine full-state intelligent early warning method provided by the foregoing embodiment, an embodiment of the present application provides a water turbine full-state intelligent early warning device, which refers to a structure schematic diagram of a water turbine full-state intelligent early warning device as shown in Figure 4 The device comprises the following parts:

[0159] The feature fusion module 402 performs data cleaning processing and multi-scale time alignment processing on the multi-modal data collected by the sensor set of the water turbine, obtains target multi-modal data, and performs feature fusion processing on the target multi-modal data based on a cross-modal attention model to obtain target fusion features.

[0160] The state memory module 404 adopts a multi-modal spatio-temporal memory network model, performs historical state memory processing on the target fusion features through a four-branch parallel memory structure to obtain a memory state set, and performs cross-modal memory fusion processing on each memory state in the memory state set through a covariance adaptive fusion model to obtain a unified memory state.

[0161] The state early warning module 406 maps the unified memory state into a full connection network model to perform mapping processing on the future state, obtains a prediction value and an uncertainty of the future state of the hydraulic turbine, and performs hierarchical early warning based on the prediction value and the uncertainty.

[0162] The above-mentioned hydraulic turbine full-state intelligent early warning device provided by the embodiments of the present application can significantly improve the accuracy of intelligent early warning.

[0163] In one implementation, when the step of performing data cleaning processing and multi-scale time alignment processing on the multi-modal data collected by the sensor set of the hydraulic turbine to obtain target multi-modal data is performed, the feature fusion module 402 is further configured to: after the multi-modal data is subjected to temperature data cleaning processing and vibration data cleaning processing, the multi-modal data is sent to a dynamic time warping model, and the optimal matching path between different modal time sequences is calculated to determine a time lag estimation value; based on the time lag estimation value, spline interpolation processing is performed to align the multi-scale time of the multi-modal data, and target multi-modal data is obtained.

[0164] In one implementation, when the step of performing feature fusion processing on the target multi-modal data based on the cross-modal attention model to obtain target fusion features is performed, the feature fusion module 402 is further configured to: perform feature extraction processing on the target multi-modal data to obtain physical prior features corresponding to each modal data; and perform feature fusion processing on each physical prior feature through the cross-modal attention model to obtain target fusion features.

[0165] In one implementation, when the step of performing history state memory processing on the target fusion features by using a multi-modal spatio-temporal memory network model through a four-branch parallel memory structure to obtain a memory state set is performed, the state memory module 404 is further configured to: use a temperature memory branch, a vibration memory branch, a pressure memory branch, and a working condition memory branch to respectively perform history state memory processing on temperature feature components, vibration feature components, pressure feature components, and working condition feature components in the target fusion features, and use a gated memory update mechanism to exponentially attenuate and accumulate the history evolution state of each modal within a corresponding preset time window to obtain a memory state set.

[0166] In one implementation, after the step of obtaining a prediction value and an uncertainty of the future state of the hydraulic turbine is performed, the state memory module 404 is further configured to: when the unit of the hydraulic turbine is replaced, collect small sample data of the start-stop process of the replaced hydraulic turbine, and based on a model-independent meta-learning framework and the small sample data, perform model fine-tuning processing on an initial model trained by an original hydraulic turbine to obtain a target model corresponding to the replaced hydraulic turbine.

[0167] In one embodiment, when the step of performing hierarchical early warning based on the predicted value and the uncertainty is performed, the state early warning module 406 is further configured to: perform an anomaly scoring process on the predicted value and the uncertainty to obtain an anomaly score, and perform a sequential test on the anomaly score to obtain an anomaly score accumulation value; when the anomaly score accumulation value reaches a trigger threshold corresponding to any one of preset three-level classification trigger conditions, perform corresponding early warning processing, and locate the root cause of the three-level early warning when the three-level early warning is reached.

[0168] In one embodiment, when the step of locating the root cause of the three-level early warning is performed, the state early warning module 406 is further configured to: obtain a causal graph and a transfer entropy of the constructed components of the hydraulic turbine, and calculate the causal strength between the bearing, the frame and the casing based on the transfer entropy and the causal graph to determine a target causal strength; based on the target causal strength, perform a similarity search process in a historical case library to match a similar historical fault set corresponding to the target causal strength, and determine a target root cause with the highest probability and a maintenance suggestion list corresponding to the target root cause according to the root cause posterior probability of each similar historical fault in the similar historical fault set.

[0169] The device provided by the embodiments of the present application has the same implementation principle and technical effects as the foregoing method embodiments. For brevity, the part not mentioned in the device embodiment part can be referred to the corresponding content in the foregoing method embodiments.

[0170] The embodiments of the present application provide a server, specifically, the server includes a processor and a storage device; the storage device stores a computer program, and the computer program performs the method of any one of the above embodiments when executed by the processor.

[0171] Figure 5 A structural diagram of a server provided by the embodiments of the present application is provided, and the server 100 includes a processor 50, a memory 51, a bus 52 and a communication interface 53, the processor 50, the communication interface 53 and the memory 51 are connected through the bus 52; the processor 50 is configured to execute the executable modules stored in the memory 51, such as a computer program.

[0172] The memory 51 may contain a high-speed random access memory (RAM, Random Access Memory) and may also include a non-volatile memory, such as at least one disk memory. The communication connection between the system network element and at least one other network element is realized through at least one communication interface 53 (which can be wired or wireless), and the Internet, a wide area network, a local area network, a metropolitan area network, etc. can be used.

[0173] The bus 52 can be an ISA bus, a PCI bus, an EISA bus, or the like. The bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, Figure 5 Only one bidirectional arrow is used to represent the bus, but it does not mean that there is only one bus or only one type of bus.

[0174] The memory 51 is configured to store a program, and the processor 50 executes the program after receiving an execution instruction. The method executed by the device for defining the flow process according to any of the foregoing embodiments of the application can be applied to the processor 50 or implemented by the processor 50.

[0175] The processor 50 can be an integrated circuit chip with processing capability. In the implementation process, each step of the foregoing method can be completed by an integrated logic circuit of hardware in the processor 50 or an instruction in the form of software. The processor 50 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), and the like; or can be a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components. The disclosed methods, steps, and logic block diagrams in the embodiments of the application can be implemented or executed. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor. The steps of the method disclosed in combination with the embodiments of the application can be directly embodied as a hardware code processor for execution, or a combination of hardware and software modules in the code processor for execution. The software module can be located in a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an electrically erasable programmable memory, a register, or other mature storage media in the art. The storage medium is located in the memory 51, and the processor 50 reads the information in the memory 51 and combines the hardware to complete the steps of the foregoing method.

[0176] The computer program product of the readable storage medium provided by the embodiments of the application includes a computer readable storage medium storing program codes, and the program codes include instructions for executing the method described in the foregoing method embodiments. For specific implementation, reference can be made to the foregoing method embodiments, which will not be described here.

[0177] If the functions are realized in the form of software function units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the present application that essentially contribute to the prior art or the parts of the technical solutions can be embodied in the form of software products. The computer software product is stored in a storage medium and includes instructions for making a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0178] Finally, it should be noted that: the above-described embodiments are only specific embodiments of the present application, which are used to illustrate the technical solutions of the present application, but not to limit them. The protection scope of the present application is not limited to this. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can modify or easily think of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed by the present application, or make equivalent replacements to some of the technical features. The modifications, changes or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for intelligent early warning of all states of a water turbine, characterized in that, The method includes: Data cleaning and multi-scale time alignment processing are performed on the multimodal data collected by the sensor set of the water turbine to obtain target multimodal data. Based on the cross-modal attention model, feature fusion processing is performed on the target multimodal data to obtain target fused features. A multimodal spatiotemporal memory network model is adopted. Through a four-branch parallel memory structure, the target fusion features are processed to perform historical state memory processing to obtain a set of memory states. Then, through a covariance adaptive fusion model, the memory states in the set of memory states are processed to perform cross-modal memory fusion to obtain a unified memory state. The unified memory state is input into a fully connected network model to map the future state, thereby obtaining the predicted value and uncertainty of the future state of the turbine, and a graded early warning is executed based on the predicted value and the uncertainty. The step of using a multimodal spatiotemporal memory network model and a four-branch parallel memory structure to process the historical state memory of the target fusion features to obtain a set of memory states includes: using temperature memory branch, vibration memory branch, pressure memory branch, and working condition memory branch to process the historical state memory of the temperature feature component, vibration feature component, pressure feature component, and working condition feature component in the target fusion features, respectively; and using a gated memory update mechanism to exponentially decay and accumulate the historical evolution state of each modality within the corresponding preset time window to obtain the set of memory states. The temperature memory branch takes temperature-related feature components as input, processes temperature signals in thermodynamic processes through gates, and uses an exponentially decaying memory kernel to capture the slow temperature rise process of components, outputting a slow timescale memory state to characterize the long-term heat accumulation effect of the equipment. The vibration memory branch takes vibration and swing-related feature components as input, processes mechanical vibration and swing signals, uses a high-frequency update mechanism combined with a frequency band attention gate to extract fault-sensitive frequency bands, and outputs a fast timescale memory state to characterize the instantaneous dynamic response of the mechanical structure. The pressure memory branch takes pressure pulsation-related feature components as input, processes pressure pulsation signals of the hydraulic system, and outputs a fast timescale memory state to characterize the instability of the hydraulic system and the propagation of pressure fluctuations. The operating condition memory branch takes operating condition parameter feature components as input, encodes the switching of operating states based on a hidden semi-Markov model, and outputs the state memory to characterize the macroscopic operating mode of the unit.

2. The intelligent early warning method for all states of a water turbine according to claim 1, characterized in that, The steps for performing data cleaning and multi-scale time alignment processing on the multimodal data collected by the sensor set of the water turbine to obtain the target multimodal data include: After the multimodal data is cleaned by temperature data and vibration data, it is sent to the dynamic time warping model, and the time delay estimate is determined by calculating the optimal matching path between different modal time series. Spline interpolation is performed based on the time delay estimate to align the multi-scale time of the multimodal data and obtain the target multimodal data.

3. The intelligent early warning method for all states of a water turbine according to claim 1, characterized in that, The step of performing feature fusion processing on the target multimodal data based on a cross-modal attention model to obtain target fused features includes: Feature extraction is performed on the target multimodal data to obtain the physical prior features corresponding to each modality; and the physical prior features are fused using a cross-modal attention model to obtain the target fused features.

4. The intelligent early warning method for all states of a water turbine according to claim 1, characterized in that, After the step of obtaining the predicted value and uncertainty of the future state of the turbine, the following is included: When the turbine unit is replaced, small sample data of the start-up and shutdown process of the replaced turbine are collected. Based on the model-independent meta-learning framework and the small sample data, the initial model obtained by training the original turbine is fine-tuned to obtain the target model corresponding to the replaced turbine.

5. The intelligent early warning method for all states of a water turbine according to claim 1, characterized in that, The steps for implementing graded early warning based on the predicted value and the uncertainty include: Anomaly scoring is performed on the predicted value and the uncertainty to obtain anomaly scores, and the anomaly scores are accumulated through sequential testing to obtain an accumulated anomaly score value. When the accumulated abnormal score reaches the trigger threshold corresponding to any level of warning in the preset three-level triggering conditions, the corresponding warning processing is executed, and when a level three warning is reached, the root cause of the level three warning is located.

6. The intelligent early warning method for all states of a water turbine according to claim 5, characterized in that, The steps for locating the root cause of the Level 3 warning include: Obtain the causal graph and transfer entropy of the components of the water turbine, and calculate the causal strength between the bearing, frame and shell based on the transfer entropy and the causal graph to determine the target causal strength; Based on the target causal strength, a similarity retrieval process is performed in the historical case database to match the set of similar historical faults corresponding to the target causal strength. Then, based on the posterior probability of the root cause corresponding to each similar historical fault in the set of similar historical faults, the target root cause with the highest probability is determined, as well as the maintenance suggestion list corresponding to the target root cause.

7. A fully intelligent early warning device for all states of a water turbine, characterized in that, The device includes: The feature fusion module performs data cleaning and multi-scale time alignment processing on the multimodal data collected by the sensor set of the water turbine to obtain target multimodal data, and performs feature fusion processing on the target multimodal data based on the cross-modal attention model to obtain target fused features; The state memory module adopts a multimodal spatiotemporal memory network model. Through a four-branch parallel memory structure, it performs historical state memory processing on the target fusion features to obtain a set of memory states. Then, through a covariance adaptive fusion model, it performs cross-modal memory fusion processing on each memory state in the set of memory states to obtain a unified memory state. The state warning module inputs the unified memory state into a fully connected network model to map the future state, obtains the predicted value and uncertainty of the future state of the turbine, and performs hierarchical warning based on the predicted value and the uncertainty. The step of using a multimodal spatiotemporal memory network model and a four-branch parallel memory structure to process the historical state memory of the target fusion features to obtain a set of memory states includes: using temperature memory branch, vibration memory branch, pressure memory branch, and working condition memory branch to process the historical state memory of the temperature feature component, vibration feature component, pressure feature component, and working condition feature component in the target fusion features, respectively; and using a gated memory update mechanism to exponentially decay and accumulate the historical evolution state of each modality within the corresponding preset time window to obtain the set of memory states. The temperature memory branch takes temperature-related feature components as input, processes temperature signals in thermodynamic processes through gates, and uses an exponentially decaying memory kernel to capture the slow temperature rise process of components, outputting a slow timescale memory state to characterize the long-term heat accumulation effect of the equipment. The vibration memory branch takes vibration and swing-related feature components as input, processes mechanical vibration and swing signals, uses a high-frequency update mechanism combined with a frequency band attention gate to extract fault-sensitive frequency bands, and outputs a fast timescale memory state to characterize the instantaneous dynamic response of the mechanical structure. The pressure memory branch takes pressure pulsation-related feature components as input, processes pressure pulsation signals of the hydraulic system, and outputs a fast timescale memory state to characterize the instability of the hydraulic system and the propagation of pressure fluctuations. The operating condition memory branch takes operating condition parameter feature components as input, encodes the switching of operating states based on a hidden semi-Markov model, and outputs the state memory to characterize the macroscopic operating mode of the unit.

8. A server, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method according to any one of claims 1 to 6.

Images