Data-driven intelligent diagnosis method for equipment faults of hydropower stations

Abstract

This invention discloses a data-driven intelligent fault diagnosis method for hydropower station equipment, belonging to the field of hydropower station equipment fault diagnosis technology. It includes: acquiring multi-source sensor monitoring data of hydropower station equipment to construct an initial monitoring data matrix, and performing time-frequency domain transformation to obtain a time-frequency joint feature map; using a first graph neural network encoder to extract global operating condition correlation features from the time-frequency joint feature map, and fusing it with the initial monitoring data matrix to generate a dynamically enhanced operating condition representation; performing a sliding window Fourier residual mapping operation on the dynamically enhanced operating condition representation to obtain a fault-sensitive residual sequence; inputting the fault-sensitive residual sequence into a sparse deep belief network to reconstruct the equipment baseline health status; calculating the Riemann manifold distance between the dynamically enhanced operating condition representation and the equipment baseline health status as the equipment health deviation; when the deviation exceeds a threshold, triggering a causal graph attention network to perform root cause localization and outputting a set of fault attribution parameters.

Description

Technical Field

[0001] This invention relates to the field of hydropower station equipment fault diagnosis technology, specifically a data-driven intelligent fault diagnosis method for hydropower station equipment. Background Technology

[0002] Hydropower station equipment operates in a complex environment, and its health status monitoring relies on the comprehensive analysis of multi-source sensor data. Traditional equipment fault diagnosis methods are usually based on single sensor thresholds or simple signal processing techniques, which are difficult to capture the spatiotemporal correlation and dynamic evolution of multi-sensor data under complex operating conditions. Constructing a monitoring data matrix and extracting features for fault judgment is a common approach. However, when faced with subtle changes in equipment status, fixed feature extraction methods are easily affected by background noise and fluctuations in operating conditions, failing to effectively amplify early fault features, resulting in insufficient sensitivity and timeliness of diagnosis. Existing technologies often use deterministic models or shallow learning methods when constructing equipment health baselines. These methods have limited ability to fit the dynamic nonlinearity and uncertainty of the system, and the reconstructed baseline often deviates from the true health status of the equipment. In addition, when quantifying the degree of deviation, existing methods often use flat space metrics such as Euclidean distance, ignoring the manifold geometry inherent in the equipment status data itself, failing to accurately reflect the essential differences in complex health status changes, and thus causing distortion in the health status assessment results.

[0003] In terms of root cause analysis, existing technologies mostly rely on simple correlation analysis or expert rules, which makes it difficult to clarify the propagation path of faults along the causal chain in complex coupled systems. This can easily lead to misjudging concurrent or derivative phenomena as the root cause of the fault, resulting in low accuracy and reliability of diagnostic conclusions.

[0004] To address the aforementioned shortcomings of existing technologies, it is necessary to solve the problems of effectively shielding complex operating conditions from interference and amplifying early fault characteristics when conducting intelligent diagnostics on hydropower station equipment, as well as how to accurately characterize the nature of state deviations in a nonlinear high-dimensional data space. At the same time, it is also necessary to solve the problem of accurately tracing the root cause of faults in complex causal relationship networks. Summary of the Invention

[0005] This invention provides a data-driven intelligent fault diagnosis method for hydropower station equipment, aiming to solve the problems of existing technologies where early fault characteristics are easily obscured by operating conditions, health baseline reconstruction has large deviations, state deviation measurements are inaccurate, and root cause tracing is inaccurate. By decoupling the influence of operating conditions from multi-source sensor data and enhancing sensitive features, and by accurately constructing the equipment's dynamic health baseline and manifold deviation measurement, the method achieves early and accurate perception of minor faults in hydropower station equipment and precise root cause localization.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a data-driven intelligent fault diagnosis method for hydropower station equipment, comprising: acquiring multi-source sensor monitoring data of hydropower station equipment at multiple historical moments, and constructing an initial monitoring data matrix; performing time-frequency domain transformation on the initial monitoring data matrix to obtain a time-frequency joint feature map; inputting the time-frequency joint feature map into a first graph neural network encoder to extract global operating condition correlation features; performing element-wise weighted fusion of the global operating condition correlation features and the initial monitoring data matrix to generate a dynamically enhanced operating condition representation; performing a sliding window Fourier residual mapping operation on the dynamically enhanced operating condition representation to obtain a fault-sensitive residual sequence; inputting the fault-sensitive residual sequence into a pre-constructed sparse deep belief network to reconstruct the equipment baseline health state layer by layer; calculating the Riemann manifold distance between the dynamically enhanced operating condition representation and the equipment baseline health state as the equipment health deviation; when the equipment health deviation exceeds a preset threshold, triggering a root cause localization mechanism based on a causal graph attention network, and outputting a set of fault attribution parameters. By performing time-frequency domain transformation on the initial monitoring data matrix to obtain a joint time-frequency feature map, the instantaneous characteristics of non-stationary signals can be fully preserved, providing rich input for subsequent correlation analysis. Element-by-element weighted fusion of global operating condition correlation features with the original monitoring data can dynamically enhance fault-sensitive components and suppress background interference based on the spatial correlation between sensors. Extracting frequency domain residual sequences using a sliding window Fourier residual mapping can highlight the spectral evolution characteristics caused by faults and reduce the interference of steady-state fluctuations on fault identification. Reconstructing the health baseline through a sparse deep belief network can adaptively learn a compact representation of the normal state of the equipment from the residual sequence, providing a highly robust reference for deviation calculation. Using Riemannian manifold distance to measure the degree of health deviation has geometric invariance and can reliably characterize changes in the true state of the equipment. When deviations exceed limits, a causal graph attention network is activated for root cause localization, which can accurately analyze the sensors or parameters causing the anomalies by combining causal relationships, improving the interpretability of the diagnosis.

[0007] As a preferred embodiment of the present invention, the specific implementation of performing time-frequency domain transformation on the initial monitoring data matrix to obtain the time-frequency joint feature map is as follows: Perform a Hilbert-Huang transform on each sensing data sequence to extract the instantaneous frequency component and instantaneous amplitude component of the sensing data sequence; multiply the instantaneous frequency component and the instantaneous amplitude component point-by-point to construct the marginal spectral energy distribution vector of the sensing data sequence; align and stack the marginal spectral energy distribution vectors of all sensing data sequences according to the monitoring time to form the time-frequency joint feature map. Preferably, perform empirical mode decomposition on each sensing data sequence to obtain a set of intrinsic mode function components of the sensing data sequence; perform a Hilbert transform on each intrinsic mode function component to analyze the instantaneous frequency function and instantaneous amplitude function of that component; superimpose the instantaneous frequency function and instantaneous amplitude function of all intrinsic mode function components in descending order of frequency to form the instantaneous frequency component and instantaneous amplitude component of the sensing data sequence, respectively. By using adaptive decomposition of Hilbert-Huang transform and Hilbert spectrum analysis, the instantaneous frequency and energy changes of non-stationary signals such as vibration of hydropower station equipment can be accurately captured, avoiding mode aliasing and making the time-frequency joint feature map more accurately reflect the dynamic characteristics of the equipment.

[0008] As a preferred embodiment of the present invention, the specific implementation of inputting the time-frequency joint feature map into the first graph neural network encoder to extract global operating condition correlation features is as follows: Using the marginal spectral energy distribution vector corresponding to each sensor data sequence in the time-frequency joint feature map as the node feature, and using the spatial distance between each sensor in the predefined physical topology of the hydropower station equipment as the edge weight, a graph structure data is constructed. In the first graph neural network encoder, a multi-head attention mechanism is used to calculate the contribution coefficient of each node's neighboring nodes to that node in the graph structure data. Based on the contribution coefficient, multiple nonlinear graph convolution operations are performed on the graph structure data to obtain an updated node feature matrix, which serves as the global operating condition correlation feature. The first graph neural network encoder is preferably a graph isomorphic network or a graph convolutional network. With the help of prior knowledge of physical topology and the multi-head attention mechanism, the graph neural network can fully integrate the potential spatial dependencies between multi-source sensor data, extract highly representative global operating condition correlation features, and enhance adaptability to complex operating conditions.

[0009] As a preferred embodiment of the present invention, the specific implementation of generating a dynamically enhanced operating condition representation by performing element-wise weighted fusion of the global operating condition correlation features and the initial monitoring data matrix is ​​as follows: the global operating condition correlation features and the initial monitoring data matrix are input into a gated fusion network, which contains two parallel convolutional layers; the first convolutional layer extracts weight coefficient tensors from the global operating condition correlation features; the second convolutional layer extracts value content tensors from the initial monitoring data matrix; and the element-wise multiplication operation is performed between the weight coefficient tensors and the value content tensors to obtain the dynamically enhanced operating condition representation. This gated fusion mechanism can adaptively adjust the contribution weights of the original monitoring data according to the strength of the global correlation features, significantly enhancing the subtle feature changes caused by faults and effectively suppressing irrelevant background information, thereby improving the sensitivity of fault detection.

[0010] As a preferred embodiment of the present invention, the specific implementation of performing a sliding window Fourier residual mapping operation on the dynamically enhanced operating condition characterization to obtain a fault-sensitive residual sequence is as follows: A sliding window of fixed length is set on the time axis; a Fourier transform is performed on the dynamically enhanced operating condition characterization within the sliding window to obtain the frequency domain amplitude vector within the window; the difference between the frequency domain amplitude vector within the window and the frequency domain amplitude vector within the sliding window at the previous moment is calculated to obtain the frequency domain residual component of that window; the sliding window is slid along the time axis to collect the frequency domain residual components of all windows, and the frequency domain residual components are arranged in chronological order to form the fault-sensitive residual sequence. The length of the sliding window is preferably adaptively determined based on the rotational frequency period of the rotating components of the hydropower station equipment. Through sliding window Fourier residual mapping, the temporal residual changes in the frequency domain can be extracted from the dynamically enhanced operating condition characterization, filtering out long-term steady-state background and highlighting the spectral evolution characteristics caused by faults; the window length is matched with the rotational frequency period to ensure that the residual sequence has higher sensitivity to typical faults of rotating machinery.

[0011] As a preferred embodiment of the present invention, the specific implementation of inputting the fault-sensitive residual sequence into a pre-constructed sparse deep belief network to reconstruct the equipment baseline health status layer by layer is as follows: the sparse deep belief network is composed of multiple stacked restricted Boltzmann machines, each of which is subject to sparse regularization constraints; using the fault-sensitive residual sequence as input data, each layer of the restricted Boltzmann machine is trained sequentially, and the hidden layer output of the previous layer is used as the input of the next layer; after completing the layer-by-layer pre-training of all layers of restricted Boltzmann machines, the entire sparse deep belief network is fine-tuned in reverse so that the data reconstructed by the output layer is aligned with the fault-sensitive residual sequence under the mean square error minimization criterion, and the reconstructed data is used as the equipment baseline health status. The sparse regularization constraints enable the network to learn more essential compact features in the residual sequence, remove redundant interference, and the reconstructed health baseline better represents the essential state of the equipment when it is fault-free. Preferably, the sparse regularization constraint of each layer of restricted Boltzmann machine is implemented using a KL divergence penalty term to control the average activation probability of hidden layer neurons and improve the robustness of the baseline model to noise.

[0012] Furthermore, the preferred method for training each layer of the Restricted Boltzmann Machine (RBM) sequentially using the fault-sensitive residual sequence as input data is as follows: A data batch is randomly sampled from the fault-sensitive residual sequence and input into the first layer of the RBM to be trained; an unsupervised learning algorithm based on contrastive divergence is executed in the first layer of the RBM to update the connection weights and biases between the visible and hidden layers; the post-training parameters of the first layer of the RBM are frozen, and the hidden layer activation probability of the RBM for the data batch is calculated, using the hidden layer activation probability as the input data for the next layer of the RBM; for each subsequent layer of the RBM, the unsupervised learning algorithm based on contrastive divergence and the hidden layer activation probability calculation are repeated until all layers of the RBM are trained. This layer-by-layer greedy pre-training combined with overall fine-tuning effectively captures the deep hierarchical features in the fault-sensitive residual sequence, making the reconstructed health baseline more accurately match the characteristics of the normal equipment state.

[0013] As a preferred embodiment of the present invention, the calculation of the Riemann manifold distance between the dynamically enhanced operating condition characterization and the equipment baseline health condition, as the specific implementation of the equipment health deviation, involves: mapping the dynamically enhanced operating condition characterization and the equipment baseline health condition onto a symmetric positive definite manifold space, respectively, to obtain a first manifold point and a second manifold point; calculating the affine-invariant Riemann distance between the first manifold point and the second manifold point in the symmetric positive definite manifold space; and determining the calculated value of the affine-invariant Riemann distance as the equipment health deviation. Using the affine-invariant Riemann distance on a symmetric positive definite manifold as a deviation measure has invariance to linear transformations and scale changes, reliably reflecting the deviation of the equipment's true state from the health baseline, and effectively reducing false alarms caused by operating condition fluctuations.

[0014] As a preferred embodiment of the present invention, when the equipment health deviation exceeds a preset threshold, the root cause localization mechanism based on a causal graph attention network is triggered, and the fault attribution parameter set is output. Specifically, after determining that the equipment health deviation exceeds the preset threshold, the residual matrix between the dynamically enhanced operating condition characterization and the equipment baseline health state is obtained as the input for root cause localization. The residual matrix is ​​input to the causal graph attention network, which internally maintains an adjacency matrix describing the causal relationships between various sensor variables. In the causal graph attention network, an attention weight is calculated for each feature dimension in the residual matrix in conjunction with the adjacency matrix. The attention weight is used to quantify the degree to which the feature dimension deviates from the baseline state. The attention weights of all feature dimensions are sorted, and several feature dimensions with the highest weight values ​​are selected. The sensor names or equipment parameter names corresponding to these feature dimensions are output as the fault attribution parameter set. By introducing a causal adjacency matrix, attention networks can locate the root cause of a fault based on the causal dependencies between variables, avoiding incorrect attribution due to spurious correlations, and outputting a set of fault parameters with high interpretability, which helps to quickly guide on-site operation and maintenance decisions.

[0015] In the above technical solution, the technical effects and advantages provided by this invention are as follows: A graph neural network is used to extract global operating condition correlation features from the joint time-frequency feature map, and these features are then fused element-wise with the initial monitoring data matrix to generate a dynamically enhanced operating condition representation. This feature extraction method utilizes prior knowledge of the equipment's physical topology, and through graph structure data modeling and attention mechanisms, automatically decouples the global correlation pattern of the equipment's operating conditions from multi-sensor data. The extracted global operating condition correlation features, as an abstract expression of the operating conditions, can adaptively adjust the signal-to-noise ratio of each sensor channel in subsequent analysis when performing gated fusion with the original data. This mechanism suppresses operating condition fluctuation components unrelated to the equipment's health status, while relatively amplifying minor abnormal changes caused by faults in the feature space. This solves the problem that fault features are masked by changes in operating conditions in traditional feature extraction, enabling subsequent fault analysis to be performed on a purer and more sensitive feature basis. A fault-sensitive residual sequence is generated from the dynamically enhanced operating condition representation through a sliding window Fourier residual mapping, and this residual sequence is then input into a sparsely regularized constrained deep belief network to reconstruct the equipment's baseline health status layer by layer. The sliding window Fourier residual mapping operation continuously captures the transient changes in the frequency domain structure of the operating condition representation. The resulting residual sequence focuses on components that deviate from the normal operating mode, providing a high signal-to-noise ratio input for baseline reconstruction. The sparse deep belief network, through multi-layer nonlinear mapping and sparsity constraints, can learn the inherent high-dimensional sparse feature representation of the equipment under healthy conditions from the residual sequence, accurately reconstructing the dynamic baseline health state. The affine-invariant Riemann distance between the dynamically enhanced operating condition representation and the baseline on a symmetric positive definite manifold is calculated as the health deviation. Utilizing the geometric properties of the data covariance structure, this metric is invariant to scaling and shearing transformations of the state space, and can capture subtle changes in equipment health state more fundamentally than Euclidean distance, thus achieving highly sensitive and accurate early warning of early faults. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of a data-driven intelligent fault diagnosis method for hydropower station equipment. Figure 2 This is a flowchart of the time-frequency joint feature map construction method; Figure 3 This is a schematic diagram of the structure for global working condition correlation feature extraction and dynamic enhanced working condition representation generation; Figure 4This is a flowchart of the process for generating fault-sensitive residual sequences using sliding window Fourier residual mapping; Figure 5 This is a flowchart of the device baseline health reconstruction process based on sparse deep belief networks; Figure 6 This is a flowchart of a device health deviation assessment and root cause localization method based on Riemannian manifold distance and causal graph attention network; Figure 7 It is a time series comparison chart of the initial monitoring data matrix and the dynamic enhanced operating condition characterization; Figure 8 It is a graph showing the variation of the frequency domain residual components within the sliding window. Figure 9 It is a curve comparing the fault-sensitive residual sequence with the equipment baseline health status; Figure 10 It is a curve showing the change in equipment health deviation over monitoring time. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] See Figure 1 This invention provides a data-driven intelligent fault diagnosis method for hydropower station equipment, comprising the following steps: acquiring multi-source sensor monitoring data of hydropower station equipment at multiple historical moments to construct an initial monitoring data matrix; performing time-frequency domain transformation on the initial monitoring data matrix to obtain a time-frequency joint feature map; inputting the time-frequency joint feature map into a first graph neural network encoder to extract global operating condition correlation features; performing element-wise weighted fusion of the global operating condition correlation features and the initial monitoring data matrix to generate a dynamically enhanced operating condition representation; performing a sliding window Fourier residual mapping operation on the dynamically enhanced operating condition representation to obtain a fault-sensitive residual sequence; inputting the fault-sensitive residual sequence into a pre-constructed sparse deep belief network to reconstruct the equipment baseline health state layer by layer; calculating the Riemann manifold distance between the dynamically enhanced operating condition representation and the equipment baseline health state as the equipment health deviation; when the equipment health deviation exceeds a preset threshold, triggering a root cause localization mechanism based on a causal graph attention network to output a set of fault attribution parameters.

[0020] Example 1: See Figure 2The implementation of time-frequency domain transformation to obtain the joint time-frequency feature map of the initial monitoring data matrix is ​​as follows: A Hilbert-Huang transform is performed on each sensor data sequence in the initial monitoring data matrix to extract the instantaneous frequency and instantaneous amplitude components of each sensor data sequence. In specific implementation, empirical mode decomposition (EMD) is performed on each sensor data sequence to obtain a set of intrinsic mode function (EMF) components for each sensor data sequence. The EMD process is executed iteratively using a sieving algorithm: all local maxima and local minima in the sensor data sequence are identified; cubic spline interpolation is used to fit the local maxima to form the upper envelope and the local minima to form the lower envelope, respectively; the mean curves of the upper and lower envelopes are calculated; the mean curves are subtracted from the sensor data sequence to obtain the intermediate sequence; and it is determined whether the intermediate sequence satisfies the EMF conditions. The intrinsic mode function (IMF) conditions are set as follows: over the entire time range of the sensing data sequence, the number of extreme points and the number of zero-crossing points in the intermediate sequence are equal or differ by no more than one, and the mean of the upper and lower envelopes is zero at any given time. If the intermediate sequence satisfies the IMF conditions, it is used as a first-order IMF component; otherwise, the intermediate sequence is used to replace the sensing data sequence, and the sieving algorithm is repeated until the conditions are met. After separating the first-order IMF components from the sensing data sequence, a first-order residual sequence is obtained. The sieving algorithm is then repeated using the first-order residual sequence as a new sensing data sequence to extract the second-order IMF components, third-order IMF components, and so on, until the residual sequence becomes a monotonic function or a constant. The decomposition then stops, resulting in a set of IMF components for the sensing data sequence. This set of IMF components contains several IMF components arranged from high frequency to low frequency. Perform a Hilbert transform on each eigenmode function component to analytically obtain the instantaneous frequency function and instantaneous amplitude function of each eigenmode function component. For the ... Each intrinsic mode function component The Hilbert transform is defined as follows: .

[0021] in, Represents a time variable. Represents the integral dummy variable. Indicates the first The intrinsic mode function components at the dummy variable time The value, This represents the improper integral over an infinite interval. This is the normalization factor for the Hilbert transform. (From...) Results of Hilbert Transform Constructing analytic signals ,in The imaginary unit is used to analyze signals. amplitude function That is, the first The instantaneous amplitude function of each eigenmode function component, analytic signal phase function Regarding time The derivative obtained That is, the first The instantaneous frequency functions of each intrinsic mode function (IMF) component are obtained by superimposing these instantaneous frequency functions in descending order of frequency to form the instantaneous frequency component of each sensing data sequence. Similarly, the instantaneous amplitude functions of each IMF component are superimposed in descending order of frequency to form the instantaneous amplitude component of each sensing data sequence. The superposition is performed at each time step. At each intrinsic mode function (IMF) component, the instantaneous frequency function values ​​are algebraically summed to obtain the instantaneous frequency component value at that moment, and the instantaneous amplitude function values ​​are algebraically summed to obtain the instantaneous amplitude component value at that moment. The instantaneous frequency and amplitude components are then multiplied point-by-point to construct the marginal spectral energy distribution vector for each sensor data sequence. Point-by-point multiplication means multiplying the instantaneous frequency component value and the instantaneous amplitude component value at each discrete-time sampling point; the product constitutes the element value of the marginal spectral energy distribution vector at that sampling point. The dimension of the marginal spectral energy distribution vector is the same as the number of time sampling points in the sensor data sequence. The marginal spectral energy distribution vectors of all sensor data sequences are aligned according to the monitoring time and stacked vertically to form a time-frequency joint feature map. The marginal spectral energy distribution vectors of all sensor data sequences have the same number of time sampling points, and each sampling point corresponds to the same monitoring time. The marginal spectral energy distribution vector of each sensor data sequence is taken as a row vector and vertically spliced ​​according to the sensor arrangement order to obtain the time-frequency joint feature map. Each row of the time-frequency joint feature map corresponds to the marginal spectral energy distribution of a sensor data sequence, and each column corresponds to the cross-sensor energy distribution of a monitoring time.

[0022] Example 2: See Figure 3 The implementation method for extracting global operating condition correlation features by inputting the time-frequency joint feature map into the first graph neural network encoder is as follows: the marginal spectral energy distribution vector corresponding to each sensor data sequence in the time-frequency joint feature map is used as the node feature, and the spatial distance between each sensor in the predefined physical topology of the hydropower station equipment is used as the edge weight to construct graph structure data.

[0023] In practical implementation, the physical topology of the hydropower station equipment is predefined based on the actual spatial layout of the equipment and the sensor installation locations. The physical topology contains several nodes, each corresponding to a sensor installation location. The edge connections within the physical topology reflect the physical connectivity or spatial adjacency between sensors. The marginal spectral energy distribution vector corresponding to each sensor data sequence in the time-frequency joint feature map is assigned to the corresponding node in the physical topology as a node feature. For any two nodes connected by an edge in the physical topology, the Euclidean distance between the sensors corresponding to the two nodes in physical space is obtained, and the reciprocal of the Euclidean distance is used as the edge weight between the two nodes. The edge weights are stored in the form of an adjacency matrix, with zero values ​​filled at the positions of unconnected nodes. The graph structure data is composed of the node feature matrix and the adjacency matrix.

[0024] In the first graph neural network encoder, a multi-head attention mechanism is used to calculate the contribution coefficients of each node's neighboring nodes to each node in the graph structure data. The first graph neural network encoder internally has multiple attention heads, each independently performing attention weight calculations. For the... Attention head, node For nodes Attention coefficient The calculation method is as follows: [The node is missing from the original text] Node feature vectors and nodes Node feature vectors After performing linear transformations on each component, the results are concatenated and then combined with a learnable attention vector. Perform an inner product and apply a non-linear activation function, with attention coefficients... The calculation formula is expressed as follows: .

[0025] in, Represents a node The node feature vectors, Represents a node The node feature vectors, Indicates the first The shared linear transformation weight matrix used by each attention head, weight matrix The dimension is set as , is the input dimension of the node feature vector. The output dimension after linear transformation. This represents a vector concatenation operation. Indicates the first Each attention head corresponds to a learnable attention vector, and the attention vector is... The dimension is , Represents a learnable attention vector transpose, This represents the activation function of the linear unit with leakage correction, where the negative slope parameter is set to 0.2. In some implementations, the node... Traversing nodes All first-order adjacent nodes. For all calculated attention coefficients By node Perform softmax normalization on all first-order adjacent nodes to obtain the normalized contribution coefficients. Normalized contribution coefficient The value ranges from 0 to 1, and the sum of the contribution coefficients of all adjacent nodes is 1. Among multiple attention heads, the average value of the normalized contribution coefficients calculated by each attention head or the coefficients are concatenated to obtain the final contribution coefficient.

[0026] Multiple nonlinear graph convolution operations are performed on the graph structure data based on the contribution coefficients to obtain an updated node feature matrix, which serves as the global working condition association feature. The number of nonlinear graph convolution operations is set to two or three layers. In each graph convolution operation, for each target node in the graph, the node feature vectors of all neighboring nodes are collected. The node feature vectors of each neighboring node are multiplied by their corresponding contribution coefficients and then summed. The sum is linearly combined with the node feature vector of the target node itself, and then transformed by a nonlinear activation function to generate the updated node feature vector of the target node. The nonlinear activation function used is the modified linear unit function. After all graph convolution layers have been operated on, each node in the graph obtains an updated node feature vector. The updated node feature vectors of all nodes are arranged in node order to form an updated node feature matrix, which is output as the global working condition association feature.

[0027] In some implementations, the first graph neural network encoder is a graph isomorphic network. The graph isomorphic network uses a summation aggregation function and a multilayer perceptron for node feature updates during graph convolution operations. The multilayer perceptron consists of two fully connected layers, with batch normalization and a modified linear unit activation function used between layers. In the graph isomorphic network, the contribution coefficients are unified with the weights of neighboring nodes, and all neighboring nodes contribute the same weight coefficients.

[0028] In other embodiments, the first graph neural network encoder is a graph convolutional network. The graph convolutional network uses a symmetric normalized Laplacian operator to aggregate adjacent node features during graph convolution operations.

[0029] The method for generating a dynamically enhanced operating condition representation by performing element-wise weighted fusion of global operating condition correlation features and initial monitoring data matrix is ​​as follows: the global operating condition correlation features and initial monitoring data matrix are input into a gated fusion network, which contains two parallel convolutional layers.

[0030] In its implementation, the first branch of the gated fusion network contains a first parallel convolutional layer. The input to this first parallel convolutional layer is the global operating condition correlation feature. This global operating condition correlation feature is a matrix, where rows correspond to sensor nodes and columns correspond to feature dimensions. The first parallel convolutional layer uses a one-dimensional convolution operation with a kernel size of 1. The kernel slides along the feature dimension to extract a weight coefficient tensor from the global operating condition correlation feature. Following the first parallel convolutional layer is a sigmoid activation function, which maps each element of the weight coefficient tensor to a range of 0 to 1. The shape of the weight coefficient tensor is the same as the shape of the global operating condition correlation feature.

[0031] The second branch of the gated fusion network contains a second parallel convolutional layer, whose input is the initial monitoring data matrix. This second parallel convolutional layer employs one-dimensional convolution operations with a kernel size of 1. The kernel slides along the time dimension to extract a value content tensor from the initial monitoring data matrix. A linear activation function or no nonlinear transformation is applied after the second parallel convolutional layer to maintain the numerical range of the value content tensor consistent with the initial monitoring data matrix. The shape of the value content tensor is the same as the shape of the initial monitoring data matrix.

[0032] The dynamically enhanced operating condition representation is obtained by performing element-wise multiplication between the weight coefficient tensor and the value content tensor. Element-wise multiplication involves performing a multiplication operation at each corresponding element position in the weight coefficient tensor and the value content tensor; the product result constitutes the element value of the dynamically enhanced operating condition representation at the corresponding position. The dynamically enhanced operating condition representation integrates the spatial weight information provided by the global operating condition correlation features and the temporal measurement information provided by the initial monitoring data matrix.

[0033] See Figure 7In the graph, the horizontal axis represents the monitoring time in seconds, and the vertical axis represents the amplitude, with a scale range of approximately -2 to 2. The graph contains two curves: the solid line with dots represents the "initial monitoring data matrix," and the dashed line with squares represents the "dynamic enhanced operating condition characterization." The solid curve shows the amplitude changes of sensor data collected by the hydropower station equipment over 10 consecutive seconds. The curve fluctuates significantly and exhibits multi-period characteristics, reflecting the original time-series signal characteristics of the equipment. The dashed curve's overall trend closely matches the solid curve, but it is smoother, with peaks and troughs at roughly the same positions. This indicates that the dynamic enhanced operating condition characterization, through element-wise weighted fusion of the initial monitoring data matrix using a gated fusion network, preserves the original signal's time-series characteristics while effectively suppressing some noise components.

[0034] As can be observed from the figure, at several key moments, the amplitudes of both the solid and dashed lines are close to their peak values, indicating that the dynamic enhanced operating condition representation can accurately reflect the global correlation characteristics of equipment operating conditions. The square markers on the dashed curve are evenly distributed across the entire time axis, indicating that the dynamic enhanced operating condition representation has undergone weight adjustments at each monitoring moment, demonstrating the effective enhancement effect of the spatial weight coefficients extracted by the gating fusion network on the original data.

[0035] Overall, this figure verifies that in Example 2, the global operating condition correlation features extracted by inputting the time-frequency joint feature map into the first neural network encoder, and the initial monitoring data matrix are fused element-wise through a gated fusion network, the resulting dynamically enhanced operating condition representation can achieve dynamic signal enhancement and noise suppression while maintaining the temporal characteristics of the original monitoring signal, further improving the accuracy and robustness of the equipment status representation.

[0036] Example 3: See Figure 4 The method for performing a sliding window Fourier residual mapping operation on the dynamic enhanced operating condition representation to obtain the fault-sensitive residual sequence is as follows: a fixed-length sliding window is set on the time axis, and a Fourier transform is performed on the dynamic enhanced operating condition representation within the sliding window to obtain the frequency domain amplitude vector within the window.

[0037] In implementation, the dynamic enhanced operating condition representation is a two-dimensional matrix. Each row of the dynamic enhanced operating condition representation corresponds to a sensor channel, and each column corresponds to a monitoring time. A sliding window is set on the time axis, with a fixed time span and a window length measured by the number of sampling points. The sliding window is placed starting from the beginning of the dynamic enhanced operating condition representation and covers several consecutive columns of sampling points. Within the time range covered by the sliding window, a Discrete Fourier Transform (DFT) is performed on the time segment of each sensor channel in the dynamic enhanced operating condition representation. The DFT maps the time-domain signal sequence to the frequency domain, outputting complex-form frequency domain coefficients. The magnitude of the complex-form frequency domain coefficients is taken to obtain the amplitude spectrum of each sensor channel within the sliding window. The amplitude spectra of all sensor channels within the sliding window are concatenated in channel order to form the frequency domain amplitude vector within the window. The dimension of the frequency domain amplitude vector within the window is equal to the product of the number of sensor channels and the number of discrete frequency points in the frequency domain.

[0038] The difference between the frequency domain amplitude vector within the sliding window and the frequency domain amplitude vector within the previous sliding window is calculated to obtain the frequency domain residual components of the sliding window. The current position of the sliding window on the time axis is denoted as the [number missing]. The first sliding window, the... The frequency domain magnitude vectors within each sliding window are denoted as follows: , No. The frequency domain magnitude vectors within each sliding window are denoted as follows: The formula for calculating the frequency domain residual components of the sliding window is as follows: .

[0039] in, Indicates the first The frequency domain amplitude vector obtained by performing a Fourier transform on the dynamic enhanced operating condition characterization within a sliding window. Indicates the first The frequency domain amplitude vector obtained by performing a Fourier transform on the dynamic enhanced operating condition characterization within a sliding window. Indicates the first The frequency domain residual components of the first sliding window; The sliding window is the first one. The position of the sliding window after moving forward one step on the time axis. When the first sliding window is the first sliding window on the timeline, there is no... , will the The frequency domain residual component of the sliding window is defined as the first... The frequency domain magnitude vector within each sliding window, i.e. .

[0040] A sliding window is used to collect the frequency domain residual components of all sliding windows along the time axis. These residual components are then arranged chronologically to form a fault-sensitive residual sequence. The sliding window slides along the time axis with a fixed step size, set to one sampling point interval. Each slide covers a new time interval. The process of performing a Fourier transform on the dynamic enhanced operating condition representation within the sliding window to obtain the frequency domain amplitude vector, and calculating the difference between the frequency domain amplitude vector within the window and the frequency domain amplitude vector at the previous time point to obtain the frequency domain residual components of the sliding window, is repeated until the sliding window exceeds the time range of the dynamic enhanced operating condition representation. The frequency domain residual components corresponding to all sliding windows are then processed. The fault-sensitive residuals are formed by stacking the sliding windows vertically in the order they appear on the time axis. The fault-sensitive residual sequence is a two-dimensional matrix, with the number of rows equal to the total number of sliding windows, and the number of columns equal to the dimension of the frequency domain amplitude vector within the window.

[0041] The window length of the sliding window is adaptively determined based on the rotational frequency cycle of the rotating components of the hydropower station equipment. In specific implementation, the rotational speed value measured by the speed sensor of the rotating components of the hydropower station equipment is obtained, with the unit of the speed value being revolutions per minute (rpm). The rotational frequency of the rotating components is calculated, which is the value obtained by dividing the speed value by 60, with the unit of the rotational frequency being Hertz (Hz). The rotational frequency represents the number of revolutions per second of the rotating components. The rotational frequency cycle is calculated based on the rotational frequency, which is the reciprocal of the rotational frequency, with the unit of the rotational frequency cycle being seconds. The window length of the sliding window is set to the integer part of the product obtained by multiplying the rotational frequency cycle by the sampling rate of the dynamic enhanced operating condition characterization. The sampling rate of the dynamic enhanced operating condition characterization is the number of sampling points collected per second by the sensor data acquisition device. When the rotational speed of the rotating components changes, the rotational frequency cycle changes accordingly. The window length of the sliding window is recalculated and adjusted each time the speed is updated, so that the window length of the sliding window always covers the number of sampling points within a complete rotational cycle of the rotating components.

[0042] See Figure 8 In the figure, the horizontal axis represents the sliding window index, ranging from 0 to 500, and the vertical axis represents the frequency domain residual amplitude. The figure shows the frequency domain residual sequence variation curves for three different frequency components, labeled as frequency component 1 (solid line, dots), frequency component 2 (dashed line, squares), and frequency component 3 (dotted line, triangles). The overall curve maintains a low residual amplitude for most of the sliding window index range, exhibiting oscillations close to zero, reflecting the stable operating state of the equipment.

[0043] Within the sliding window index range of approximately 90 to 110, frequency component 1 exhibits a significant peak, with the residual amplitude rapidly rising to approximately 1.7. This indicates a significant deviation between the dynamic enhanced operating condition characterization corresponding to this frequency component and the baseline health condition, suggesting a potential fault-sensitive signal. Within the sliding window index range of approximately 210 to 260, the residual amplitude of frequency component 2 increases significantly, reaching a peak of approximately 2.1, significantly higher than other ranges. This indicates that the fault-sensitive residual of this frequency component is enhanced during this period, potentially reflecting the equipment fault evolution process. Within the sliding window index range of approximately 370 to 400, the residual amplitude of frequency component 3 also increases significantly, with a peak value approaching 2.0, indicating abnormal fluctuations in frequency component 3 during this stage.

[0044] The residual peaks of the three frequency components do not overlap in time, and their amplitudes are significantly higher than those during the background stabilization period. This demonstrates the sensitive response capability of the sliding window Fourier residual mapping operation to different frequency components, effectively capturing the abnormal characteristics of multiple frequency bands during equipment operation. The abrupt changes and peak intervals of the residual amplitudes provide crucial fault-sensitive information for subsequent fault identification and root cause analysis. The fluctuation pattern of the overall curve corresponds to the local changes in the dynamic enhanced operating condition characterization in the sliding window frequency domain, verifying the effectiveness of the sliding window Fourier residual mapping operation in Example 3.

[0045] Example 4: See Figure 5 The method for reconstructing the baseline health status of a device by inputting the fault-sensitive residual sequence into a pre-constructed sparse deep belief network (SDN) is as follows: the SDN is composed of multiple stacked Restricted Boltzmann Machines (RBMs), each of which incorporates sparse regularization constraints. In specific implementation, the SDN consists of three stacked RBMs. The number of visible layer units in the first RBM is equal to the dimension of the fault-sensitive residual sequence, and the number of hidden layer units is set to half the dimension of the fault-sensitive residual sequence. The number of visible layer units in the second RBM is equal to the number of hidden layer units in the first RBM, and the number of hidden layer units is set to half the number of hidden layer units in the first RBM. The number of visible layer units in the third Restricted Boltzmann Machine (RBM) is equal to the number of hidden layer units in the second RBM, and the number of hidden layer units in the third RBM is set to half that of the second RBM. The three RBMs are stacked to form an encoding structure that progressively compresses the high-dimensional input to a low-dimensional latent space. The hidden layer of the third RBM serves as the bottleneck layer of the sparse deep belief network.

[0046] Each Restricted Boltzmann Machine (RBM) incorporates a sparse regularization constraint during training, implemented using a KL divergence penalty term. For a single-layer RBM, its training objective function consists of the sum of the log-likelihood term and the KL divergence penalty term. The expression for the KL divergence penalty term is as follows: .

[0047] in, This represents the KL divergence penalty term; Represents the sparsity penalty coefficient. The range of values ​​for is real numbers greater than 0. The settings are determined by cross-validation on the dataset, in one implementation scheme The value is set to 0.1 to balance reconstruction error and sparsity constraints; This indicates the number of hidden layer units in the current restricted Boltzmann machine; Represents the sparsity parameter, The value of is a real number between 0 and 1. The value of is 0.05, which is based on keeping the average activation probability of the hidden layer units at a low level of 5%, thereby forcing the hidden layer units to be in a suppressed state for most of the time and respond only to a few specific modes. Indicates the first The average activation probability of each hidden layer unit on a batch of training data. The calculation method involves inputting batches of training data into a restricted Boltzmann machine, and then statistically analyzing the results of the first batch. The proportion of samples where each hidden layer unit is active. The base of the logarithmic term is the natural constant.

[0048] Using fault-sensitive residual sequences as input data, each layer of the Restricted Boltzmann Machine (RBM) is trained sequentially, and the hidden layer output of the previous RBM is used as the input of the next RBM.

[0049] In practice, a data batch is randomly sampled from the fault-sensitive residual sequence and input into the first layer of the restricted Boltzmann machine to be trained. The data batch size is set to 128 samples, with each sample being a row in the fault-sensitive residual sequence, corresponding to the frequency domain residual component of a sliding window.

[0050] An unsupervised learning algorithm based on contrastive divergence is executed in the first-layer Restricted Boltzmann Machine (RBM) to update the connection weights and biases between the visible and hidden layers. The contrastive divergence-based unsupervised learning algorithm employs a one-step contrastive divergence algorithm. The specific update process is as follows: data batches are used as input to the visible layer states; the hidden layer activation probabilities are calculated; and the binary hidden layer states are sampled from these probabilities. The visible layer states are reconstructed using the binary hidden layer states, and the hidden layer activation probabilities in the reconstructed visible layer states are calculated. Based on the expected values ​​of the visible and hidden layer states, and the difference between the expected values ​​of the reconstructed visible and hidden layer states, the connection weight matrix, the visible layer bias vector, and the hidden layer bias vector are updated. The update amount of the connection weight matrix is ​​proportional to the cross product of the visible layer states and the hidden layer activation probabilities minus the cross product of the reconstructed visible layer states and the reconstructed hidden layer activation probabilities. During training, the learning rate is set to 0.01, the momentum term coefficient is set to 0.5, and the weight decay coefficient is set to 0.0002.

[0051] Freeze the trained parameters of the first-layer Restricted Boltzmann Machine (RBM) and calculate the hidden activation probabilities of the first-layer RBM for each data batch. The hidden activation probabilities are obtained by inputting the data batch into the first-layer RBM and performing a forward propagation using the trained connection weights and biases. Use the hidden activation probabilities of the first-layer RBM as the input data for the next layer RBM.

[0052] For the second-layer Restricted Boltzmann Machine (RBM), the unsupervised learning algorithm based on contrastive divergence and the hidden activation probability calculation are repeatedly performed. The original input data is replaced with the hidden activation probabilities output by the first-layer RBM. In the second-layer RBM, the connection weights and biases are updated using the contrastive divergence algorithm. The parameters of the second-layer RBM are frozen, and the hidden activation probabilities of the second-layer RBM are calculated as the input to the third-layer RBM.

[0053] For the third-layer Restricted Boltzmann Machine (RBM), the same unsupervised learning algorithm based on contrastive divergence is executed to update the connection weights and biases of the third-layer RBM, freeze its parameters, and calculate its hidden layer activation probabilities. At this point, the training of all layers of RBM is complete.

[0054] After layer-by-layer pre-training of all layers of Restricted Boltzmann Machines (RBMs), the entire sparse deep belief network is fine-tuned in reverse. The three stacked RBMs are unfolded along the encoder and decoder directions to form a symmetrical autoencoder network structure. The encoder part consists of the encoding weights of the first, second, and third RBMs, sequentially connected, mapping the input layer by layer to the bottleneck layer hidden representation. The decoder part consists of the decoding weights of the third, second, and first RBMs, sequentially connected, restoring the bottleneck layer hidden representation layer by layer back to the original input dimension. The decoder outputs the reconstructed data. The inverse fine-tuning aims to minimize the mean squared error (MSE), with the loss function defined as the mean squared error between the reconstructed data and the input fault-sensitive residual sequence. All weights and biases of the entire sparse deep belief network are used as trainable parameters. Backpropagation updates are performed using a stochastic gradient descent optimization algorithm with a learning rate of 0.001 and 100 iterations. After backpropagation fine-tuning, the reconstructed data is used as the equipment baseline health status, with the baseline dimension matching the dimension of the fault-sensitive residual sequence.

[0055] See Figure 9 In the figure, the horizontal axis represents the sample index, with a value range of 0 to 500, and the vertical axis represents the amplitude, with an amplitude range of approximately -0.5 to 2.5. The figure shows the amplitude curves of the fault-sensitive residual sequence (dashed line with dots) and the equipment baseline health status (solid line) as the sample index changes.

[0056] Both the fault-sensitive residual sequence and the equipment baseline health status exhibit periodic fluctuations and have largely consistent overall waveforms, indicating that the sparse deep belief network effectively reconstructs the fault-sensitive residual sequence. Specifically, within the sample index range of 0 to 500, the equipment baseline health status curve maintains smooth fluctuations, with amplitudes mostly concentrated between -0.3 and 1.4. While maintaining a similar fluctuation trend, the fault-sensitive residual sequence's amplitude curve exhibits relatively large local oscillations, with peaks significantly deviating from the baseline at some points. For example, at sample indices of approximately 130 and 370, the peak amplitude of the fault-sensitive residual sequence is significantly higher than the baseline health status, reaching an amplitude range of 2.3 to 2.5.

[0057] This figure illustrates the process of reconstructing the fault-sensitive residual sequence layer by layer using a sparse deep belief network to generate a baseline of equipment health status. The smooth fluctuations of the equipment baseline health status reflect the baseline characteristics of the equipment in a healthy state, while the obvious peaks in the fault-sensitive residual sequence at certain times reflect possible fault signals. By comparing the waveforms of the two, abnormal changes in equipment status can be effectively identified, providing a reliable basis for subsequent health deviation calculations and fault root cause localization.

[0058] Example 5: See Figure 6 The Riemannian distance between the dynamic enhanced operating condition characterization and the equipment baseline health condition is calculated as the realization of the equipment health deviation. The dynamic enhanced operating condition characterization and the equipment baseline health condition are mapped to a symmetric positive definite manifold space respectively to obtain the first manifold point and the second manifold point.

[0059] In practical implementation, the dynamic enhanced operating condition characterization is represented as a two-dimensional matrix, where the number of rows corresponds to the number of sensor channels, and the number of columns corresponds to the number of time sampling points. The equipment baseline health status is a two-dimensional matrix with the same shape as the dynamic enhanced operating condition characterization. The dynamic enhanced operating condition characterization is mapped to a symmetric positive definite manifold space by calculating the sample covariance matrix of the dynamic enhanced operating condition characterization. The formula for calculating the sample covariance matrix is ​​as follows: .

[0060] in, This represents the sample covariance matrix corresponding to the dynamically enhanced operating condition characterization. Let A be the first manifold point in the space of a symmetric positive definite manifold. This represents the dynamic enhanced operating condition characterization matrix. Each line represents a time series of one sensor channel. Each column is a cross-channel vector for a time sampling point; This represents the mean vector of the dynamic enhanced operating condition characterization matrix. The The element is the dynamically enhanced operating condition characterization matrix. The vector is obtained by subtracting the average of all columns in a row from itself and then repeatedly calculating it. During the calculation, the mean of the corresponding row is subtracted from each row of the dynamic enhanced working condition characterization matrix. This represents the number of columns in the dynamic enhanced operating condition characterization matrix, i.e., the number of time sampling points; This represents the matrix transpose operation; This represents the covariance normalization factor. In This represents the correction value for the degrees of freedom.

[0061] The method for mapping the equipment baseline health condition to a symmetric positive definite manifold space is as follows: Calculate the sample covariance matrix of the equipment baseline health condition. The calculation method of the sample covariance matrix is ​​the same as that of the sample covariance matrix of the dynamic enhanced operating condition characterization. Subtract the row mean of the equipment baseline health condition matrix, multiply it by its transpose, and divide by the number of columns minus one to obtain the sample covariance matrix corresponding to the equipment baseline health condition. The sample covariance matrix corresponding to the equipment baseline health condition is the second manifold point in the symmetric positive definite manifold space.

[0062] In a symmetric positive definite manifold space, the affine invariant Riemann distance between a first manifold point and a second manifold point is calculated. The affine invariant Riemann distance measures the geometric distance between two positive definite symmetric matrices on a symmetric positive definite manifold. The affine invariant Riemann distance is calculated as follows: eigenvalue decomposition is performed on the product of the inverse matrix of the first manifold point and the product of the second manifold point, yielding a set of real eigenvalues. The natural logarithm of these eigenvalues ​​is taken, the sum of squares is calculated, and the square root is taken. When the first and second manifold points are identical, the affine invariant Riemann distance is zero; the greater the difference between the first and second manifold points, the larger the affine invariant Riemann distance. The calculated affine invariant Riemann distance is defined as the equipment health deviation.

[0063] When the equipment health deviation exceeds a preset threshold, a root cause localization mechanism based on a causal graph attention network is triggered, outputting a set of fault attribution parameters. The implementation is as follows: after determining that the equipment health deviation exceeds the preset threshold, the residual matrix between the dynamically enhanced operating condition representation and the equipment baseline health state is obtained as input for root cause localization. In specific implementation, the preset threshold is determined based on the statistical distribution of equipment health deviation under historical healthy operating conditions of the hydropower station equipment. The preset threshold is set to the 95th or 99th percentile of the equipment health deviation under historical healthy operating conditions. The residual matrix is ​​obtained by subtracting the dynamically enhanced operating condition representation matrix from the equipment baseline health state matrix at each corresponding element position. The shape of the residual matrix is ​​the same as that of the dynamically enhanced operating condition representation matrix. Each element in the residual matrix represents the deviation between the actual monitored value of the corresponding sensor channel and the health baseline value at the corresponding time.

[0064] The residual matrix is ​​input into the causal graph attention network, which internally maintains an adjacency matrix describing the causal relationships between the sensing variables.

[0065] In its implementation, the core architecture of the causal graph attention network includes a causal discovery module and multiple stacked attention propagation layers. The causal discovery module employs a constraint-based causal structure learning algorithm, such as the PC algorithm, to perform independence and conditional independence tests on the observation data of all sensor variables during historical healthy operation periods, outputting a causal adjacency matrix of a directed acyclic graph structure. The causal adjacency matrix is ​​a square matrix where the number of rows and columns equals the number of sensor variables. Line number An element in the column with a value of 1 indicates a sensor variable. Sensor variables The direct cause is that a value of 0 indicates a sensor variable. Not a sensor variable The direct cause.

[0066] The attention propagation layer is set to two layers. Each attention propagation layer contains a multi-head attention mechanism and a feedforward network sublayer. The number of attention heads in the multi-head attention mechanism is set to four. In each attention propagation layer, the causal adjacency matrix is ​​used as the attention mask, and attention weights are calculated only for sensor variable pairs with a value of 1 in the causal adjacency matrix. The attention weights between sensor variable pairs with a value of 0 in the causal adjacency matrix are forced to zero. By stacking two attention propagation layers, the causal graph attention network can propagate indirect causal effects based on direct causal relationships.

[0067] In the causal graph attention network, attention weights are calculated for each feature dimension in the residual matrix using the causal adjacency matrix. These attention weights quantify the degree to which each feature dimension deviates from the baseline state. Specifically, each column of the residual matrix is ​​treated as a feature dimension vector, and each feature dimension vector corresponds to the residual sequence of a sensor channel across all time sampling points. All sensor channel feature dimension vectors are used as input node features of the causal graph attention network. In the first attention propagation layer, a linear projection transformation is performed on each feature dimension vector to generate a query vector, key vector, and value vector. The inner product of the query vector and key vector is used, along with a causal adjacency matrix mask, to calculate the original attention weight values ​​between sensor variables. Softmax normalization is applied to the original attention weight values ​​to obtain normalized attention weights. These normalized attention weights are then weighted and aggregated with the value vectors to obtain the updated feature dimension representation. The above attention calculation process is repeated in the second attention propagation layer to obtain the final feature dimension representation.

[0068] The output of the second attention propagation layer is subjected to global average pooling of all feature dimensions, followed by linear projection and softmax operation to obtain the attention weight corresponding to each feature dimension. The attention weight is a scalar with a value between 0 and 1, and the sum of the attention weights of all feature dimensions is 1.

[0069] Sort the attention weights of all feature dimensions in descending order of value, select three feature dimensions with the highest attention weight values, and output the sensor names or device parameter names corresponding to the three selected feature dimensions as the fault attribution parameter set.

[0070] See Figure 10 In the figure, the horizontal axis represents the monitoring time of the equipment in hours, ranging from 0 to 100 hours. The vertical axis represents the equipment health deviation, which is the Riemannian distance between the dynamic enhanced operating condition characterization and the equipment baseline health status. The value ranges from approximately 0 to 4.2. The solid curve in the figure represents the trend of the equipment health deviation with the monitoring time, and the dashed horizontal line represents the preset threshold with a value of 2.

[0071] The curve trend shows that the equipment health deviation exhibits a slow upward trend in the initial monitoring period (0 to 30 hours), remaining below 1.5, which is below the preset threshold, indicating that the equipment is in a normal healthy state. Subsequently, in the 30 to 45 hour range, the deviation gradually approaches the preset threshold, and the trend becomes more gradual. After 45 hours, the equipment health deviation continues to rise, and at approximately 50 hours, it exceeds the preset threshold of 2 for the first time, indicating that the equipment health status has begun to deviate from the normal baseline, potentially indicating a risk of failure. After exceeding the threshold, the deviation remains above 2 and shows an oscillating upward trend, especially in the 60 to 100 hour range, where the deviation fluctuations intensify and gradually increase overall, reaching a maximum of approximately 4.1, significantly exceeding the preset threshold, indicating that the equipment health status has significantly deviated from the normal baseline, and the risk of failure is further aggravated.

[0072] The solid black dots marked on the curve, distributed at different time points, highlight key numerical nodes of equipment health deviation, facilitating observation of the phased changes in deviation. Overall, the graph reflects the gradual evolution of equipment health deviation calculated based on Riemannian manifold distance from low to high over monitoring time. A preset threshold serves as an anomaly trigger threshold, clearly distinguishing between healthy and abnormal equipment states. This embodiment effectively quantifies the difference between the dynamically enhanced operating condition characterization and the equipment baseline health state using Riemannian manifold distance, achieving dynamic monitoring and early warning functions for equipment health deviation.

[0073] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A data-driven intelligent fault diagnosis method for hydropower station equipment, characterized in that, Includes the following steps: Multi-source sensor monitoring data of hydropower station equipment at multiple historical moments are acquired to construct an initial monitoring data matrix; the initial monitoring data matrix is ​​transformed in the time and frequency domain to obtain a time-frequency joint feature map; the time-frequency joint feature map is input into a first graph neural network encoder to extract global operating condition related features; The global operating condition correlation features are fused with the initial monitoring data matrix element by element to generate a dynamically enhanced operating condition representation; a sliding window Fourier residual mapping operation is performed on the dynamically enhanced operating condition representation to obtain a fault-sensitive residual sequence. The fault-sensitive residual sequence is input into a pre-constructed sparse deep belief network to reconstruct the baseline of the device's baseline health status layer by layer. Calculate the Riemannian distance between the dynamic enhanced operating condition characterization and the equipment baseline health condition, as the equipment health deviation. When the device health deviation exceeds a preset threshold, a root cause localization mechanism based on a causal graph attention network is triggered, and a set of fault attribution parameters is output.

2. The data-driven intelligent fault diagnosis method for hydropower station equipment according to claim 1, characterized in that, The specific implementation of performing time-frequency domain transformation on the initial monitoring data matrix to obtain the time-frequency joint feature map is as follows: Perform a Hilbert-Huang transform on each sensor data sequence in the initial monitoring data matrix to extract the instantaneous frequency component and instantaneous amplitude component of the sensor data sequence; multiply the instantaneous frequency component and the instantaneous amplitude component point by point to construct the marginal spectral energy distribution vector of the sensor data sequence; align the marginal spectral energy distribution vectors of all sensor data sequences according to the monitoring time and stack them vertically to form the time-frequency joint feature map.

3. The data-driven intelligent fault diagnosis method for hydropower station equipment according to claim 1, characterized in that, The specific implementation of inputting the time-frequency joint feature map into the first graph neural network encoder to extract global operating condition related features is as follows: taking the marginal spectral energy distribution vector corresponding to each sensor data sequence in the time-frequency joint feature map as the node feature, and taking the spatial distance between each sensor in the predefined physical topology of the hydropower station equipment as the edge weight, to construct graph structure data; In the first graph neural network encoder, a multi-head attention mechanism is used to calculate the contribution coefficient of each node's neighboring nodes to the node in the graph structure data; based on the contribution coefficient, multiple nonlinear graph convolution operations are performed on the graph structure data to obtain an updated node feature matrix, which serves as the global working condition association feature.

4. The data-driven intelligent fault diagnosis method for hydropower station equipment according to claim 3, characterized in that, The first graph neural network encoder is a graph isomorphic network or a graph convolutional network.

5. The data-driven intelligent fault diagnosis method for hydropower station equipment according to claim 1, characterized in that, The specific implementation of generating a dynamically enhanced operating condition representation by performing element-wise weighted fusion of the global operating condition correlation features and the initial monitoring data matrix is ​​as follows: the global operating condition correlation features and the initial monitoring data matrix are input into a gated fusion network, which contains two parallel convolutional layers; the weight coefficient tensor is extracted from the global operating condition correlation features through the first convolutional layer; the value content tensor is extracted from the initial monitoring data matrix through the second convolutional layer; and the weight coefficient tensor and the value content tensor are multiplied element-wise to obtain the dynamically enhanced operating condition representation.

6. The data-driven intelligent fault diagnosis method for hydropower station equipment according to claim 1, characterized in that, The specific implementation of performing a sliding window Fourier residual mapping operation on the dynamically enhanced operating condition characterization to obtain the fault-sensitive residual sequence is as follows: a sliding window of fixed length is set on the time axis, and a Fourier transform is performed on the dynamically enhanced operating condition characterization within the sliding window to obtain the frequency domain amplitude vector within the window; the difference between the frequency domain amplitude vector within the window and the frequency domain amplitude vector within the sliding window at the previous time step is calculated to obtain the frequency domain residual components of the window. The sliding window is slid along the time axis to collect the frequency domain residual components of all windows, and the frequency domain residual components are arranged in chronological order to form the fault-sensitive residual sequence.

7. The data-driven intelligent fault diagnosis method for hydropower station equipment according to claim 6, characterized in that, The length of the sliding window is adaptively determined based on the rotation frequency cycle of the rotating components of the hydropower station equipment.

8. The data-driven intelligent fault diagnosis method for hydropower station equipment according to claim 1, characterized in that, The specific implementation of inputting the fault-sensitive residual sequence into a pre-constructed sparse deep belief network to reconstruct the equipment baseline health status layer by layer is as follows: The sparse deep belief network is composed of multiple stacked restricted Boltzmann machines, each of which is subject to sparse regularization constraints; the fault-sensitive residual sequence is used as input data, and each layer of the restricted Boltzmann machine is trained sequentially, with the hidden layer output of the previous layer's restricted Boltzmann machine used as the input of the next layer's restricted Boltzmann machine; after completing the layer-by-layer pre-training of all layers of restricted Boltzmann machines, the entire sparse deep belief network is fine-tuned in reverse so that the data reconstructed from the output layer is aligned with the fault-sensitive residual sequence under the mean square error minimization criterion, and the reconstructed data is used as the equipment baseline health status.

9. The data-driven intelligent fault diagnosis method for hydropower station equipment according to claim 8, characterized in that, The sparse regularization constraint of each layer of the restricted Boltzmann machine in the sparse deep belief network is implemented using the KL divergence penalty term.

10. The data-driven intelligent fault diagnosis method for hydropower station equipment according to claim 1, characterized in that, The Riemannian manifold distance between the dynamically enhanced operating condition characterization and the equipment baseline health condition is calculated as the specific implementation of the equipment health deviation: the dynamically enhanced operating condition characterization and the equipment baseline health condition are respectively mapped to a symmetric positive definite manifold space to obtain a first manifold point and a second manifold point; In the symmetric positive definite manifold space, the affine invariant Riemann distance between the first manifold point and the second manifold point is calculated; the calculated value of the affine invariant Riemann distance is determined as the device health deviation.

Images