Multi-sensor cooperative partial discharge waveform identification method and system

Abstract

The invention discloses a multi-sensor cooperative partial discharge waveform identification method and system, and the method comprises the steps: extracting an envelope line of an oscillation waveform through Hilbert transformation for a collected partial discharge signal, carrying out the Logistic function fitting and exponential decay function fitting of each sensor signal, and carrying out the recognition of the partial discharge waveform through the Logistic function fitting and exponential decay function fitting. Taking the characteristic parameters extracted by each sensor as node characteristics in a graph structure, establishing an edge connection relation in a graph according to the maximum value starting time difference and the similarity of signals between the sensors, constructing a graph neural network model, performing training and reasoning by utilizing the graph structure, and finally outputting a recognition result of whether a partial discharge event occurs or not; the system comprises a signal conversion module, a feature extraction module, a data fusion module and a partial discharge diagnosis module. According to the invention, the accuracy and adaptability of partial discharge signal identification are improved, and the on-line monitoring requirement of the transformer in a complex electromagnetic environment is met.

Description

Technical Field

[0001] The present invention relates to the field of online monitoring of primary equipment in power systems, and in particular to a multi-sensor coordinated partial discharge waveform recognition method and system. Background Art

[0002] Partial discharge (PD) refers to non-penetrating discharges limited to a localized area within the insulation system of electrical equipment. It is a key precursor to insulation degradation in high-voltage electrical equipment. In HVDC transmission and transformation equipment, such as converter transformers, PD is often caused by manufacturing defects, operational aging, or environmental stress, manifesting as weak but reproducible electromagnetic waves, acoustic emission signals, and high-frequency current disturbances.

[0003] As core equipment in HVDC transmission systems, the reliability of the converter transformer's insulation system is directly related to the safe and stable operation of the entire DC transmission project. Due to its high operating voltage levels, complex operating conditions, and long-term exposure to AC / DC composite electric fields, the identification of partial discharge signals places higher demands on it: not only high-sensitivity acquisition capabilities, but also strong anti-interference capabilities and multi-source information fusion mechanisms.

[0004] Among existing partial discharge detection technologies, commonly used methods include: threshold-triggered pulse peak detection, which determines whether a partial discharge event has occurred by setting an amplitude or energy threshold; frequency domain analysis, which uses fast Fourier transform (FFT) or wavelet transform (WT) to extract the spectral characteristics of the partial discharge signal; pattern recognition, which combines traditional machine learning algorithms (such as support vector machine SVM and random forest RF) for classification; and multi-sensor collaborative detection, which uses UHF electromagnetic sensors, HFCT current sensors, and AE acoustic sensors to jointly determine partial discharge events.

[0005] However, these methods face several practical challenges: Low edge feature extraction accuracy: Traditional methods typically rely on fixed windows to extract rising and falling edges, making it difficult to accurately capture key features in PD signals, such as steep rises and slow falls, resulting in a high false positive rate. Imperfect multi-sensor fusion mechanisms: Existing studies have often used result-level voting or feature splicing to process multi-source signals, lacking modeling of physical consistency and temporal synchronization between sensors, and thus failing to effectively improve recognition robustness. Poor anti-interference capabilities: In actual operating environments, converter transformer PD signals are often accompanied by complex electromagnetic interference and environmental noise. Due to the inherent weakness and transient nature of PD signals, traditional PD identification methods are prone to false positives or missed positives when faced with interference from high-frequency noise, transient equipment operation, and external electromagnetic coupling.

[0006] Therefore, it is urgent to propose a new method that can effectively extract key edge features in partial discharge waveforms and combine multi-sensor data for intelligent fusion. Summary of the Invention

[0007] Purpose of the invention: The purpose of the present invention is to provide a multi-sensor collaborative partial discharge waveform recognition method and system, aiming to solve the problems of low waveform feature extraction accuracy, imperfect multi-sensor fusion mechanism and poor system robustness in the existing technology.

[0008] Technical solution: The multi-sensor collaborative partial discharge waveform recognition method of the present invention comprises the following steps:

[0009] Step 1: Install multiple sensors on the transformer box wall to collect partial discharge signals inside the running transformer. The sensors include ultra-high frequency sensors, high frequency sensors, and acoustic sensors.

[0010] Step 2: Perform Hilbert transform on the collected signals to extract the envelope of the oscillation waveform;

[0011] Step 3: Apply the Logistic function to the rising edge area in the envelope of each sensor for nonlinear fitting, and extract the Logistic function steepness and start time as edge features;

[0012] Step 4: Apply an exponential decay function to the falling edge region of the envelope of each sensor for nonlinear fitting, and extract the time constant as the energy dissipation feature;

[0013] Step 5: Use the feature parameters extracted by each sensor as node features in the graph structure;

[0014] Step 6: Calculate the edge weight based on the initial time difference and similarity of the signals between sensors, obtain the edge attributes between nodes, and establish the edge connection relationship in the graph structure;

[0015] Step 7: Build a graph neural network model and use the graph structure for training and reasoning to achieve collaborative fusion of multi-source sensor information;

[0016] Step 8: Output the identification result of whether it is a partial discharge event.

[0017] Furthermore, the step 3 uses the Logistic function for nonlinear fitting, and the expression of the Logistic function is:

[0018]

[0019] Where L is the platform height, k is the edge steepness, and t0 is the starting time.

[0020] Furthermore, the step 4 uses an exponential decay function for nonlinear fitting, and the expression of the exponential decay function is:

[0021]

[0022] Where A is the initial amplitude and τ is the time constant of the falling edge.

[0023] Furthermore, the node features in the graph structure of step 5 are composed of parameters extracted by the logistic function and the exponential decay function, including steepness, start time and time constant.

[0024] Furthermore, the step of calculating the edge weight in step 6 includes:

[0025] Step 6.1: Calculate the start time difference between different sensor signals

[0026] Step 6.2: If the time difference is less than the threshold, create an edge between the two corresponding nodes;

[0027] Step 6.3: The edge weight is dynamically assigned according to the time difference. The smaller the time difference, the larger the edge weight, indicating a higher consistency between the two.

[0028] Furthermore, the edge weight in step 6.2 not only considers the time difference, but also combines the correlation score between sensors. The correlation score is derived from the statistical analysis of previous historical data and is calculated as follows:

[0029]

[0030] Among them, y m (t) ,y n (t) represents the signal amplitude collected by sensors m and n at time t, represents the mean value of the signals of sensors m and n, S mn Represents the signal correlation between sensors m and n, and its value range is [-1, 1].

[0031] Furthermore, step 6.2 introduces an edge attribute mechanism into the graph neural network, combining the time difference and similarity into an edge attribute vector:

[0032] e mn =[Δt mn , S mn ]

[0033] And input it into the graph neural network layer that supports edge attributes for training and inference.

[0034] Furthermore, in step 7, the graph neural network uses a graph isomorphism network framework to model the graph structure, and the graph isomorphism network framework uses the GINEConv layer as the graph convolution layer.

[0035] The update rules of the GINEConv layer are as follows:

[0036]

[0037] in, is the embedding representation of the mth node in layer l; N(m) is the set of neighbors of node m; e mn is the edge attribute vector; φ (l) Is a learnable edge feature mapping function; MLP l is a shared multilayer perceptron; e (l) is a trainable scaling factor.

[0038] Furthermore, step 8 introduces a global average pooling operation into the output end of the graph neural network to obtain a graph-level embedding vector, and inputs it into a classifier to determine whether a partial discharge event occurs.

[0039] The multi-sensor coordinated partial discharge waveform recognition system of the present invention comprises:

[0040] The signal conversion module is used to perform Hilbert transform on the collected signals and extract the envelope of the oscillation waveform;

[0041] The feature extraction module is used to apply a logistic function to the rising edge region of the envelope of each sensor for nonlinear fitting, extracting the logistic function steepness and start time as edge features; and to apply an exponential decay function to the falling edge region of the envelope of each sensor for nonlinear fitting, extracting the time constant as the energy dissipation feature;

[0042] The data fusion module is used to build a graph neural network model and use the graph structure for training and reasoning to achieve the collaborative fusion of multi-source sensor information;

[0043] The partial discharge diagnosis module is used to introduce the output of the graph neural network into the global average pooling operation to obtain a graph-level embedding vector, and input it into the classifier to determine whether a partial discharge event occurs.

[0044] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: the present invention uses function fitting to extract key parameters with physical significance, avoids the problem of information loss, and improves the accuracy of edge recognition; introduces the edge attribute mechanism into the graph neural network model, so that the model can consider time synchronization and sensor correlation at the same time, realizes feature-level fusion in the true sense, breaks through the limitations of traditional "result voting" or "feature splicing", and makes the final recognition result more stable and consistent, which is suitable for online monitoring tasks of transformers under complex electromagnetic environments on site. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A logic flow chart of the multi-sensor collaborative partial discharge waveform recognition method implemented in the present invention;

[0046] Figure 2 This is an effect diagram of implementing curve fitting and extracting waveform features in the present invention;

[0047] Figure 3 A graph structure diagram for implementing feature node modeling in the present invention;

[0048] Figure 4 This is the technical roadmap for implementing partial discharge waveform recognition in the present invention. DETAILED DESCRIPTION

[0049] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0050] like Figure 1 As shown, the multi-sensor collaborative partial discharge waveform recognition method of the present invention includes the following steps:

[0051] Step 1: Install multiple sensors on the transformer box wall to collect partial discharge signals inside the running transformer. The sensors include ultra-high frequency sensors, high frequency sensors, and acoustic sensors.

[0052] Ultra-high frequency sensor (UHF): used to detect electromagnetic wave signals;

[0053] High frequency current sensor (HFCT): used to detect partial discharge induced current in the winding;

[0054] Acoustic sensor (AE): used to detect ultrasonic signals caused by partial discharge.

[0055] The three types of sensors are deployed in a distributed manner based on the actual equipment structure and monitoring requirements to form a multi-channel acquisition system. For example, in one specific embodiment, a total of 14 sensor channels are deployed: 4 UHF, 4 HFCT, and 6 AE sensors, to synchronously acquire multi-source information of partial discharge signals.

[0056] The original signals collected by each sensor are time-domain oscillation waveforms, which exhibit non-stationary and nonlinear characteristics, and have varying degrees of noise interference and electromagnetic coupling effects.

[0057] Step 2: Perform Hilbert transform on the collected signals to extract the envelope of the oscillation waveform;

[0058] Hilbert transform is performed on the original oscillation signal collected by each sensor to construct an analytical signal, and its modulus value is extracted as the envelope of the signal.

[0059] Specifically, for the signal y(t) collected by any sensor, its analytical signal is defined as:

[0060] z(t)=y(t)+j·H[y(t)]

[0061] Where H[y(t)] is the Hilbert transform result of the signal, and z(t) is the complex analytical signal.

[0062] Then calculate the modulus of the analytical signal:

[0063]

[0064] The envelope reflects the overall energy change trend of the signal, eliminates the influence of high-frequency oscillation, and facilitates subsequent edge identification and function fitting.

[0065] Step 3: Apply the Logistic function to the rising edge area in the envelope of each sensor for nonlinear fitting, and extract the Logistic function steepness and start time as edge features;

[0066] After completing the envelope extraction, the main rising edge area of ​​each sensor signal is identified.

[0067] Specifically, the gradient analysis method or threshold determination method is used to identify the interval where the signal amplitude rises rapidly from the background noise to the peak value. Then the logistic function is applied to this area for nonlinear least squares fitting:

[0068]

[0069] Among them, L represents the platform height, k represents the edge steepness, and t0 represents the starting time of Logistic fitting, such as Figure 2 The blue fitted line is shown.

[0070] These parameters have clear physical meanings and can be used to characterize the characteristics of the rising phase of a partial discharge signal. Furthermore, the initial parameters are automatically estimated based on the signal's maximum value and timeline information, improving the algorithm's stability and adaptability.

[0071] Step 4: Apply an exponential decay function to the falling edge region of the envelope of each sensor for nonlinear fitting, and extract the time constant as the energy dissipation feature;

[0072] Identify the falling edge region of each sensor signal, that is, the process where the signal slowly falls from the peak to the bottom noise. Apply the exponential decay function to this region for nonlinear fitting:

[0073]

[0074] Among them, A represents the initial amplitude, and τ represents the time constant of the falling edge, such as Figure 2 The green fitted line is shown.

[0075] This time constant can be used to subsequently fusion the important input features of the model.

[0076] Step 5: Use the feature parameters extracted by each sensor as node features in the graph structure;

[0077] All sensor acquisition channels are regarded as nodes in the graph structure, and each node corresponds to a sensor channel. Specifically, the characteristics of each node are composed of the steepness k of the logistic function fitting, the starting time t0, and the time constant τ of the exponential decay function fitting.

[0078] Step 6: Calculate the edge weight based on the initial time difference and similarity of the signals between sensors, obtain the edge attributes between nodes, and establish the edge connection relationship in the graph structure, such as Figure 3 As shown in the figure, the edge connection relationship in the graph is established according to the starting time difference of the signals between the sensors. The specific method is to extract the starting time t0 of the Logistic function and calculate the starting time difference between any two sensors m and n. When the time difference is less than the set threshold (1ms), the two signals are considered to represent the same partial discharge, and an edge is established between the two corresponding nodes. The weight of this edge not only considers the time difference, but also the correlation score between the sensors. The correlation score is derived from the statistical analysis of previous historical data and is calculated as follows:

[0079]

[0080] Among them, y m (t) ,y n (t) represents the signal amplitude collected by sensors m and n at time t, represents the mean value of the signals of sensors m and n, S mn Represents the signal correlation between sensors m and n, with a value range of [-1, 1]. Introducing the edge attribute mechanism in the graph neural network, the time difference and similarity are combined into an edge attribute vector:

[0081] e m =[Δt mn , S mn ]

[0082] This is then fed into a graph neural network layer that supports edge attributes for training and inference. Edge weights are dynamically assigned based on the time difference. The smaller the time difference, the greater the edge weight, indicating greater consistency between the two.

[0083] Step 7: Build a graph neural network model and use the graph structure for training and reasoning to achieve collaborative fusion of multi-source sensor information;

[0084] like Figure 4 As shown, after completing the graph structure modeling, a graph neural network model is constructed to achieve the coordinated fusion of multi-source sensor information and output the recognition results of partial discharge events. In an embodiment of the present invention, the GINEConv (Graph Isomorphism Network with Edge features) layer is used as the core component of the graph neural network. This layer is an extended version of the Graph Isomorphism Network (GIN) that can accept and process edge attribute information and is very suitable for the scenario of fusing time difference and similarity score in the present invention.

[0085] The update rules of the GINEConv layer are as follows:

[0086]

[0087] in, is the embedding representation of the mth node in layer l; N(m) is the set of neighbors of node m; e mn is the edge attribute vector, including time difference and relevance score; (l) Is a learnable edge feature mapping function; MLP l Is a shared multi-layer perceptron used to aggregate neighbor information; ∈ (l) is a trainable scaling factor.

[0088] Through the above formula, GINEConv can simultaneously consider the node's own state and the information propagated on the edge, thereby achieving more refined graph-level modeling.

[0089] The multi-layer GNN propagation process is as follows:

[0090] 1. First layer GINEConv:

[0091] Input: original node features (such as logistic and exponential fitting parameters);

[0092] Output: Node embedding after preliminary aggregation of neighbor information;

[0093] 2. Second layer GINEConv:

[0094] Input: node embedding + edge attributes of the previous layer;

[0095] Output: high-dimensional representation after further information propagation;

[0096] 3. Subsequent layers:

[0097] GINEConv layers can be stacked based on actual scenarios to improve the model's ability to understand graph structures.

[0098] Each layer of GINEConv updates the node representation, allowing each sensor channel to gradually perceive the status of its neighboring sensors, and combines the time difference and correlation score in the edge attributes to form a unified graph structure representation.

[0099] In the last layer of the graph neural network, global mean pooling is used to obtain a unified embedding vector at the graph level:

[0100]

[0101] Where N is the total number of sensor nodes; is the node embedding output by the last layer of the graph neural network; g is the graph-level embedding vector, which represents the comprehensive information of the entire graph.

[0102] Step 8: Output the identification result of whether it is a partial discharge event.

[0103] Then, the graph embedding vector obtained in step 7 is input into the classifier, and the probability value of whether the current event is a partial discharge event is output:

[0104] P(PD)=σ(W2·ReLU(W1·g+b1)+b2)

[0105] Where W1 and W2 are the classifier's weight matrices; b1 and b2 are bias terms; and σ is a sigmoid function that outputs a probability value (0 to 1). When the output probability is greater than a set threshold (e.g., 0.5), it is considered a partial discharge event; otherwise, it is considered a false alarm or background noise.

Claims

1. A multi-sensor collaborative partial discharge waveform recognition method, characterized in that: The steps include: Step 1: Install multiple sensors on the transformer box wall to collect partial discharge signals inside the running transformer. The sensors include ultra-high frequency sensors, high frequency sensors, and acoustic sensors. Step 2: Perform Hilbert transform on the collected signals to extract the envelope of the oscillation waveform; Step 3: Apply the Logistic function to the rising edge area in the envelope of each sensor for nonlinear fitting, and extract the Logistic function steepness and start time as edge features; Step 4: Apply an exponential decay function to the falling edge region of the envelope of each sensor for nonlinear fitting, and extract the time constant as the energy dissipation feature; Step 5: Use the feature parameters extracted by each sensor as node features in the graph structure; Step 6: Calculate the edge weight based on the initial time difference and similarity of the signals between sensors, obtain the edge attributes between nodes, and establish the edge connection relationship in the graph structure; Step 7: Build a graph neural network model and use the graph structure for training and reasoning to achieve collaborative fusion of multi-source sensor information; Step 8: Output the identification result of whether it is a partial discharge event.

2. The multi-sensor coordinated partial discharge waveform recognition method according to claim 1, characterized in that: The step 3 uses the Logistic function for nonlinear fitting, and the expression of the Logistic function is: Where L is the platform height, k is the edge steepness, and t0 is the starting time.

3. The multi-sensor collaborative partial discharge waveform recognition method according to claim 1, characterized in that: The step 4 uses an exponential decay function for nonlinear fitting, and the expression of the exponential decay function is: Where A is the initial amplitude and T is the time constant of the falling edge.

4. The multi-sensor collaborative partial discharge waveform recognition method according to claim 1, characterized in that: The node features in the graph structure of step 5 are composed of parameters extracted by the logistic function and the exponential decay function, including steepness, start time and time constant.

5. The multi-sensor coordinated partial discharge waveform recognition method according to claim 1, characterized in that: The construction of the edge connection relationship in step 6 includes: Step 6.1: Calculate the start time difference between different sensor signals Step 6.2: If the time difference is less than the threshold, create an edge between the two corresponding nodes; Step 6.3: The edge weight is dynamically assigned according to the time difference. The smaller the time difference, the larger the edge weight, indicating a higher consistency between the two.

6. The multi-sensor coordinated partial discharge waveform recognition method according to claim 5, characterized in that: The edge weights in step 6.2 not only consider the time difference, but also the correlation scores between sensors. The correlation scores are derived from the statistical analysis of previous historical data and are calculated as follows: Among them, y m (t) ,y n (t) represents the signal amplitude collected by sensors m and n at time t, represents the mean value of the signals of sensors m and n, S mn Represents the signal correlation between sensors m and n, and its value range is [-1, 1].

7. The multi-sensor coordinated partial discharge waveform recognition method according to claim 1, characterized in that: Step 6.2 introduces an edge attribute mechanism into the graph neural network, combining the time difference and similarity into an edge attribute vector: e mn =[Δt mn ,S mn ] And input it into the graph neural network layer that supports edge attributes for training and inference.

8. The multi-sensor coordinated partial discharge waveform recognition method according to claim 1, characterized in that: In step 7, the graph neural network uses a graph isomorphism network framework to model the graph structure. The graph isomorphism network framework uses the GINEConv layer as the graph convolution layer. The update rules of the GINEConv layer are as follows: in, is the embedding representation of the mth node in layer l; N(m) is the set of neighbors of node m; e mn is the edge attribute vector; φ (l) Is a learnable edge feature mapping function; MLP l is a shared multilayer perceptron; ∈ (l) is a trainable scaling factor.

9. The multi-sensor coordinated partial discharge waveform recognition method according to claim 1, characterized in that: The step 8 introduces the global average pooling operation into the output end of the graph neural network to obtain a graph-level embedding vector, and inputs it into the classifier to determine whether a partial discharge event occurs.

10. A multi-sensor coordinated partial discharge waveform recognition system, characterized in that: include: The signal conversion module is used to perform Hilbert transform on the collected signals and extract the envelope of the oscillation waveform; A feature extraction module is used to apply a Logistic function to the rising edge area in the envelope of each sensor for nonlinear fitting, and extract the Logistic function steepness and start time as edge features; An exponential decay function is applied to the falling edge region of the envelope of each sensor for nonlinear fitting, and the time constant is extracted as the energy dissipation feature. The data fusion module is used to build a graph neural network model and use the graph structure for training and reasoning to achieve the collaborative fusion of multi-source sensor information; The partial discharge diagnosis module is used to introduce the output of the graph neural network into the global average pooling operation to obtain a graph-level embedding vector, and input it into the classifier to determine whether a partial discharge event occurs.

Images