CN-121980382-A - Power equipment operation data anomaly detection method

CN121980382ACN 121980382 ACN121980382 ACN 121980382ACN-121980382-A

Abstract

The invention provides an abnormal detection method of power equipment operation data, which belongs to the technical field of power operation detection based on computer data processing, and comprises the steps of collecting multi-source data in an operation state of the power equipment, preprocessing, inputting a designed depth feature extraction module based on multi-physical field signal characteristic analysis, designing a depth feature extraction network facing different signal characteristics, automatically focusing abnormal fragments from an original waveform, mining space-time evolution rules from a sensor network with spatial distribution, decoupling sparse discharge mode primitives with definite physical meaning from partial discharge pulses, weighting and splicing the extracted features to form global state characterization, inputting a mixed supervision variation self-encoder with training completion, and outputting class probability and equipment state labels. The invention realizes the depth fusion of equipment panoramic state sensing and multi-source information, improves the depth representation and discrimination capability of complex abnormal modes, and realizes dynamic self-adaptive abnormal judgment and fault primary screening diagnosis.

Inventors

ZHU JINGLIAN
ZHANG HONGQIAO
LU HAO
LIU YALI
WEI LIANGCHAO
WU BIN
WANG XINGMIN
WANG LIXUE

Assignees

保定华创电气有限公司

Dates

Publication Date: 20260505
Application Date: 20260119

Claims (9)

1. The method for detecting the abnormal operation data of the power equipment is characterized by comprising the following steps of: S1, acquiring three-phase current time sequence data, three-phase voltage time sequence data, temperature time sequence data, vibration sound wave time sequence data and partial discharge pulse sequence data under the running state of power equipment; S2, preprocessing the acquired data, and inputting the data into a depth feature extraction module based on multi-physical field signal characteristic analysis; The power equipment is modeled into a graph structure, the temperature time sequence data and the vibration sound wave time sequence data are used as node characteristics, and the physical field space-time correlation characteristic vector is extracted through a graph space-time synchronous coding network; And S3, carrying out weighted splicing on the electric quantity depth feature vector, the physical field space-time correlation feature vector and the discharge mode feature vector to form a global state representation, inputting a trained hybrid supervision variation self-encoder, and outputting category probability and equipment state labels.
2. The method for detecting abnormal operation data of power equipment according to claim 1, further comprising the steps of detecting abnormal dynamic threshold, specifically: for an input device state sample, extracting a new global state representation of the input device state sample And input the encoder-decoder part of the hybrid supervisory variation self-encoder to calculate a new global state representation Reconstruction errors of (a) ; Introducing a conditional normalization flow model, wherein the model adopts a flow structure based on affine coupling layers, and is formed by stacking 8 coupling layers, and each coupling layer inputs vectors Split into two parts Through a conditional information A sub-network for input, wherein the condition information The statistics of the equipment working condition label are obtained by three-layer MLP coding to generate a scaling factor Translation factor For the purpose of , Keep unchanged, let For the following , First with a scaling factor Element-wise multiplication of the exponents is performed, and then a translation factor is added Obtaining , For the first output vector to be present, Is the second output vector And Splicing to obtain output This output serves as an input to the next coupling layer; In online detection, given current condition information Calculating new device state samples in current condition information New sample error of (2) And set a dynamic threshold 99% Quantile, when And triggering an abnormal alarm, outputting a device state label corresponding to the abnormality through the multiple abnormality classification heads, and otherwise, judging that the device state label is normal.
3. A method for detecting abnormal operation data of power equipment according to claim 2, wherein the equipment working condition label is divided according to the percentage of real-time active power and rated power of the equipment, wherein less than 10% is defined as no load, 10% to 50% is defined as light load, 50% to 100% is defined as rated load, and more than 100% is defined as overload.
4. The method for detecting the operation data abnormality of the electric power equipment according to claim 1, wherein the equipment state label is a two-stage system, the first stage is normal or abnormal, the second stage is subdivided into five types of overheat abnormality, partial discharge abnormality, mechanical loosening abnormality, three-phase unbalance abnormality and harmonic abnormality under the condition of abnormality, and specific labeling rules are as follows: When the temperature of any monitoring point in the temperature time sequence data continuously exceeds the upper limit of the safe operation of the equipment and unbalance exists between the three-phase current time sequence data and the three-phase voltage time sequence data, the temperature time sequence data is marked as overheat abnormal, when a phase distribution spectrogram of the partial discharge pulse sequence data shows typical internal discharge or creeping discharge mode characteristics, the temperature time sequence data is marked as partial discharge abnormal, when an abnormal peak value occurs at the mechanical natural frequency of the frequency spectrum of the vibration sound wave time sequence data and is irrelevant to load change, the temperature time sequence data is marked as mechanical loosening abnormal, when negative sequence components of the three-phase current time sequence data and the three-phase voltage time sequence data exceed a set threshold value, the temperature time sequence data is marked as three-phase unbalance abnormal, when the total harmonic distortion rate of the three-phase current time sequence data and the three-phase voltage time sequence data exceeds the set threshold value, the temperature time sequence data is marked as harmonic abnormal, and the data which do not accord with any abnormal rules and are in a stable operation period are marked as normal.
5. The method for detecting abnormal operation data of the power equipment according to claim 1, wherein the depth separable convolution attention network in S2 comprises a convolution-attention block stacking structure and a global average pooling layer, wherein the network body is formed by stacking four convolution-attention blocks, and each block comprises a depth separable convolution layer, a channel-time double-attention module, a batch normalization layer and GeLU activation functions; The convolution-attention block adopts a one-dimensional depth separable convolution layer to replace standard convolution, the channel-time dual-attention module comprises a parallel channel attention sub-module and a time attention sub-module, wherein the channel attention sub-module generates weight of each channel through global average pooling and a full connection layer, the time attention sub-module learns time step weight through one-dimensional convolution and focuses on key characteristics of three-phase current waveform transient mutation and three-phase voltage transient mutation, and three-phase current time sequence characteristics and three-phase voltage time sequence characteristics output by the last convolution-attention block can sensitively represent deep waveform characteristic representation of various electrical anomalies through global average pooling layer output Wherein And the depth characteristic vector of the electrical quantity extracted from the three-phase current time sequence data and the three-phase voltage time sequence data is represented.
6. The method for detecting abnormal operation data of power equipment according to claim 1, wherein in S2, the power equipment is modeled as a graph structure, specifically: According to the physical structure topology of the power equipment and the distribution position of the sensor, each monitoring point is defined as a graph node, the node is characterized by preprocessed temperature time sequence data and vibration sound wave time sequence data acquired by the monitoring point, the connecting edge weight among the nodes is determined by the prior relation between the physical distance and heat conduction/vibration transmission, the weight integrates the physical distance among the monitoring points and the coupling attenuation effect in the two physical processes of heat conduction and vibration transmission, the correlation strength of the heat conduction is mainly determined by the heat conduction performance of the component materials where the node is located, the correlation strength of the vibration transmission is determined by the rigidity characteristic of the mechanical connecting structure among the components, and the final edge weight is a weighted comprehensive measurement of the distance attenuation effect, the heat conduction characteristic and the structural rigidity characteristic.
7. The method for detecting abnormal operation data of power equipment according to claim 6, wherein the space-time synchronous coding network of the diagram in the step S2 comprises a diagram convolution coding layer, a time sequence causal convolution layer and a diagram aggregation layer, and is used for extracting the space dependence and time evolution characteristics among signals of multiple monitoring points; Firstly, inputting the characteristics of all nodes in a spatial relationship graph into a graph convolution coding layer, aggregating neighbor node information on a graph structure by using a two-layer graph convolution network, learning the spatial dependency relationship among the nodes, and outputting the temperature-vibration acoustic wave spatial enhancement characteristics of each node Wherein temperature-vibration acoustic wave space enhancement features Representing node characteristics which contain space association information after graph convolution coding; secondly, the temperature-vibration sound wave space enhancement characteristic of each node is input into a time sequence causal convolution layer at the same time, and the temperature-vibration sound wave space enhancement characteristic of each node is obtained Splicing the time series data with the original standardized time series data, taking the time series data as the time series input of the node, adopting a time convolution network with an expansion causal convolution structure by the layer, capturing the long-time series causal dependency relationship by overlapping multiple layers of expansion convolutions, and outputting the temperature-vibration sound wave space-time fusion characteristic of each node Wherein temperature-vibration acoustic wave spatiotemporal fusion features Representing node characteristics fused with space association and long-time evolution information; finally, the temperature-vibration sound wave space-time fusion characteristics of all nodes are obtained The characteristics are aggregated through the pooling operation to obtain the comprehensive characteristic representation for representing the multi-physical-field space-time evolution mode of the whole equipment Wherein And representing the physical field time-space correlation characteristic vector extracted from the temperature time sequence data and the vibration sound wave time sequence data.
8. The method for detecting abnormal operation data of power equipment according to claim 1, wherein the self-encoder of the structured sparse discharge mode in S2 comprises the following specific steps: preprocessing partial discharge pulse sequence data into a phase-amplitude-frequency statistical map for representing discharge statistical characteristics Wherein The dimension of the representation map is input into a structured sparse discharge mode self-encoder, and the encoder part consists of two fully-connected layers, namely a first layer to be input into Mapping to 256-dimensional hidden space and using ReLU activation function, mapping hidden features to 32-dimensional potential space by second layer, outputting potential characterization vector of discharge pulse Potential characterization vector of discharge pulse Pre-divided into Groups, each group comprising 4 consecutive features, introducing a group sparse regularization term in the total loss function of the network training The group sparse regularization term forces the network to learn a sparse representation, i.e. only a few feature groups are activated for reconstructing the input, each active group corresponds to a physically well defined discharge pattern primitive, the decoder section is symmetrical to the encoder, also consists of two fully connected layers, responsible for potential characterization of vectors from the discharge pulses Reconstructing an input map ; Minimizing reconstruction losses from an encoder by partial discharge pulses Group sparsity loss Training the sum, and after training, outputting potential characterization vectors of discharge pulse by the encoder I.e. a desired sparse discharge characteristic representation with a well-defined physical meaning Wherein Representing low-dimensional discharge pattern feature vectors with a group sparse structure extracted from partial discharge pulse train data.
9. The method for detecting abnormal operation data of power equipment according to claim 1, wherein the hybrid supervisory variation self-encoder comprises an encoder, a decoder and a classification head, wherein the encoder is a three-layer full-connection network, and the input is a global state representation Sequentially passing through two hidden layers with dimensions of 256 and 128, and generating a mean vector of potential spatial distribution by a final output layer And logarithmic variance vector The decoder is a symmetrical structure of the encoder, and the input is distributed from potential space The potential variable of the middle sampling is output to reconstruct the characteristic through two hidden layers with dimensions of 128 and 256 The classifying head consists of two fully connected layers, and the input of the first layer is the average value vector of the output of the encoder The dimension is 64, the second layer output dimension is the generation category probability of each device state label, and the device state label is finally output by the Softmax activation function.

Description

Power equipment operation data anomaly detection method Technical Field The invention belongs to the technical field of electric power operation detection based on computer data processing, and particularly relates to an electric power equipment operation data abnormality detection method. Background The electric power equipment is a foundation stone for safe and stable operation of the power grid, and the health condition of the electric power equipment is directly related to the reliability of power supply and the order of social production and life. With the continuous expansion of the power grid scale and the continuous improvement of the intelligent level, the structure of the power equipment is increasingly complex, the running environment is changeable, and the traditional mode of depending on regular inspection and post-maintenance is difficult to meet the requirements of early warning and accurate diagnosis of potential faults. Hidden defects such as partial discharge, overheating, mechanical looseness and the like in equipment can gradually evolve into serious electrical or mechanical faults if the hidden defects cannot be found in time, even chain reactions are initiated, and large-scale power failure and economic loss are caused. The new generation of information technology represented by the Internet of things and artificial intelligence is deeply changing the operation and maintenance management mode of industrial equipment. The core of the research and development of the advanced power equipment abnormality detection method is to deeply fuse multi-source sensor data distributed over key nodes of equipment with intelligent algorithms such as deep learning and the like, so as to realize real-time insight and intelligent judgment on the running state of the equipment. The method can automatically extract depth features representing the health state of equipment from massive signals such as current, voltage, temperature, vibration and the like, and learn complex boundaries between normal modes and various abnormal modes. The intelligent model transformation method not only can realize early capture and preliminary classification of the latent faults, provides accurate decision support for operation and maintenance personnel, but also can promote intelligent model transformation of operation and maintenance of the power equipment. The existing power data anomaly detection method comprises the following steps: The invention patent of China with the application number of CN202510923942. X discloses a power equipment data anomaly detection method and system based on LSTM-COF, and relates to the technical field of power equipment state monitoring, wherein the method comprises the following steps of collecting historical data and constructing a three-dimensional data matrix; predicting equipment parameters at extreme temperature through an LSTM model, compressing features, quantifying covariance deviation among the features by combining a correlation anomaly factor algorithm, detecting anomaly points, outputting a visual result, and integrating data collection, data preprocessing, data prediction, detection model generation and visual modules by a system. The Chinese patent application No. CN202111575119.8 is characterized in that electric power operation data of each measuring point of each electric power equipment in an automatic master station system to be measured are collected, then the data are cleaned to obtain effective offline data samples, then the effective offline data samples are subjected to dimension reduction, a time sequence sample sequence is obtained through calculation and is input into an improved cyclic neural network for training, an electric power data abnormality detection model is obtained through training, abnormal data are detected through the trained electric power data abnormality detection model, abnormal electric power data are clustered through an improved clustering algorithm, finally an abnormal alarm interval is set through an adaptive setting method, and an alarm is given when the abnormal data exceed the abnormal alarm interval. However, the above-described method and the prior art have the following limitations: 1) The data mode is single, and the information of multiple physical fields cannot be fully utilized. The existing method focuses on analyzing electric quantity or single type data, and fails to effectively synchronize and fuse multiple physical field monitoring signals such as temperature, vibration, sound wave, partial discharge and the like, so that the existing method is insensitive to complicated electromechanical thermal coupling faults in equipment, and the health state of the equipment is difficult to comprehensively describe; 2) The features represent shallow layers and have limited capability of characterizing complex anomaly modes. In the prior art, the characteristic dimension reduction is carried out by a traditional signal processing or simpl