CN-122020443-A - Transformer substation intelligent inspection robot anomaly identification method and system based on multi-mode fusion and deep learning

CN122020443ACN 122020443 ACN122020443 ACN 122020443ACN-122020443-A

Abstract

The invention belongs to the technical field of intelligent inspection and anomaly identification of substation equipment, and particularly relates to a substation intelligent inspection robot anomaly identification method and system based on multi-mode fusion and deep learning. The method aims to solve the problems of single-mode limitation, poor working condition suitability and the like in abnormal identification of substation equipment. The method comprises the steps of synchronously collecting multi-mode time sequence data through a patrol robot, embedding GPS positions and unique IDs of equipment to complete scene marks, extracting depth features of all modes after preprocessing and time alignment, outputting self-adaptive weights through a dynamic weight fusion network by combining working condition parameters to generate combined multi-mode features, synchronously capturing space static features and time dynamic features through a space-time double-flow depth learning network, identifying abnormal states and positioning abnormal positions. The system comprises a core module for data acquisition, preprocessing, feature extraction and the like, supports online model updating, adapts to complex dynamic working conditions of the transformer substation, and improves inspection efficiency and anomaly identification accuracy.

Inventors

GONG ZHENZHOU
HUANG XIN
ZHU JIANWU
ZHENG CHENQUAN
YU ZHONGSHU
WANG JIAXIN
WANG WEIZHANG
PENG MINGCONG

Assignees

国网江西省电力有限公司南昌供电分公司

Dates

Publication Date: 20260512
Application Date: 20251212

Claims (10)

1. The substation intelligent inspection robot anomaly identification method based on multi-mode fusion and deep learning is characterized by comprising the following steps of: S1, synchronously acquiring multi-mode time sequence data of target equipment of a transformer substation by a patrol robot, wherein the multi-mode time sequence data comprises visible light images, infrared images, audio signals and partial discharge signals, and embedding real-time GPS position information and unique ID of the target equipment into corresponding time sequence data to complete data scene marking; S2, preprocessing and time alignment are carried out on the multi-mode time sequence data, wherein the preprocessing comprises image enhancement, signal noise reduction and format conversion, and depth feature vectors of all modes are extracted respectively; S3, inputting the depth feature vector of each mode and current working condition parameters acquired in real time into a dynamic weight fusion network, outputting adaptive fusion weights of each mode adapting to the current working condition by the dynamic weight fusion network, and generating joint multi-mode features after weighting and fusing the depth feature vector by the weights; And S4, based on the combined multi-mode characteristics, combining the embedded GPS position information in the S1 with the unique ID of the target equipment, identifying the abnormal state, type and abnormal position of the target equipment through a deep learning network, and outputting an identification result containing confidence.
2. The substation intelligent patrol robot anomaly identification method based on multi-mode fusion and deep learning of claim 1 is characterized in that the preprocessing in step S2 specifically comprises the following steps: Image enhancement, namely adopting a limiting contrast self-adaptive histogram equalization algorithm to a visible light image, setting parameters to be 8 multiplied by 8 pixels in block size and a contrast limiting factor of 2.0, adopting a temperature linear stretching algorithm to an infrared image, and adopting an original temperature value Mapping to [0,255] gray space; The signal noise reduction is that the audio signal adopts a wavelet threshold noise reduction method, the wavelet base adopts db4, the decomposition layer number is 5, and the threshold function is a soft threshold function: wherein , Is the standard deviation of the noise, which is the standard deviation of the noise, The partial discharge signal adopts moving average filtering, and the window size is 50 sampling points; Format conversion, namely uniformly converting the preprocessed visible light/infrared image into RGB three-channel tensor, converting the audio signal into a Mel spectrogram, converting the partial discharge signal into a time-frequency chart, and uniformly converting all mode data into a time sequence tensor format , For the length of the time sequence, Is a single frame feature dimension.
3. The substation intelligent patrol robot anomaly identification method based on multi-mode fusion and deep learning of claim 1 is characterized in that the deep feature vector extraction in step S2 specifically comprises the following steps: The visible light image feature extraction adopts ResNet-50 network, removes the last full-connection layer, retains convolution basis output and feature vector dimension ; The infrared image feature extraction adopts a lightweight MobileNetV network, and feature vector dimensions ; The audio signal feature extraction adopts a pre-training VGGish network and outputs feature vector dimension ; The partial discharge signal feature extraction adopts a one-dimensional convolutional neural network, and feature vector dimension is output ; Each modal depth feature vector is defined as Wherein Is the first Single frame feature vector of time instant.
4. The substation intelligent patrol robot anomaly identification method based on multi-mode fusion and deep learning of claim 1 is characterized in that the construction of the dynamic weight fusion network in the step S3 comprises the following steps: The network input is depth characteristic vector of each mode Sensing parameters of current working condition Wherein Including environmental parameters, temperature Humidity of Load factor for the operating parameters of the plant Voltage of 4-Dimensional vector; The network structure is divided into two layers: The characteristic splicing layer splices the mode characteristics into joint input according to time steps ; Attention weight calculating layer, which adopts multi-head self-attention mechanism and head number The attention score of each modality is calculated : Wherein For input of A query, a key, a value matrix after linear transformation, Is a unimodal feature dimension; outputting the adaptive fusion weight of each mode , And satisfy the following 。
5. The substation intelligent patrol robot anomaly identification method based on multi-mode fusion and deep learning as claimed in claim 1, wherein the weighted fusion in step S3 generates joint multi-mode characteristics The calculation formula of (2) is as follows: Wherein the method comprises the steps of Is a time-averaged vector of the unimodal feature, Modality for fusing network output for dynamic weights Is a weight of (2).
6. The substation intelligent patrol robot anomaly identification method based on multi-mode fusion and deep learning of claim 1 is characterized in that in step S4, the deep learning network adopts a space-time double-flow network architecture and comprises the following steps: Spatial branching by joint multi-modal features Spatial feature extraction for input through a 2-layer fully connected network ; Time branching by original time sequence characteristics of each mode Extracting time-dependent features for input via a bi-directional LSTM network ; Fusion classification layer to be And (3) with After splicing, inputting the full-connection layer, and activating and outputting abnormal probability distribution through Softmax, wherein the abnormal state comprises normal state Superheated and heated Loosening of machinery Partial discharge Abnormal sound Class 5; The network loss function employs cross entropy loss with weights: Wherein the method comprises the steps of In order to obtain the number of samples, For the sample Is a real tag of the (c) in the (c), In order to predict the probability of a probability, For the class weight to be a class weight, , The difference between the number of normal samples and the number of abnormal samples is balanced.
7. The substation intelligent patrol robot anomaly identification method based on multi-mode fusion and deep learning according to claim 1 is characterized in that the anomaly position determination method in step S4 is as follows: For visible/infrared image features 、 Generating a thermodynamic diagram 、 , Is the image height and width; For audio/partial discharge signal features 、 Locating abnormal energy concentration periods by time-frequency analysis 、 ; Combining GPS location information: unique ID with target device: corresponding preset part coordinate set Mapping the thermodynamic diagram peak coordinates or the spatial positions corresponding to the abnormal time periods to the equipment component, and outputting the abnormal positions.
8. The substation intelligent patrol robot anomaly identification method based on multi-mode fusion and deep learning of claim 1 is characterized by further comprising online updating of the deep learning network: defining network performance assessment metrics , For a time period of The number of samples in the sample set is, In order to indicate the function, Is a predictive tag; When (when) Triggering on-line update and reference accuracy Allow the threshold to drop ; The updating function adopts an elastic weight integration algorithm, and the loss function adds a regular term: Wherein the method comprises the steps of In order for the cross-entropy loss to occur, For the regularization coefficient(s), As a matrix of importance of the parameters, 、 For current and old network parameters, the network is finely tuned by the newly collected labeling sample, and the learning rate is set as 。
9. The substation intelligent inspection robot abnormality recognition system based on multi-mode fusion and deep learning according to claim 1 is characterized by comprising a data acquisition module, a preprocessing and feature extraction module, a dynamic weight fusion module and an abnormality recognition and positioning module, wherein the modules are sequentially in communication connection and have the following specific functions: The data acquisition module is mounted on the inspection robot and used for synchronously acquiring visible light images, infrared images, audio signals and partial discharge signals of the substation target equipment, acquiring real-time GPS position information, embedding the GPS position information and the unique ID of the target equipment into corresponding time sequence data and finishing data scenerization marking; the preprocessing and feature extraction module is used for receiving the multi-mode time sequence data output by the data acquisition module, executing preprocessing operation firstly, and then respectively extracting depth feature vectors of all modes through a preset network; The dynamic weight fusion module is used for receiving the depth feature vectors of all modes output by the preprocessing and feature extraction module, acquiring current working condition parameters at the same time, outputting self-adaptive fusion weights of all modes through a built-in dynamic weight fusion network, and carrying out weighted fusion on the time average feature vectors of all modes based on the weights to generate joint multi-mode features; The anomaly identification and positioning module is internally provided with a space-time double-flow deep learning network, receives the combined multi-mode characteristics, the original time sequence characteristics of each mode, the embedded GPS position information and the unique ID of the target equipment, extracts the spatial characteristics through a spatial branch full-connection network, extracts the time dependent characteristics through a time branch bidirectional LSTM network, outputs the anomaly state, the anomaly type and the corresponding confidence coefficient of the target equipment through a fusion classification layer, and simultaneously combines the GPS position and the equipment preset part coordinate set through image thermodynamic diagram generation and signal time-frequency analysis to position the anomaly specific position.
10. The substation intelligent patrol robot anomaly identification system based on multi-mode fusion and deep learning according to claim 1 is characterized by further comprising a data storage module, a communication module, a visual interaction module and a model online updating module, and has the following specific functions: The data storage module is in communication connection with the data acquisition module and the anomaly identification and positioning module and is used for storing the multi-mode time sequence data, the preprocessing result, the depth feature vector, the combined multi-mode feature and the anomaly identification result and the position information containing the confidence coefficient after the scene marking; The communication module is used for realizing data transmission among the data acquisition module, the preprocessing and feature extraction module, the dynamic weight fusion module and the abnormality identification and positioning module by adopting a wireless communication technology, uploading an abnormality identification result to a transformer substation monitoring center in real time and receiving a control instruction of the monitoring center; the visual interaction module is arranged in the monitoring center and used for visually displaying the multi-mode acquisition data, the abnormal state, the abnormal type, the abnormal position and the confidence coefficient of the target equipment and supporting the historical data inquiry and the abnormal record tracing; And the model online updating module is in communication connection with the anomaly identification and positioning module and the data storage module, calculates the operation performance index of the space-time double-flow deep learning network in real time, and triggers online updating when the index is lower than the reference accuracy and the descending amplitude exceeds 0.05.

Description

Transformer substation intelligent inspection robot anomaly identification method and system based on multi-mode fusion and deep learning Technical Field The invention belongs to the technical field of intelligent inspection and anomaly identification of substation equipment, and particularly relates to a substation intelligent inspection robot anomaly identification method and system based on multi-mode fusion and deep learning. Background The transformer substation is used as a core hub of the power system transmission, transmission and distribution links, and the running state of equipment directly determines the safety and stability of power supply. On one hand, the manual inspection relies on experience of operation and maintenance personnel, labor intensity is high, inspection efficiency is low, inspection omission and false inspection are easy to occur under the influence of severe environments such as high temperature, heavy rain, night and artificial subjective factors, early potential abnormality of equipment cannot be found in time, on the other hand, the abnormal types of transformer substation equipment are various (such as overheat, mechanical looseness, partial discharge, abnormal sound and the like), different abnormal characteristic performances are scattered in multiple dimensions such as appearance, temperature, sound, electric signals and the like, and the manual inspection is difficult to comprehensively capture multidimensional abnormal information. The intelligent inspection robot is gradually applied to a transformer substation scene to solve the limitation of manual inspection, but the prior art still has obvious short plates, namely, a plurality of intelligent inspection systems rely on single-mode data to perform abnormal recognition, the single-mode data can only capture local information, the whole coverage of different types of anomalies cannot be realized, the integrity of the abnormal recognition is insufficient, a part of multi-mode fusion scheme adopts a fixed weight fusion strategy, the dynamic change characteristic of transformer substation working conditions is ignored, the fixed weight cannot be adapted to complex working conditions, the discrimination of fusion characteristics is insufficient, the existing abnormal recognition model is characterized in that the single-space characteristics or time characteristics are captured in multiple sides, the static state (such as appearance damage and temperature distribution) and the dynamic change trend (such as continuous temperature rise and periodic occurrence of discharge signals) of equipment are difficult to consider, the recognition accuracy is low, and meanwhile, the abnormal positioning is limited to equipment-level positioning and cannot be mapped to specific parts accurately, so that quick handling of operation and maintenance personnel is not facilitated. In summary, the existing intelligent inspection technology for the transformer substation has obvious defects in aspects of multi-mode information fusion suitability, abnormal feature capture comprehensiveness, identification positioning accuracy, model dynamic adaptability and the like, and a technical scheme capable of integrating multi-mode complementary information, adapting dynamic working conditions, accurately identifying abnormal states and positions and having continuous optimization capability is needed to improve the intelligent level and operation and maintenance reliability of inspection for the transformer substation. Disclosure of Invention The method aims to solve the problems that single-mode data is not fully covered, fixed weights cannot be adapted to dynamic working conditions, abnormal characteristics are not fully captured, positioning accuracy is insufficient, a model lacks self-adaptive updating capability and the like in the existing intelligent substation inspection technology, and reliable intelligent technical support is provided for safe and stable operation of substation equipment by synchronously acquiring multi-mode time sequence data, embedding scene information, optimizing preprocessing and depth characteristic extraction processes, constructing a working condition-aware dynamic weight fusion network, adopting a space-time double-flow architecture to jointly capture static and dynamic characteristics of equipment, combining multidimensional information to realize abnormal accurate positioning and online model updating, comprehensively improving the accuracy, robustness and scene suitability of abnormal identification, reducing missed inspection, simplifying inspection and maintenance processes, improving inspection efficiency. In order to achieve the above object, the present invention adopts the following technical scheme: the substation intelligent inspection robot anomaly identification method based on multi-mode fusion and deep learning comprises the following steps: S1, synchronously acquiring multi-mode time sequence data of