CN-121982097-A - Equipment fault source positioning method and device based on multi-mode information fusion

CN121982097ACN 121982097 ACN121982097 ACN 121982097ACN-121982097-A

Abstract

The application discloses a method and a device for positioning equipment fault sources based on multi-mode information fusion, wherein the method comprises the steps of collecting data by a distribution room inspection robot carrying an infrared thermal imager, a visible light camera and a self-adaptive lifting mechanism, preprocessing the collected infrared images, constructing a distribution room equipment data set, registering the images of the image pairs based on an improved whale optimization algorithm, fusing the registered images of the image pairs by a deep learning fusion network containing a coordinate attention module, obtaining fused images, and positioning fault sources based on a relative temperature difference and Monte Carlo method according to the fused images. The application realizes the accurate identification and grade judgment of the fault source through a four-stage process of multi-mode data acquisition, image accurate registration, high-quality image fusion and automatic fault location.

Inventors

LIU HANGYU
CUI JIACHEN
TIAN ZHILIN
SHI KEJIAN
TIAN YE
YANG LUYU
GU TAIYU
LI HAIFENG
Shen Zilun
Ji Xinzhe
WANG ZEBIN
SUN JIAZHENG
WANG ZIANG
LIU JINGLU
Li Menhua
XU XINGYANG
ZHANG JIAJIA
ZHANG XIMING
ZHU YIDONG
ZHANG ZHI
WANG JIN
MA SIYUAN
CUI JIA
ZHANG NING
Bian Gechen

Assignees

国网辽宁省电力有限公司电力科学研究院
国网辽宁省电力有限公司本溪供电公司
沈阳工业大学

Dates

Publication Date: 20260505
Application Date: 20251203

Claims (11)

1. The equipment fault source positioning method based on multi-mode information fusion is characterized by comprising the following steps of: carrying out data acquisition by adopting a distribution room inspection robot carrying an infrared thermal imager, a visible light camera and a self-adaptive lifting mechanism, preprocessing acquired infrared images, and constructing a distribution room equipment data set, wherein the distribution room equipment data set comprises a plurality of image pairs formed by infrared images and visible light images which are in one-to-one correspondence; performing multi-modal image registration on each image pair based on an improved whale optimization algorithm to obtain registered image pairs; a depth learning fusion network containing a coordinate attention module is adopted to fuse the registered image pairs, so that a fused image is obtained; and according to the fusion image, fault source positioning is carried out based on a relative temperature difference and a Monte Carlo method.
2. The method of claim 1, wherein the preprocessing of the acquired infrared image comprises: the infrared image progress is subjected to gray processing by adopting a temperature value mapping formula, wherein the temperature value mapping formula is as follows (1): (1); Wherein, the Is the coordinates in the infrared gray scale obtained after pretreatment Pixel value at T The temperature value of the corresponding position in the infrared thermal image is T min 、T max , which is the minimum temperature value and the maximum temperature value of the infrared thermal image respectively, and L is the number of gray levels.
3. The method of claim 2, wherein different gray mapping slopes and slight fault temperature differences of the vacuum circuit breaker are preset, a gray range corresponding to a first preset temperature difference range is compressed by formula (2), and a gray range corresponding to a second preset temperature difference range is expanded to enhance fault discrimination, wherein formula (2) is represented as follows: (2); Wherein, the For outputting the gray scale deviation enhancement result, representing the gray scale deviation characteristic value of the image after the piecewise function processing at the coordinates (x, y), For highlighting details or deviations of the image, representing gray-scale deviations of the image at coordinates (x, y) 、、 For the purpose of connecting The value range of (2) is divided into three sections and different enhancement strategies are adopted for different gray level deviation families; g1, G2, G3 provide the underlying gain contribution for each segment interval.
4. The method of claim 1, wherein the traditional whale optimization algorithm is modified in the following manner to yield a modified whale optimization algorithm: adopting Tent chaotic mapping to replace random initialization, and improving initial solution distribution uniformity; the existing linear decreasing convergence factor is replaced by a nonlinear convergence factor based on a Sigmoid function, so that the self-adaptive balance of global search and local exploration is realized; And a Gaussian variation mechanism of a firework algorithm is introduced to perform variation on the excellent solutions in the iterative process, so that premature convergence of the population is avoided.
5. The method of claim 4, wherein the Tent chaotic map formula is formula (3): (3); Wherein, the It can be understood as a "state value" at a certain moment in the system; I.e. the state of the system at the next moment; the nonlinear convergence factor formula is formula (4): (4); Wherein, the An adaptive transition from global exploration to local convergence is implemented for a nonlinear convergence factor, T is the current iteration number, and tmax is the maximum iteration number; wherein the dynamic convergence factor Formula (5): (5); the gaussian variation formula is formula (6): (6); Wherein, the Is a new individual after the disturbance and is a new individual after the disturbance, Representing a new solution generated after the mutation operation, The method is used for controlling the disturbance intensity and determining the influence degree of random disturbance on the original; Is the range of the search space, a and b are used to normalize the random perturbations to a reasonable search range, The algorithm can be explored in the solution space immediately.
6. The method of claim 1, wherein performing multi-modal image registration on each of the image pairs based on a modified whale optimization algorithm results in registered image pairs, comprising: Extracting edge features of the infrared gray level image and the visible light image by adopting a Canny edge detection operator respectively; taking normalized mutual information or structural similarity index as an objective function, and solving rigid affine transformation parameters between the infrared gray level image and the visible light image through the improved whale optimization algorithm; And carrying out space transformation on the infrared gray level image based on the rigid affine transformation parameters obtained by solving, so as to realize registration with the visible light image.
7. The method of claim 6, wherein the alignment is performed based on a modified SSIM, the modified SSIM being represented by formula (7): (7); Wherein, the And The dynamic parameters calculated for the gray scale range differences, Respectively registering pixel matrixes of the infrared image and the visible light image; Respectively, images A gray average value in the calculation region; Respectively, images Gray variance within the calculation region.
8. The method of claim 6, wherein the dynamic parameters of gray scale difference calculation And Is calculated according to the following method: For each feature map M, initial activity level graph thereof Is calculated by an application norm represented by equation (16): (16); Wherein, the An activity level graph of the ith feature, The calculation of LM norm is based on the result of local measurement (Local Measurement) of the ith feature at each pixel point of the feature map Quantifying the activity; smoothing activity level graphs using block-based averaging operations Obtaining a final activity level map as shown in formula (17): (17); Where r determines the size of the block, A smoothed activity level graph, r, a neighborhood radius, determines a smoothed range, The total number of pixels in the neighborhood; The fused feature map is calculated by weighting the sum of the feature maps, the weights being obtained by normalizing the activity level map for each location to the sum of the activity level maps for all feature maps, as in equation (18): (18); Wherein, the The final feature after fusion is a weighted combination of multi-source features, k is the total number of features participating in fusion, An mth class feature representation of the ith feature; (19); Wherein, the Is the weight after the normalization, Weights of ith feature, activity level map smoothed by the weight Normalizing to obtain the final product.
9. The method of claim 1, wherein the multi-modal image fusion network with coordinate attention comprises a feature extraction module, a feature fusion module, and an image reconstruction module; The characteristic extraction module comprises an input layer, a convolution-coordinate attention module and a dense connecting block which are sequentially connected, wherein the convolution-coordinate attention module is preferably composed of a 3X 3 convolution layer and a CAB module, the CAB module enhances the extraction capacity of infrared temperature characteristics and visible light texture characteristics through encoding space position information, and the dense connecting block comprises 3X 3 convolution layers, enhances the deep characteristic extraction capacity through characteristic multiplexing and reduces information loss; The feature fusion module comprises a texture feature screening block, a neighborhood flat sliding block and a feature fusion block which are connected in sequence, wherein the texture feature block screens the obvious information in the infrared temperature feature and the visible light texture feature based on a 1-norm fusion strategy; And the image reconstruction module gradually reduces the number of characteristic channels by adopting 4 attention blocks containing coordinates, and introduces a composite loss function constraint reconstruction process to ensure that the fused image has both temperature and texture information.
10. The method of claim 1, wherein fault source localization comprises: And carrying out probability fault judgment on the random samples, wherein the formula of each group of samples is as follows, and the formula of each group of samples is as follows: (31); Wherein: In order to be a relative temperature difference, In order to measure the temperature 1 of the material, In order to measure the temperature 2 of the wafer, The method comprises the steps of counting corresponding fault probability distribution at different temperatures for the environmental temperature; based on the temperature information of the fusion image, calculating a relative temperature difference value of equipment to be monitored by adopting a relative temperature method; and determining the fault level and the position based on a preset three-level fault threshold according to the relative temperature difference value.
11. The utility model provides a device fault source positioner based on multimodal information fuses which characterized in that includes: The system comprises an acquisition and preprocessing unit, a power distribution room equipment data set and a control unit, wherein the acquisition and preprocessing unit is used for acquiring data by adopting a power distribution room inspection robot carrying an infrared thermal imager, a visible light camera and a self-adaptive lifting mechanism, preprocessing acquired infrared images and constructing the power distribution room equipment data set, and the power distribution room equipment data set comprises a plurality of image pairs which are formed by infrared images and visible light images in one-to-one correspondence; The registration unit is used for carrying out multi-mode image registration on each image pair based on an improved whale optimization algorithm to obtain registered image pairs; The fusion unit is used for fusing the registered image pairs by adopting a deep learning fusion network containing a coordinate attention module to obtain a fused image; and the positioning unit is used for positioning a fault source based on the relative temperature difference and the Monte Carlo method according to the fusion image.

Description

Equipment fault source positioning method and device based on multi-mode information fusion Technical Field The application relates to the technical field of industrial equipment fault diagnosis, in particular to an equipment fault source positioning method and device based on multi-mode information fusion. Background Along with the development of science and technology and the upgrading of requirements in the fields of national defense, electric power and the like, the complexity, the comprehensiveness and the intelligent degree of large-scale systems such as complex industrial sites (such as distribution rooms), aerospace, electric power systems and the like are obviously improved, and the requirements on the stability and the reliability of the systems are also improved. The system has the characteristics of long research and development period, high technical content and huge value, and if faults occur in the operation process and the faults are not detected and treated in time, equipment damage, production interruption and even safety accidents are easily caused, so that the fault diagnosis technology (including fault detection, isolation, identification and positioning) becomes a core support for guaranteeing the safe operation of the system. Traditional fault diagnosis often relies on manual inspection or single sensor monitoring, has quite outstanding limitations, has higher manual dependency degree and lower efficiency, needs operation and maintenance personnel to acquire sensor data or perform manual inspection on site, is greatly influenced by subjective experience, has higher requirements on professional ability of personnel, and is limited in part of equipment due to structure or cost, only a small number of sensors are installed, even no sensors are installed, so that fault information acquisition is not timely and complete. The single-mode detection capability is limited, and the state of the device cannot be comprehensively estimated only by means of a single detection technology, such as visible light imaging and single sensor signals, for example, the visible light image can capture the details of the appearance of the device, such as the shape and the position of a component, but the temperature abnormality cannot be perceived, and the temperature abnormality is a core characteristic of early thermal failure, and the single sensor signals, such as vibration and current, are easily interfered by noise, so that the specific failure component is difficult to determine. The application of the nondestructive testing technology presents a trend, in order to solve the defects of the traditional method, the nondestructive testing technology such as X-ray, ultrasonic wave, photoacoustic, vortex, infrared thermal imaging and the like is rapidly developed, wherein the infrared thermal imaging technology is a core tool for detecting abnormal equipment temperature because of the characteristics of non-invasiveness, safety, rapidness, sensitivity and the like, and can identify abnormal areas according to equipment surface temperature distribution under the condition of not interrupting equipment operation, and capture early-stage thermal faults such as the conditions of loosening of a wiring terminal of a distribution equipment, local overheating caused by line overload and the like. The complementarity of the infrared image (reflecting temperature information) and the visible light image (reflecting detail information) makes the infrared image (reflecting temperature information) and the visible light image a key multi-mode data source for equipment fault diagnosis. However, because of imaging principle (infrared: thermal radiation; visible light: light reflection), shooting angle and sensor parameter difference, parallax exists (spatial position mismatch), and ghost and artifact can be generated if direct fusion is performed, image registration (multi-mode image spatial alignment is realized) is a precondition of subsequent fusion and fault positioning. The region-based method is to calculate image similarity (such as normalized cross correlation) through a sliding window to realize alignment, has the advantages of robustness to noise and deformation, has the advantages of high calculation complexity and high registration accuracy caused by local change, has the advantages of extracting significant features (such as corner points, edges and contours, such as SIFT and SURF algorithms) of the image and matching, has the advantages of quick calculation and adaptive rotation/scaling transformation, has the advantages of relying on feature extraction accuracy, large feature difference of a multi-mode image (such as infrared and visible light) and easy occurrence of matching error, and has the advantages of adapting to complex scenes and high registration accuracy by automatically learning registration features and transformation parameters through CNN and an end-to-end network based on deep learning, and