KR20230109713A - Smart inspection method and system for substation facilities - Google Patents

Abstract

The present invention relates to a smart detection method and system for substation equipment, wherein the method extracts temperature matrix data based on an infrared thermal image, multi-values processes the temperature matrix, extrudes the temperature matrix, and generates an artificial intelligence model. It includes building and performing smart inspection of substation equipment based on the model. The present invention reduces the number of calculations for the unique characteristics of infrared thermal images of substation facilities in the process of data preprocessing, feature extraction, and model creation, increases the complexity of extraction in model feature extraction, and increases the pooling sensitivity in the pooling process. The amount of training and computation was reduced while maintaining the loss of sensitivity due to the reduction and finally not losing the model accuracy.

Description

Smart inspection method and system for substation facilities

The present invention belongs to the technical field of smart power systems, and particularly relates to a smart detection method and system for substation equipment.

Current power equipment detection methods are mainly based on infrared photographs, which artificially identify the types of equipment photographed by the naked eye observation method. It is more difficult to identify artificially.

Overall, existing methods have problems such as slow speed and low detection accuracy. The present invention reduces the number of unique features of thermal images of substation facilities in the process of data preprocessing, feature extraction, and model creation, and extracts By increasing the complexity and increasing the pooling sensitivity during the pooling process, the loss of sensitivity due to dimensionality reduction was maintained, and the amount of training and computation was reduced without losing model accuracy.

In order to solve the above-mentioned problems of the prior art, the present invention proposes a smart detection method and system for substation equipment, and the smart detection method for substation equipment extracts temperature matrix data based on an infrared thermal image S1; Step S2 of multi-value processing of the temperature matrix; Step S3 of protruding the temperature matrix; and a step S4 of constructing an artificial intelligence model and performing smart detection of substation equipment based on the model.

Further, the step S4 is specifically: building an artificial intelligence model, the input of the artificial intelligence model is the preprocessed and projected temperature matrix, and the output is the location and type of power equipment.

Further, the power equipment includes lightning arresters, circuit breakers, current transformers, bushings, voltage transformers, GIS bushings, insulation switches, insulators, clamps, transformers, capacitors, reactors, ceiling bushings, power cables and oil conservators.

Further, a sample data set is formed by selecting 1000 temperature data for each facility.

Further, in the process of training the artificial intelligence model, the learning rate was set to , and after 200 cycles, the learning rate is set to lower.

Further, a total of 2500 rounds of vaginal belt training were performed, and the model's mAP = 89.98% after 1200 steps of training per vaginal belt.

In the substation equipment smart detection system, it includes a server and one or more client terminals, the client terminals take images and upload the captured images to the server to obtain detection results, and the server executes the substation equipment smart detection method. do.

Further, the server is a cloud server.

A substation facility smart detection device comprising: a storage unit for storing an application; and a processing unit electrically coupled to the input unit and the storage unit, and configured to execute the smart detection method of substation equipment.

The storage medium used for smart detection of substation equipment is characterized in that it is used to store commands of a method for smart detection of substation equipment.

The beneficial effects of the present invention are specifically as follows: (1) A matrix can contain as much valid data as possible through position rotation and adjustment without losing calculation accuracy while effectively reducing the amount of calculation through multi-value processing, ( 2) Through preprocessing of multiple values of the matrix and protrusion processing and feature extraction based on matrix element values, the computational requirements for the prediction model are reduced without reducing the accuracy and recall rate at the same time, (3) By dynamically setting the filtering area, , By distinguishing the size of the sub-region and the small-size pooling region, the sensitivity of the model can be adjusted according to the detection result, and the balance between the validity of the model and the amount of available resources can be balanced, (4) AI model and the previous multi-value processing. With the combination of extrusion and extrusion processing, the sensitivity and accuracy of the model can be improved by extracting more image features by reducing the amount of subsequent computation and training based on extrusion processing and setting several convolution kernels in the model.

1 is a schematic diagram of a smart detection method for substation equipment of the present invention.
Figure 2 is a schematic diagram of the operation mode of the artificial intelligence model for substation equipment identification of the present invention.

In the following, the present invention will be described in detail through the accompanying drawings and specific examples, wherein the schematic examples and descriptions are only for explaining the present invention and do not limit the present invention.

In the present invention, instead of taking an infrared thermal image as an input, the infrared thermal image temperature matrix of the equipment to be inspected collected by the equipment is used as an input for smart detection, where each element value of the matrix is the actual temperature of the corresponding pixel point of the infrared thermal image. represents; After preprocessing the temperature matrix data, feature extraction is performed first, and information on the location and type of power facilities is obtained from the infrared thermal image through an artificial intelligence detection model. It is determined whether the power facility of the corresponding type has failed due to excessive temperature.

Preferably, the power equipment includes lightning arresters, circuit breakers, current transformers, bushings, voltage transformers, GIS bushings, insulation switches, insulators, clamps, transformers, capacitors, reactors, ceiling bushings, power cables and oil conservators.

Preferably, a data set is formed by selecting 1000 temperature data for each type of facility;

Alternatively, the artificial intelligence model is a Faster R-CNN model using Resnet50 as a backbone feature extraction network.

The smart detection method for substation equipment described in the present invention specifically includes the following steps.

Step S1: Extract temperature matrix data based on the infrared thermal image; Specifically, an infrared thermal image of the equipment to be inspected is obtained, and each element value of the temperature matrix represents an actual temperature value of a corresponding pixel in the infrared thermal image;

Step S2: multivalue processing of temperature matrix; Specifically, one or several multi-value intervals are acquired, the element value of each element of the temperature matrix is compared with the multi-value interval, and the element value is set to a corresponding fixed value of the multi-value interval in which it is located, where: The multi-value interval of corresponds to one fixed value, and the larger the value of the multi-value interval, the larger the fixed value.

Normal preprocessing removes inconsistent data, and does not start from the computational load and model characteristics, but in practice, to identify this particular type of power facility, successive temperature matrix elements cause unnecessary massive redundant computation of artificial intelligence, and the computational complexity of this computational quantity The increase does not increase the calculation accuracy, and the existing binarization process loses a lot of information, so multi-value processing effectively reduces the amount of calculation and at the same time does not lose calculation accuracy, and through rotation and adjustment of the position, the matrix can take the valid data one step further. Since the type of substation equipment is limited, the form of the image shown is also relatively limited, so the present invention reduces the amount of computation required for the predictive model without reducing the accuracy and recall rate through pre-processing before detection.

Step S3: Protrusion processing for the temperature matrix; Specifically, determine the adjustment reference position of the temperature matrix, adjust the temperature matrix based on the adjustment reference position, make the target object more prominent through such adjustment, in other words, as much valid data as possible in the matrix through adjustment is included, and since the type of substation equipment is limited, the form of the image shown is also relatively limited. Therefore, the present invention, through multi-value preprocessing of the matrix and extraction of salient features based on matrix element values, if accuracy and recall rate are not reduced At the same time, it reduces the amount of computation required for the predictive model.

Determination of the reference position for adjusting the temperature matrix includes the following steps in detail. Step A1: Determine the central position of the temperature matrix, wherein the central position of the temperature matrix is a specific position of the temperature matrix selected according to the size of the temperature matrix, and there are one or more central positions.

Preferably, the temperature matrix is divided into a plurality of identical or different sub-regions, and a central position of the sub-regions is selected as a central position of the temperature matrix, and the number of the sub-regions is directly proportional to the size of the temperature matrix.

Step A2: obtain the central position of the temperature matrix, determine one or more feature lines containing the central position of the matrix, the feature line composed of one or more temperature matrix elements including the element at the central position; ;

The crystal includes one or several feature lines at the central position of the matrix, specifically, the plurality of feature lines are divided and installed in a specific way, preferably, when there are several feature lines, the plurality of feature lines spaced at a fixed angle, for example, the sub-region is a 3*3 matrix; When is the center position, the feature line at the horizontal position is , and the feature line at a position outside the 45 degree angle is , and three feature lines can be set in the corresponding sub area.

Preferably, the feature line is a feature line that is symmetrical/relatively symmetrical with respect to the center position.

Step A3: Acquire one feature line and matrix element sequentially and process in step A4.

Step A4: judge whether the matrix element of the feature line exhibits local symmetry; When the determination of all feature lines at the central position of one matrix is completed, it is determined whether the central position is a symmetric central position based on the recorded feature line, symmetric radius, and symmetric order. If so, the corresponding symmetric central position is recorded , otherwise skip to step A5.

Determining whether the matrix elements on the feature line exhibit local symmetry is specifically, determining whether the symmetric elements on the feature line centered on the central position show similarity and similar maximum length, and if there is similarity, local symmetry is determined. is determined to represent , and the maximum length is used as the radius of symmetry;

Preferably, if the symmetrical elements are the same or the difference is within a critical range, the two symmetrical elements are considered to be similar, and the consecutive length of the symmetrical elements showing the similarity is taken as the maximum similarity length;

Determining whether the center position is the center of symmetry with the recorded feature line, symmetry radius, and symmetry number is specifically, symmetry area Determine , where N is the number of symmetries, is the recorded ith feature line, silver is the symmetrical radius of , and the ratio of the symmetrical area AC and the size of the temperature matrix is greater than the area threshold, it is determined that the central position is the symmetrical central position, where the area threshold is a preset value, and through the area threshold, it is further value-free detection. Filters the target and generated invalid computations;

Step A5: Proceed to step A3 to continue processing of the next center position until all center positions have been processed, and when all center positions have been processed, an adjustment reference position is determined based on the symmetric center position.

To determine the adjustment reference position based on the symmetric central position, specifically, according to the size of the symmetrical area of the symmetric central position, one or more symmetric central positions are used as the adjustment reference position and the corresponding symmetric radius;

Preferably, when there is a significant difference between the symmetric areas of the plurality of symmetric central positions, the symmetric central positions with large symmetric areas are left and the symmetric central positions with small symmetric areas are deleted, and then a plurality of consecutive symmetric central positions are formed. The fusion method determines the new symmetric center position, and the fusion method takes the position closest to the average value of the matrix elements of the symmetric center position as the new symmetric center position, and the farthest distance between the center of the new symmetric position and the fused symmetric position and the most The symmetric radius of the center of the symmetric position corresponding to the far distance is taken as the symmetric radius of the center of the new symmetric position, and the average value is the average value of the matrix elements of the plurality of consecutive symmetric central positions, there is no significant difference, and the plurality of symmetric central positions When is discontinuous, the plural symmetric center positions are taken as adjustment reference positions, and the symmetric radius does not change, where the discontinuities are not located in adjacent sub-regions.

The adjustment of the temperature matrix based on the adjustment reference position is specifically made by taking the adjustment reference position as the central position, taking the symmetrical radius as the minimum radius, holding the matrix element values within the minimum radius, and optionally setting the matrix elements outside the minimum radius. value is zeroed, and the selective zeroing is to keep the matrix element if the matrix element is greater than the zeroing threshold, otherwise zero the matrix element, preferably set the zeroing threshold to a relatively large value, and set the relatively large multivalue located in the section

Preferably, when there are multiple adjustment reference positions, when selectively zeroing out, if processing is held for the same matrix element, holding processing is selected.

The above step also includes an S3EX step: reducing the size of the extruded temperature matrix, and when the maximum or minimum row or column of the temperature matrix is all 0 values, the row or column is deleted, that is, the size is reduced here. This does not mean reducing the size of the internal elements of the matrix, but rather reducing the size of the matrix without changing the adjacent correlations of the matrix elements.

Step S4: Build an artificial intelligence model, the input of the artificial intelligence model is the preprocessed and projected temperature matrix, and the output is the location and type of power equipment.

In the construction of the artificial intelligence model, specifically, the artificial intelligence model is set as a neural network model, and the neural network model includes a convolution layer, a pooling layer, and a fully connected layer, wherein the convolution layer is a temperature matrix The matrix features are extracted and input, the pooling layer performs a ferromanipulation on the acquired matrix features to prevent excessive matching, the fully connected layer outputs the detection result, and the artificial intelligence model performs the above-described multi-value processing and extrusion Combined with processing, the amount of subsequent computation and training based on extrusion processing is reduced, and setting multiple convolution kernels in the model extracts more image features, improving the sensitivity and accuracy of the model.

Preferably, when performing step S3EX, when working on the model, feature diagrams of different sizes may be obtained, and a feature pooling layer of interest may be set to solve the problem of different feature sizes, and each layer may be passed through the corresponding layer. A region obtains a fixed-dimensional feature vector.

The convolution layer obtains a matrix feature corresponding to the convolution kernel by filtering each region of the temperature matrix using a convolution kernel, and multiple matrix features are obtained through multiple convolutions with multiple convolution kernels. obtained, and through the processing of steps S2 and S3, the amount of calculation of convolution operation is greatly reduced.

Preferably, the size of the region filtered by the convolution kernel is equal to the size of the sub-region divided in step S3.

Preferably, the size of the region filtered by the convolution kernel and the size of the sub-region divided in step S3 are dynamically set, and the size is modified and recalled according to the output result of the artificial intelligence model, and the size of the output result The region size and/or sub-region size is adjusted according to sensitivity or accuracy.

In the prior art, the size of the pooling area and the size of the temperature matrix are generally the same, but in the case of the artificial intelligence model mentioned above, multiple convolution kernels are used to extract multi-angle matrix features, so setting pooling with the same sensitivity This part of information needs to be covered, and if the size of the direct pooling area is equal to the size of the temperature matrix area, the adaptability is poor. The present invention proposes a small size pooling method, which reduces the pooling sensitivity, and the small size area is specifically , the temperature matrix , the size of the pooling area is set to M1*N1, M1*N1<M*N (e.g. M1<M,N1<N), the pooling step length is 1, and the number of pooling areas NP is , then one region of the temperature matrix For , the corresponding pooling area is APk.

Preferably, the pooling is median pooling, and the median of pooling is ego;

Preferably, the small-size region is a dynamic small-size region, performing correction and recall of the small-size region after each layer of the feature model, performing correction and recall of the small-size region based on the output result, and similarly , The size of the convolution filter area is a dynamic size area, and after each layer of the feature model, correction and recall of the small size area are performed, while the size of the filter area is modified, where the modification is either the filter area or the small size area. is a parameter of , in other words, the dynamic application of the area was performed using the convolution kernel and the recalled feature, and as a result, the calculation speed was remarkably improved.

However, random random adjustment may not improve the performance, so the adjustment between the size of the segmented sub-region and the filtered region and the pooled small-size region has a correlation, and the size of the selectively partitioned sub-region and the filtered region. The adjustment directions are identical, and the adjustment directions between the divided sub-region, the size of the filtered region, and the pooled small-size region are opposite in a situation where the amount of computation is limited, and the adjustment directions between them are the same when the amount of computation is not limited. do.

The small-sized pooling domain setting can obtain local information of several sub domains by feeding the input matrix data only once, and does not lose the effective multidimensional information acquired by the convolution layer while improving the feature representation capability.

Preferably, the pooling layer selects salient matrix features through down-sampling dark-axis matrix features;

Preferably, the artificial intelligence model extracts features and performs network training between stages through backpropagation and random layer descent.

That is, in the process of data preprocessing, feature extraction, and model creation, the intrinsic characteristics of the thermal image of substation facilities are quantitatively reduced, and the complexity of extraction increases in model feature extraction. This reduces the amount of training and computation without losing the accuracy of the model.

In the process of training the artificial intelligence model, the learning rate in the first 200 times set to , and after 200 cycles, the learning rate is , and a total of 2500 times of vaginal belt training, 1200 steps per vaginal belt, and model AP = 89.98%, and after completion of training, the same experimental environment is configured, and 400 test data of each type of power equipment are selected. After testing, and calculating the accuracy and recall rate of each type of equipment, the results are as follows:

A logic module of each of the examples above using a general-purpose processor, digital signal processor (DSP), ASIC, FPGA or other PLD, discrete gate or transistor logic, discrete hardware device, or any other combination designed to perform the functions described herein. , to realize modules and circuits. A general-purpose processor may be a microprocessor, but in the alternative, the processor may be a commercially available processor, controller, microcontroller, or state machine. A processor may be realized as a combination of computing equipment, for example, a combination of a DSP and a microprocessor, a plurality of microprocessors, one or several microprocessors cooperating with a DSP core, or any other such configuration. The steps of a method or algorithm described herein may be directly embodied in hardware, in a software module executed by a processor, or in a combination of the two. A software module can exist in any type of tangible storage medium, and examples of storage media that can be used include random access memory (RAM), read-only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, and hard drives. A storage medium may be coupled to the processor such that the processor can read information from and write information to the storage medium, including disks, removable disks, CD-ROMs, and the like; alternatively, the storage medium is one unit with the processor. A software module can be a single instruction or multiple instructions, can be distributed in several different code segments, different programs, and can span several storage media.

The function may be realized by hardware, software, firmware, or any combination thereof, and when implemented by software, the function may be stored in an actual computer-readable medium as one or several instructions, and may be read by the computer. A computer-readable medium includes a computer-readable storage medium, and a computer-readable storage medium can be any available storage medium that can be accessed by a computer, and as a non-limiting example, such computer-readable medium includes RAM, ROM, EEPROM, CD-ROM or other CD-ROM disk storage device, magnetic disk or other magnetic disk storage device, or any other medium accessible to the computer and the expected program code that hosts or stores instructions or data structures; and Propagated signals are not included within the scope of computer readable storage media, and computer readable media also includes communication media, including any medium that facilitates transfer of a computer program from one place to another. , the connection may be, for example, a communication medium. For example, if the software is transmitted from a website, server, or other remote source using coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technology such as infrared, radio, or microwave, the definition of communication medium is This includes coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio, and microwave. The scope of computer readable storage media should also include combinations of the above, and alternatively or alternatively, the functions described herein may be performed by one or more hardware logic modules, e.g., may be used. Exemplary types of hardware logic modules include FPGAs, ASICs, ASSPs, SOCs, CPLDs, and the like.

Accordingly, a computer program product may perform the operations presented herein, and such a computer program product may be, for example, a computer-readable tangible medium having tangible storage (and/or coding) and instructions thereof, and the corresponding instructions executes the operation by one or several processors. A computer program product may include a packaging material.

The software or instructions may be transmitted over a transmission medium, for example, the software may be transmitted over a website using a transmission medium such as coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technology such as infrared, radioelectric, or microwave. , may be transmitted from a server or other remote source.

In addition, modules and/or other suitable means for performing the methods and techniques described herein may be suitably downloaded and/or otherwise obtained by a user terminal and/or base station, e.g., such an apparatus may be Connecting to a server to facilitate transmission of means for performing the methods described herein, or the various methods described herein may be provided through a storage module (e.g. RAM, ROM, physical storage medium such as a CD or floppy disk). The user terminal and/or the base station acquires various methods when connected to the device or providing a storage component to the device. In addition, any other suitable technique for providing the methods and techniques described herein to a device may be used.

Since what has been described above is only a relatively preferred embodiment of the present invention, all equivalent changes or modifications according to the structure, characteristics, and principles described in the scope of the patent application of the present invention are included in the scope of the patent application of the present invention.

Claims (10)

Step S1 of extracting temperature matrix data based on the infrared thermal image;
Step S2 of multi-value processing of the temperature matrix;
Step S3 of protruding the temperature matrix;
A smart detection method for substation equipment, characterized in that it comprises a step S4 of constructing an artificial intelligence model and performing smart detection of substation equipment based on the model. The substation equipment according to claim 1, characterized in that step S4 specifically: builds an artificial intelligence model, inputs of the artificial intelligence model are preprocessed and projected temperature matrices, and outputs are the location and type of power equipment. of the smart detection method. According to claim 2, the power equipment includes lightning arresters, circuit breakers, current transformers, bushings, voltage transformers, GIS bushings, insulation switches, insulators, clamps, transformers, capacitors, reactors, ceiling bushings, power cables and oil conservators. A smart detection method for substation equipment, characterized in that for doing. The smart detection method of substation facilities according to claim 3, characterized in that a sample data set is formed by selecting 1000 temperature data for each facility. The method of claim 4, wherein in the process of training the artificial intelligence model, the learning rate is set to , and after 200 cycles, the learning rate is Smart detection method of substation equipment, characterized in that set to lower to. The smart detection method of substation equipment according to claim 5, characterized in that a total of 2500 times of quality belt training is performed, and the AP of the model is made = 89.98% after 1200 steps of training per quality belt. It includes a server and one or more client terminals, wherein the client terminal takes an image and uploads the captured image to the server to obtain a detection result, wherein the server is the substation equipment of any one of claims 1 to 6. A substation equipment smart detection system characterized by using a smart detection method. The substation equipment smart detection system according to claim 7, wherein the server is a cloud server. a storage unit for storing applications; and
A substation equipment smart detection device comprising: a processing unit electrically coupled to the input unit and the storage unit, and configured to execute the method of smart detection of substation equipment according to claims 1 to 6. A storage medium for smart detection of substation equipment, characterized in that it stores instructions of the smart detection method of substation equipment according to claims 1 to 6.