WO2022267375A1

WO2022267375A1 - Method and system for intelligent detection of substation device

Info

Publication number: WO2022267375A1
Application number: PCT/CN2021/136949
Authority: WO
Inventors: 梁川; 常娜; 朱怡良; 刘学臻; 石焕江
Original assignee: 浙江天铂云科光电股份有限公司
Priority date: 2021-06-22
Filing date: 2021-12-10
Publication date: 2022-12-29
Also published as: KR20230109713A; CN113408426A; CN113408426B

Abstract

The present invention relates to a method and system for intelligent detection of a substation device. The method comprises: extracting data of a temperature matrix on the basis of an infrared thermogram; performing multi-valued processing on the temperature matrix; performing saliency processing on the temperature matrix; and building an artificial intelligence model, and performing intelligent detection on a substation device on the basis of the model. By means of the present invention, during the processes of data pre-processing, feature extraction and model creation, calculation amount reduction is performed for a natural characteristic of an infrared thermogram of a substation device, the extraction complexity is increased during model feature extraction, the pooling sensitivity is cooperatively improved during a pooling process, and a sensitivity loss caused by dimension reduction is maintained, such that a training amount and a calculation amount are ultimately reduced without losing model precision.

Description

A method and system for intelligent detection of substation equipment

technical field

The invention belongs to the technical field of intelligent power systems, and in particular relates to an intelligent detection method and system for substation equipment.

Background technique

At present, the detection method of electric power equipment is mainly based on infrared pictures, and adopts visual observation method to artificially identify the type of equipment photographed. A false color, which increases the difficulty of human identification.

technical problem

On the whole, the existing methods have problems such as slow speed and low detection accuracy; the present invention reduces the number of natural characteristics of thermal images of substation equipment in the process of data preprocessing, feature extraction and model creation, while In the feature extraction of the model, the complexity of the extraction is increased, and the pooling sensitivity is improved during the pooling process, and the sensitivity loss caused by dimensionality reduction is maintained. Finally, the amount of training and calculation is reduced without losing the accuracy of the model.

technical solution

In order to solve the above problems in the prior art, the present invention proposes an intelligent detection method for substation equipment and its system. The intelligent detection method for substation equipment includes step S1, extracting temperature matrix data based on infrared thermal images; step S2, for Multi-valued processing of the temperature matrix; step S3, performing saliency processing on the temperature matrix; step S4, building an artificial intelligence model, and performing intelligent detection of substation equipment based on the model.

Further, the step S4 is specifically to build an artificial intelligence model, the input of the artificial intelligence model is the temperature matrix that has been preprocessed and highlighted, and the output is the location and category of the electric equipment.

Further, the power equipment includes: lightning arresters, circuit breakers, current transformers, bushings, voltage transformers, GIS bushings, isolating switches, insulators, clamps, transformers, capacitors, reactors, wall bushings, power Cables and oil conservator etc.

Further, 1000 pieces of temperature data are selected for each type of equipment to form a sample data set.

Further, in the process of training the above artificial intelligence model, in the first 200 iterations, the learning rate is set to

, after 200 iterations, the learning rate setting is reduced to

.

Further, a total of 2,500 iterative trainings were performed, with 1,200 steps per iteration. After training, the mAP of the model was 89.98%.

An intelligent detection system for substation equipment includes a server and one or more client terminals. The client terminal takes images and uploads the captured images to the server to obtain detection results; the server is used to execute the intelligent detection method for substation equipment.

Further, the server is a cloud server.

A substation equipment intelligent detection device includes a storage unit configured to store an application program; and a processing unit electrically coupled to an input unit and the storage unit, the processing unit configured to execute the substation equipment intelligent detection method.

A storage medium for intelligent detection of substation equipment, characterized in that the storage medium is used to store instructions for an intelligent detection method for substation equipment.

Beneficial effect

The beneficial effects of the present invention specifically include: (1) through multi-valued processing, the calculation amount can be effectively reduced without losing calculation accuracy; and through position rotation and adjustment, the matrix can further contain as much valid data as possible; (2) Through the multi-valued preprocessing of the matrix and the saliency processing and feature extraction based on the matrix element values, the calculation requirements for the prediction model are reduced without reducing the precision and recall; (3) The setting of dynamic adaptation Filtering areas, dividing sub-areas and small-size pooling areas can adjust the sensitivity of the model according to the detection results, and balance the effectiveness of the model and the amount of available resources; (4) artificial intelligence models and the previous multi-value The combination of processing and saliency processing reduces the amount of subsequent calculation and training on the basis of saliency features, and setting multiple convolution kernels in the model can extract more image features to improve the sensitivity and accuracy of the model.

Description of drawings

Fig. 1 is a schematic diagram of the intelligent detection method for substation equipment of the present invention.

Fig. 2 is a schematic diagram of the working mode of the artificial intelligence model for identifying substation equipment according to the present invention.

Embodiments of the present invention

The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments, wherein the schematic embodiments and descriptions are only used to explain the present invention, but are not intended to limit the present invention.

The present invention uses the temperature matrix in the infrared thermal image of the detected equipment collected by the device as the input of intelligent detection instead of the infrared thermal image as the input, wherein: the value of each element in the matrix represents the infrared thermal image Corresponding to the actual temperature value of the pixel point; after preprocessing the temperature matrix data, first perform feature extraction, and obtain the position and category information of the electrical equipment in the infrared thermal image through the artificial intelligence detection model, and then filter out the location and category information of the infrared thermal image The category and location of the most central power equipment in the figure are output, which is used to judge whether this type of power equipment is faulty due to overheating.

Preferably, the electrical equipment includes: lightning arresters, circuit breakers, current transformers, bushings, voltage transformers, GIS bushings, isolating switches, insulators, clamps, transformers, capacitors, reactors, wall bushings, power Cables and oil conservator etc.

Preferably, 1000 pieces of temperature data are respectively selected for each type of equipment to form a data set.

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

The intelligent detection method for substation equipment according to the present invention specifically includes the following steps.

Step S1, extracting temperature matrix data based on the infrared thermal image; specifically: obtaining the infrared thermal image of the detected device, and the value of each element in the temperature matrix represents the actual temperature value of the corresponding pixel in the infrared thermal image.

Step S2, multivaluing the temperature matrix; specifically: obtaining one or more multivalued intervals, comparing the element value of each element in the temperature matrix with the multivalued interval, and setting the element value to The fixed value corresponding to the multivalued interval that falls in, wherein each multivalued interval corresponds to a fixed value, and the larger the value of the multivalued interval, the larger the fixed value.

Common preprocessing is often to remove inconsistent data, without starting with the amount of calculation and model characteristics. In fact, for the identification of this specific type of electrical equipment, continuous temperature matrix elements will cause a large amount of redundant calculation of the artificial intelligence model. , such an increase will not increase the calculation accuracy, and the traditional binarization process will obviously lose too much information; through multi-value processing, the calculation amount can be effectively reduced without losing calculation accuracy; and through position rotation and adjustment, it can be Further, the matrix contains as much valid data as possible; since the types of substation equipment are limited, the image forms presented are also relatively limited. The present invention reduces the precision rate and recall rate while not reducing the precision rate and recall rate through preprocessing before detection. The computational requirements of the forecasting model are reduced.

Step S3: temperature matrix salience processing; specifically: determine the temperature matrix adjustment reference position; adjust the temperature matrix based on the adjustment reference position; make possible target objects more prominent through such adjustment; that is, make the adjustment make The matrix contains as much valid data as possible; since the types of substation equipment are limited, the image form it presents is also relatively limited. The present invention uses matrix multivalued preprocessing and matrix element value-based saliency feature extraction. While reducing the precision rate and recall rate, it also reduces the computational requirements for the prediction model.

The determination of the temperature matrix adjustment reference position specifically includes the following steps, step A1: determine the central position of the temperature matrix; wherein: the central position of the temperature matrix is a specific position in the temperature matrix selected according to the size of the temperature matrix; the central position is a or more.

Preferably: the temperature matrix is divided into multiple same or different sub-regions, and the central position of the sub-region is selected as the central position of the temperature matrix; the number of the sub-regions is positively correlated with the size of the temperature matrix.

Step A2: Obtain a central position of the temperature matrix; determine one or more characteristic lines including the central position of the matrix; the characteristic line is composed of one or more elements of the temperature matrix including the elements at the central position.

The determination includes one or more characteristic lines at the center position of the matrix, specifically: dividing and setting a plurality of characteristic lines in a specific way; preferably: when there are multiple characteristic lines, the plurality of characteristic lines are separated by a fixed distance Angle; for example: the sub-area is a 3*3 matrix, if

is the center position, then the feature line at the horizontal position contains

, the characteristic line deviated from the 45-degree angle position contains

;Three feature lines can be set for this subarea.

Preferably, the feature line is a feature line that is symmetrical/relatively symmetrical about the central position.

In step A3, a characteristic line and its matrix elements are obtained in sequence, and the processing proceeds to step A4.

Step A4: Determine whether the matrix elements on the feature line exhibit local symmetry; if so, record the feature line, symmetry radius, and increased symmetry times; skip to step A3 and continue to judge the next feature line until all feature lines Lines are all judged; when all feature lines at the center of a matrix are judged, determine whether the center position is a symmetrical center position based on the recorded feature line, symmetry radius, and symmetry times; if so, record the symmetric center position; otherwise Go to step A5.

Said judging whether the matrix elements on the feature line exhibit local symmetry is specifically: determining whether the symmetrical elements on the feature line centered on the central position exhibit similarity, and the maximum length of the similarity; if there is similarity, determine Local symmetry is assumed, and the maximum length is taken as the radius of symmetry.

Preferably: when the symmetric elements are equal or the difference is within a threshold range, it is considered that two symmetric elements exhibit similarity; the continuous length of the symmetric elements exhibiting similarity is taken as the maximum length of similarity.

The determination of whether the center position is a symmetrical center position based on the recorded characteristic line, symmetry radius and symmetry times is specifically: determining the symmetry area

; where N is the number of symmetry,

is the i-th feature line recorded,

Is the symmetric radius of Li; if the ratio of the symmetric area AC and the size of the temperature matrix is greater than the area threshold, then determine the center position as the symmetric center position; where: the area threshold is a preset value; further filter worthless detection objects and the above by the area threshold Resulting invalid calculation.

Step A5: Jump to step A3 to continue the processing of the next center position until all center positions are processed; if all center positions are processed, then determine the adjustment reference position based on the symmetrical center position.

The determination of the adjustment reference position based on the symmetry center position specifically includes: selecting one or more symmetry center positions as the adjustment reference position and its corresponding symmetry radius according to the size of the symmetry area of the symmetry center position.

Preferably: when there is a significant difference between the symmetric areas of multiple symmetric center positions, retain the symmetric center position with a large symmetric area and delete the symmetric center position with a small symmetric area; then, determine a new The symmetric center position; the fusion method is to use the position of the matrix element closest to the mean at the symmetric center position as the new symmetric center position, and add the furthest distance between the new symmetric position center and the fused symmetric position center to the farthest The symmetric radius from the center of the corresponding symmetric position is taken as the symmetric radius of the new symmetric center; the mean value is the mean value of the matrix elements at the multiple consecutive symmetric center positions; when there is no significant difference, and the multiple symmetric center positions are not When continuous, the plurality of symmetrical center positions are used as adjustment reference positions, and the symmetrical radius remains unchanged; wherein: the discontinuous is not located in adjacent sub-regions.

The adjustment of the temperature matrix based on the adjustment reference position is specifically: taking the adjustment reference position as the center position, taking the symmetrical radius as the minimum radius, retaining the matrix element values within the minimum radius, and selectively zeroing the minimum The matrix element value outside the radius; the selective zeroing is: when the matrix element is greater than the zeroing threshold, keep the matrix element, otherwise, the matrix element is set to zero; preferably: the zeroing threshold is set to A relatively large value within a large multi-valued interval.

Preferably: when there are multiple adjustment reference positions, when selectively returning to zero, as long as there is a hold process for the same matrix element, the hold process is selected.

Said steps also include step S3EX: reducing the size of the temperature matrix processed through salience; when the maximum or minimum row or column of the temperature matrix is all zero values, delete the row or column; that is to say, the reduced size here is not Reduce the size of the internal elements of the matrix, thereby reducing the size of the matrix without changing the adjacent relationship of the matrix elements.

Step S4: Building an artificial intelligence model, the input of the artificial intelligence model is the temperature matrix that has been pre-processed and highlighted; the output is the location and category of the electric equipment.

The artificial intelligence model is set up, specifically: the artificial intelligence model is set as a neural network model, and the neural network model includes a convolutional layer, a pooling layer and a fully connected layer; wherein: the convolutional layer extracts the matrix features in the input temperature matrix The pooling layer performs dimensionality reduction operations on the obtained matrix features to prevent overfitting; the fully connected layer outputs the detection results; the artificial intelligence model cooperates with the previous multi-value processing and saliency processing, and the Basically, the amount of subsequent calculation and training is reduced, and the multi-convolution kernel of the model can extract more image features to improve the sensitivity and accuracy of the model.

Preferably: when step S3EX is performed, feature maps of different sizes will also be obtained when performing operations in the model, and the feature pooling layer of interest is set to solve the problem of different feature scales; after this layer, each region can be obtained A feature vector of a fixed dimension.

The convolution layer uses a convolution kernel to filter each region of the temperature matrix to obtain matrix features corresponding to the convolution kernel; there are multiple convolution kernels, and multiple matrix features are obtained through multiple convolution kernels; After the processing of steps S2 and S3, the calculation amount of the convolution operation here is greatly reduced.

Preferably: the size of the region filtered by the convolution kernel is the same as the size of the sub-regions divided in step S3.

Preferably: the size of the area filtered by the volume set and the size of the sub-area divided by step S3 can be dynamically set, and its size is corrected and regressed according to the output result of the artificial intelligence model; according to the sensitivity or accuracy of the output result, the adjusted The above region size and/or subregion size.

In the prior art, the size of the pooling area and the size of the temperature matrix are generally the same, but for the above-mentioned artificial intelligence model, which uses multiple convolution kernels to extract matrix features in an all-round way, it is necessary to set up the same sensitive pooling To cover this part of the information, directly setting the pooling area equal to the size of the temperature matrix area lacks adaptability. The present invention proposes a small-size pooling method, which can reduce the pooling sensitivity; the small-size area is specifically: for temperature matrix

, set the size of the pooling area to M1*N1, and M1*N1<M*N (for example: M1<M, N1<N); the pooling step size is 1, then the number of pooling areas NP is

; then for a region of the temperature matrix

For example, its corresponding pooling area is APk.

Preferably: the pooling is a median pooling; the median of the pooling

.

Preferably: the small-sized area is a dynamic small-sized area. After the operation of each layer of the feature model, the correction and regression of the small-sized area are performed, and the correction and regression of the size of the small-sized area is performed based on the output result; similar: setting volume The size of the product filter area is a dynamic size area. After the operation of each layer of the feature model, the size of the filter area is corrected while the correction and regression of the small-size area are performed; here, the parameters of the filter area or small-size area are corrected; That is, the convolution kernel and its regressive characteristics are used to dynamically adapt the region, so that the calculation speed has been significantly improved.

However, disordered random adjustments may not bring about an increase in capacity. Therefore, there is a correlation between the divided sub-region and the size of the filtered region, and the adjustment of the pooled small-size region; optional: divided The adjustment direction of the size of the sub-area and the filtered area is the same; and in the case of a limited amount of calculation, the adjustment direction between the divided sub-area and the size of the filtered area is opposite to that of the pooled small-size area , and the adjustment direction between them is the same when the amount of calculation is not limited.

Through the small-size pooling area setting, it is only necessary to feed forward the input matrix data once to obtain the local information of multiple sub-regions, and improve the feature expression ability without losing the effective multi-dimensional information obtained by the convolutional layer.

Preferably: the pooling layer compresses matrix features by downsampling, and selects significant matrix features.

Preferably: the artificial intelligence model extracts features, and uses backpropagation and stochastic gradient descent to perform end-to-end network training.

That is to say, in the process of data pre-processing, feature extraction and model creation, the natural characteristics of the thermal image of the substation equipment are reduced in quantity, and the complexity of the extraction is increased in the model feature extraction, which is coordinated in the pooling process. Improve the pooling sensitivity, maintain the sensitivity loss caused by dimensionality reduction, and finally reduce the amount of training and calculation without losing the accuracy of the model.

In the process of training the above artificial intelligence model, in the first 200 iterations, the learning rate is set to

, after 200 iterations, the learning rate setting is reduced to

; A total of 2500 iterations of training, each iteration 1200 steps, after training, the model's AP=89.98%; The precision and recall for each type of device are as follows.

.

In general-purpose processors, digital signal processors (DSPs), ASICs, field-programmable gate arrays (FPGAs), or other programmable logic devices (PLDs), discrete gates designed to perform the functions described herein may be utilized. or transistor logic, discrete hardware components, or any combination thereof to implement or perform the various illustrated logic blocks, modules, and circuits described; the general-purpose processor may be a microprocessor, but alternatively, the processor may be any commercially available Available processors, controllers, microcontrollers, or state machines; processors can also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, multiple microprocessors, one or more a microprocessor or any other such configuration; the steps integrating the methods or algorithms described in this disclosure may be embedded directly in hardware, in a software module executed by the processor, or in a combination of the two; the software module may exist in any form Some examples of storage media that may be used include random access memory (RAM), read only memory (ROM), flash memory, EPROM memory, EEPROM memory, registers, hard disks, removable disks, CDs - ROM, etc.; the storage medium may be coupled to the processor such that the processor can read information from and write information to the storage medium; in the alternative, the storage medium may be integral to the processor; the software module may A single instruction or many instructions, and may be distributed over several different code segments, among different programs, and across multiple storage media.

The functions described may be implemented in hardware, software, firmware or any combination thereof; if implemented in software, the functions may be stored as one or more instructions on a tangible computer-readable medium; computer-readable media include computer-readable storage medium; a computer-readable storage medium can be any available storage medium that can be accessed by a computer; by way of example and not limitation, such a computer-readable medium can include RAM, ROM, EEPROM, CD-ROM, or other optical disk storage, magnetic disk storage, or Other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer; otherwise, propagated signals are not included within the scope of computer-readable storage media ; computer-readable media also includes communication media, including any medium that facilitates the transfer of a computer program from one place to another; a connection can be, for example, a communication medium; for example, if the software uses a coaxial cable, fiber optic cable, twisted pair, Digital Subscriber Line (DSL), or wireless technologies such as infrared, radio, and microwave to transmit from a web site, server, or other remote source, the coaxial cable, fiber optic cable, twisted pair, DSL, or such as Wireless technologies such as infrared, radio, and microwave are included in the definition of communication media. Combinations of the above should also be included within the scope of computer-readable media; alternatively or alternatively, the functions described herein may be at least partially performed by one or more hardware logic components; for example, the hardware logic components that may be used Illustrative types of include field programmable gate array (FPGA), application specific integrated circuit (ASIC), application specific standard product (ASSP), system on chip (SOC), complex programmable logic device (CPLD), and the like.

Thus, a computer program product can perform the operations set forth herein; for example, such a computer program product can be a computer-readable tangible medium having instructions tangibly stored (and/or encoded) thereon, which can be read by one or Multiple processors execute to perform the operations described herein. A computer program product may include packaging materials.

Software or instructions may also be transmitted via transmission media; for example, transmission media such as coaxial cables, fiber optic cables, twisted pair wires, digital subscriber lines (DSL), or wireless technologies such as infrared, radio, or microwaves may be transmitted from a website, server Or other remote source transfer software.

In addition, modules and/or other suitable means for performing the methods and techniques described herein may be downloaded and/or otherwise obtained by user terminals and/or base stations as appropriate; for example, such devices may be coupled to server to facilitate the transfer of the means for performing the methods described herein; alternatively, the various methods described herein may be provided via storage means such as RAM, ROM, physical storage media such as CDs or floppy disks, Such that a user terminal and/or a base station can obtain various methods when coupled to the device or providing storage means to the device. In addition, any other suitable technique for providing the methods and techniques described herein to a device may be utilized.

The above is only a preferred embodiment of the present invention, so all equivalent changes or modifications made according to the structure, features 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

A method for intelligent detection of substation equipment, characterized in that the method comprises:

Step S1: Extract temperature matrix data based on the infrared thermal image;

Step S2: Multivalued the temperature matrix;

Step S3: Perform saliency processing on the temperature matrix;

Step S4: Build an artificial intelligence model, and perform intelligent detection of substation equipment based on the model.
The intelligent detection method for substation equipment according to claim 1, wherein said step S4 is specifically: building an artificial intelligence model, the input of said artificial intelligence model is a temperature matrix through preprocessing and significant processing; the output is The location and type of electrical equipment.
The intelligent detection method for substation equipment according to claim 2, wherein the power equipment includes: lightning arresters, circuit breakers, current transformers, bushings, voltage transformers, GIS bushings, isolating switches, insulators, and clamps , transformers, capacitors, reactors, wall bushings, power cables and oil conservator etc.
The intelligent detection method for substation equipment according to claim 3, wherein 1000 pieces of temperature data are respectively selected for each type of equipment to form a sample data set.
The intelligent detection method for substation equipment according to claim 4, wherein, in the process of training the above-mentioned artificial intelligence model, in the first 200 iterations, the learning rate is set to
, after 200 iterations, the learning rate setting is reduced to
.
The intelligent detection method for substation equipment according to claim 5, characterized in that, a total of 2500 iterative trainings, 1200 steps per iteration, make the AP of the model = 89.98% after training;
An intelligent detection system for substation equipment, characterized in that it includes: a server and one or more client terminals, the client terminal takes images, and uploads the images taken to the server to obtain detection results; the server is used to implement claims The intelligent detection method for substation equipment described in any one of 1-6.
The intelligent detection system for substation equipment according to claim 7, wherein the server is a cloud server.
An intelligent detection device for substation equipment, characterized in that it includes:

a storage unit configured to store an application; and

A processing unit, electrically coupled to an input unit and the storage unit, configured to execute the intelligent detection method for substation equipment described in claims 1-6.
A storage medium for intelligent detection of substation equipment, characterized in that the storage medium is used to store instructions for executing the intelligent detection method for substation equipment according to claims 1-6.