CN115995051A - Substation equipment fault period identification method based on minimum residual error square sum method - Google Patents

Abstract

The invention discloses a substation equipment fault period identification method based on a minimum residual error sum-of-squares method; through ISP algorithm accumulation, by utilizing a deep learning intelligent algorithm and combining the basis of security and protection big data, analysis samples of more than 10 tens of thousands of different climate environments in the world are collected for learning, so that a high-precision smoke and fire recognition algorithm based on deep learning is obtained, the smoke and fire recognition rate is better, and the environmental adaptability to different regions in the world is better. Meanwhile, a frame interpolation method and a background interpolation method are comprehensively considered, a motion area is detected more accurately, then the detected motion target area is sent to a deep learning model for detection, the calculated amount is reduced through data dimension reduction processing, the real-time performance of detection is guaranteed, the calculation efficiency is improved by using an acceptance structure and a global pooling layer in a deep learning link, the degradation problem of a deep network is solved by using a Resnet residual block, characteristics can be extracted better, parameters are reduced, and the efficiency and the accuracy of a detection process are guaranteed.

Description

Substation equipment fault period identification method based on minimum residual error square sum method

Technical Field

The invention relates to the field of fire disaster identification of power systems, in particular to a substation equipment fault period identification method based on a minimum residual error square sum method.

Background

The power transformation link is a link of high occurrence of disaster accidents in the power grid, and the disaster accidents of power transformation equipment are usually accompanied by smoking and firing. The fire is firstly started, and the effective detection of the smoke can timely kill the fire. Since the first phenomenon of fire is smoke, only smoke is often generated at the early stage of a fire disaster of the power transformation equipment, and no flame exists. The accurate and reliable identification of the fire smoke is an important part of the fire alarm system of the transformer substation. When the fire is in smoke, the smoke is not static but in a moving state, and a moving object in the video can be detected to extract a suspected smoke target. However, detection and identification of smoke cannot be performed at present, and misinformation often occurs in fire identification.

In practical application, because the smoke has the characteristic of slow movement, and the variability of the scene changes along with weather, illumination and the like is strong, the detection effect of the traditional motion detection algorithms such as an optical flow method, a frame interpolation method, a background interpolation method and the like is poor. The traditional optical flow method adopts a fuzzy motion calculation method, deduces an image movement track according to the upper frame and the lower frame, automatically produces a new blank frame, has larger error and low recognition accuracy. Although the frame interpolation method has low sensitivity to illumination change and strong dynamic environment adaptability, complete smoke objects cannot be extracted, smoke in a slow motion state cannot be identified, and the method can only be applied to simple real-time motion detection. Although the background interpolation method can better acquire the complete detection object, the real-time performance is good, the method has the defects of low robustness, the changed dynamic scene can greatly influence the detection result, and the reliability of smoke detection is inferior to the reliability of the former two algorithms.

For example, a "a smoke detection method, a smoke detection system, a computer device, and a computer-readable storage medium", which are disclosed in chinese patent literature, bulletin number CN108764264a; comprising the following steps: acquiring a picture data set of smoke and chimney, and adding a position label and a category label to the picture data set; inputting a picture data set with a position label and a category label as a training sample set into a CNN model for training to obtain a plurality of trained CNN models, and calculating a loss function while training; and selecting an optimal CNN model from a plurality of trained CNN models according to the loss function to perform smoke detection on the real-time video so as to obtain a smoke detection result. According to the smoke detection method, the smoke and the chimney are identified in the video with large data volume, so that the smoke can be timely detected in the real-time video, the accuracy of smoke detection can be improved, and the detection speed of the smoke can be improved. However, the method is still affected by the environment, so that the smoke is difficult to detect in the early stage of occurrence, the background environment and the smoke are difficult to separate by means of multi-model training alone, and real-time smoke movement detection is realized.

Disclosure of Invention

The invention aims at the problems of high smoke detection difficulty and low recognition rate caused by the fact that smoke is easy to change along with the change of scenes such as weather, illumination and the like; the method for identifying the fault period of the substation equipment based on the least residual error square sum method is provided; through ISP algorithm accumulation, by utilizing a deep learning intelligent algorithm and combining the basis of security and protection big data, analysis samples of more than 10 tens of thousands of different climate environments in the world are collected for learning, so that a high-precision smoke and fire recognition algorithm based on deep learning is obtained, the smoke and fire recognition rate is better, and the environmental adaptability to different regions in the world is better. Meanwhile, a frame interpolation method and a background interpolation method are comprehensively considered, a motion area is detected more accurately, then the detected motion target area is sent to a deep learning model for detection, the calculated amount is reduced through data dimension reduction processing, the real-time performance of detection is guaranteed, the calculation efficiency is improved by using an acceptance structure and a global pooling layer in a deep learning link, the degradation problem of a deep network is solved by using a Resnet residual block, characteristics can be extracted better, parameters are reduced, and the efficiency and the accuracy of a detection process are guaranteed.

The technical problems of the invention are mainly solved by the following technical proposal:

a substation equipment fault period identification method based on a minimum residual error sum-of-squares method comprises the following steps:

s1: collecting a smoke image dataset: collecting smoke analysis samples of different climatic environments, and establishing a smoke image data set, namely a training set;

s2, preprocessing a smoke image data set: the data dimension is reduced after the image is subjected to graying treatment, and main data characteristics are reserved;

s3: training a neural network: training a convolutional neural network by combining the characteristics of the preprocessed smoke image data with an improved InceptionResnetv2 network model;

s4: acquiring an original video stream, and performing video motion detection on the original video stream: detecting a moving object in a video through a frame interpolation method and a background interpolation method, and extracting a suspected smoke target;

s5: preprocessing a moving object image: the data dimension is reduced after the moving target image is subjected to graying treatment, and main data characteristics are reserved;

s6: and importing the moving target image into a neural network after training is completed, and outputting a detection result.

Through ISP algorithm accumulation, by utilizing a deep learning intelligent algorithm and combining the basis of security and protection big data, analysis samples of more than 10 tens of thousands of different climate environments in the world are collected for learning, so that a high-precision smoke and fire recognition algorithm based on deep learning is obtained, the smoke and fire recognition rate is better, and the environmental adaptability to different regions in the world is better. Meanwhile, a frame interpolation method and a background interpolation method are comprehensively considered, a motion area is detected more accurately, then the detected motion target area is sent to a deep learning model for detection, the calculated amount is reduced through data dimension reduction processing, the real-time performance of detection is guaranteed, the calculation efficiency is improved by using an acceptance structure and a global pooling layer in a deep learning link, the degradation problem of a deep network is solved by using a Resnet residual block, characteristics can be extracted better, parameters are reduced, and the efficiency and the accuracy of a detection process are guaranteed.

Preferably, the step S2 specifically includes: carrying out graying treatment on the smoke image data set by adopting an empirical value method; the high-dimensional data is projected into the low-dimensional space by a principal component analysis algorithm. The method comprises the steps of carrying out gray processing on image data by using an empirical value method, and then carrying out dimension reduction processing on the data by using a PCA algorithm to reduce influence of irrelevant features, so that the calculation amount of the image in the training, testing and detecting processes is reduced, and the calculation efficiency is improved.

Preferably, the principal component analysis algorithm projects the high-dimensional data into the low-dimensional space specifically includes:

s31, setting a smoke image gray value vector x _i The dataset of the training process is { x } ₁ ，x ₂ ，x ₃ ，...x _N The data set average vector X and the covariance matrix cov (X, Y) of the data set are calculated by:

s32, obtaining the characteristic vector u according to the covariance matrix cov (X, Y) _i And corresponding eigenvalue lambda _i ；

S33, after the characteristic values are arranged from large to small, filtering the characteristic values to be smaller than lambda _d To obtain a transformation matrix U= (U) composed of principal components ₁ ，u ₂ ，...u _d )；

S34, projecting the smoke image data to the subspace to obtain a low-dimensional vector y with dimension of d multiplied by 1; wherein y= UTx; a set of low-dimensional vectors L is obtained as a training dataset for smoke recognition, where U has dimensions mxd and x-x has dimensions mx1.

Principal Component Analysis (PCA), a process of implementing dimension reduction on an image, projects high-dimensional data into a low-dimensional space, and can represent high-dimensional information using fewer dimensions while preserving data features that contribute more.

Preferably, training the convolutional neural network by the improved innoresnetv 2 network model specifically comprises:

using the acceptance structure: the admission structure connects different convolution layers in parallel; the acceptance structure uses filters with different sizes to splice the respective processing results into a deeper matrix;

using the Resnet residual block: the Resnet residual network connects the input and the output of a convolution layer by utilizing a shortcut to form a residual block;

global pooling layer is used instead of fully connected layer: and using a global pooling layer to replace a full-connection layer at the last layer of the network, and carrying out global average pooling on all the output characteristic graphs to obtain output.

The computational efficiency is increased by using the acceptance structure and the global pooling layer, and the degradation problem of the deep network is solved by using the Resnet residual block.

Preferably, when the residual function F (x) is the same as the input vector size, the residual block formula is as follows:

y＝F(x，{W _i })+x；

wherein x and y are input and output, respectively, F (x, { W _i -a residual function);

when the residual function F (x) is different from the input vector size, the residual block formula is as follows:

y＝F(x，{W _i })+W _s ；

the above adopts linear projection W _s And performing dimension matching.

The residual block formula is selected through input, and the use of a residual network maximally solves the degradation problem caused by the increase of network depth.

Preferably, the step S4 specifically includes:

s61, performing three-frame interpolation processing on all n frames in the continuous video clips with n frames, namely performing two-by-two frame interpolation between the ith frame and 3 frames of i-1 frames and i+1 frames, wherein the frame interpolation formula is as follows:

wherein K is _i The result is the frame inserted result; t is a threshold value, and the result of the two-by-two interpolation of the ith frame and the front and rear frames is subjected to AND operation to obtain a three-frame interpolation result D of the ith frame _i ；

S62, inserting and adding three frames of a plurality of frames in a time domain to obtain a result D, wherein the formula is as follows:

s63, taking average of n frames in the continuous video stream as a background B, extracting a kth frame of the video, performing background interpolation operation on the kth frame of the video, and adopting the following formula:

wherein the background model is updated with time, the background update formula is as follows:

/>

wherein, beta is the background learning rate;

s64, performing OR operation on the three-frame interpolation result D and the background interpolation result to obtain a motion detection result.

Since smoke is not stationary but in motion in the event of a fire forming smoke, moving objects in the video can be detected to extract suspected smoke objects. The selection of the threshold T has a great influence on the frame interpolation result, and the threshold T is determined by using an Ostu method. The method can filter out the influence of environmental changes such as background light on motion detection, reduce the influence of the environment on smoke detection as much as possible, and simultaneously can perform motion detection in real time, and has better robustness.

The beneficial effects of the invention are as follows:

through ISP algorithm accumulation, by utilizing a deep learning intelligent algorithm and combining the basis of security and protection big data, analysis samples of more than 10 tens of thousands of different climate environments in the world are collected for learning, so that a high-precision smoke and fire recognition algorithm based on deep learning is obtained, the smoke and fire recognition rate is better, and the environmental adaptability to different regions in the world is better. Meanwhile, a frame interpolation method and a background interpolation method are comprehensively considered, a motion area is detected more accurately, then the detected motion target area is sent to a deep learning model for detection, the calculated amount is reduced through data dimension reduction processing, the real-time performance of detection is guaranteed, the calculation efficiency is improved by using an acceptance structure and a global pooling layer in a deep learning link, the degradation problem of a deep network is solved by using a Resnet residual block, characteristics can be extracted better, parameters are reduced, and the efficiency and the accuracy of a detection process are guaranteed.

Drawings

FIG. 1 is an improvement of the concept architecture used by neural networks;

FIG. 2 is a detailed internal architecture of the improved neural network of the present invention;

FIG. 3 is a schematic flow chart of the algorithm of the present invention.

Detailed Description

A substation equipment failure period identification method based on a minimum residual error square sum method is divided into two parts of training a neural network and detecting smoke, and an overall flow of an algorithm is shown in fig. 3. In the figure, the upper part is the training process of the network, the data set is sent into the network model for training after being preprocessed, and finally the network model after data training is obtained. The lower part is smoke video detection, after the original video stream is subjected to motion detection, a moving target is sent into a trained network for recognition, and if the smoke image is detected, the smoke image is marked in the original image and then is alarmed.

Since the first phenomenon of fire is smoke, only smoke is often generated at the early stage of a fire disaster of the power transformation equipment, and no flame exists. The accurate and reliable identification of the fire smoke is an important part of a fire alarm system of the transformer substation, and is an important part of the fire alarm system of the transformer substation; meanwhile, the power network safety construction is deepened continuously due to the fact that various power grid equipment auxiliary management platforms are built, various fire signals can enter the power private network through the auxiliary platforms, and the possibility is created for operation and maintenance personnel to discover fire at the first time.

In the detection process, the most important parts include:

video motion detection:

since smoke is not stationary but in motion in the event of a fire forming smoke, moving objects in the video can be detected to extract suspected smoke objects. The embodiment combines the frame interpolation and the background interpolation method to accurately and effectively detect the motion area.

Firstly, carrying out three-frame interpolation processing on all n frames in a continuous video fragment with n frames, namely, carrying out two-by-two frame interpolation between an ith frame and 3 frames of an i-1 frame and an i+1 frame, wherein the frame interpolation formula is as follows:

wherein K is _i The result is the frame inserted result; t is a threshold value, and the selection of the threshold value has a great influence on a frame interpolation result, and the threshold value is determined by using an Ostu method. And performing AND operation on the results of the two-by-two interpolation of the ith frame and the front and rear frames to obtain a three-frame interpolation result Di of the ith frame.

And inserting and adding three frames of the multi-frame in one time domain to obtain a result D, wherein the formula is as follows:

taking average of n frames in continuous video stream as background B, then extracting kth frame of video to make background interpolation operation, and the formula is as follows:

the background model needs to be updated continuously along with time, and the following formula is a formula for updating the background, wherein beta is a background learning rate.

And finally, performing OR operation on the three-frame interpolation result D and the background interpolation result, wherein the result is the motion detection result of the embodiment. And then the detected moving target area is fed into a deep learning model for detection.

2. Data preprocessing:

deep learning completes training of a model through feature analysis of data, usually, one image data contains a large amount of information, only main features affecting the final learning effect are included, and the large amount of features cause data redundancy, so that the operation efficiency is greatly reduced.

An original characteristic of a video frame image with a size of 500×500 is m=250000, which generates huge operand in the application process and seriously affects the operation efficiency. The system aims to monitor and alarm smoke in real time and has high requirement on algorithm instantaneity, so that in order to reduce the operation amount of images in the training, testing and detecting processes and improve the operation efficiency, the image data is subjected to gray processing by using an empirical value method and then subjected to dimension reduction processing by using a PCA algorithm, so that influence of irrelevant features is reduced. Principal Component Analysis (PCA), a process of implementing dimension reduction on an image, projects high-dimensional data into a low-dimensional space, and can represent high-dimensional information using fewer dimensions while preserving data features that contribute more.

Let a smoke image gray value vector x _i The data set in the training process is { x } ₁ ，x ₂ ，x ₃ ，...x _N The data set average vector X and the covariance matrix cov (X, Y) of the data set are calculated by the following equation.

Feature vector u of covariance matrix is obtained through calculation _i And corresponding eigenvalue lambda _i Then the characteristic values are arranged from big to small and then are smaller than lambda _d Only larger eigenvalues are reserved in eigenvalue discarding to obtain a transformation matrix U= (U) composed of principal components ₁ ，u ₂ ，...u _d ) U is a reduced dimension subspace, the dimension of which is Mxd, and then the smoke sample image is projected to the subspace to obtain a low dimension vector y with dimension d x 1: y= UTx; the resulting set of low-dimensional vectors L is used as a training dataset for smoke recognition.

In the smoke detection process, for one input test sample x, the deviation between the input test sample x and an average sample in a data set is obtained, and the projection vector y of the input test sample x in the main characteristic subspace is obtained by the following formula:

the dimension of U is Mxd, the dimension of x-x is Mx1, and the dimension of y is dx1.

For a 500×500 image, the M dimension is 500×500=250000 dimensions, taking 300 principal components, i.e., 300 feature vectors, the y dimension of the final projected coefficient vector is reduced to 300 dimensions.

3. Building a convolution network model:

convolutional Neural Networks (CNNs) are a deep learning architecture that is widely used in artificial intelligence, text processing, image recognition, and the like. The main process of the convolutional neural network for smoke recognition is to train the network by using the existing smoke image, further extract the smoke image characteristics, and the process is called the learning process of the network because the process is similar to the memory learning process of human beings.

According to the invention, the smoke characteristics are extracted through the training of the preprocessed smoke image data set on the convolutional neural network, and smoke detection is performed by combining with the InceptionResnetv2 network model, so that the extraction training time of the smoke characteristics is greatly shortened under the support of multiple GPUs based on TensorFlow.

Aiming at the problems of degradation, low efficiency, large resource consumption and the like of a deep network, the neural network is improved based on an InceptionResnetv2 network model.

The method has the advantages that the accuracy of feature extraction is improved by increasing the number of convolution layers, the phenomenon of network degradation caused by deepening training errors and increasing test errors of a network is avoided, the consumption of resources is ensured to be as low as possible, the calculation efficiency is improved by using an acceptance structure and a global pooling layer, and the degradation problem of a deep network is solved by using a Resnet residual block. The admission structure connects different convolution layers in parallel; the method comprises the steps that filters with different sizes are used by the acceptance, and the processing results are spliced into a deeper matrix; the acceptance structure increases the network depth and width, can better extract the characteristics, reduces the parameters and ensures the calculation efficiency.

Fig. 1 shows an acceptance module structure.

The Resnet residual network connects the input and output of a convolution layer by using shortcut to form a residual block, and when the residual function F (x) is the same as the input vector, the residual block formula is expressed by using the following formula:

y＝F(x，{W _i })+x；

wherein x and y are input and output, respectively, F (x, { W _i -x) is a residual function, when the residual function F (x) is different from the input vector size, the residual block formula is expressed using the following formula:

y＝F(x，{W _i })+W _s ；

in the above, linear projection W is used _s And performing dimension matching. The use of the residual network maximally solves the degradation problem caused by the increase of the network depth. Using a global pooling layer to replace a full-connection layer at the last layer of the network to carry out global average pool on all the output characteristic graphsThe output is obtained, the parameters can be greatly reduced by eliminating the full-connection layer, and the overfitting and the overconsumption of resources caused by overmany parameters of the full-connection layer are avoided.

Fig. 2 shows the improved network structure, where partially repeated residual blocks are compressed for more visual presentation due to the deep depth of the model. The whole model architecture consists of 40 receptionResnet modules, and 233 convolution layers are arranged in the whole model. The receptionResnet model can rapidly and accurately extract more features, so that the efficiency and the accuracy in the training and detection processes are further ensured.

Claims (6)

1. The substation equipment fault period identification method based on the least residual error square sum method is characterized by comprising the following steps of:

s1: collecting a smoke image dataset: collecting smoke analysis samples of different climatic environments, and establishing a smoke image data set, namely a training set;

s2, preprocessing a smoke image data set: the data dimension is reduced after the image is subjected to graying treatment, and main data characteristics are reserved;

s3: training a neural network: training a convolutional neural network by combining the characteristics of the preprocessed smoke image data with an improved InceptionResnetv2 network model;

s4: acquiring an original video stream, and performing video motion detection on the original video stream: detecting a moving object in a video through a frame interpolation method and a background interpolation method, and extracting a suspected smoke target;

s5: preprocessing a moving object image: the data dimension is reduced after the moving target image is subjected to graying treatment, and main data characteristics are reserved;

s6: and importing the moving target image into a neural network after training is completed, and outputting a detection result.

2. The substation equipment failure period identification method based on the minimum residual square sum method according to claim 1, wherein the step S2 specifically includes: carrying out graying treatment on the smoke image data set by adopting an empirical value method; the high-dimensional data is projected into the low-dimensional space by a principal component analysis algorithm.

3. The substation equipment failure period identification method based on the least squares residual error method according to claim 1, wherein the principal component analysis algorithm projects high-dimensional data to a low-dimensional space specifically comprises:

s31, setting a smoke image gray value vector x _i The dataset of the training process is { x } ₁ ，x ₂ ，x ₃ ，...x _N The data set average vector X and the covariance matrix cov (X, Y) of the data set are calculated by:

s32, obtaining the characteristic vector u according to the covariance matrix cov (X, Y) _i And corresponding eigenvalue lambda _i ；

S33, after the characteristic values are arranged from large to small, filtering the characteristic values to be smaller than lambda _d To obtain a transformation matrix U= (U) composed of principal components ₁ ，u ₂ ，...u _d )；

S34, projecting the smoke image data to the subspace to obtain a low-dimensional vector y with dimension of d multiplied by 1; wherein y= UTx; a set of low-dimensional vectors L is obtained as a training dataset for smoke recognition, where U has dimensions mxd and x-x has dimensions mx1.

4. The substation equipment failure period identification method based on the least squares residual error method according to claim 1, wherein training the convolutional neural network through the improved receptionresnetv 2 network model specifically comprises:

using the acceptance structure: the admission structure connects different convolution layers in parallel; the acceptance structure uses filters with different sizes to splice the respective processing results into a deeper matrix;

using the Resnet residual block: the Resnet residual network connects the input and the output of a convolution layer by utilizing a shortcut to form a residual block;

global pooling layer is used instead of fully connected layer: and using a global pooling layer to replace a full-connection layer at the last layer of the network, and carrying out global average pooling on all the output characteristic graphs to obtain output.

5. The method for identifying the fault period of the substation equipment based on the least squares residual error method according to claim 4, wherein when the residual error function F (x) is the same as the input vector, the residual error block formula is as follows:

y＝F(x，{W _i })+x；

wherein x and y are input and output, respectively, F (x, { W _i -a residual function);

when the residual function F (x) is different from the input vector size, the residual block formula is as follows:

y＝F(x，{W _i })+W _s ；

the above adopts linear projection W _s And performing dimension matching.

6. The substation equipment failure period identification method based on the minimum residual square sum method according to claim 1, wherein the step S4 specifically includes:

s61, performing three-frame interpolation processing on all n frames in the continuous video clips with n frames, namely performing two-by-two frame interpolation between the ith frame and 3 frames of i-1 frames and i+1 frames, wherein the frame interpolation formula is as follows:

wherein K is _i The result is the frame inserted result; t is a threshold value, and the result of the i-th frame and the two-frame insertion of the front frame and the rear frame is ANDCalculating to obtain a three-frame interpolation result D of the ith frame _i ；

S62, inserting and adding three frames of a plurality of frames in a time domain to obtain a result D, wherein the formula is as follows:

s63, taking average of n frames in the continuous video stream as a background B, extracting a kth frame of the video, performing background interpolation operation on the kth frame of the video, and adopting the following formula:

wherein the background model is updated with time, the background update formula is as follows:

wherein, beta is the background learning rate;

s64, performing OR operation on the three-frame interpolation result D and the background interpolation result to obtain a motion detection result.

Images