CN107067026A - Electrical equipment fault detection method based on deep neural network - Google Patents

Abstract

The invention discloses a kind of electrical equipment fault detection method based on deep neural network, including collection infared spectrum, the spectrum data storehouse of standardization is built；Many layer depth neutral nets are set up to classify to test with infared spectrum；Set up target detection framework RCNN and equipment subregion is carried out with infared spectrum to point class testing；According to infared spectrum defect analysis criterion, temperature pre-warning rule in subregion is set；According to temperature pre-warning rule, obtain equipment safety analytical conclusions and realize early warning.The method is applied to that the tedious works such as artificial apparatus identification subregion can be greatly reduced after grid equipment detection, improves operating efficiency.

Description

Fault detection method of power equipment based on deep neural network

technical field

The invention relates to a power equipment fault detection method based on a deep neural network, which belongs to the field of image recognition.

Background technique

With the in-depth construction of UHV AC-DC hybrid large power grid, new generation of smart substations, smart transmission lines and other projects, new major complete sets of power grid equipment and smart equipment continue to emerge, posing new challenges to power grid equipment safety and intelligent operation and maintenance. With the help of new technologies such as the Internet of Things and big data, it is necessary to propose intelligent equipment operation and maintenance technology, and gradually build an informatized, visualized and intelligent equipment operation and maintenance management and control system to meet the higher security operation and maintenance requirements of future power grid equipment.

Video image detection and recognition technology has been widely used in power equipment fault detection and channel monitoring, but as far as the actual application effect is concerned, on the one hand, there are problems such as poor availability of equipment monitoring data, false alarms and missing alarms, and weak guidance of condition maintenance. ; On the other hand, due to various reasons, a large amount of equipment map data collected has not been fully utilized, resulting in waste of resources.

In recent years, artificial neural networks have developed to the stage of deep learning, that is, deep neural networks. Deep learning attempts to use a series of algorithms that contain complex structures or multiple processing layers composed of multiple nonlinear transformations to perform high-level abstraction on data. Its powerful expressive ability enables it to achieve the best results in various machine learning tasks. The performance on video classification also outperforms other methods so far.

Deep learning uses the idea of hierarchical abstraction, and high-level concepts are learned through low-level concepts. This hierarchical structure is usually constructed using a greedy layer-by-layer training algorithm, and effective features that are helpful for machine learning are selected from it. Many deep learning algorithms appear in the form of unsupervised learning, so these algorithms can be applied to Unlabeled data that other algorithms cannot match. This type of data is more abundant and easier to obtain than labeled data. This has become an important advantage of deep learning. Typical unsupervised learning tasks include: dimensionality reduction, which projects input data into a low-dimensional space to achieve more meaningful distance representation or visualization, such as PCA (Principle Component Analysis); clustering , find the similarity between samples and classify them into corresponding categories, such as k-means; density estimation, learn a distribution that generates training data X, such as GMM (Gaussian Mixture Models).

Another advantage of deep learning is that it replaces manual feature acquisition with efficient algorithms for unsupervised or semi-supervised feature learning and hierarchical feature extraction. It combines feature extraction and classification of traditional learning methods, thus making The features obtained through learning have the best effect on classification. Referring to the fault map in the fault image database, it can give safety analysis conclusions and realize intelligent early warning.

The above-mentioned advantages of deep neural network provide a new way for the research of perception technology of equipment and channel state based on intelligent image recognition, but there is no power equipment fault detection method based on deep neural network.

Contents of the invention

In order to solve the above technical problems, the present invention provides a method for detecting faults of electrical equipment based on a deep neural network.

In order to achieve the above object, the technical scheme adopted in the present invention is:

A fault detection method for power equipment based on a deep neural network, including

Collect infrared spectra and build a standardized spectral database; the collected infrared spectra include infrared spectra for testing and infrared spectra for modeling;

Establish a multi-layer deep neural network to classify the infrared spectra for testing in the normalized spectral database;

Establish the target detection framework RCNN to partition the infrared spectrum for the classification test;

According to the infrared spectrum defect analysis criteria, set the temperature warning rules in the partition;

According to the temperature early warning rules, the equipment safety analysis conclusions are obtained and the early warning is realized.

The process of constructing a standardized atlas database is as follows: analyze the collected infrared atlas, select the target features representing the image, define the device representation type in the infrared atlas, and construct a standardized atlas database.

Infrared spectra collected at multiple different viewpoints were filtered based on visual and relevant context before building a normalized spectral database.

The process of building a multi-layer deep neural network is,

Divide the infrared spectrum for modeling in the spectrum database into a training set and a test set;

Construct a multi-layer deep neural network system, input the training set, set the corresponding input activation a ^{x ,l} for each training sample x, carry out forward propagation of data, and calculate the layer parameters of each layer. The specific formula is as follows:

z ^x,l =w ^l a ^x,l-1 +b ^l

a ^x,l-1 = σ(z ^x,l-1 )

Among them, z ^{x, l} is the output of the l-layer neural network unit, z ^{x, l-1} is the output of the l-1 layer neural network unit, w ^l is the weight of the l-layer neural network unit, and a ^{x, l-1} is The input of the l-layer neural network unit, b ^l is the bias of the l-layer neural network unit, σ is the sigmoid function, and l is the number of layers of the neural network;

Calculate the output error δ ^x,l , through the backpropagation algorithm, by the gradient descent rule for each l according to with To update the weight and configuration, so that the value of the cost function tends to 0, where η, m are the learning rate of the neural network and the total number of samples in the training set;

In the iterative process, observe the change of the loss value to judge the convergence, adjust the learning rate, and generate a multi-layer deep neural network model including the parameters of each layer.

The process of device partitioning is,

Build a target detection framework, and use the RCNN target retrieval method to extract the area and position of the backbone equipment in the infrared atlas for classification testing;

Use the Label parting method to partition the extracted backbone device inside the device.

According to the defect analysis criteria of the infrared spectrum, the pixel points in the partition are extracted, and the extracted pixel points are converted into temperatures according to the conversion rules of pixel points and temperature components, and the temperature warning rules in the partition are set according to the temperature.

If it is a one-way device, calculate the average temperature in the zone and generate a single device temperature warning rule; if it is a three-item device, calculate the temperature difference between items in the zone to generate a three-item device temperature warning rule.

The beneficial effects achieved by the present invention: the method realized by the present invention classifies and partitions the images of the power grid by means of a neural network, which can greatly reduce the work of manually classifying and partitioning the images taken by the power grid, reduce cumbersome manual intervention, and improve the efficiency of the power grid. Staff productivity.

Description of drawings

Fig. 1 is a flowchart of the present invention.

Figure 2 is the first infrared spectrum after partitioning.

Figure 3 is the first infrared spectrum after partitioning.

detailed description

The present invention will be further described below in conjunction with the accompanying drawings. The following examples are only used to illustrate the technical solution of the present invention more clearly, but not to limit the protection scope of the present invention.

As shown in Figure 1, the power equipment fault detection method based on deep neural network includes the following steps:

Step 1: collect infrared spectra, including infrared spectra for testing and modeling, and construct a standardized spectral database.

The specific process is as follows:

11) Research on non-contact infrared spectrum collection, infrared spectrum collected at multiple different viewpoints (generally greater than 6 different viewpoints), to ensure the diversity of equipment angles in the collected infrared spectrum, and based on vision and related background. The infrared spectrum is filtered to lay the foundation for the construction of a standardized spectral library;

12) Analyze the collected infrared spectrum, select the target features representing the image, define the device representation type in the infrared spectrum, and construct a standardized spectrum database.

Step 2, establish a multi-layer deep neural network to classify the infrared spectra for testing in the normalized spectral database.

The process of building a multi-layer deep neural network is:

21) The infrared spectrum for modeling in the spectrum database is divided into a training set and a test set;

22) Construct a multi-layer deep neural network system, input the training set, set the corresponding input activation ^ax,l for each training sample x, carry out forward propagation of data, and calculate the layer parameters of each layer, the specific formula is as follows:

z ^x,l =w ^l a ^x,l-1 +b ^l

a ^x,l-1 = σ(z ^x,l-1 )

Among them, z ^{x, l} is the output of the l-layer neural network unit, z ^{x, l-1} is the output of the l-1 layer neural network unit, w ^l is the weight of the l-layer neural network unit, and a ^{x, l-1} is The input of the l-layer neural network unit, b ^l is the bias of the l-layer neural network unit, σ is the sigmoid function, and l is the number of layers of the neural network;

23) Calculate the output error δ ^x,l , through the backpropagation algorithm, and use the gradient descent rule for each l according to with To update the weight and configuration, so that the value of the cost function tends to 0, where η, m are the learning rate of the neural network and the total number of samples in the training set;

24) During the iterative process, observe the change of the loss value to judge the convergence, adjust the learning rate, and generate a multi-layer deep neural network model including the parameters of each layer.

The above multi-layer deep neural network system uses typical AlexNet and CaffeNet to train and test the infrared spectrum; among them, the definition of training parameters in the process of CaffeNet training infrared spectrum has a great influence on the final convergence and accuracy of the network; for momentum The definition of parameters comes from the inertia in Newton's first law. The basic idea is that when there is a flat area in the error surface, the gradient descent algorithm SGD can learn at a faster speed, and its weight changes as W is the weight of the neural network, and E is the cost function of the neural network; the process is a process of continuously changing W to achieve the global optimum based on gradient derivation; the definition of the lr_mult parameter is the learning rate, and the learning rate for setting the bias It is twice the learning rate of weight; the definition of dacay_mult parameter is weight attenuation. In order to avoid the over-fitting that often occurs in deep learning models, it is necessary to add a normative item λ to the cost function (cost function) to limit Overfitting, the formula becomes That is, regular terms are added to the original SGD process, and for the cost function, the cross entropy cost function is used Instead of the quadratic cost function, where C is the value of the cost function, which is used to adjust the parameters in the network during the SGD process, a is the actual input of the training sample x after being processed by the input activation, and the sum is over all training inputs It is carried out, and y is the corresponding target output, which solves the problem of slow learning after multiple iterations of the quadratic cost function.

Step 3, establish a target detection framework RCNN to perform device partitioning on the classified testing infrared spectrum.

The specific process of device partitioning is as follows:

31) Build a target detection framework, and use the RCNN target retrieval method to extract the area and position of the backbone equipment in the infrared atlas for the classification test;

32) Carry out internal partitioning of the extracted backbone device by means of Label parting.

The above process mainly uses the method of Fast RCNN and part labeling to establish a method for device image partitioning. Fast RCNN is mainly used for rectangular calibration of the overall outline of the device, and part labeling is to further partition the interior of the device.

Fast RCNN transforms the positioning problem in the infrared spectrum into a regression problem (regression problem), and finally adopts the strategy of Recognition using region. In each test process, a vector machine (SVM) is trained, and a score is defined for each partition according to the output feature class judgment, and this process is handled by accepting or rejecting a partition. Finally, generate a map change detection model, extract the main device partition in the infrared map, combine the method of target segmentation, and refer to the method part labeling mentioned in Hypercolumns for Object Segmentation and Fine-grained Localization to achieve further partitioning inside the device. As shown in Figures 2 and 3, RO1, RO2, and RO3 are partitions.

Step 4, according to the defect analysis criteria of the infrared spectrum, set the temperature warning rules in the partition.

According to the defect analysis criteria of the infrared spectrum, the pixel points in the partition are extracted, and the extracted pixel points are converted into temperatures according to the conversion rules of pixel points and temperature components, and the temperature warning rules in the partition are set according to the temperature.

If it is a one-way device, calculate the average temperature in the zone and generate a single device temperature warning rule; if it is a three-item device, calculate the temperature difference between items in the zone to generate a three-item device temperature warning rule.

Step 5, according to the temperature warning rules, get the equipment safety analysis conclusion and realize the warning.

The above method realizes the application of image intelligent perception in power detection. Through the non-contact technology of map collection, a standardized image database is constructed, and a multi-layer deep neural network is established to realize the classification of infrared images of typical power equipment. Then, through the latest The target detection framework RCNN realizes the intelligent partitioning of the infrared spectrum, refers to the infrared defect analysis criteria, gives the safety analysis conclusion of large-scale power equipment and realizes intelligent early warning. The realization of the present invention classifies and partitions the infrared images of the power grid through the algorithm of the neural network, which greatly reduces the amount of manual operations of the power grid staff, improves the efficiency of the power grid staff, and reduces cumbersome time.

The above is only a preferred embodiment of the present invention, it should be pointed out that for those of ordinary skill in the art, without departing from the technical principle of the present invention, some improvements and modifications can also be made. It should also be regarded as the protection scope of the present invention.

Claims (7)

1. The power equipment fault detection method based on deep neural network, is characterized in that: comprising Collect infrared spectra and build a standardized spectral database; the collected infrared spectra include infrared spectra for testing and infrared spectra for modeling; Establish a multi-layer deep neural network to classify the infrared spectra for testing in the normalized spectral database; Establish the target detection framework RCNN to partition the infrared spectrum for the classification test; According to the infrared spectrum defect analysis criteria, set the temperature warning rules in the partition; According to the temperature early warning rules, the equipment safety analysis conclusions are obtained and the early warning is realized. 2. The power equipment fault detection method based on deep neural network according to claim 1, characterized in that: the process of constructing a standardized map database is: analyzing the infrared map collected, selecting the target feature representing the image, defining the infrared The type of equipment in the map is represented, and a standardized map database is constructed. 3. the power equipment fault detection method based on deep neural network according to claim 2, is characterized in that: in the infrared atlas collected at a plurality of different viewpoints, before constructing the normalized atlas database, based on vision and relevant background to collection Infrared spectrum was filtered. 4. the power equipment fault detection method based on deep neural network according to claim 1, is characterized in that: the process of setting up multilayer deep neural network is, Divide the infrared spectrum for modeling in the spectrum database into a training set and a test set; Construct a multi-layer deep neural network system, input the training set, set the corresponding input activation a ^x,l for each training sample x, carry out the forward propagation of data, and calculate the layer parameters of each layer. The specific formula is as follows: z ^x,l =w ^l a ^x,l-1 +b ^l a ^x,l-1 = σ(z ^x,l-1 ) Among them, z ^{x, l} is the output of the l-layer neural network unit, z ^{x, l-1} is the output of the l-1 layer neural network unit, w ^l is the weight of the l-layer neural network unit, and a ^{x, l-1} is The input of the l-layer neural network unit, b ^l is the bias of the l-layer neural network unit, σ is the sigmoid function, and l is the number of layers of the neural network; Calculate the output error δ ^x,l , through the backpropagation algorithm, by the gradient descent rule for each l according to with To update the weight and configuration, so that the value of the cost function tends to 0, where η, m are the learning rate of the neural network and the total number of samples in the training set; In the iterative process, observe the change of the loss value to judge the convergence, adjust the learning rate, and generate a multi-layer deep neural network model including the parameters of each layer. 5. the electric equipment fault detection method based on deep neural network according to claim 1, is characterized in that: the process of equipment division is, Build a target detection framework, and use the RCNN target retrieval method to extract the area and position of the backbone equipment in the infrared atlas for classification testing; Use the Label parting method to partition the extracted backbone device inside the device. 6. The power equipment fault detection method based on deep neural network according to claim 1, characterized in that: according to the infrared spectrum defect analysis criteria, extract the pixels in the partition, according to the conversion rules of pixels and temperature components, extract The pixel points are converted into temperature, and the temperature warning rules in the partition are set according to the temperature. 7. The power equipment fault detection method based on deep neural network according to claim 6, characterized in that: if it is a one-way equipment, then calculate the average temperature in the partition to generate a single equipment temperature warning rule, if it is a three-item equipment , then calculate the temperature difference between items in the partition, and generate three early warning rules for equipment temperature.