A method for defect identification of power transmission and transformation equipment based on artificial intelligence

Abstract

The present invention relates to a method for identifying defects in power transmission and transformation equipment based on artificial intelligence, and belongs to the field of identifying defects in power transmission and transformation equipment. The method comprises the following steps: step 1, data collection and labeling; step 2, data expansion; step 3, data preprocessing; step 4, feature extraction; step 5, training classifier; step 6, identification and classification of defects in power transmission and transformation equipment. The method of the present invention accurately and effectively identifies and classifies defects in power transmission and transformation equipment, and provides guarantee for the stable operation of the power system.

Description

Technical Field

[0001] The present invention belongs to the field of power transmission and transformation equipment defect identification, and in particular relates to a power transmission and transformation equipment defect identification method based on artificial intelligence. Background Art

[0002] Power transmission and transformation equipment plays a vital role in the power system. Both transmission equipment and transformation equipment are important components of the power grid. Transmission equipment mainly includes transmission lines, power equipment, etc. It is the main way of power transmission. Because it spans a wide area and is often exposed to various natural environments and harsh climatic conditions, it is prone to various defects in transmission equipment. Transformation equipment mainly includes transformers and other equipment, which are responsible for the conversion of power at different voltage levels. After decades of rapid development, the number of power transmission and transformation equipment has increased and the distribution has become wider. At the same time, with the increase in the number of years of operation, the defects of power transmission and transformation equipment have gradually increased. In order to ensure the safe and stable operation of the power system, timely and accurate discovery and location of potential line problems has become a key task. Traditional inspection methods for power transmission and transformation equipment often rely on manual inspections or simple mechanical equipment, which is not only inefficient, but may also lead to incomplete inspections or omissions due to human factors. With the advancement of science and technology, although technologies such as high-definition cameras and drones have been introduced, due to the limitations of data processing and analysis technology, a large amount of collected image data has not been fully utilized. In addition, traditional data processing and model training methods are prone to fall into local optimality when faced with complex and changeable power transmission and transformation equipment images, resulting in poor recognition results. Therefore, an efficient and accurate power transmission and transformation equipment defect recognition technology is urgently needed to meet the above challenges and improve the operational safety and stability of the power system.

[0003] Chinese invention patent CN202310656311.2 proposes a transmission line monitoring system based on image recognition, including a monitoring module, a positioning module, a communication module, an environmental information acquisition module, an edge computing module and a power supply module; the monitoring module is used to monitor the transmission line; the positioning module is used to obtain the location information of the system; the communication module is used to complete the information transmission between the system and the user end; the environmental information acquisition module is used to obtain the surrounding environment information of the location of the system, and the edge computing module is used to analyze the captured transmission line image and environmental information to monitor the abnormal conditions of the transmission line; the power supply module is used to power the system; while monitoring the abnormalities of the real-time transmission line, it can also give early warning of possible abnormalities of the transmission line in the future.

[0004] Chinese invention patent CN202010624342.6 proposes a method for panoramic reconstruction of substation equipment inspection images. This invention aims to solve the problem that a large number of inspection images obtained by various inspection methods of the substation cannot fully present the panoramic information of the equipment. According to the complex, angular and regular appearance of the equipment in the substation, a method for panoramic reconstruction of substation equipment inspection images is provided. This method uses an improved SURF feature detection algorithm to extract image features based on the characteristics of the structure and appearance of the substation equipment, and combines multiple matching methods to complete efficient matching between inspection images, and finally achieves a good image fusion effect.

[0005] Chinese invention patent CN202310738565.9 proposes a method for identifying transmission tower corrosion based on a fully convolutional network, which includes the following steps: Step 1: Obtain an image of a transmission tower under actual service conditions; Step 2: Perform image preprocessing on the captured transmission tower image to establish a transmission tower corrosion identification data set; Step 3: Build a transmission tower corrosion identification network model based on a fully convolutional network, select multiple sets of hyperparameters for training, and select the optimal hyperparameters as the training result; Step 4: Input the transmission tower corrosion image into the model for identification, and determine the area of ​​corrosion in the image. The application of this technical solution can quickly identify the corrosion area of ​​the transmission tower.

[0006] The existing technology has the following deficiencies:

[0007] 1. Data expansion problem: Existing technologies may rely too much on traditional means for data expansion and fail to make full use of generative adversarial networks, especially the dynamic characteristics of echo state networks for data expansion, resulting in the generated samples being single or not realistic enough.

[0008] 2. Fixed learning rate: Traditional models may use a fixed or manually adjusted learning rate without dynamically adjusting the learning rate based on the output of the model, which may lead to an inefficient training process or slow model convergence.

[0009] 3. Simple data preprocessing: Existing technologies may only use simple image processing or feature extraction methods without utilizing multi-scale kernelized image transformation and entropy weight-based fusion methods, which may miss or ignore some important image features.

[0010] 4. Easy to fall into local optimum: Traditional neural networks may rely mainly on gradient descent or its variants for optimization, which can easily lead to local optimum rather than global optimum, resulting in poor performance of the model.

[0011] 5. Static topology structure: Traditional self-organizing map classification algorithms may use static or fixed topology structures, and do not make dynamic adjustments based on the diversity of power transmission and transformation equipment defects and the characteristics of the data, resulting in inaccurate classification results. Summary of the invention

[0012] The purpose of the present invention is to provide a method for identifying defects in power transmission and transformation equipment based on artificial intelligence. The neural network based on elastic fluctuation optimization has a higher probability of finding the global optimal solution in the feature extraction and model training stages, thereby improving the accuracy of the model. By enhancing the topological self-organizing map classification algorithm, the defects of power transmission and transformation equipment can be more accurately and effectively identified and classified, thereby providing a guarantee for the stable operation of the power system.

[0013] A method for identifying defects in power transmission and transformation equipment based on artificial intelligence comprises the following steps:

[0014] Step 1: Data collection and annotation;

[0015] Step 2: Data expansion;

[0016] Step 3: Data preprocessing;

[0017] Step 4: Feature extraction;

[0018] Step 5: Train the classifier;

[0019] Step 6: Identify and classify defects in power transmission and transformation equipment.

[0020] Preferably, the step one specifically includes: collecting and labeling image data of power transmission and transformation equipment, the data format is RGB image, the size of each image is W×H×3, wherein W and H represent the width and height of the image respectively, and 3 represents the three RGB channels.

[0021] Preferably, the image data of the power transmission and transformation equipment comes from a drone or a high-definition camera;

[0022] Preferably, the step 2 is specifically: generating samples based on a generative adversarial network of echo states, comprising the following steps:

[0023] S2.1 initialization, that is, initializing the weights and structures of the generator G, the discriminator D, and the echo state network ESN;

[0024] S2.2 Echo state mapping, input the image data into the echo state network ESN, and map the input image data I to the echo state E. This mapping is done by inputting the weight matrix W in and the recurrent connection weight matrix W of the echo state network res Completed, represented by:

[0025] E = tanh(W in ×I+W res ×E t-1 )

[0026] Among them, Et-1 is the echo state of the previous time step; tanh() is the hyperbolic tangent activation function; I is the input image data, the size is m×n, where m is the height of the image and n is the width of the image; E is the state matrix of the echo state, the size is p×q;

[0027] S2.3 uses the generator G to generate new image data based on the echo state. The generator uses the echo state E and the output weight matrix W out Generate a new image I gen , expressed as:

[0028] I gen =G(E×W out )

[0029] Among them, W out is the output weight matrix of ESN; G() is the generator function;

[0030] S2.4 The discriminator D evaluates both the real image data and the generated image data. The discriminator evaluates the real image and the generated image and outputs the identification result D real and D gen , expressed as:

[0031] D real =D(I)

[0032] D gen =D(I gen )

[0033] Where D() is the discriminator function;

[0034] S2.5 Optimization and feedback: Based on the output of the discriminator D, the loss is calculated and used to update the weight of the generator G. At the same time, the difference between the real data and the generated data is used to update the output weight of the ESN.

[0035] The entropy of the discriminator output is defined as:

[0036] H(D)=-D real log(D real )-(1-D real )log(1-D real )

[0037] Among them, H(D) is the entropy output by the discriminator, which is used to measure the quality of the generated image;

[0038] In order to dynamically adjust the learning rates of the generator and discriminator, the following rules are defined:

[0039] α=α0e -κH(D)

[0040] β=β0(1-e-κH(D) )

[0041] Among them, α0 and β0 are the initial learning rates; κ is a regulation factor;

[0042] The loss function optimization process is:

[0043] For the generator:

[0044]

[0045] For the discriminator:

[0046]

[0047] Among them, L D is the loss of the discriminator; L G is the loss of the generator; is the gradient symbol, which is calculated by multiple iterative changes; according to the above loss, the weights of the generator and discriminator, as well as the output weight W of ESN are updated using the gradient descent method out ;

[0048] S2.6 performs iterative training until a termination condition is met, wherein the termination condition is reaching a preset maximum number of iterations;

[0049] S2.7 uses the trained generator G and ESN to jointly generate a large amount of image data as a result of data expansion.

[0050] Preferably, the step three specifically includes: performing data preprocessing on the expanded training data set, including:

[0051] S3.1 Scale selection, determine the scale range;

[0052] S3.2 Entropy calculation: The entropy value of the image at each scale is calculated to represent the amount of information at that scale. For an image at a certain scale s, the entropy is calculated as:

[0053]

[0054] Where H(s) is the image entropy at scale s; p(i|s) is the probability that the pixel value is i at scale s; L is the number of gray levels of the image;

[0055] S3.3 Entropy normalization: normalize the entropy values ​​of all scales to ensure that their sum is 1. The normalization method is expressed as:

[0056]

[0057] Among them, w(s) is the entropy weight under scale s; S is the total number of selected scales;

[0058] S3.4 Image kernelization: Use the Gaussian kernel function to map the image at each scale to a high-dimensional feature space. The kernel value K(x,y) of the Gaussian kernel function is defined as:

[0059]

[0060] Where x and y are two pixels in the image; σ is the width parameter of the Gaussian kernel function;

[0061] For an image at scale s, calculate the kernel value between each pixel and its neighboring pixels, and then add these kernel values ​​to get the new pixel value;

[0062] S3.5 weighted fusion uses the normalized entropy value as the weight to perform weighted fusion of kernelized images at all scales. For a certain pixel position p, its value I in the fused image ′ (p) is:

[0063]

[0064] Among them, I s (p) is the value of the image at pixel position p under scale s;

[0065] S3.6 performs feature extraction and uses the scale-invariant feature transform (SIFT) method to extract key points and descriptors on the fused high-dimensional image.

[0066] Preferably, in the SIFT algorithm, a 128-dimensional descriptor is extracted for each key point. If a total of N key points are extracted, the length of the final one-dimensional vector L is ′ , providing input data for the subsequent feature extraction of power transmission and transformation equipment defect identification:

[0067] L ′ =N×128.

[0068] Preferably, the step 4 specifically comprises: using a neural network to extract features from the preprocessed data,

[0069] Define the structure of the neural network. When extracting features, the multi-layer neural network used has L hidden layers. th There are hidden layers, and the weights and biases of the layers are represented by W (l) and b (l) ;

[0070] Define the parameter space of the neural network, and assume that the parameter space of the neural network is Any point P i A set of weights w i and a set of biases b i Defined as:

[0071]

[0072] Where N is the number of points in the parameter space;

[0073] The neural network loss function is defined as L(w,b), which is a fluctuation loss function, expressed as:

[0074]

[0075] Among them, y and They are the target output and actual output of the feature input to the preset Softmax classifier after neural network feature extraction, N1 is the number of input samples, and t is the number of current training iterations; α L and β L is the amplitude coefficient of the cosine and sine waves; ω and φ are the frequencies of the cosine and sine waves;

[0076] Define elastic constants and fluctuation speeds, including elastic constants κ and fluctuation speeds v, which are associated with the weights w and biases b of the network;

[0077] Define the elastic fluctuation update rules. The elastic fluctuation update process is the process of updating the neural network parameters. Define a set of elastic fluctuation update rules as follows:

[0078]

[0079] Among them, Δt is the time step; Δw and Δb are the change parameters of weight and bias; w i+1 and b i+1 are the updated weights and biases;

[0080] The oscillation adjustment mechanism is defined. The oscillation adjustment is the adjustment strategy of the elastic constant κ and the fluctuation velocity v, which can be expressed as:

[0081]

[0082] Among them, κ0 and v0 are the initial elastic constant and fluctuation speed, β z and γ z is the adjustment factor.

[0083] Preferably, in the training process of the neural network, the training steps are as follows:

[0084] 1. Initialize w, b, κ, and v, and set the time step Δt;

[0085] 2. Calculate the initial loss function L(w,b);

[0086] 3. For each iteration i;

[0087] 3.1. Update w and b using the elastic fluctuation update rule. For the lth th The weight of the layer W (l) and bias b (l) , using the following elastic fluctuation update rule:

[0088]

[0089] 3.2. Apply the oscillation adjustment mechanism to update κ and v. The update of the oscillation adjustment mechanism for each layer l is expressed as:

[0090]

[0091] in, and It is the first th The initial elastic constants and fluctuation velocities of the layers;

[0092] 3.3. Recalculate the loss function L(w,b);

[0093] 4. Repeat iterations 3.1 to 3.3 until the training reaches convergence and the algorithm terminates;

[0094] The convergence is judged in such a way that the change ΔL of the loss value does not change after a preset number of iterations, that is, convergence is achieved; the calculation basis of the change ΔL of the loss value is that during the iteration process, the state transition of the algorithm between any iteration i and i+1 is considered; the change ΔL of the loss function is expressed as:

[0095] ΔL=L(w i+1 ,b i+1 )-L(w i ,b i )

[0096] According to the elastic fluctuation update rule and Taylor series expansion, we have:

[0097]

[0098] Substituting the update rule into the above equation, we get:

[0099]

[0100] Because ΔL is negative or zero, the algorithm will ensure that the loss function L(w,b) decreases or remains unchanged;

[0101] After the algorithm terminates, it means that the training of the feature extraction neural network model is completed.

[0102] Preferably, the step five is specifically:

[0103] Input the extracted features into the classifier for classification;

[0104] Initialization operation, that is, initializing the weights and topological structure of SOM; assigning a weight vector w to each neuron n n , the dimension is the same as the input data dimension, and the weights are randomly initialized to small values ​​in the range [0,1], expressed as:

[0105] w n =[w n1 ,w n2 ,...,w nd ]

[0106] Where d is the dimension of the data;

[0107] The one-dimensional vector data after feature extraction is input into the SOM network. Assume that the input one-dimensional vector is x and the dimension is d, which is expressed as:

[0108] x=[x1,x2,...,x d ]

[0109] For each input sample, find the best matching neuron BMU and calculate the Euclidean distance D(n) of each neuron:

[0110]

[0111] Select the neuron with the smallest distance as the BMU:

[0112]

[0113] Topology adjustment: dynamically adjust the topology of SOM according to the density and characteristics of the defect data of the power transmission and transformation equipment, and set the characteristic value ρ n Represents the data density near neuron n, and the update method is expressed as:

[0114] ρ n =α×h(BMU,n)+(1-α)×ρ n

[0115] Among them, α is the learning rate, h() is the neighborhood function;

[0116] Weight update, local perception adjustment, according to the local perception ability of each neuron, adjust its weight, introduce the perception function P(n), expressed as:

[0117]

[0118] Among them, β is the sensitivity parameter, ρ threshold is the density threshold, and the weights of neurons in the BMU and its neighborhood are updated. The way to update the weights of neurons is expressed as:

[0119] w n+1 '=w n +η×P(n)×(xw n )

[0120] Among them, w n is the weight before update, w n+1 ' is the updated weight, η is the weight adjustment coefficient;

[0121] Historical information feedback: Based on the previous historical information, the weights are further fine-tuned, and the fine-tuning amount δw n It is expressed as:

[0122] δw n =γ×(w n,prev -w n )

[0123] Among them, γ is the historical information feedback coefficient; w n,prev is the historical weight, then the method of updating the weight according to the historical information feedback is expressed as:

[0124] w n+1 ”=w n+1 '+δw n

[0125] Among them, w n+1 ” is the weight updated according to historical information feedback;

[0126] Adaptive topological connectivity adjustment, for the nth neuron, define a magnetization index M n It represents its ability to attract other neurons and is initialized to 0. Whenever the nth neuron is selected as a BMU, its magnetization index is updated according to the following formula:

[0127]

[0128] Where λ is an attenuation coefficient between 0 and 1, ∈ ′ is a small value to prevent division by zero, M n+1 is the updated magnetization index;

[0129] The degree of connectivity between neurons is determined by the relative magnitude of the magnetization index. The connection strength C between two neurons n and m is nm Defined as:

[0130]

[0131] When C nm Exceeding a threshold C threshold, then it is considered that there is a strong connection between neuron n and neuron m; otherwise, it is a weak connection; then in the process of weight update, the weight update formula is:

[0132] w n+1 '=w n+1 +η×C BMU,n ×h(BMU,n)×(xw n )

[0133] Among them, w n+1 ' is the updated weight after the adaptive topology connectivity is adjusted;

[0134] The judgment of the termination condition is that when the average weight change in several consecutive rounds is less than the preset threshold ∈ or the maximum number of iterations is reached, the training is stopped.

[0135] Preferably, the step six is ​​specifically:

[0136] The above model is trained using training samples, and the trained model is used to identify and classify defects in power transmission and transformation equipment.

[0137] Beneficial technical effects:

[0138] 1. Data richness: By using echo-state-based generative adversarial networks for data augmentation, more and higher-quality training samples can be obtained, which helps improve the performance of the model.

[0139] 2. Optimized data preprocessing: Multi-scale kernelized image transformation based on entropy weight can better extract the features of power transmission and transformation equipment, which is helpful for subsequent identification and classification.

[0140] 3. Avoid local optimality: The neural network based on elastic fluctuation optimization has a higher probability of finding the global optimal solution during the feature extraction and model training stages, thereby improving the accuracy of the model.

[0141] 4. Efficient defect identification: By enhancing the topology of the self-organizing map classification algorithm, the defects of power transmission and transformation equipment can be identified and classified more accurately and effectively, providing guarantee for the stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0142] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0143] Figure 1 Schematic diagram of the neural network structure of an embodiment of the present invention.

[0144] Figure 2 Schematic diagram of model training according to an embodiment of the present invention. DETAILED DESCRIPTION

[0145] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0146] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or device including the element.

[0147] Embodiment 1

[0148] The present invention proposes a method for identifying defects in power transmission and transformation equipment based on artificial intelligence, the main steps of which are as follows:

[0149] Step 1: Data collection and annotation

[0150] Image data collection and annotation are performed on power transmission and transformation equipment. The data format is RGB image. The size of each image is W×H×3, where W and H represent the width and height of the image respectively, and 3 represents the three RGB channels.

[0151] Optionally, the image data of the power transmission and transformation equipment of the present invention comes from a drone or a high-definition camera, which is obtained when inspecting the power transmission and transformation equipment, and these images capture the real-time status of the power transmission and transformation equipment under various environments and weather conditions.

[0152] Step 2: Data expansion

[0153] It is understandable that the collection and annotation of images is time-consuming and labor-intensive, and the insufficient amount of data makes it difficult for the artificial intelligence model to converge quickly. The present invention proposes a generative adversarial network based on echo states for sample generation, thereby achieving data expansion. The traditional generative adversarial network (GAN) uses a generator and a discriminator for adversarial training. The goal of the generator is to generate data similar to the real data set, while the goal of the discriminator is to distinguish between real data and generated data. The echo state-based generative adversarial network model uses the dynamic characteristics of the echo state network (ESN) to capture the temporal and spatial dependencies in the image data, making the generated image data richer and more diverse. The present invention refers to the echo phenomenon in nature and simulates multiple "reflections" of data in the network, so that the data produces rich nonlinear changes in the network.

[0154] Preferably, the process of generating samples based on the echo state generative adversarial network includes the following steps:

[0155] S2.1 Initialization, that is, initializing the weights and structure of the generator G, discriminator D, and echo state network ESN.

[0156] S2.2 Echo state mapping, the image data is input to the echo state network ESN, and its sparse, fixed recurrent connections are used to generate echo states that capture the spatiotemporal dependencies in the image. The input image data I is mapped to the echo state E. This mapping is done by inputting the weight matrix W in and the recurrent connection weight matrix W of the echo state network res Completed, represented by:

[0157] E = tanh(W in ×I+W res ×E t-1 )

[0158] Among them, E t-1 is the echo state of the previous time step; tanh() is the hyperbolic tangent activation function; I is the input image data, with a size of m×n, where m is the height of the image and n is the width of the image; E is the state matrix of the echo state, with a size of p×q.

[0159] S2.3 uses the generator G to generate new image data based on the echo state. The generator uses the echo state E and the output weight matrix W out Generate a new image I gen , expressed as:

[0160] I gen =G(E×W out )

[0161] Among them, W outis the output weight matrix of ESN; G() is the generator function.

[0162] S2.4 The discriminator D evaluates both the real image data and the generated image data. The discriminator evaluates the real image and the generated image and outputs the identification result D real and D gen , which can be expressed as:

[0163] D real =D(I)

[0164] D gen =D(I gen )

[0165] Among them, D() is the discriminator function.

[0166] S2.5 Optimization and Feedback, based on the output of the discriminator D, calculate the loss and use this loss to update the weight of the generator G; at the same time, use the difference between the real data and the generated data to update the output weight of the ESN so that it can better capture the dynamic characteristics of the image. The loss is an improved cross entropy loss function.

[0167] Specifically, in order to enhance the performance of ESGAN, the present invention proposes an entropy-based dynamic adjustment mechanism, which can automatically adjust the training rates of the generator and the discriminator, so that the network can learn quickly in the early stage and pay more attention to details in the later stage, thereby avoiding overfitting.

[0168] The entropy of the discriminator output is defined as:

[0169] H(D)=-D real log(D real )-(1-D real )log(1-D real )

[0170] Among them, H(D) is the entropy output by the discriminator, which is used to measure the quality of the generated image.

[0171] Furthermore, the entropy H(D) reflects the difference between the generated image and the real image. When the generated image gets closer to the real image, the entropy decreases.

[0172] In order to dynamically adjust the learning rates of the generator and discriminator, the following rules are defined:

[0173] α=α0e -κH(D)

[0174] β=β0(1-e -κH(D) )

[0175] Among them, α0 and β0 are the initial learning rates; κ is a regulating factor. When H(D) increases, it means that the difference between the generated image and the real image is large. At this time, the learning rate of the generator is increased and the learning rate of the discriminator is reduced. Conversely, the learning rate of the generator is reduced and the learning rate of the discriminator is increased.

[0176] According to the above dynamic adjustment mechanism, the loss function optimization process is:

[0177] For the generator:

[0178]

[0179] For the discriminator:

[0180]

[0181] Among them, L D is the loss of the discriminator; L G is the loss of the generator; is the gradient symbol, which is calculated by multiple iterative changes; according to the above loss, the weights of the generator and discriminator, as well as the output weight W of ESN are updated using the gradient descent method out .

[0182] S2.6 performs iterative training until a termination condition is met, wherein the termination condition is reaching a preset maximum number of iterations.

[0183] S2.7 uses the trained generator G and ESN to jointly generate a large amount of image data as a result of data expansion.

[0184] Step 3: Data preprocessing

[0185] The expanded training data set is subjected to data preprocessing. The present invention proposes a data preprocessing method based on entropy weighted multi-scale kernelized image transformation (EMKIT) to fuse the information of the image at multiple scales. The fusion is not a simple superposition of information, but a weighting is assigned to different scales by calculating the entropy at each scale, and then weighted fusion is performed. In addition, in order to further extract the local features of the image, the present invention uses the kernelization technique to map the original image into a high-dimensional feature space. In this high-dimensional space, the local features of the image can be further amplified and can provide more information about the defects of the power transmission and transformation equipment.

[0186] S3.1 Scale selection, determining a scale range, for example, from 1 to 5, each scale represents a different pixel neighborhood size. Further preferably, for each scale, a bilinear interpolation method is used to resample the original image.

[0187] S3.2 Entropy calculation: The entropy value of the image at each scale is calculated to represent the amount of information at that scale. For an image at a certain scale s, the entropy is calculated as:

[0188]

[0189] Where H(s) is the image entropy at scale s; p(i|s) is the probability that the pixel value is i at scale s; L is the number of gray levels of the image. For 8-bit images, L = 256.

[0190] S3.3 Entropy normalization, normalize the entropy values ​​of all scales to ensure that their sum is 1. In order to convert the entropy values ​​at different scales into weights, the normalization method is expressed as:

[0191]

[0192] Among them, w(s) is the entropy weight at scale s; S is the total number of selected scales.

[0193] S3.4 Image kernelization: Use the Gaussian kernel function to map the image at each scale to a high-dimensional feature space. The kernel value K(x,y) of the Gaussian kernel function is defined as:

[0194]

[0195] Among them, x and y are two pixels in the image; σ is the width parameter of the Gaussian kernel function, which determines the "width" of the kernel function.

[0196] For an image at scale s, the kernel value between each pixel and its neighboring pixels is calculated, and then these kernel values ​​are added together to obtain the new pixel value.

[0197] S3.5 weighted fusion uses the normalized entropy value as the weight to perform weighted fusion of kernelized images at all scales. For a certain pixel position p, its value I in the fused image ′ (p) is:

[0198]

[0199] Among them, I s (p) is the value of the image at pixel position p under scale s.

[0200] S3.6 performs feature extraction. On the fused high-dimensional image, SIFT (Scale Invariant Feature Transform) method is used to extract key points and descriptors. Assume that in the SIFT algorithm, a 128-dimensional descriptor is extracted for each key point. If a total of N key points are extracted, the length of the final one-dimensional vector L is ′ , providing input data for the subsequent feature extraction of power transmission and transformation equipment defect identification:

[0201] L ′ =N×128.

[0202] Step 4: Feature extraction

[0203] The present invention uses a neural network to extract features from preprocessed data. Traditional neural networks usually use a gradient descent optimization algorithm during training. During feature extraction, the gradient descent optimization algorithm is prone to fall into local optimality, resulting in poor feature extraction effect. Inspired by the elastic fluctuation principle, the elastic fluctuation principle is based on the natural phenomenon that elastic objects (such as springs, rubber bands, etc.) produce fluctuations after being acted upon by external forces and eventually tend to a steady state. This phenomenon can be used as a reference method for parameter optimization. Therefore, the present invention proposes a neural network based on elastic fluctuation optimization. Different from the gradient-based optimization algorithm, it can effectively avoid falling into the local optimal solution during neural network training.

[0204] Define the structure of the neural network. When extracting features, the multi-layer neural network used has L hidden layers. In a specific embodiment, L is 3. th There are hidden layers, and the weights and biases of the layers are represented by W (l) and b (l) ,like Figure 1 As shown:

[0205] Define the parameter space of the neural network, and assume that the parameter space of the neural network is Any point P i It can be represented by a set of weights w i and a set of biases b i Definition, then, is expressed as:

[0206]

[0207] P i =(w i ,b i )

[0208] Where N is the number of points in parameter space.

[0209] The neural network loss function is defined as L(w,b), which is a fluctuation loss function, expressed as:

[0210]

[0211] Among them, y and They are the target output and actual output of the feature input to the preset Softmax classifier after neural network feature extraction, N1 is the number of input samples, and t is the number of current training iterations; α L and β Lis the amplitude coefficient of the cosine and sine waves; ω and φ are the frequencies of the cosine and sine waves. Different from the traditional cross entropy loss function, when the volatility loss function replaces the conventional loss function in the elastic volatility optimization algorithm, the network will experience different "pressure" and "tension" states during the training process, which helps the network to explore the parameter space more effectively and avoid local optimal solutions.

[0212] The elastic constant and the fluctuation speed are defined, including the elastic constant κ and the fluctuation speed v, which are associated with the weight w and bias b of the network.

[0213] Define the elastic fluctuation update rules. The elastic fluctuation update process is the process of updating the neural network parameters. Define a set of elastic fluctuation update rules as follows:

[0214]

[0215] w i+1 =w i +Δw

[0216] b i+1 =b i +Δb

[0217] Among them, Δt is the time step; Δw and Δb are the change parameters of weight and bias; w i+1 and b i+1 are the updated weights and biases.

[0218] The oscillation adjustment mechanism is defined. The oscillation adjustment is the adjustment strategy of the elastic constant κ and the fluctuation velocity v, which can be expressed as:

[0219]

[0220] Among them, κ0 and v0 are the initial elastic constant and fluctuation speed, β z and γ z is the adjustment factor.

[0221] During the training process of the neural network, the training steps are as follows:

[0222] 1. Initialize w, b, κ, and v, and set the time step Δt.

[0223] 2. Calculate the initial loss function L(w,b).

[0224] 3. For each iteration i:

[0225] 3.1. Update w and b using the elastic fluctuation update rule. For the lth th The weight of the layer W (l) and bias b (l) , using the following elastic fluctuation update rule:

[0226]

[0227] Here, Δt is the time step.

[0228] 3.2. Apply the oscillation adjustment mechanism to update κ and v. The update of the oscillation adjustment mechanism for each layer l is expressed as:

[0229]

[0230] in, and It is the first th Initial elastic constants and fluctuation velocities of the layer.

[0231] 3.3. Recalculate the loss function L(w,b).

[0232] 4. Repeat iterations 3.1 to 3.3 until the training reaches convergence and the algorithm terminates.

[0233] The convergence is judged in the following way: if the change ΔL of the loss value does not change after a preset number of iterations, convergence is achieved. The calculation basis of the change ΔL of the loss value is to consider the state transition of the algorithm between any iteration i and i+1 during the iteration process. The change ΔL of the loss function can be expressed as:

[0234] ΔL=L(w i+1 ,b i+1 )-L(w i ,b i )

[0235] According to the elastic fluctuation update rule and Taylor series expansion, we have:

[0236]

[0237] Substituting the update rule into the above equation, we get:

[0238]

[0239] Because ΔL is negative or zero, the algorithm will ensure that the loss function L(w,b) decreases or remains unchanged.

[0240] After the algorithm terminates, it means that the training of the feature extraction neural network model is completed.

[0241] Step 5: Train the classifier

[0242] The extracted features are input into the classifier for classification. The present invention proposes an enhanced topology self-organizing map classification algorithm (ET-SOM). The traditional SOM is an unsupervised learning algorithm, which is often used for data clustering and visualization. In order to better adapt to the morphological diversity of power transmission and transformation equipment defects, the present invention designs a dynamically adjusted topological structure. This structure can not only self-adjust based on the density of the data, but also make targeted adjustments based on the specific patterns of power transmission and transformation equipment defects.

[0243] Specifically, preferably, the initialization operation is to initialize the weights and topological structure of the SOM; a weight vector w is assigned to each neuron n. n , the dimension is the same as the input data dimension, and the weights are randomly initialized to small values ​​ranging between 0 and 1, which can be expressed as:

[0244] w n =[w n1 ,w n2 ,...,w nd ]

[0245] Here, d is the dimension of the data.

[0246] The one-dimensional vector data after feature extraction is input into the SOM network. Assume that the input one-dimensional vector is x, and the dimension is 1, d, which can be expressed as:

[0247] x=[x1,x2,...,x d ]

[0248] For each input sample, find the best matching neuron BMU and calculate the Euclidean distance D(n) of each neuron:

[0249]

[0250] Select the neuron with the smallest distance as the BMU:

[0251]

[0252] Topology adjustment: dynamically adjust the topology of SOM according to the density and characteristics of the defect data of the power transmission and transformation equipment, and set the characteristic value ρ n Represents the data density near neuron n, and the update method is expressed as:

[0253] ρ n =α×h(BMU,n)+(1-α)×ρ n

[0254] Among them, α is the learning rate and h() is the neighborhood function.

[0255] Weight update, local perception adjustment, based on the principle of chemical reaction, according to the local perception ability of each neuron, adjust its weight, introduce the perception function P(n), expressed as:

[0256]

[0257] Among them, β is the sensitivity parameter, ρ threshold is the density threshold, and the weights of neurons in the BMU and its neighborhood are updated. The way to update the weights of neurons is expressed as:

[0258] w n+1 '=w n +η×P(n)×(xw n )

[0259] Among them, w n is the weight before update, w n+1 ' is the updated weight, and η is the weight adjustment coefficient.

[0260] Historical information feedback: Based on the previous historical information, the weights are further fine-tuned, and the fine-tuning amount δw n It can be expressed as:

[0261] δw n =γ×(w n,prev -w n )

[0262] Among them, γ is the historical information feedback coefficient; w n,prev is the historical weight, then the method of updating the weight according to the historical information feedback is expressed as:

[0263] w n+1 ”=w n+1 '+δw n

[0264] Among them, w n+1 ” is the weight updated based on historical information feedback.

[0265] Adaptive topological connectivity adjustment: The present invention allows the topological connectivity between neurons to be dynamically adjusted, thereby better capturing the nonlinear distribution of data. For the nth neuron, a magnetization index M is defined n It represents its ability to attract other neurons and is initialized to 0. Whenever the nth neuron is selected as a BMU, its magnetization index is updated according to the following formula:

[0266]

[0267] Where λ is an attenuation coefficient between 0 and 1, ∈ ′ is a small value to prevent division by zero, Mn+1 is the updated magnetization index.

[0268] The degree of connectivity between neurons is determined by the relative magnitude of the magnetization index. The connection strength C between two neurons n and m is nm Defined as:

[0269]

[0270] When C nm Exceeding a threshold C threshold , then it is considered that there is a strong connection between neuron n and neuron m; otherwise, it is a weak connection; then in the process of weight update, the weight update formula is:

[0271] w n+1 '=w n+1 +η×C BMU,n ×h(BMU,n)×(xw n )

[0272] Among them, w n+1 ' is the updated weight after the adaptive topology connectivity is adjusted.

[0273] By introducing the magnetization index and adaptive topological connectivity, the algorithm is allowed to dynamically adjust the connection relationship between neurons to better capture the intrinsic structure and distribution of the data, further improving the performance of the classifier.

[0274] The judgment of the termination condition is that when the average weight change is less than the preset threshold ∈ or the maximum number of iterations is reached in several consecutive rounds, the training is stopped. Preferably, it can be expressed as:

[0275]

[0276] That is, if Δw avg <∈, then stop.

[0277] Step 6: Identification and classification of power transmission and transformation equipment defects

[0278] Based on the previous steps, the above model is trained using training samples, and the trained model is used to identify and classify defects in power transmission and transformation equipment. Defect recognition of power transmission and transformation equipment image data aims to timely and accurately discover and locate potential power transmission and transformation equipment problems, and provide guarantee for the stable operation of the power system.

[0279] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0280] It should be noted that the serial numbers of the embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

[0281] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, and any combination of embodiments or schemes are also included in the patent protection scope of the present invention.

Claims

1. A method for identifying defects in power transmission and transformation equipment based on artificial intelligence, characterized in that: The steps include: Step 1: Data collection and annotation; Step 2: Data expansion; Step 3: Data preprocessing; Step 4: Feature extraction; Step 5: Train the classifier; Step 6: Identify and classify defects in power transmission and transformation equipment.

2. The method according to claim 1, characterized in that The step 1 is specifically as follows: collecting and labeling image data of power transmission and transformation equipment, the data format is RGB image, the size of each image is W×H×3, where W and H represent the width and height of the image respectively, and 3 represents the three channels of RGB.

3. The method according to claim 2, characterized in that The image data of power transmission and transformation equipment comes from drones or high-definition cameras.

4. The method according to claim 1, characterized in that The step 2 is specifically: generating samples based on the generative adversarial network of the echo state, including the following steps: S2.1 initialization, that is, initializing the weights and structures of the generator G, the discriminator D, and the echo state network ESN; S2.2 Echo state mapping, input the image data into the echo state network ESN, and map the input image data I to the echo state E. This mapping is done by inputting the weight matrix W in and the recurrent connection weight matrix W of the echo state network res Completed, represented by: E=tanh(W in ×I+W res ×E t-1 ) Among them, E t-1 is the echo state of the previous time step; tanh() is the hyperbolic tangent activation function; I is the input image data, the size is m×n, where m is the height of the image and n is the width of the image; E is the state matrix of the echo state, the size is p×q; S2.3 uses the generator G to generate new image data based on the echo state. The generator uses the echo state E and the output weight matrix W out Generate a new image I gen , expressed as: I gen =G(E×W out ) Among them, W out is the output weight matrix of ESN; G() is the generator function; S2.4 The discriminator D evaluates both the real image data and the generated image data. The discriminator evaluates the real image and the generated image and outputs the identification result D real and D gen , expressed as: D real =D(I) D gen =D(I gen ) Where D() is the discriminator function; S2.5 Optimization and feedback: Based on the output of the discriminator D, the loss is calculated and used to update the weight of the generator G. At the same time, the difference between the real data and the generated data is used to update the output weight of the ESN. The entropy of the discriminator output is defined as: H(D)=-D real log(D real )-(1-D real )log(1-D real ) Among them, H(D) is the entropy output by the discriminator, which is used to measure the quality of the generated image; In order to dynamically adjust the learning rates of the generator and discriminator, the following rules are defined: α＝α0e -κH(D) β＝β0(1-e -κH(D) ) Among them, α0 and β0 are the initial learning rates; κ is a regulation factor; The loss function optimization process is: For the generator: For the discriminator: Among them, L D is the loss of the discriminator; L G is the loss of the generator; is the gradient symbol, which is calculated by multiple iterative changes; according to the above loss, the weights of the generator and discriminator, as well as the output weight W of ESN are updated using the gradient descent method out ; S2.6 performs iterative training until a termination condition is met, wherein the termination condition is reaching a preset maximum number of iterations; S2.7 uses the trained generator G and ESN to jointly generate a large amount of image data as a result of data expansion.

5. The method according to claim 1, characterized in that The step three is specifically: preprocessing the expanded training data set, including: S3.1 Scale selection, determine the scale range; S3.2 Entropy calculation: The entropy value of the image at each scale is calculated to represent the amount of information at that scale. For an image at a certain scale s, the entropy is calculated as: Where H(s) is the image entropy at scale s; p(i|s) is the probability that the pixel value is i at scale s; L is the number of gray levels of the image; S3.3 Entropy normalization: normalize the entropy values ​​of all scales to ensure that their sum is 1. The normalization method is expressed as: Among them, w(s) is the entropy weight under scale s; S is the total number of selected scales; S3.4 Image kernelization: Use the Gaussian kernel function to map the image at each scale to a high-dimensional feature space. The kernel value K(x,y) of the Gaussian kernel function is defined as: Where x and y are two pixels in the image; σ is the width parameter of the Gaussian kernel function; For an image at scale s, calculate the kernel value between each pixel and its neighboring pixels, and then add these kernel values ​​to get the new pixel value; S3.5 weighted fusion uses the normalized entropy value as the weight to perform weighted fusion of kernelized images at all scales. For a certain pixel position p, its value I in the fused image ′ (p) is: Among them, I s (p) is the value of the image at pixel position p under scale s; S3.6 performs feature extraction and uses the scale-invariant feature transform (SIFT) method to extract key points and descriptors on the fused high-dimensional image.

6. The method according to claim 5, characterized in that Assume that in the SIFT algorithm, a 128-dimensional descriptor is extracted for each key point. If a total of N key points are extracted, the length of the final one-dimensional vector is L ′ , providing input data for the subsequent feature extraction of power transmission and transformation equipment defect identification: L ′ =N×128。 7. The method according to claim 1, characterized in that The step 4 specifically includes: using a neural network to extract features from the preprocessed data, Define the structure of the neural network. When extracting features, the multi-layer neural network used has L hidden layers. th There are hidden layers, and the weights and biases of the layers are represented by W (l) and b (l) ; Define the parameter space of the neural network, and assume that the parameter space of the neural network is Any point P i A set of weights w i and a set of biases b i Defined as: P i =(w i ,b i ) Where N is the number of points in the parameter space; The neural network loss function is defined as L(w,b), which is a fluctuation loss function, expressed as: Among them, y and They are the target output and actual output of the feature input to the preset Softmax classifier after neural network feature extraction, N1 is the number of input samples, and t is the number of current training iterations; α L and β L is the amplitude coefficient of the cosine and sine waves; ω and φ are the frequencies of the cosine and sine waves; Define elastic constants and fluctuation speeds, including elastic constants κ and fluctuation speeds v, which are associated with the weights w and biases b of the network; Define the elastic fluctuation update rules. The elastic fluctuation update process is the process of updating the neural network parameters. Define a set of elastic fluctuation update rules as follows: In i+1 =in i +Δw b i+1 =b i +Δb Among them, Δt is the time step; Δw and Δb are the change parameters of weight and bias; w i+1 and b i+1 are the updated weights and biases; The oscillation adjustment mechanism is defined. The oscillation adjustment is the adjustment strategy of the elastic constant κ and the fluctuation velocity v, which can be expressed as: Among them, κ0 and v0 are the initial elastic constant and fluctuation speed, β z and γ z is the adjustment factor.

8. The method according to claim 7, characterized in that During the training process of the neural network, the training steps are as follows:

1. Initialize w, b, κ, and v, and set the time step Δt; 2. Calculate the initial loss function L(w,b); 3. For each iteration i; 3.

1. Update w and b using the elastic fluctuation update rule. For the lth th The weight of the layer W (l) and bias b (l) , using the following elastic fluctuation update rule: 3.

2. Apply the oscillation adjustment mechanism to update κ and v. The update of the oscillation adjustment mechanism for each layer l is expressed as: in, and It is the first th The initial elastic constants and fluctuation velocities of the layers; 3.

3. Recalculate the loss function L(w,b); 4. Repeat iterations 3.1 to 3.3 until the training reaches convergence and the algorithm terminates; The convergence is judged in such a way that the change ΔL of the loss value does not change after a preset number of iterations, that is, convergence is achieved; the calculation basis of the change ΔL of the loss value is that during the iteration process, the state transition of the algorithm between any iteration i and i+1 is considered; the change ΔL of the loss function is expressed as: ΔL=L(w i+1 ,b i+1 )-L(w i ,b i ) According to the elastic fluctuation update rule and Taylor series expansion, we have: Substituting the update rule into the above equation, we get: Because ΔL is negative or zero, the algorithm will ensure that the loss function L(w,b) decreases or remains unchanged; After the algorithm terminates, it means that the training of the feature extraction neural network model is completed.

9. The method according to claim 1, characterized in that The step five is specifically as follows: Input the extracted features into the classifier for classification; Initialization operation, that is, initializing the weights and topological structure of SOM; assigning a weight vector w to each neuron n n , the dimension is the same as the input data dimension, and the weights are randomly initialized to small values ​​in the range [0,1], expressed as: In n =[in n1 ,In n2 ,...,In nd ] Where d is the dimension of the data; The one-dimensional vector data after feature extraction is input into the SOM network. Assume that the input one-dimensional vector is x and the dimension is d, which is expressed as: x=[x1,x2,...,x d ] For each input sample, find the best matching neuron BMU and calculate the Euclidean distance D(n) of each neuron: Select the neuron with the smallest distance as the BMU: Topology adjustment: dynamically adjust the topology of SOM according to the density and characteristics of the defect data of the power transmission and transformation equipment, and set the characteristic value ρ n Represents the data density near neuron n, and the update method is expressed as: r n =α×h(BMU,n)+(1-α)×ρ n Among them, α is the learning rate, h() is the neighborhood function; Weight update, local perception adjustment, according to the local perception ability of each neuron, adjust its weight, introduce the perception function P(n), expressed as: Among them, β is the sensitivity parameter, ρ threshold is the density threshold, and the weights of neurons in the BMU and its neighborhood are updated. The way to update the weights of neurons is expressed as: w n+1 '=w n +η×P(n)×(x-w n ) Among them, w n is the weight before update, w n+1 ' is the updated weight, η is the weight adjustment coefficient; Historical information feedback: Based on the previous historical information, the weights are further fine-tuned, and the fine-tuning amount δw n It is expressed as: δw n =γ×(in n,prev -In n ) Among them, γ is the historical information feedback coefficient; w n,prev is the historical weight, then the method of updating the weight according to the historical information feedback is expressed as: In n+1 ”=in n+1 '+δw n Among them, w n+1 ” is the weight updated according to historical information feedback; Adaptive topological connectivity adjustment, for the nth neuron, define a magnetization index M n It represents its ability to attract other neurons and is initialized to 0. Whenever the nth neuron is selected as a BMU, its magnetization index is updated according to the following formula: Where λ is an attenuation coefficient between 0 and 1, ∈ ′ is a small value to prevent division by zero, M n+1 is the updated magnetization index; The degree of connectivity between neurons is determined by the relative magnitude of the magnetization index. The connection strength C between two neurons n and m is nm Defined as: When C nm Exceeding a threshold C threshold , then it is considered that there is a strong connection between neuron n and neuron m; otherwise, it is a weak connection; then in the process of weight update, the weight update formula is: w n+1 '=w n+1 +η×C BMU,n ×h(BMU,n)×(x-w n ) Among them, w n+1 ' is the updated weight after the adaptive topology connectivity is adjusted; The judgment of the termination condition is that when the average weight change in several consecutive rounds is less than the preset threshold ∈ or the maximum number of iterations is reached, the training is stopped.

10. The method according to claim 1, characterized in that The step six is ​​specifically as follows: The above model is trained using training samples, and the trained model is used to identify and classify defects in power transmission and transformation equipment.

Images