Electric power material defect screening method and system based on image and voiceprint technology

Abstract

The invention relates to the technical field of image processing, in particular to an electric power material defect screening method and system based on an image and voiceprint technology. The method comprises the following steps: forming a training set; training a neural network model by using the training set to obtain a defect screening model; classifying the candidate frames, and respectively establishing Gaussian models for the features of the operation sound corresponding to the images where the candidate frames are located; collecting a target image and operation sound of the electric power material, and obtaining the probability of belonging to each Gaussian model; a target Gaussian model is screened out, the quotient of the size of the candidate frame with the highest recognition accuracy and the average size corresponding to the target Gaussian model serves as a scale factor, and the target image is multiplied by the scale factor to obtain a transformation image; and inputting the transformed image into the defect screening model to obtain a defect detection result. The method can greatly improve the accuracy and recognition efficiency of an electric power material defect recognition result, and improves the automation degree of electric power material defect detection.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology. More specifically, the present invention relates to a method and system for screening defects in electric power materials based on image and voiceprint technology. Background Art

[0002] The power system is the infrastructure of modern society, and its reliability and safety directly affect the normal operation of society and economic development. Any defects in power materials may lead to serious power failures and even cause large-scale power outages. Traditional power material defect detection mainly relies on manual inspection and simple instrument detection. These methods are often inefficient, inaccurate, and easily affected by human factors, making it difficult to meet the needs of modern power systems for high reliability and high efficiency. With the development of image processing technology and voiceprint analysis technology, the application of these technologies in power material defect detection has become possible.

[0003] In the prior art, defect detection of power materials is usually performed by collecting images of power materials, and using image processing technology combined with defect detection models to identify defects on the outer surface of power materials. However, since there are many defect areas on the outer surface of power materials, and the defect areas are of different sizes, the defect detection model has different recognition accuracy for defects of different sizes, and there may be errors in the recognition of defects of sizes with low recognition result accuracy. For example, the electronic transformer defect detection method disclosed in the Chinese patent application document with application publication number CN115950887A uses the collected image data of the electronic transformer combined with the defect detection model to detect the defect type of the transformer, but this method does not take into account the difference in the recognition accuracy of the defect detection model for defects of different sizes, resulting in certain errors in the transformer defect detection method. Summary of the invention

[0004] In order to solve the technical problem of errors in the prior art when using image processing technology and neural network models to detect defects in power materials, the present invention provides solutions in the following aspects.

[0005] In a first aspect, the present invention provides a method for screening defects in electric power materials based on image and voiceprint technology, comprising:

[0006] Acquire historical data, including images of power materials, and use candidate boxes to mark defect locations and defect types in the images to form a training set; use the training set to train a neural network model to obtain a defect screening model; the power materials include transformers;

[0007] The historical data also includes the running sound corresponding to the image; classifying the candidate frames, classifying the candidate frames with similar sizes into the same category, calculating the average size of the candidate frames of each category, and establishing Gaussian models for the features of the running sound corresponding to the image where each candidate frame is located;

[0008] Collecting target images and operating sounds of electric power materials, extracting features of the operating sounds, and inputting the features into each of the Gaussian models to obtain the probability that the features belong to each of the Gaussian models;

[0009] The target Gaussian model whose probability is greater than the preset probability threshold is screened out, and the quotient of the candidate box size with the highest recognition accuracy of the defect screening model and the average size corresponding to the target Gaussian model is used as a scale factor. The target image is multiplied by the scale factor to obtain a transformed image; if the size of the transformed image is larger than the target image, the transformed image is input into the defect screening model to obtain a defect detection result.

[0010] The beneficial effects are as follows: the image processing technology can efficiently identify appearance defects, while the voiceprint analysis technology can detect mechanical and electrical anomalies inside the equipment. The power material defect screening method based on image and voiceprint technology of the present invention not only identifies defects based on the collected images when identifying defects in power materials, but also further considers that the operating sound of power materials when there are defects is different from the operating sound of power materials under normal conditions. The approximate size of the candidate box corresponding to the defective part is identified through the characteristics of the operating sound of the power materials, and then the collected image is scaled so that the size of the defective part in the image is close to the optimal size that is conducive to the defect screening model, thereby greatly improving the accuracy of the defect identification result; the defects of power materials are automatically identified by the neural network model instead of manual identification, thereby improving the speed of defect detection and reducing labor costs and time consumption. Combined with high-resolution image processing technology and precise voiceprint analysis, potential defects of power materials can be discovered earlier and more accurately, thereby avoiding small problems from turning into major faults. Due to the use of the neural network model, the defect detection algorithm can be continuously optimized to improve the intelligence level of the detection system.

[0011] In one embodiment, a clustering algorithm is used when classifying the candidate boxes, including: when clustering, selecting several candidate boxes as initial cluster centers, and taking the difference between the size of the candidate box and the size of the candidate box at the cluster center as the distance from the candidate box to the cluster center, and the calculation expression is:

[0012]

[0013] Among them, L represents the distance between the candidate box and the jth cluster center, h i represents the height of the i-th candidate box, w iIndicates the width of the i-th candidate box, h j represents the height of the candidate box of the jth cluster center, w j Indicates the width of the candidate box of the jth cluster center.

[0014] The beneficial effect is that by adopting a clustering algorithm to classify candidate frames, the classification efficiency and classification accuracy can be greatly improved, thereby improving the efficiency and accuracy of power material defect screening.

[0015] In one embodiment, the characteristics of the sound include the vibration frequency and vibration amplitude of the sound.

[0016] The beneficial effect is that the vibration frequency and amplitude of the operating sound of power materials are different for different defects. The vibration frequency and amplitude of the operating sound are comprehensively considered when constructing the Gaussian model, so that the candidate box category of the defect corresponding to the operating sound can be identified more accurately.

[0017] In one embodiment, establishing a Gaussian model for the features of the running sound corresponding to the image where a certain type of candidate box is located includes:

[0018] The whole of the images where each candidate box in this type of candidate box is located is recorded as an image set;

[0019] Performing Fourier transform or wavelet transform on the sound of the electric power material corresponding to each image in the image set, so as to obtain the vibration frequency and vibration amplitude of the operation sound of each image;

[0020] A Gaussian model is established for all the vibration frequencies and vibration amplitudes obtained.

[0021] In one embodiment, the probability threshold is set to 0.4.

[0022] In one embodiment, a method for obtaining the candidate box size with the highest recognition accuracy of a defect screening model is as follows: respectively input the image of each candidate box in each type of candidate box into the defect screening model to obtain a defect recognition result, calculate the recognition accuracy of each type of candidate box based on the recognition result, and select the average size of the candidate box of the type with the highest recognition accuracy as the candidate box size with the highest recognition accuracy of the defect screening model.

[0023] Its beneficial effect is: by calculating the recognition accuracy of defects corresponding to each type of candidate frame respectively, and taking the average size of the candidate frame of the category with the highest recognition accuracy as the candidate frame size with the highest recognition accuracy of the power material defect screening model, the candidate frame size with the highest recognition accuracy can be obtained quickly and accurately.

[0024] In one embodiment, it also includes: in response to the size of the transformed image being smaller than the target image, the target image is cut into multiple sub-images, the size of the sub-image is equal to the size of the candidate box with the highest recognition accuracy of the screening model, and the cut multiple sub-images are input into the screening model to obtain defect detection results.

[0025] In one embodiment, cutting the target image into a plurality of sub-images comprises:

[0026] A sliding window of a preset size is set, and sliding is performed from the upper left corner of the image according to a preset step length to obtain images corresponding to the sliding windows at various positions, wherein the size of the sliding window is equal to the size of the candidate box with the highest recognition accuracy of the screening model.

[0027] In one embodiment, the cut multiple sub-images are input into the screening model, the screening model outputs the defect type corresponding to each sub-image, and the defect type with the highest frequency among all the defect types output by the screening model is used as the defect detection result.

[0028] The beneficial effects are as follows: different sizes of images input into the defect screening model may lead to different defect detection results, and the detection results corresponding to the same image may include multiple defect types; by taking the most frequently occurring defects as the defect detection results of power materials, the accuracy of the detection results can be improved.

[0029] In a second aspect, the present invention provides an electric power material defect screening system based on image and voiceprint technology, comprising a memory and a processor, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the electric power material defect screening method based on image and voiceprint technology of the present invention is implemented. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood. In the accompanying drawings, several embodiments of the present invention are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0031] Figure 1 is a flowchart schematically illustrating a method for screening defects in electric power materials based on image and voiceprint technology according to an embodiment of the present invention;

[0032] Figure 2 is a flow chart schematically illustrating a method for establishing a Gaussian model according to an embodiment of the present invention;

[0033] Figure 3 is a schematic diagram schematically illustrating target image size according to an embodiment of the present invention;

[0034] Figure 4 It is a schematic diagram schematically showing the structure of a power material defect screening system based on image and voiceprint technology according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0036] The specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0037] Example of a method for screening defects in electric power materials based on image and voiceprint technology:

[0038] like Figure 1 As shown, the power material defect screening model training method of the present invention includes:

[0039] S101. Training a power material defect screening model, specifically: obtaining historical data, including images of power materials, using candidate boxes to mark defect locations and defect types in the images to form a training set; using the training set to train a neural network model to obtain a power material defect screening model; the power materials include transformers.

[0040] The images of electric power materials at multiple historical moments include images of electric power materials with various defects and images of electric power materials without defects. Images of electric power materials without defects can be marked as defect-free, and there is no need to use candidate boxes to mark the defect locations. Among them, an image of an electric power material with defects has only one defect. The electric power material can be a transformer, a relay protection device or other electric power equipment. If the electric power material is a transformer, the types of defects it has include: bushing fault and insulator fault. Each image of the electric power material is taken with the same shooting equipment, and the shooting angle and the distance between the shooting equipment and the electric power material are the same.

[0041] For insulators, after long-term operation of the transformer, there may be dirt on their surface, and the insulation performance of the insulators will decrease, resulting in abnormal operation of the transformer; for bushings, defects such as deformation, scratches, bubbles and cracks may appear, which will affect their insulation and protection performance and cause abnormal operation of the transformer.

[0042] In one embodiment, when a candidate box is used to mark the defective position in an image, the defective part of the transformer is selected using the candidate box and text content corresponding to the defect type is added to the candidate box. For example, if the sleeve of the transformer is faulty, the candidate box will select the entire sleeve area in the image, and the size of the candidate box is close to the size of the sleeve area.

[0043] During the training process, the model's loss function uses cross entropy loss, and the model's evaluation index uses accuracy. The optimal model is selected based on the model's evaluation index.

[0044] S102, establishing a Gaussian model, specifically: the historical data also includes the running sound corresponding to the image; classifying the candidate frames, classifying the candidate frames of similar sizes into the same category, calculating the average size of the candidate frames of each category, and establishing a Gaussian model for the features of the running sound corresponding to the image where each candidate frame is located.

[0045] The operating sound corresponding to the image refers to the sound made when the corresponding electric power material is in operation when the image is collected.

[0046] The input of the established Gaussian model is the feature of the sound, and the output is the probability that the feature of the sound belongs to the Gaussian model.

[0047] There are many ways to classify the candidate boxes. In this embodiment, the minimum width, maximum width, minimum height and maximum height of all candidate boxes can be counted, and then the candidate box sizes can be divided into several size intervals. For example, assuming that the minimum width and maximum width are 5 and 10 respectively, and the minimum height and maximum height are 3 and 5 respectively, three size intervals can be divided, namely, the size interval of width 5 to 6 and height 3; the size interval of width 7 to 8 and height 4; the size interval of width 9 to 10 and height 5. In other embodiments, clustering algorithms can also be used to classify candidate boxes.

[0048] S103, obtaining the probability that the operating sound belongs to each Gaussian model, specifically: collecting the target image and operating sound of the power material, extracting the features of the operating sound, and inputting the features into each Gaussian model to obtain the probability that it belongs to each Gaussian model.

[0049] S104, obtaining defect detection results, specifically: screening out target Gaussian models with probabilities greater than a preset probability threshold, taking the quotient of the candidate frame size of the defect screening model with the highest recognition accuracy and the average size corresponding to the target Gaussian model as a scale factor, multiplying the target image by the scale factor to obtain a transformed image; if the size of the transformed image is larger than the target image, inputting the transformed image into the power material defect screening model to obtain the defect detection results.

[0050] In this embodiment, the probability threshold is set to 0.4, and other appropriate values ​​may be used in other embodiments.

[0051] If there are multiple Gaussian models with a probability greater than a preset probability threshold that are screened out, the collected image is resized according to the average size corresponding to each Gaussian model to obtain transformed images of different sizes. The transformed images of different sizes are input into the defect screening model respectively, and the defect with the highest frequency in the detection results is taken as the power material defect detection result.

[0052] Different sizes of images input into the power material defect screening model may lead to different defect detection results, and the detection results corresponding to the same image may include multiple defect types; by taking the most frequently occurring defects as the power material defect detection results, the accuracy of the detection results can be improved.

[0053] The expression for calculating the size of the transformed image is:

[0054]

[0055] Among them, new scale Indicates the size of the transformed image, pre scale Indicates the size of the target image, scale maxp Indicates the candidate box size with the highest recognition accuracy of the power material defect screening model, Indicates the average size corresponding to the target Gaussian model.

[0056] In this embodiment, if the size of the transformed image is smaller than the size of the corresponding target image, the transformed image can be directly input into the defect screening model to obtain the defect detection result; in other embodiments, if the size of the transformed image is smaller than the size of the corresponding target image, the target image can also be cut into multiple sub-images, each sub-image has a size equal to the candidate frame size with the highest recognition accuracy of the defect screening model, and the cut sub-images are input into the power material defect screening model to obtain the defect detection result. The image information corresponding to all sub-images combined together must cover all the image information in the image before cutting.

[0057] Image processing technology can efficiently identify appearance defects, while voiceprint analysis technology can detect mechanical and electrical anomalies inside the equipment. The power material defect screening method based on image and voiceprint technology of the present invention not only identifies defects based on the collected images when identifying defects in power materials, but also further considers that the operating sound of power materials when there are defects is different from the operating sound of power materials under normal conditions. The approximate size of the candidate box corresponding to the defective part is identified through the characteristics of the operating sound of the power materials, and then the collected image is scaled so that the size of the defective part in the image is close to the optimal size that is conducive to the defect screening model, thereby greatly improving the accuracy of the defect identification result; by using the neural network model to automatically identify the defects of power materials instead of manual identification, the defect detection speed is improved, and the labor cost and time consumption are reduced. Combined with high-resolution image processing technology and precise voiceprint analysis, potential defects of power materials can be discovered earlier and more accurately, thereby avoiding small problems from turning into major failures. Due to the use of the neural network model, the defect detection algorithm can be continuously optimized to improve the intelligence level of the detection system.

[0058] It can be seen from the above embodiments that a clustering algorithm or other methods may be used to classify candidate boxes. In one embodiment, a clustering algorithm is used to classify candidate boxes, including: when clustering, a number of candidate boxes are selected as initial cluster centers, and the difference between the size of the candidate box and the size of the candidate box at the cluster center is used as the distance from the candidate box to the cluster center, and the calculation expression is:

[0059]

[0060] Among them, L represents the distance between the candidate box and the jth cluster center, h i represents the height of the i-th candidate box, w i Indicates the width of the i-th candidate box, h j represents the height of the candidate box of the jth cluster center, w j Indicates the width of the candidate box of the jth cluster center.

[0061] In this embodiment, the clustering algorithm may adopt the K-Means clustering algorithm, and other clustering algorithms may also be adopted in other embodiments.

[0062] By adopting clustering algorithm to classify candidate frames, the classification efficiency and classification accuracy can be greatly improved, thereby improving the efficiency and accuracy of power material defect screening.

[0063] In one embodiment, the characteristics of the sound include the vibration frequency and the vibration amplitude of the sound.

[0064] The vibration frequency and amplitude of the operating sound of power materials are different for different defects. The vibration frequency and amplitude of the operating sound are comprehensively considered when constructing the Gaussian model, so that the candidate box category of the defect corresponding to the operating sound can be identified more accurately.

[0065] like Figure 2 As shown, in one embodiment, establishing a Gaussian model for the features of the running sound corresponding to the image where a certain type of candidate box is located includes:

[0066] S201, record the whole of the images where each candidate box in the candidate box is located as an image set;

[0067] S202, processing the operation sound to obtain the vibration frequency and vibration amplitude of the operation sound, specifically: performing Fourier transform or wavelet transform on the operation sound corresponding to each image in the image set, thereby obtaining the vibration frequency and vibration amplitude of the operation sound of each image;

[0068] S203: Establish a Gaussian model for all obtained vibration frequencies and vibration amplitudes.

[0069] For example, this type of candidate frame includes three candidate frames, and the images where the three candidate frames are located are image A, image B, and image C, respectively. Then the image set is a set consisting of image A, image B, and image C. Fourier transform or wavelet transform is performed on the running sound corresponding to image A, the running sound corresponding to image B, and the running sound corresponding to image C, respectively, so as to obtain the vibration frequency and vibration amplitude of the running sound of image A, the running sound of image B, and the running sound of image C; and a Gaussian model is established for all the obtained vibration frequencies and vibration amplitudes.

[0070] In one embodiment, a method for obtaining the candidate box size with the highest recognition accuracy of a defect screening model is as follows: input the image of each candidate box in each type of candidate box into the defect screening model respectively to obtain a defect recognition result, calculate the recognition accuracy of each type of candidate box based on the recognition result, and select the average size of the candidate box of the type with the highest recognition accuracy as the candidate box size with the highest recognition accuracy of the defect screening model.

[0071] Assume that there are three types of candidate frames, namely the first type of candidate frames, the second type of candidate frames and the third type of candidate frames. The first type of candidate frames includes candidate frames D and E; the second type of candidate frames includes candidate frames F and G; the third type of candidate frames includes candidate frames H and I. Then the images corresponding to candidate frames D, E, F, G, H and I are input into the defect screening model respectively. If the image defect recognition result corresponding to candidate frame D is wrong, the image defect recognition result corresponding to candidate frame G is wrong, and the image defect recognition results corresponding to other candidate frames are correct; then the recognition accuracy of the defects corresponding to the first type of candidate frames is 50%, the recognition accuracy of the defects corresponding to the second type of candidate frames is 50%, and the recognition accuracy of the defects corresponding to the third type of candidate frames is 100%. Then calculate the average size of candidate frames H and I as the candidate frame size with the highest recognition accuracy of the power material defect screening model.

[0072] By calculating the recognition accuracy of defects corresponding to each type of candidate frame respectively, and taking the average size of the candidate frame of the category with the highest recognition accuracy as the candidate frame size with the highest recognition accuracy of the power material defect screening model, the candidate frame size with the highest recognition accuracy can be obtained quickly and accurately.

[0073] In one embodiment, cutting the target image into multiple sub-images includes: setting a sliding window of a preset size, sliding from the upper left corner of the image according to a preset step length, obtaining the image corresponding to the sliding window at each position, and the size of the sliding window is equal to the candidate frame size with the highest recognition accuracy of the power material defect screening model. For example: assuming that the preset step length is 10 pixels and the size of the sliding window is 20×20, the initial position of the sliding window is an area with a width of 20 and a height of 20 including the upper left corner of the image, and the sliding window first slides to the right according to the preset step length, and then slides down one step length when sliding to the rightmost side of the image, and then slides to the left again; slides down one step length when sliding to the leftmost side of the image, and then slides to the right again; whenever the sliding window slides to the left edge or right edge of the image, the sliding window is slid down one step length, and then the sliding window is slid in the horizontal direction; until the sliding window slides over all the pixels of the image. The sliding path of the sliding window is an S-shaped path.

[0074] By using a sliding window to crop the target image, it can be ensured that the cropped sub-image contains all the image information of the target image on the basis that the size of the sub-image is an optimal size.

[0075] In another embodiment, if the target image can be segmented into several sub-images whose sizes are equal to the candidate frame size with the highest recognition accuracy of the defect screening model, there is no need to set a sliding window, and the target image is directly segmented. Figure 3As shown, if the width and height of the target image are respectively twice the width and height of the candidate box size with the highest recognition accuracy of the defect screening model, the target image can be cut from the center point along the horizontal and vertical directions to obtain four sub-images of area a, area b, area c and area d.

[0076] In one embodiment, the multiple cut sub-images are input into the power material defect screening model, and the screening model outputs the defect type corresponding to each sub-image, and the defect type with the highest frequency among all the defect types output by the screening model is used as the defect detection result. The calculation expression of the defect detection result is:

[0077]

[0078] Where, pre represents the defect detection result, Q t It indicates the number of times the cropped power material image is input into the model and the prediction type is defect t, Q 总 Indicates the total number of cropped power material images. express The value of t when it reaches its maximum value.

[0079] Different sizes of images input into the power material defect screening model may lead to different defect detection results, and the detection results corresponding to the same image may include multiple defect types; by taking the most frequently occurring defects as the power material defect detection results, the accuracy of the detection results can be improved.

[0080] Example of power material defect screening system based on image and voiceprint technology:

[0081] The present invention also provides a power material defect screening system based on image and voiceprint technology. Figure 4 As shown, the power material defect screening system based on image and voiceprint technology includes a processor and a memory, and the memory stores computer program instructions. When the computer program instructions are executed by the processor, the power material defect screening method based on image and voiceprint technology according to the first aspect of the present invention is implemented.

[0082] The power material defect screening system based on image and voiceprint technology also includes other components familiar to those skilled in the art, such as a communication bus and a communication interface, whose settings and functions are known in the art and will not be described in detail here.

[0083] In the present invention, the aforementioned memory may be any tangible medium containing or storing a program that may be used by or in combination with an instruction execution system, apparatus or device. For example, a computer-readable storage medium may be any suitable magnetic storage medium or magneto-optical storage medium, such as a resistive random access memory RRAM (Resistive Random Access Memory), a dynamic random access memory DRAM (Dynamic Random Access Memory), a static random access memory SRAM (Static Random-Access Memory), an enhanced dynamic random access memory EDRAM (Enhanced Dynamic Random Access Memory), a high-bandwidth memory HBM (High-Bandwidth Memory), a hybrid memory cube HMC (Hybrid Memory Cube), etc., or any other medium that may be used to store the required information and may be accessed by an application, a module, or both. Any such computer storage medium may be part of a device or accessible or connectable to a device. Any application or module described in the present invention may be implemented using computer-readable / executable instructions that may be stored or otherwise maintained by such a computer-readable medium.

[0084] In the description of this specification, "plurality" or "several" means at least two, such as two, three or more, etc., unless otherwise clearly and specifically defined.

[0085] Although this specification has shown and described a number of embodiments of the present invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will conceive of many modifications, changes and alternatives without departing from the ideas and spirit of the present invention. It should be understood that in the practice of the present invention, various alternatives to the embodiments of the present invention described herein may be employed.

Claims

1. A method for screening defects in power materials based on image and voiceprint technology, characterized in that: include: Obtain historical data, including images of power materials, and use candidate boxes to mark the defect locations and defect types in the images to form a training set; The training set is used to train a neural network model to obtain a defect screening model; the power material includes a transformer; The historical data also includes the running sound corresponding to the image; classifying the candidate frames, classifying the candidate frames with similar sizes into the same category, calculating the average size of the candidate frames of each category, and establishing Gaussian models for the features of the running sound corresponding to the image where each candidate frame is located; Collecting target images and operating sounds of electric power materials, extracting features of the operating sounds, and inputting the features into each of the Gaussian models to obtain the probability that the features belong to each of the Gaussian models; The target Gaussian model whose probability is greater than the preset probability threshold is screened out, and the quotient of the candidate box size with the highest recognition accuracy of the defect screening model and the average size corresponding to the target Gaussian model is used as a scale factor. The target image is multiplied by the scale factor to obtain a transformed image; if the size of the transformed image is larger than the target image, the transformed image is input into the defect screening model to obtain a defect detection result.

2. The method for screening defects of electric power materials based on image and voiceprint technology according to claim 1, characterized in that: A clustering algorithm is used when classifying the candidate boxes, including: when clustering, several candidate boxes are selected as initial cluster centers, and the difference between the size of the candidate box and the size of the candidate box at the cluster center is used as the distance from the candidate box to the cluster center, and the calculation expression is: Among them, L represents the distance between the candidate box and the jth cluster center, h i represents the height of the i-th candidate box, w i Indicates the width of the i-th candidate box, h j represents the height of the candidate box of the jth cluster center, w j Indicates the width of the candidate box of the jth cluster center.

3. The method for screening defects of electric power materials based on image and voiceprint technology as claimed in claim 1, characterized in that: The characteristics of the sound include the vibration frequency and vibration amplitude of the sound.

4. The method for screening defects of electric power materials based on image and voiceprint technology as claimed in claim 2, characterized in that: The Gaussian model is established for the features of the running sound corresponding to the image where a certain type of candidate box is located, including: The whole of the images where each candidate box in this type of candidate box is located is recorded as an image set; Performing Fourier transform or wavelet transform on the sound of the electric power material corresponding to each image in the image set, so as to obtain the vibration frequency and vibration amplitude of the operation sound of each image; A Gaussian model is established for all the vibration frequencies and vibration amplitudes obtained.

5. The method for screening defects of electric power materials based on image and voiceprint technology as claimed in claim 1, characterized in that: The value of the probability threshold is 0.

4.

6. The method for screening defects of electric power materials based on image and voiceprint technology as claimed in claim 1, characterized in that: The method for obtaining the candidate frame size with the highest recognition accuracy of the defect screening model is as follows: respectively inputting the image of each candidate frame in each type of candidate frame into the defect screening model to obtain the defect recognition result, calculating the recognition accuracy of each type of candidate frame according to the recognition result, and selecting the average size of the candidate frame of the type with the highest recognition accuracy as the candidate frame size with the highest recognition accuracy of the defect screening model.

7. The method for screening defects of electric power materials based on image and voiceprint technology according to any one of claims 1 to 6, characterized in that: Also includes: In response to the size of the transformed image being smaller than the target image, the target image is cut into multiple sub-images, the size of the sub-images being equal to the size of the candidate box with the highest recognition accuracy of the screening model, and the cut multiple sub-images are input into the screening model to obtain defect detection results.

8. The method for screening defects of electric power materials based on image and voiceprint technology as claimed in claim 7, characterized in that: The cutting of the target image into a plurality of sub-images comprises: A sliding window of a preset size is set, and sliding is performed from the upper left corner of the image according to a preset step length to obtain images corresponding to the sliding windows at various positions, wherein the size of the sliding window is equal to the size of the candidate box with the highest recognition accuracy of the screening model.

9. The method for screening defects of electric power materials based on image and voiceprint technology according to claim 7, characterized in that: The cut multiple sub-images are input into the screening model, and the screening model outputs the defect type corresponding to each sub-image, and the defect type with the highest frequency among all the defect types output by the screening model is taken as the defect detection result.

10. A power material defect screening system based on image and voiceprint technology, comprising a memory and a processor, wherein the memory stores computer program instructions, characterized in that: When the computer program instructions are executed by the processor, the method for screening defects of electric power materials based on image and voiceprint technology as described in any one of claims 1 to 9 is implemented.

Images