Layered optimization pruning method and device for electric power multi-mode defect detection and medium

Abstract

The invention relates to a hierarchical optimization pruning method and device for electric power multi-modal defect detection and a medium, and the method comprises the steps: collecting multi-modal data in the defect detection of electric power equipment, inputting the multi-modal data into a pre-training defect detection model, carrying out the feature vector extraction of each modal, and carrying out the perception strategy based on a Hessian matrix, and obtaining the weight importance of each modal; calculating the L1 norm of each channel weight in the pre-trained defect detection model, evaluating the channel importance of the defect detection model, determining a redundant channel according to the L1 norm and cutting off the redundant channel, comparing the weight importance with a preset threshold, marking the weight importance lower than the preset threshold as a pruning object, generating a pruning mask matrix, and performing pruning on the pruning object; performing pruning operation on the pre-trained defect detection model through the pruning mask matrix; and sparse matrix representation optimization is carried out on the pruned model, and retraining is carried out. Compared with the prior art, the method has the advantages of enhanced adaptability, enhanced stability, high resource utilization rate and the like.

Description

Technical Field

[0001] The present invention relates to the field of deep learning, and in particular to a hierarchical optimization pruning method, device, and medium for power multimodal defect detection. Background Art

[0002] With the rapid development of intelligent and digital transformation in the power industry, the safe and stable operation of power equipment has placed higher demands on real-time defect detection technology. In intelligent power operation and maintenance scenarios, equipment defect detection relies on multimodal data such as infrared images, acoustic signals, and visible light images to accurately identify issues such as abnormal equipment temperatures, mechanical vibration faults, and surface damage. However, existing deep learning models still face significant challenges in practical application.

[0003] Due to their high complexity and large number of parameters, traditional deep learning models often suffer from high computational resource consumption and slow inference speed in power equipment defect detection. Especially on edge devices or embedded systems, strict hardware resource constraints, such as memory and computing power, make efficient deployment of complex models difficult, resulting in real-time detection that cannot meet on-site operation and maintenance requirements. Furthermore, the heterogeneous nature of multimodal data, such as the low contrast of infrared images, the spectral diversity of acoustic signals, and background interference in visible light images, further exacerbates the difficulty of model optimization. While existing lightweight technologies can reduce model complexity in general scenarios, they often lack adaptability to the specific multimodal data processing requirements of the power industry, resulting in loss of key features or high levels of redundant computation, compromising detection accuracy and efficiency. Currently, model optimization methods for power equipment defect detection mostly focus on single-modal data processing, making it difficult to effectively integrate the correlations between multimodal features. These methods require targeted optimization strategies to balance model performance and efficiency, limiting the practicality and scalability of lightweight models in complex power scenarios.

[0004] Chinese patent application CN115223049A discloses a knowledge distillation and quantization technology for the compression of large edge computing models in power scenarios. Through the knowledge distillation and quantization method for the compression of large edge computing models in power scenarios, the student model can fully learn the knowledge of each dimension of the teacher model, thereby improving the recognition accuracy in the variable scenarios of power transmission lines. It is used for inspection of high-voltage overhead transmission lines. However, the application does not optimize the physical feature differences of multimodal data in power scenarios, and does not deeply optimize the real-time reasoning efficiency of edge devices. Therefore, how to balance the computing efficiency and accuracy of the network while taking into account the characteristics of multimodal data and adapting to the needs of power professionals is a technical problem that needs to be solved. Summary of the Invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned existing technologies and provide a hierarchical optimization pruning method, equipment and medium for power multimodal defect detection. By implementing separate pruning strategies for the multimodal characteristics of power defect detection, the shortcomings of existing general compression technologies in adaptability to multimodal data are solved. Through Hessian matrix-aware pruning and sparse optimization, key features are retained while reducing the amount of calculation, avoiding the loss of accuracy that may be caused by quantization and distillation.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] According to one aspect of the present invention, a hierarchical optimization pruning method for power multimodal defect detection is provided, which specifically comprises the following steps:

[0008] S1. Collect multimodal data for power equipment defect detection, input it into a pre-trained defect detection model, and extract feature vectors of each modality. The multimodal data includes infrared images, sound signals, and visible light images.

[0009] S2. For the eigenvectors of each modality, the weight importance of each modality is obtained based on the perception strategy of the Hessian matrix;

[0010] S3. Calculate the L1 norm of each channel weight in the pre-trained defect detection model, evaluate the importance of the defect detection model channel, determine redundant channels based on the L1 norm and prune them, then compare the weight importance with a preset threshold, mark the weights with importance lower than the preset threshold as pruning targets, generate a pruning mask matrix, assign redundant weights to 0 in the pruning mask matrix, and retain the remaining weights. Prune the pre-trained defect detection model using the pruning mask matrix;

[0011] S4. Optimize the sparse matrix representation of the pruned model, store the pruned model parameters in a compressed sparse row or column format, and retrain the pruned defect detection model.

[0012] Furthermore, in S2, the pruning objects of the infrared image are determined by retaining the thermal feature sensitive channels related to temperature anomalies; the pruning objects of the sound signal are determined by retaining the weights related to the high-frequency vibration features through spectrum analysis to identify the vibration anomalies of the power equipment; and the pruning objects of the visible light image are determined by pruning the channels that are not related to the equipment surface area.

[0013] Furthermore, the generation of the Hessian matrix in S2 specifically includes: generating the second-order derivative matrix of the loss function with respect to the model parameters, that is, the Hessian matrix, obtaining the sensitivity of the model parameters to the loss function through Taylor expansion, and calculating the weight importance by combining the square value of the parameter and the diagonal elements of the Hessian.

[0014] The Hessian matrix is ​​the rate of change of the loss function in the parameter space. The expression of the Hessian matrix H is:

[0015]

[0016] Where θ is the model parameter and L(θ) is the loss function. The smaller the elements in the Hessian matrix, the lower the sensitivity of the model parameters to the loss function, that is, the lower the pruning risk. The expression of the Taylor expansion is:

[0017]

[0018] Where ΔL(θ) is the change in the loss function L(θ) when the parameter θ changes slightly by Δθ. The first term of the Taylor expansion is the linear effect of pruning on the loss, and the second term captures the nonlinear risk of pruning.

[0019] The expression of the weight importance I is:

[0020]

[0021] Among them, H ii is the diagonal element of the Hessian matrix, representing the parameter θ i The second derivative of is the square value of the model parameter, indicating the importance of the model parameter itself.

[0022] Furthermore, the perception strategy based on the Hessian matrix in S2 is that the calculation of the weight importance of each modality is set separately according to the network characteristics of processing each modality. In the calculation of the weight importance of the infrared image, the Hessian matrix is ​​approximated as a Gauss-Newton matrix, and the expression is:

[0023] H≈J T J,

[0024] Where J is the Jacobian matrix of the loss function with respect to the weights.

[0025] In the calculation of the weighted importance of the sound signal, the weighted spectral energy is calculated, the spectral coefficients are extracted based on the discrete Fourier transform, the high-order frequency features are retained according to the energy threshold, and the weighted importance is obtained through the Hessian matrix. The expression of the weighted spectral energy E(W) is:

[0026]

[0027] in, It is the coefficient in the spectrum after discrete Fourier transform of convolution kernel W.

[0028] In the calculation of the weight importance of the visible light image, the diagonal elements in the Hessian matrix are selected, and the expression is:

[0029]

[0030] Among them, H ii is the diagonal element in the Hessian matrix, L(θ) is the loss function, θ i are model parameters.

[0031] According to a second aspect of the present invention, an electronic device is provided, comprising a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the method when executing the program.

[0032] According to a third aspect of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the method described above is implemented.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] (1) Improved adaptability of multimodal data: A hierarchical pruning strategy is designed based on the physical characteristics of multimodal data in power equipment defect detection, namely infrared images, sound signals, and visible light images. Infrared images retain thermal feature-sensitive channels, sound signals retain high-frequency vibration-related weights, and visible light images remove background-irrelevant channels. By combining the domain characteristics of each modality, the problem of key feature loss or redundant calculation caused by lack of targeted optimization is avoided, significantly improving the detection accuracy and efficiency of the model in power scenarios.

[0035] (2) Enhanced pruning accuracy and stability: By introducing the Hessian matrix perception strategy and spectral energy analysis, the impact of weights on model performance is comprehensively evaluated. Compared with the traditional pruning method that only relies on the L1 norm, the present invention combines second-order derivative sensitivity analysis with frequency domain feature screening to accurately identify and retain weights that are critical to power equipment defect detection, such as temperature anomaly channels and high-frequency vibration characteristics, reducing the risk of mispruning and improving the rationality and effectiveness of pruning.

[0036] (3) Resource utilization and real-time optimization: The pruned model parameters are stored in compressed sparse row or column format, which greatly reduces storage overhead and computational complexity. Fine-tuning is performed based on historical equipment failure datasets to restore local performance losses caused by pruning, ensuring efficient deployment of the model on edge devices. While maintaining high accuracy, the model inference speed is significantly improved, meeting the stringent real-time detection requirements of intelligent power operation and inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1Flowchart of the hierarchical optimization pruning method for power multimodal defect detection;

[0038] Figure 2 Schematic diagram of the hierarchical optimization pruning process for power multimodal defect detection. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0040] Example 1

[0041] like Figure 1 As shown, this embodiment provides a hierarchical optimization pruning method for power multimodal defect detection, and the specific steps include:

[0042] S1. Collect multimodal data for power equipment defect detection, input it into a pre-trained defect detection model, and extract feature vectors of each modality. The multimodal data includes infrared images, sound signals, and visible light images.

[0043] S2. For the eigenvectors of each modality, the weight importance of each modality is obtained based on the perception strategy of the Hessian matrix;

[0044] S3. Calculate the L1 norm of each channel weight in the pre-trained defect detection model, evaluate the importance of the defect detection model channel, determine redundant channels based on the L1 norm and prune them, then compare the weight importance with a preset threshold, mark the weights with importance lower than the preset threshold as pruning targets, generate a pruning mask matrix, assign redundant weights to 0 in the pruning mask matrix, and retain the remaining weights. Prune the pre-trained defect detection model using the pruning mask matrix;

[0045] S4. Optimize the sparse matrix representation of the pruned model, store the pruned model parameters in a compressed sparse row or column format, and retrain the pruned defect detection model.

[0046] In S2, the multimodal data features combined with the power equipment data are layered by modality to determine the pruning objects, and the feature vectors of each modal data are found through feature extraction. Among them, for infrared images, thermal feature-sensitive channels related to temperature anomalies, such as local overheating or low temperature anomalies, are retained to ensure the integrity of key features. For sound signals, through spectral analysis, weights related to high-frequency vibration features are retained, such as the spectral regions corresponding to mechanical looseness or arc noise of power equipment, to improve the ability to identify vibration anomalies of power equipment. For visible light images, channels irrelevant to the surface area of ​​the equipment are pruned, and channels related to the surface texture of the equipment, such as scratches and rust, are retained to improve detection efficiency. By processing each modality separately, the pruning objects are obtained in a targeted manner, the accuracy and efficiency of pruning are improved, and the actual performance of the lightweight model is significantly improved.

[0047] During the pruning process, a Hessian matrix-based perception strategy is introduced, and second-order derivatives are used to analyze the sensitivity of weights to the loss function. This ensures that pruning has minimal impact on model performance and improves the rationality and effectiveness of pruning. An efficient sparse storage format is used in the sparse matrix representation after optimization pruned models, optimizing the pruned model to improve inference efficiency, reduce storage overhead, and achieve efficient utilization of computing resources.

[0048] Specifically, the generation of the Hessian matrix in S2 includes: generating the second-order derivative matrix of the loss function with respect to the model parameters, namely the Hessian matrix, obtaining the sensitivity of the model parameters to the loss function through Taylor expansion, and calculating the weight importance by combining the square value of the parameter and the diagonal elements of the Hessian. In the pruning task, the Hessian matrix is ​​used to measure the model parameters θ i That is, the smaller the second-order derivative of the parameter, the lower its sensitivity to the loss function and the lower the risk of pruning. The larger the second-order derivative of the parameter, the more significant its impact on the loss function and pruning may lead to performance degradation.

[0049] The Hessian matrix is ​​the rate of change of the loss function in the parameter space. The expression of the Hessian matrix H is:

[0050]

[0051] Where θ is the model parameter and L(θ) is the loss function. The smaller the elements in the Hessian matrix, the less sensitive the model parameters are to the loss function, that is, the lower the pruning risk. The Taylor expansion expression is:

[0052]

[0053] Here, ΔL(θ) is the change in the loss function L(θ) for a small change in parameter θ by Δθ. The first term of the Taylor expansion represents the linear effect of pruning on the loss and is used to quickly assess weight importance. The second term captures the nonlinear risk of pruning and avoids accidentally pruning critical features.

[0054] The expression of weight importance I:

[0055]

[0056] Among them, H ii is the diagonal element of the Hessian matrix, representing the parameter θ i The second derivative of is the square value of the model parameter, indicating the importance of the model parameter itself.

[0057] The Hessian matrix-based perception strategy calculates the importance of each modality's weight based on the network characteristics used to process it. Infrared images have low contrast, and the feature differences between the target and background are not obvious, resulting in small gradient changes. The Gauss-Newton approximation makes gradient calculations more stable by ignoring non-square terms in the second-order derivative, reducing the impact of noise on the optimization process and making pruning optimization more stable, making it suitable for the robustness requirements of infrared images. Since acoustic images contain frequency features, spectral energy is calculated and pruning is performed based on energy thresholds. Visible light images use the Hessian matrix diagonal approximation to remove background redundancy, allowing the model to focus computing resources on analyzing device surface details.

[0058] In the calculation of the weight importance of infrared images, the Hessian matrix is ​​approximated as a Gauss-Newton matrix, and the expression is:

[0059] H≈J T J,

[0060] Where J is the Jacobian matrix of the loss function with respect to the weights.

[0061] In the calculation of the weighted importance of the sound signal, the weighted spectral energy is calculated, the spectral coefficients are extracted based on the discrete Fourier transform, the high-order frequency features are retained according to the energy threshold, and the weighted importance is obtained through the Hessian matrix. The expression of the weighted spectral energy E(W) is:

[0062]

[0063] in, It is the coefficient in the spectrum after discrete Fourier transform of convolution kernel W.

[0064] In the calculation of the weight importance of the visible light image, the diagonal elements in the Hessian matrix are selected, and the expression is:

[0065]

[0066] Among them, H ii is the diagonal element in the Hessian matrix, L(θ) is the loss function, θ i are model parameters.

[0067] Through differentiated pruning strategies for three modalities and collaborative deployment between the cloud and the edge, the real-time and accuracy of power equipment defect detection can be significantly improved.

[0068] After pruning is completed, the pruned model weight matrix is ​​converted to compressed sparse row or compressed sparse column format. The compressed sparse row format records the row index, column index and numerical value of non-zero values, which is suitable for matrix operations accessed by row. The compressed sparse column format records the column index, row index and numerical value of non-zero values, which is suitable for matrix operations accessed by column. Only the numerical value of non-zero weights and their position information are stored, which greatly reduces the size of the model file. A historical equipment failure dataset involving multi-modal samples of infrared images, sound signals, and visible light images is used as fine-tuning data. The historical equipment failure dataset must cover typical defect scenarios. If the fault detection performance of the model in a specific mode decreases after pruning, the pruning threshold is dynamically adjusted. For devices at the target edge, the loading and calculation process of the sparse matrix is ​​optimized.

[0069] Compared with the existing network lightweighting methods, the pruning strategy of this embodiment, which is based on modal hierarchical optimization, improves the model efficiency in combination with data characteristics. Specifically, through the Hessian-aware pruning strategy, the sensitivity of each weight can be accurately evaluated to avoid performance degradation caused by excessive pruning. In addition, the use of diagonal approximation or gradient square approximation methods greatly reduces the computational overhead, making it suitable for large-scale deep networks. At the same time, combined with domain characteristics, different pruning strategies are adopted for important feature areas of different modes. Therefore, the method of this embodiment is of great significance in improving the efficiency of power equipment defect detection, reducing operation and maintenance costs, and ensuring equipment safety, and provides reliable technical support for the intelligent and digital transformation of the power industry.

[0070] This embodiment applies the above solution, such as Figure 2 As shown, firstly, the infrared image, visible light image and sound signal are input into the pre-trained network model, and the multimodal feature vector of the pre-trained model is extracted;

[0071] Then calculate the weight gradient and second-order derivative, for each weight θ i , calculate the gradient and the Hessian diagonal elements H ii ;

[0072] Then, the importance of the weights is evaluated. Based on the different characteristics of visible light, infrared, and sound, corresponding methods are selected (i.e., diagonal approximation processing for visible light images, Gauss-Newton approximation processing for infrared images, and spectral energy pruning for sound signals) to calculate the importance of the weights.

[0073] Combined with the pre-set pruning threshold, according to the importance distribution, a suitable threshold τ is selected, and weights with weight importance lower than τ are marked as pruning objects;

[0074] Generate pruning mask, construct mask matrix M, assign redundant weights to 0, and retain the rest of the weights;

[0075] Post-pruning fine-tuning retrains the pruned model to restore performance.

[0076] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described module can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0077] Example 2

[0078] The electronic device of the present invention includes a central processing unit (CPU), which can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM, and RAM are connected to each other via a bus. An input / output (I / O) interface is also connected to the bus.

[0079] Multiple components in the device are connected to the I / O interface, including: input units, such as a keyboard, mouse, etc.; output units, such as various types of displays, speakers, etc.; storage units, such as magnetic disks, optical disks, etc.; and communication units, such as network cards, modems, wireless communication transceivers, etc. The communication unit allows the device to exchange information / data with other devices via computer networks such as the Internet and / or various telecommunications networks. The processing unit performs the various methods and processes described above, such as the method of the present invention. For example, in some embodiments, the method of the present invention can be implemented as a computer software program that is tangibly contained in a machine-readable medium, such as a storage unit. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device via ROM and / or the communication unit. When the computer program is loaded into RAM and executed by the CPU, one or more steps of the method of the present invention described above can be performed. Alternatively, in other embodiments, the CPU can be configured to perform the method of the present invention by any other suitable means (e.g., by means of firmware).

[0080] The functions described above herein may be performed, at least in part, by one or more hardware logic components. For example, and without limitation, exemplary types of hardware logic components that may be used include: field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chip (SOCs), complex programmable logic devices (CPLDs), and the like.

[0081] The program code for implementing the method of the present invention can be written in any combination of one or more programming languages. Such program code can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0082] In the context of the present invention, machine-readable medium can be a tangible medium that can contain or store a program for use with an instruction execution system, device or equipment or used in combination with an instruction execution system, device or equipment. Machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. More specific examples of machine-readable storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0083] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A hierarchical optimization pruning method for power multimodal defect detection, characterized in that: The specific steps include: S1. Collect multimodal data for power equipment defect detection, input it into a pre-trained defect detection model, and extract feature vectors of each modality. The multimodal data includes infrared images, sound signals, and visible light images. S2. For the eigenvectors of each modality, the weight importance of each modality is obtained based on the perception strategy of the Hessian matrix; S3. Calculate the L1 norm of each channel weight in the pre-trained defect detection model, evaluate the importance of the defect detection model channel, determine redundant channels based on the L1 norm and prune them, then compare the weight importance with a preset threshold, mark the weights with importance lower than the preset threshold as pruning targets, generate a pruning mask matrix, assign redundant weights to 0 in the pruning mask matrix, and retain the remaining weights. Prune the pre-trained defect detection model using the pruning mask matrix; S4. Optimize the sparse matrix representation of the pruned model, store the pruned model parameters in a compressed sparse row or column format, and retrain the pruned defect detection model.

2. A hierarchical optimization pruning method for power multimodal defect detection according to claim 1, characterized in that: The generation of the Hessian matrix in S2 specifically includes: generating the second-order derivative matrix of the loss function with respect to the model parameters, that is, the Hessian matrix, obtaining the sensitivity of the model parameters to the loss function through Taylor expansion, and calculating the weight importance by combining the square value of the parameter and the diagonal elements of the Hessian.

3. The hierarchical optimization pruning method for power multimodal defect detection according to claim 2, characterized in that: The Hessian matrix is ​​the rate of change of the loss function in the parameter space. The expression of the Hessian matrix H is: Where θ is the model parameter and L(θ) is the loss function. The smaller the elements in the Hessian matrix, the lower the sensitivity of the model parameters to the loss function, that is, the lower the pruning risk. The expression of the Taylor expansion is: Where ΔL(θ) is the change in the loss function L(θ) when the parameter θ changes slightly by Δθ. The first term of the Taylor expansion is the linear effect of pruning on the loss, and the second term captures the nonlinear risk of pruning.

4. The hierarchical optimization pruning method for power multimodal defect detection according to claim 2, characterized in that: The expression of the weight importance I is: Among them, H ii is the diagonal element of the Hessian matrix, representing the parameter θ i The second derivative of is the square value of the model parameter, indicating the importance of the model parameter itself.

5. The hierarchical optimization pruning method for power multimodal defect detection according to claim 1, characterized in that: The perception strategy based on the Hessian matrix in S2 is that the calculation of the weight importance of each modality is set separately according to the network characteristics of processing each modality. In the calculation of the weight importance of the infrared image, the Hessian matrix is ​​approximated as a Gauss-Newton matrix, and the expression is: H≈J T J, Where J is the Jacobian matrix of the loss function with respect to the weights.

6. The hierarchical optimization pruning method for power multimodal defect detection according to claim 5, characterized in that: In the calculation of the weighted importance of the sound signal, the weighted spectral energy is calculated, the spectral coefficients are extracted based on the discrete Fourier transform, the high-order frequency features are retained according to the energy threshold, and the weighted importance is obtained through the Hessian matrix. The expression of the weighted spectral energy E(W) is: in, It is the coefficient in the spectrum after discrete Fourier transform of convolution kernel W.

7. The hierarchical optimization pruning method for power multimodal defect detection according to claim 5, characterized in that: In the calculation of the weight importance of the visible light image, the diagonal elements in the Hessian matrix are selected, and the expression is: Among them, H ii is the diagonal element in the Hessian matrix, L(θ) is the loss function, θ i are model parameters.

8. The hierarchical optimization pruning method for power multimodal defect detection according to claim 1, characterized in that: The pruning objects in S3, wherein the pruning objects of the infrared image are determined to retain thermal feature sensitive channels related to temperature anomalies; The pruning objects of the sound signal are determined by retaining the weights related to the high-frequency vibration characteristics through spectrum analysis to identify abnormal vibration of the power equipment; The pruning objects of the visible light image are determined by pruning channels of areas irrelevant to the device surface.

9. An electronic device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images