Intelligent equipment inspection method and device based on deep learning multiple scales and medium

Abstract

The invention relates to the technical field of power equipment detection, in particular to an intelligent equipment inspection method and device based on deep learning multiple scales and a medium. Comprising three parts: data preprocessing, iterative training of an equipment intelligent inspection model by using a training sample, and inspection of equipment by using the trained equipment intelligent inspection model. The equipment intelligent inspection model utilizes a feature extraction encoder network and a multi-scale feature decoder network to enhance the adaptability of the model in spatial transformation and the ability of the model in spatial feature extraction, and at the same time, the non-linear expression ability of the network is enhanced. The feature extraction encoder network uses information of different positions in a deformable convolution sampling sequence, and in the multi-scale feature decoder network, feature information loss in an up-sampling process is reduced through a fusion module, so that the accuracy and efficiency of electrical equipment defect detection are enhanced; and valuable reference information is provided for preventive maintenance and fault prediction of the power system.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment detection, and specifically to a method, device and medium for intelligent equipment inspection based on deep learning and multi-scale. Background Art

[0002] Traditionally, inspection work mainly relies on manual operations, such as on-site observation and recording of equipment status by inspectors. However, this traditional technology has several significant disadvantages: first, manual inspection is inefficient, requires a lot of time and manpower, has a long inspection cycle, and is difficult to detect and handle equipment failures in a timely manner; second, the quality of inspection is difficult to guarantee, and the accuracy of inspection results is uncertain due to human factors (such as fatigue, negligence, etc.) and the external environment; finally, in some high-risk or complex environments, manual inspection may bring safety hazards.

[0003] With the continuous development of technology, especially the rise of deep learning technology, equipment inspection methods based on deep learning have gradually been applied. This method uses deep learning algorithms to process and analyze equipment images, sounds or sensor data to identify the operating status and potential faults of the equipment. Its significant advantage is that it can efficiently and accurately process large amounts of data, improve inspection efficiency, reduce human errors, and achieve real-time monitoring and early warning of potential faults. This is particularly important for key areas such as the power industry, where equipment failures may lead to serious consequences such as power grid outages and safety accidents. However, the current intelligent inspection technology for industrial equipment still needs to improve its recognition accuracy when dealing with subtle defects such as loose screws, tiny cracks, and missing parts. These tiny targets are often difficult to be accurately captured by existing systems due to their small size, low contrast, or interference from complex backgrounds. Summary of the invention

[0004] In view of the defects of the prior art, the present invention provides a method, device and medium for intelligent equipment inspection based on deep learning and multi-scale, aiming to improve the accuracy of small target detection on power equipment.

[0005] In order to solve the technical problem, the technical solution adopted by the present invention is: a multi-scale intelligent inspection method for equipment based on deep learning, comprising the following steps: S01, data collection and preprocessing, collecting power equipment images, using annotation tools to annotate the targets to be detected in the power equipment images, preprocessing the annotated power equipment images to obtain training images, and collecting the training images in a training sample set; S02. Construct a deep learning multi-scale equipment intelligent inspection model. The deep learning multi-scale equipment intelligent inspection model includes an encoder and a decoder. The encoder is composed of a plurality of stacked convolution modules. Each convolution module includes a deformable convolution layer, a three-dimensional convolution layer and an activation layer. The deformable convolution layer uses affine transformation to recalculate the position of the convolution kernel in the feature map, so that the encoder can simultaneously consider the information at different positions in the image sequence when extracting features; the decoder includes multiple layers of timing prediction modules, upsampling modules, fusion modules and a deconvolution module of different scales. The timing prediction module and the upsampling module of the first layer are respectively connected to the output of the encoder, the timing prediction modules and upsampling modules of other layers are connected to the output of the fusion module of the previous layer, the fusion module is connected to the output of the timing prediction module and the upsampling module of the layer, the input of the deconvolution module is connected to the output of the fusion module of the last layer, and the output of the deconvolution module is the inspection result of the deep learning multi-scale equipment intelligent inspection model; S03, model training, sampling training sample sets to train the deep learning multi-scale equipment intelligent inspection model to obtain a trained equipment intelligent inspection model; S04. Collect images of power equipment in real time and input them into the trained equipment intelligent inspection model for prediction to obtain inspection results.

[0006] Furthermore, the convolution module of the encoder processes the image as follows: , in , Represent the feature graphs at time t-1 and t respectively, , represents the size of the convolution kernel area, Represents a splicing operation, Indicates the current sampling point The offset on the feature map, by the convolutional subnetwork Generated from the feature maps at adjacent times t-1 and t, To calculate the offset The convolution operation, Indicates the calculation formula of deformable convolution for the feature map at time t-1, D represents the convolution kernel area, which defaults to 3×3 here; , They represent the kernel functions used for the feature graph at time t and the feature graph at time t-1, respectively. Indicates the current sampling point The corresponding other areas involved in the convolution calculation, Indicates the calculation formula of using ordinary convolution for the feature map at time t, represents the activation function, , Indicatesl The feature map at time t-1 and the feature map at time t input by the layer convolution module, Respectively represent l The output of the convolutional module is l+1 The feature map at time t input to the layer convolution module.

[0007] Furthermore, the time series prediction module combines bilinear interpolation and depthwise separable convolution operations. The bilinear interpolation method doubles the length and width of the image, and then uses depthwise separable convolution to replace ordinary convolution. At the same time, the batch normalization layer and ReLU activation function are used to enhance the learning and generalization capabilities of the model. The calculation process of the time series prediction module is: , Where x represents the input data, and scale is the upsampling factor, which means that the spatial dimension of the input data is doubled. represents the upsampling operation, represents the depthwise convolution operation, represents the point-by-point convolution operation, represents the batch normalization operation, is the activation function, Represents the output of the time series prediction module; The depth-wise separable convolution operation utilizes the characteristics of image data space and depth dimensions that can be separably and independently calculated. It reduces the number of parameters by decomposing the convolution operation. First, convolution calculation is performed independently on each channel, that is, depth-wise convolution without changing the number of channels. Then a standard convolution operation is performed, that is, point-by-point convolution of the feature map is generated by changing the number of channels.

[0008] Furthermore, the upsampling module includes a transposed convolution layer and a ReLU activation layer. The calculation process of the upsampling module is: , in is a two-dimensional transposed convolution operation, is a nonlinear activation function, is a two-dimensional convolution operation, is the output of the upsampling module.

[0009] Furthermore, the fusion module fuses the outputs of the time series prediction module and the upsampling module, and the calculation process is as follows: , in Represents a splicing operation, represents a two-dimensional convolution operation, is the activation function, Represent the outputs of the timing prediction module and the upsampling module respectively.

[0010] Furthermore, The offset is calculated using a bilinear interpolation operation.

[0011] Furthermore, bilinear interpolation is used in the upsampling operation to double the spatial dimension of the input data.

[0012] Furthermore, there are 3 layers of time series prediction modules, upsampling modules, and fusion modules of different scales.

[0013] The present invention also discloses a device for intelligent inspection of equipment based on deep learning multi-scale, comprising a processor and a memory storing program instructions, wherein the processor is configured to execute the method for intelligent inspection of equipment based on deep learning multi-scale as described above when running the program instructions.

[0014] The present invention also discloses a storage medium storing program instructions, which, when running, execute the above-mentioned deep learning multi-scale equipment intelligent inspection method.

[0015] Beneficial effects of the present invention: 1. Feature extraction encoder network Feature extraction Encoder is composed of multiple convolution modules ConvBlock stacked together, each module contains a deformable convolution layer, a common three-dimensional convolution layer and an activation layer. This combination not only enriches the feature extraction capability of the model, but also enhances the nonlinear expression capability of the network through the introduction of activation functions Sigmoid and Tanh, so that the model can handle more complex detection tasks and accurately locate different types of target defect images in complex scene images.

[0016] 2. The introduction of deformable convolutional layers is an important innovation of the model. By allowing the position of the convolution kernel in the feature map to be recalculated according to the affine transformation, the model can simultaneously consider the information at different positions in the sequence when predicting, thereby significantly enhancing the ability to extract features in spatial positions and effectively improving the detection accuracy of objects at different positions in the image.

[0017] 3. The deep learning multi-scale equipment intelligent inspection model can effectively process target features of different scales by designing a multi-scale feature decoder network Multi-Scale Decoder. This design enhances the adaptability of the model in spatial transformation when processing target images, making the model more robust and accurate when processing small targets of power equipment with different sizes and shapes. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a structural diagram of the equipment intelligent inspection model. DETAILED DESCRIPTION

[0019] In order to be able to understand the features and technical contents of the embodiments of the present disclosure in more detail, the implementation of the embodiments of the present disclosure is described in detail below in conjunction with the accompanying drawings. The attached drawings are for reference only and are not used to limit the embodiments of the present disclosure. In the following technical description, for the convenience of explanation, a full understanding of the disclosed embodiments is provided through multiple details. However, one or more embodiments can still be implemented without these details. In other cases, to simplify the drawings, well-known structures and devices can be simplified for display.

[0020] Example 1 This embodiment discloses a multi-scale intelligent equipment inspection method based on deep learning, which is suitable for processing small target detection and can improve the accuracy of small target detection on power equipment, including the following steps: S01. Data collection and preprocessing: Before the input data enters the network, it needs to be preprocessed. Directly obtain the power equipment image, or extract the power equipment image from the collected video, use the annotation tool to annotate the target to be detected in the power equipment image, preprocess the annotated power equipment image to obtain the training image, and collect the training image in the training sample set.

[0021] The target image of the power equipment is directly captured by the camera, or the target image is extracted from the collected video data. The specific method is to intercept the video every 30 frames to generate the target image of the power equipment with a resolution of 480×480. Then, these target images are preprocessed, including cropping, flipping and scaling, to generate images for training. Finally, these preprocessed training images are aggregated into a training sample set.

[0022] S02. Build a deep learning multi-scale equipment intelligent inspection model, such as Figure 1 As shown in the figure, the deep learning multi-scale equipment intelligent inspection model includes a feature extraction encoder Feature extraction Encoder and a multi-scale feature decoder Multi-Scale Decoder. The encoder is composed of multiple convolution modules ConvBlock stacked together. Each convolution module ConvBlock includes a deformable convolution layer, a three-dimensional convolution layer and an activation layer. The deformable convolution layer uses affine transformation to recalculate the position of the convolution kernel in the feature map, so that the encoder can simultaneously consider the information at different positions in the image sequence when extracting features.

[0023] The decoder includes multiple layers of temporal prediction modules of different scales, namely Temporal Prediction Block, upsampling module UpsamplingBlock, fusion module FusionBlock and a layer of deconvolution module ConvTranspose2d. The temporal prediction module and upsampling module of the first layer are respectively connected to the output of the encoder, the temporal prediction modules and upsampling modules of other layers are connected to the output of the fusion module of the previous layer, the fusion module is connected to the output of the temporal prediction module and upsampling module of this layer, the input of the deconvolution module is connected to the output of the fusion module of the last layer, and the output of the deconvolution module is the inspection result of the deep learning multi-scale equipment intelligent inspection model.

[0024] S03, model training, sampling training sample sets to train the deep learning multi-scale equipment intelligent inspection model to obtain a trained equipment intelligent inspection model; S04. Collect images of power equipment in real time and input them into the trained equipment intelligent inspection model for prediction to obtain inspection results.

[0025] In this embodiment, the core of the deep learning multi-scale equipment intelligent inspection model is the feature extraction encoder network Feature extraction Encoder and the multi-scale feature decoder network Multi-Scale Decoder to enhance the adaptability of the model in spatial transformation and its ability in spatial feature extraction, while enhancing the nonlinear expression ability of the network. Compared with the traditional convolutional neural network, this model can handle small target detection and further improve the detection accuracy of small target defects.

[0026] One problem with the end-to-end object detection model is that it tends to ignore the underlying details and location information of the image, making it difficult to effectively identify and detect small objects. The root of this problem lies in two aspects: one is the lack of contextual information, and the other is the lack of semantic information in shallow features. In order to more effectively extract the features of small objects in the image, the core of the feature extraction encoder Feature extraction Encoder in this embodiment is composed of multiple convolutional modules ConvBlock stacked together, each convolutional module ConvBlock includes a deformable convolution layer, a normal three-dimensional convolution layer and an activation layer, where the activation layer uses Sigmoid and Tanh activation functions respectively.

[0027] The convolution module allows the model to consider information at different positions in the sequence when making predictions, enhancing the model's ability to extract features in spatial positions. Deformable convolution recalculates the position of the convolution kernel in the feature map using affine transformation, thereby achieving adaptive learning of the model. The specific operation is to treat the spatial position of each layer at time t the same as ordinary convolution, while the corresponding position at time t-1 is calculated by deformable convolution. The calculation process is: , in , Represent the feature graphs at time t-1 and t respectively, , represents the size of the convolution kernel area, Represents a splicing operation, Indicates the current sampling point The offset on the feature map, by the convolutional subnetwork Generated from the feature maps at adjacent times t-1 and t, To calculate the offset The convolution operation, Indicates the calculation formula of deformable convolution for the feature map at time t-1, D represents the convolution kernel area, which defaults to 3×3 here; , They represent the kernel functions used for the feature graph at time t and the feature graph at time t-1, respectively. Indicates the current sampling point The corresponding other areas involved in the convolution calculation, Indicates the calculation formula of using ordinary convolution for the feature map at time t, represents the activation function, , Indicates l The feature map at time t-1 and the feature map at time t input by the layer convolution module, Respectively represent l The output of the convolutional module is l+1 The feature map at time t input to the layer convolution module.

[0028] In this embodiment, the multi-scale feature decoder network Multi-Scale Decoder mainly includes three layers of temporal prediction modules of different scales, Temporal Prediction Block, upsampling module UpsamplingBlock, fusion module FusionBlock and a layer of deconvolution operation. The task backend decoder needs to implement upsampling to restore the compressed features to the size of the input data. Considering that information will be lost during the encoder calculation process, this paper processes the multi-scale features of the middle layer of the network into fusion features through the fusion module to reduce information loss and improve the efficiency of target detection.

[0029] In the design of the time series prediction module, we combine bilinear interpolation and depthwise separable convolution operations. First, bilinear interpolation doubles the length and width of the data. Then, a depthwise separable convolution network is used to replace ordinary convolutions. At the same time, batch normalization layers and ReLU activation functions are used to enhance the learning and generalization capabilities of the model.

[0030] The specific operation is to use a depth-separable convolution operation, and its calculation process is as follows: , Where x represents the input data, and scale is the upsampling factor, which means that the spatial dimension of the input data is doubled. represents the upsampling operation, represents the depthwise convolution operation, represents the point-by-point convolution operation, represents the batch normalization operation, is the activation function, Represents the output of the time series prediction module; The depthwise separable convolution operation utilizes the characteristics of image data space and depth dimensions that can be separated and independently calculated, and reduces the number of parameters by decomposing the convolution operation, thereby optimizing the model structure. First, each channel is independently convolved, that is, the depthwise convolution of the corresponding number of channels (Depthwise Convolution), and then the standard convolution operation is performed, that is, the pointwise convolution (Pointwise Convolution) that changes the number of channels to generate the feature map.

[0031] In a convolutional neural network, feature extraction of an image often causes a change in its size. In order to restore the original resolution of the feature map, an upsampling operation is usually required. The upsampling module Upsampling-Block of this embodiment consists of a transposed convolution layer and a subsequent ReLU activation layer, and its calculation process is as follows: , in is a two-dimensional transposed convolution operation, is a nonlinear activation function, is a two-dimensional convolution operation, is the output of the upsampling module.

[0032] The fusion module effectively integrates multi-scale information, realizes the mutual complementation of features at different scales, and provides richer feature information for target prediction at the task backend. The calculation process is as follows: in Represents a splicing operation, represents a two-dimensional convolution operation, is the activation function, Represent the outputs of the timing prediction module and the upsampling module respectively.

[0033] In this embodiment, the model training process is the same as the existing training method and will not be repeated here.

[0034] This embodiment uses a codec architecture to build a multi-scale intelligent equipment inspection model based on deep learning. The input of the model is a target image of power equipment with a resolution of 480×480, and the output feature map is the final target detection result. The feature extraction encoder network Feature extraction Encoder in the model uses information at different positions in the deformable convolution sampling sequence, and the multi-scale feature decoder network Multi-Scale Decoder reduces the loss of feature information in the upsampling process through the fusion module, which is particularly important for improving the detection accuracy of small target defects.

[0035] Through this specific implementation, the flexible adaptability of deep learning multi-scale feature extraction and decoding and the strong ability of deformable convolutional layers in spatial feature extraction can provide an accurate and reliable small target defect detection model for power equipment. This not only enhances the accuracy and efficiency of power equipment defect detection, but also provides valuable reference information for preventive maintenance and fault prediction of power systems.

[0036] Example 2 The embodiment of the present disclosure provides a device for intelligent inspection of equipment based on deep learning multi-scale, including a processor and a memory. Optionally, the device may also include a communication interface and a bus. Among them, the processor, the communication interface, and the memory can communicate with each other through the bus. The communication interface can be used for information transmission. The processor can call the logic instructions in the memory to execute the device intelligent inspection method based on deep learning multi-scale of the above embodiment.

[0037] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product.

[0038] The memory, as a computer-readable storage medium, can be used to store software programs and computer executable programs, such as program instructions / modules corresponding to the method in the embodiment of the present disclosure. The processor executes the functional application and data processing by running the program instructions / modules stored in the memory, that is, the device intelligent inspection method based on deep learning multi-scale in the above embodiment is implemented.

[0039] The memory may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required for at least one function; the data storage area may store data created according to the use of the terminal device, etc. In addition, the memory may include a high-speed random access memory and may also include a non-volatile memory.

[0040] Example 3 An embodiment of the present disclosure provides a computer-readable storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured to execute the above-mentioned multi-scale intelligent inspection method for equipment based on deep learning.

[0041] The computer-readable storage medium mentioned above may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0042] The technical solution of the embodiment of the present disclosure can be embodied in the form of a software product, which is stored in a storage medium and includes one or more instructions for enabling a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiment of the present disclosure. The aforementioned storage medium may be a non-transient storage medium, including: a USB flash drive, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a disk or an optical disk, and other media that can store program codes, or a transient storage medium.

[0043] The above description and the accompanying drawings fully illustrate the embodiments of the present disclosure so that those skilled in the art can practice them. Other embodiments may include structural, logical, electrical, process and other changes. The embodiments represent only possible changes. Unless explicitly required, separate components and functions are optional, and the order of operation may vary. The parts and features of some embodiments may be included in or replace the parts and features of other embodiments. Moreover, the words used in this application are only used to describe the embodiments and are not used to limit the scope of protection. As used in the description in the text, unless the context clearly indicates, the singular forms of "a", "an" and "the" are intended to include plural forms as well. Similarly, the term "and / or" as used in this application refers to any and all possible combinations of one or more associated listings. In addition, when used in this application, the term "comprise" and its variants "comprises" and / or comprising refer to the presence of stated features, wholes, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or groups thereof. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the presence of other identical elements in the process, method or device including the elements. In this article, each embodiment may focus on the differences from other embodiments, and the same and similar parts between the embodiments may refer to each other. For the methods, products, etc. disclosed in the embodiments, if they correspond to the method part disclosed in the embodiments, then the relevant parts can refer to the description of the method part.

[0044] Those skilled 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, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software may depend on the specific application and design constraints of the technical solution. The technicians may 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 embodiments of the present disclosure. The technicians may clearly understand that for the convenience and simplicity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments, and will not be repeated here.

Claims

1. A multi-scale intelligent inspection method for equipment based on deep learning, characterized by: The following steps are involved: S01, data collection and preprocessing, collecting power equipment images, using annotation tools to annotate the targets to be detected in the power equipment images, preprocessing the annotated power equipment images to obtain training images, and collecting the training images in a training sample set; S02. Construct a deep learning multi-scale equipment intelligent inspection model. The deep learning multi-scale equipment intelligent inspection model includes an encoder and a decoder. The encoder is composed of a plurality of stacked convolution modules. Each convolution module includes a deformable convolution layer, a three-dimensional convolution layer and an activation layer. The deformable convolution layer uses affine transformation to recalculate the position of the convolution kernel in the feature map, so that the encoder can simultaneously consider the information at different positions in the image sequence when extracting features; the decoder includes multiple layers of timing prediction modules, upsampling modules, fusion modules and a deconvolution module of different scales. The timing prediction module and the upsampling module of the first layer are respectively connected to the output of the encoder, the timing prediction modules and upsampling modules of other layers are connected to the output of the fusion module of the previous layer, the fusion module is connected to the output of the timing prediction module and the upsampling module of the layer, the input of the deconvolution module is connected to the output of the fusion module of the last layer, and the output of the deconvolution module is the inspection result of the deep learning multi-scale equipment intelligent inspection model; S03, model training, sampling training sample sets to train the deep learning multi-scale equipment intelligent inspection model to obtain a trained equipment intelligent inspection model; S04. Collect images of power equipment in real time and input them into the trained equipment intelligent inspection model for prediction to obtain inspection results.

2. The device intelligent inspection method based on deep learning multi-scale according to claim 1 is characterized by: The convolution module of the encoder processes the image as follows: ， in , Represent the feature graphs at time t-1 and t respectively, , represents the size of the convolution kernel area, Represents a splicing operation, Indicates the current sampling point The offset on the feature map, by the convolutional subnetwork Generated from the feature maps at adjacent times t-1 and t, To calculate the offset The convolution operation, It represents the calculation formula of deformable convolution for the feature map at time t-1, and D represents the convolution kernel area; , They represent the kernel functions used for the feature graph at time t and the feature graph at time t-1, respectively. Indicates the current sampling point The corresponding other areas involved in the convolution calculation, Indicates the calculation formula of using ordinary convolution for the feature map at time t, represents the activation function, , Indicates l The feature map at time t-1 and the feature map at time t input by the layer convolution module, Respectively represent l The output of the convolutional module is l+1 The feature map at time t input to the layer convolution module.

3. The device intelligent inspection method based on deep learning multi-scale according to claim 1 is characterized by: The time series prediction module combines bilinear interpolation and depthwise separable convolution operations. The bilinear interpolation method doubles the length and width of the image, and then uses depthwise separable convolution to replace ordinary convolution. At the same time, the batch normalization layer and ReLU activation function are used to enhance the learning and generalization capabilities of the model. The calculation process of the time series prediction module is as follows: ， Where x represents the input data, and scale is the upsampling factor, which means that the spatial dimension of the input data is doubled. represents the upsampling operation, represents the depthwise convolution operation, represents the point-by-point convolution operation, represents the batch normalization operation, is the activation function, Represents the output of the time series prediction module; The depth-wise separable convolution operation utilizes the characteristics of image data space and depth dimensions that can be separably and independently calculated. It reduces the number of parameters by decomposing the convolution operation. First, convolution calculation is performed independently on each channel, that is, depth-wise convolution without changing the number of channels. Then a standard convolution operation is performed, that is, point-by-point convolution of the feature map is generated by changing the number of channels.

4. The device intelligent inspection method based on deep learning multi-scale according to claim 1 is characterized by: The upsampling module includes a transposed convolution layer and a ReLU activation layer. The calculation process of the upsampling module is: ， in is a two-dimensional transposed convolution operation, is a nonlinear activation function, is a two-dimensional convolution operation, is the output of the upsampling module.

5. The device intelligent inspection method based on deep learning multi-scale according to claim 1 is characterized by: The fusion module fuses the outputs of the time series prediction module and the upsampling module. The calculation process is as follows: ， in Represents a splicing operation, represents a two-dimensional convolution operation, is the activation function, Represent the outputs of the timing prediction module and the upsampling module respectively.

6. The device intelligent inspection method based on deep learning multi-scale according to claim 2 is characterized by: The offset is calculated using a bilinear interpolation operation.

7. The device intelligent inspection method based on deep learning multi-scale according to claim 3 is characterized by: The upsampling operation uses bilinear interpolation to double the spatial dimension of the input data.

8. The device intelligent inspection method based on deep learning multi-scale according to claim 1 is characterized by: It has 3 layers of time series prediction modules of different scales, upsampling modules, and fusion modules.

9. A multi-scale intelligent inspection device for equipment based on deep learning, comprising a processor and a memory storing program instructions, characterized in that: The processor is configured to execute the deep learning multi-scale equipment intelligent inspection method according to any one of claims 1 to 8 when running the program instructions.

10. A storage medium storing program instructions, characterized in that: When the program instructions are run, they execute the deep learning multi-scale equipment intelligent inspection method according to any one of claims 1 to 8.

Images