Instrument remote inspection method based on computer vision

Abstract

The invention discloses an instrument remote inspection method based on computer vision, and the method comprises the following steps: 1, collecting an input image of an instrument, and obtaining a sector index; 2, respectively inputting the local images of the instruments and meters into a teacher network and a student network in an improved OfficientAD model; 3, based on a cross-domain joint feature construction mechanism, generating a cross-domain Transform guidance graph, and outputting student multi-scale features; 4, generating student prediction features and teacher target features through the online projection sub-network, the online prediction sub-network and the target projection sub-network of the BYOL dual-network structure; 5, generating an abnormal feature map; and 6, outputting a remote inspection result of the instrument. According to the invention, the requirements of high efficiency, precision and intelligence of automatic inspection of the remote intelligent instrument in an industrial scene are met.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, industrial automation and intelligent inspection technology, and in particular to a remote inspection method for instruments and meters based on computer vision. Background Technology

[0002] With the development of industrial automation and intelligent manufacturing, a large number of instruments and meters have been deployed in key industrial sites such as petrochemicals, power energy, and metallurgy to monitor equipment operating status and process parameters in real time. However, these instruments and meters are mostly installed in widely distributed, complex, and even high-risk working environments. The traditional method of relying on manual inspection is not only inefficient but also poses significant safety hazards in dangerous areas such as high temperature, high pressure, toxic or flammable environments. At the same time, due to the uneven skill levels of personnel and inconsistent recording methods, manual reading of instrument readings is also prone to misreading, missed detections, and information lag.

[0003] To improve the real-time performance and security of on-site monitoring, some enterprises have begun to explore acquiring instrument images through industrial camera terminals and using computer vision methods for identification and interpretation. Currently, mainstream visual inspection methods largely rely on standard convolutional neural networks for target detection and classification, such as Faster R-CNN, the YOLO series, and RetinaNet. However, these methods are mostly designed for natural images and cannot adequately model the circumferential structural characteristics of instrument images with regular structures and numerous circular components (such as dials, pointers, and scale rings). Furthermore, existing visual anomaly detection methods often employ image reconstruction or feature residual mechanisms, which struggle to provide highly reliable responses to anomalies in unsupervised or low-sample scenarios, particularly performing poorly when detecting fine-grained anomalies such as detached pointers, missing digits, or blurred scales.

[0004] While the knowledge distillation framework with a teacher-student structure is widely used in industrial image detection and representation learning, it generally lacks effective path hierarchical and cross-domain guidance mechanisms in handling the fusion of high-frequency details and low-frequency structural features. This results in the detection model's limited ability to perceive abnormal changes in local areas of the instrument, making it difficult to meet the robustness and generalization requirements of industrial applications.

[0005] Therefore, how to provide a remote inspection method for instruments and meters based on computer vision is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a remote inspection method for instruments and meters based on computer vision. This method integrates a structure-aware detection network and an improved anomaly recognition model. It extracts the structural regions of the instrument by constructing a target detection network with circumferential perception capabilities, and combines this with the EfficientAD model, which possesses cross-domain graph neural network mechanisms and Transformer mapping capabilities, for anomaly recognition. This invention fully utilizes computer vision, graph convolutional neural networks, multi-head attention mechanisms, and deep feature fusion methods. It details the entire process of achieving high-precision remote inspection and automatic anomaly recognition of instruments and meters in hazardous or widely distributed industrial environments. It possesses advantages such as strong structural feature modeling capabilities, high robustness in anomaly detection, and strong adaptability to complex industrial environments.

[0007] A remote inspection method for instruments based on computer vision according to an embodiment of the present invention includes the following steps:

[0008] Step 1: Deploy camera terminals at the industrial site to capture input images from instruments and send them to a remote processing device. A target detection network with circular convolution performs target detection on the input images and obtains sector indexes.

[0009] Step 2: Obtain the corresponding local image of the instrument based on the sector index, and input the local image of the instrument into the teacher network and student network of the improved EfficientAD model respectively; in the student network, input the local image of the instrument into the orientation feature enhancement module, and form an orientation response feature map through multi-directional convolution kernels; process the orientation response feature map through the detail feature extraction path to obtain high-frequency detail features; process the local image of the instrument through the structural feature extraction path to obtain low-frequency structural features;

[0010] Step 3: Based on the cross-domain joint feature construction mechanism, construct a joint graph structure for high-frequency detail features and low-frequency structural features, perform graph convolution update, complete spatial location reconstruction, generate cross-domain graph guidance graph and cross-domain Transformer guidance graph, and output student multi-scale features by setting a feature fusion network;

[0011] Step 4: In the feature difference calculation module, the alignment mapping of student multi-scale features is completed through the multi-scale feature mapping sub-network. The predicted features of students and the target features of teachers are generated through the online projection sub-network, online prediction sub-network and target projection sub-network of the BYOL dual network structure, respectively, and a feature difference map is constructed.

[0012] Step 5: In the feature energy anomaly enhancement module, perform local spatial domain energy analysis and frequency domain energy analysis on the feature difference map to generate anomaly feature maps;

[0013] Step 6: In the anomaly scoring generation module, a residual map is generated based on the anomaly feature map, and anomalies are determined in the local images of the instruments based on a preset threshold, and the remote inspection results of the instruments are output.

[0014] Optionally, the improved EfficientAD model includes a teacher network, a student network, a directional feature enhancement module, a feature fusion network, a feature difference calculation module, a feature energy anomaly enhancement module, and an anomaly score generation module, wherein the student network includes a detailed feature extraction path and a structural feature extraction path.

[0015] Optionally, step one specifically includes:

[0016] Camera terminals are installed at predetermined locations in the industrial site. The field of view of the camera terminals covers the instruments and meters. The camera terminals are used to collect on-site images containing the instruments and meters as input images.

[0017] In the remote processing device, the input image is sent to the remote processing device via a wired network;

[0018] The input image is fed into a target detection network with circumferential convolutional kernels in a remote processing device;

[0019] The target detection network with circumferential convolutional kernels includes a backbone feature extraction network, a polar coordinate transformation branch, a circle center localization head, a radius prediction head, and a sector partitioning head;

[0020] Using the input image as input, the backbone feature extraction network performs convolution operations, downsampling operations, and feature mapping to generate the first feature map;

[0021] In the polar coordinate transformation branch, the first feature map is transformed into a polar coordinate feature map with the origin of multiple candidate circle center positions according to the preset polar coordinate sampling rules;

[0022] The preset polar coordinate sampling rule is as follows: taking the center position of each candidate circle in the input image as the origin, the radial direction is divided into [variables] at a preset radius sampling interval. A sampling radius is used to divide the angular direction from 0 to 360 degrees at a preset angular sampling interval. For each discrete angle, the first feature map is sampled according to the radial index and the angle index to form a sample of size [value missing]. The polar coordinate feature map.

[0023] A convolutional layer with a circumferential convolutional kernel is set in the polar coordinate feature map, and convolution operation is performed along the polar coordinate angle direction to obtain a circumferential feature map that characterizes the circular structure of the instrument panel.

[0024] The circumferential feature map is input into the center localization head. Through the prediction structure containing two convolutional layers and one normalization layer, the center confidence map is output for each candidate center position, and the center coordinates of the instrument panel are determined based on the position with the highest confidence.

[0025] The circumferential feature map is input into the radius prediction head, and the radius prediction structure, which includes two convolutional layers, outputs the radius parameter with the center of the circle as a reference.

[0026] The circumferential feature map is input into the sector partitioning head. The classification features in the angular direction are obtained by convolution operation along the angular direction. Based on the preset angular discretization rule, the range from 0 to 360 degrees is divided into equal angles. The output sector category sequence of length N is used as the sector index of the N circumferential sectors obtained by partitioning along the circumferential direction.

[0027] Optionally, step two specifically includes:

[0028] Based on the center coordinates, radius parameters, and sector index, local images of instruments and meters corresponding to each sector are cropped from the input image, and the cropped local images of instruments and meters are combined into an image data set.

[0029] Simultaneously input local images of each instrument in the image dataset to both the teacher network and the student network;

[0030] In the student network, local images of instruments are input to the orientation feature enhancement module;

[0031] The directional feature enhancement module includes a multi-directional convolution kernel, which performs multi-directional convolution operations on local images of instruments to generate corresponding directional response feature maps.

[0032] The multi-directional convolution kernel is set according to a preset set of directions, which includes 0-degree direction, 45-degree direction, 90-degree direction, 135-degree direction and 180-degree direction;

[0033] The directional response feature maps are concatenated along the channel dimension, and the concatenated feature maps are transformed by a one-dimensional convolutional layer to obtain the input feature map.

[0034] The detail feature extraction path includes a detail feature extraction network consisting of three convolutional layers and two downsampling layers. The number of channels in the three convolutional layers are C1, C2 and C3, respectively, and C2 is greater than C1 and C3 is greater than C2.

[0035] High-frequency detail features are obtained by performing layer-by-layer convolution and downsampling on the input feature map through a detail feature extraction network;

[0036] The structural feature extraction path includes a structural feature extraction network consisting of a convolutional layer and a downsampling layer, wherein the number of channels in the convolutional layer is C4, and C4 is less than C1;

[0037] Low-frequency structural features are obtained by convolution and downsampling local images of instruments and meters using a structural feature extraction network.

[0038] In the teacher network, local images of instruments are used as input, and features are extracted layer by layer through a pre-trained convolutional neural network backbone, outputting corresponding multi-scale teacher features at multiple scales.

[0039] Optionally, the cross-domain joint feature construction mechanism is specifically as follows:

[0040] In the student network, high-frequency detail features and low-frequency structural features are aligned in spatial dimension and then used as the first domain input features and the second domain input features, respectively.

[0041] A spatial domain node set is constructed using the first domain input features, and a frequency domain node set is constructed using the second domain input features;

[0042] Based on the spatial domain node set and the frequency domain node set in spatial location, cross-domain connection edges are established to form a joint graph structure containing spatial domain nodes, frequency domain nodes and cross-domain edges;

[0043] Perform graph convolution update operation on the joint graph structure to obtain the node features after cross-domain update;

[0044] A feature reconstruction mechanism based on spatial location mapping is adopted to reconstruct the updated node features in the corresponding spatial locations to generate a cross-domain graph guidance graph;

[0045] The feature reconstruction mechanism based on spatial location mapping writes the updated node features across domains into a two-dimensional feature map matrix of a preset size according to the corresponding spatial coordinates set for spatial domain nodes and frequency domain nodes when establishing the joint graph structure.

[0046] For locations not covered by nodes, local interpolation is performed through a reconstructed network containing convolutional and interpolation layers to generate a complete cross-domain graph guide graph.

[0047] The cross-domain graph guidance graph and the first domain input features are concatenated along the channel dimension to form a query vector. The second domain input features are used as key vectors and value vectors, respectively, and then input into a cross-domain Transformer structure containing multi-head attention computation layers and convolutional layers.

[0048] The similarity calculation between the query vector and the key vector is performed through a multi-head attention computing layer, and the value vector is linearly mapped according to the similarity to output cross-domain mapping features;

[0049] Cross-domain mapped features are transformed through convolutional layers to generate cross-domain Transformer guidance maps;

[0050] The cross-domain graph guidance map and the cross-domain Transformer guidance map are concatenated along the channel dimension and input into a feature fusion network consisting of three convolutional layers, one downsampling layer, and one upsampling layer. Convolution, downsampling, and upsampling operations are performed sequentially to generate fused feature maps with different spatial resolutions. These fused feature maps are then used as multi-scale features for students.

[0051] Optionally, step four specifically involves:

[0052] In the feature difference calculation module, a multi-scale feature mapping sub-network is set up for multi-scale teacher features and multi-scale student features, with each scale corresponding to the other.

[0053] The student multi-scale features at each scale are input into the corresponding multi-scale feature mapping sub-network, and spatial resolution alignment and channel mapping are performed to obtain the student mapping features.

[0054] The student mapping features are input into the online projection subnetwork and the online prediction subnetwork respectively to obtain the student prediction features at the corresponding scale.

[0055] The multi-scale teacher features of the corresponding scale are input into the target projection sub-network to obtain the teacher target features of the corresponding scale.

[0056] The online projection subnetwork, online prediction subnetwork, and target projection subnetwork are derived from the BYOL dual-network structure.

[0057] Based on the differences between student predicted features and teacher target features at each scale, a corresponding multi-scale feature difference map is generated. The multi-scale feature difference map is then fused according to a preset fusion rule to obtain a feature difference map that represents the difference relationship between the teacher network and the student network.

[0058] Optionally, step five specifically includes:

[0059] In the spatial domain of the feature energy anomaly enhancement module, the feature difference map is divided into M local regions in the spatial dimension according to the preset window size;

[0060] In each local region, the corresponding local spatial domain energy value is calculated based on the amplitude and distribution of the feature values ​​in the current local region, thus obtaining a local spatial domain energy map that corresponds one-to-one with the feature difference map in spatial location.

[0061] The local spatial domain energy map is normalized, and a spatial energy weight map is generated according to the preset spatial energy mapping relationship.

[0062] In the frequency domain of the feature energy anomaly enhancement module, a frequency domain transformation is performed on the feature difference map along the channel dimension to convert the feature difference map into the corresponding frequency domain feature map;

[0063] In the frequency domain feature map, the frequency domain features are divided into frequency bands according to the preset frequency band division rules, and the corresponding frequency domain energy value is calculated based on the statistical results of the frequency domain feature amplitude in each frequency band to generate a frequency domain energy distribution map;

[0064] Normalize the frequency domain energy distribution map and generate a frequency domain energy weight map according to the preset frequency domain energy mapping relationship;

[0065] The spatial energy weight map and the frequency domain energy weight map are fused in spatial location according to a preset fusion rule to obtain a fused energy weight map.

[0066] The fused energy weight map and feature difference map are weighted at their corresponding spatial locations and channel dimensions. The feature responses in the feature difference map are then subjected to energy enhancement processing to generate an anomaly feature map.

[0067] Optionally, step six specifically includes:

[0068] In the anomaly scoring generation module, the anomaly feature map is aligned with the local images of each instrument in terms of sector index and spatial location to obtain the anomaly response matrix of each sector.

[0069] The abnormal response values ​​in the abnormal response matrix are normalized, and sector-level residual maps are generated according to the preset residual coding rules.

[0070] Based on the sector-level residual map, statistical calculations are performed on the abnormal response values ​​of each sector within each instrument within a preset spatial range to obtain sector-level anomaly scoring parameters. These parameters are then aggregated according to the instrument dimension to obtain instrument-level anomaly scoring parameters.

[0071] The instrument-level anomaly scoring parameter is compared with a preset threshold. When the instrument-level anomaly scoring parameter is greater than or equal to the preset threshold, the corresponding instrument is marked as an abnormal state. When the instrument-level anomaly scoring parameter is less than the preset threshold, the corresponding instrument is marked as a normal state.

[0072] Based on the status markers of each instrument, generate and output remote inspection result data containing instrument identification information and status information.

[0073] The beneficial effects of this invention are:

[0074] This invention constructs a remote inspection method for instruments based on computer vision, introducing a target detection network with a circumferential convolution structure. This network performs circular perceptual convolution in polar coordinates, significantly enhancing its ability to detect and represent circumferential targets such as instrument dials. Through this network, the center, radius, and circumferential sector distribution index of each instrument's dial can be obtained with high precision, thereby achieving sector-level spatial modeling that better reflects the physical layout of the instrument and providing a more accurate and structured input foundation for subsequent local analysis.

[0075] In the anomaly detection phase, the improved EfficientAD model incorporates several structural optimizations. The student network employs a "dual-path structure," consisting of a detail feature extraction path and a structural feature extraction path, to respectively mine local high-frequency anomaly details and overall low-frequency structural information. Combined with a directional feature enhancement module, it extracts directional anomaly features such as cracks, deviations, and occlusions in the instrument through multi-directional convolutional kernels, enhancing the model's ability to perceive subtle anomalies.

[0076] This invention proposes a cross-domain graph convolution guidance mechanism and a cross-domain Transformer mapping mechanism. By constructing spatial and frequency domain node sets and their joint graph structure in the student network, cross-semantic feature fusion is achieved through graph convolution updates and feature reconstruction. Furthermore, cross-domain attention mapping is performed using the Transformer structure, which significantly improves the collaborative modeling ability between high- and low-frequency features. This results in the construction of more expressive and discriminative fusion features, effectively improving the accuracy and robustness of anomaly detection.

[0077] In the feature difference analysis and anomaly enhancement stage, local energy calculation methods in the spatial and frequency domains are combined to extract the energy saliency regions of the feature difference map. Multidimensional weighted amplification is then performed by fusing energy weight maps, ensuring stable and precise anomaly response capabilities even in industrial scenarios with complex lighting, background interference, or local blurring. Finally, a dual-level scoring aggregation mechanism based on sectors and instruments outputs accurate and clearly defined anomaly state judgment results.

[0078] This invention significantly improves detection accuracy and robustness of anomaly identification through perceptual optimization of the target detection structure and multi-level improvement of the anomaly detection model. It also enhances the deep expressive power of feature fusion and the practicality of the model, meeting the requirements of efficiency, accuracy and intelligence for automatic inspection of remote intelligent instruments in industrial scenarios. Attached Figure Description

[0079] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0080] Figure 1This is an overall flowchart of a remote inspection method for instruments and meters based on computer vision proposed in this invention;

[0081] Figure 2 This is a schematic diagram of a target detection network structure with circular convolution for a computer vision-based remote inspection method for instruments and meters proposed in this invention.

[0082] Figure 3 This is a schematic diagram of the improved EfficientAD model structure of a computer vision-based remote inspection method for instruments and meters proposed in this invention. Detailed Implementation

[0083] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0084] refer to Figures 1-3 A remote inspection method for instruments and meters based on computer vision includes the following steps:

[0085] Step 1: Deploy camera terminals at the industrial site to capture input images from instruments and send them to a remote processing device. A target detection network with circular convolution performs target detection on the input images and obtains sector indexes.

[0086] Step 2: Obtain the corresponding local image of the instrument based on the sector index, and input the local image of the instrument into the teacher network and student network of the improved EfficientAD model respectively; in the student network, input the local image of the instrument into the orientation feature enhancement module, and form an orientation response feature map through multi-directional convolution kernels; process the orientation response feature map through the detail feature extraction path to obtain high-frequency detail features; process the local image of the instrument through the structural feature extraction path to obtain low-frequency structural features;

[0087] Step 3: Based on the cross-domain joint feature construction mechanism, construct a joint graph structure for high-frequency detail features and low-frequency structural features, perform graph convolution update, complete spatial location reconstruction, generate cross-domain graph guidance graph and cross-domain Transformer guidance graph, and output student multi-scale features by setting a feature fusion network;

[0088] Step 4: In the feature difference calculation module, the alignment mapping of student multi-scale features is completed through the multi-scale feature mapping sub-network. The predicted features of students and the target features of teachers are generated through the online projection sub-network, online prediction sub-network and target projection sub-network of the BYOL dual network structure, respectively, and a feature difference map is constructed.

[0089] Step 5: In the feature energy anomaly enhancement module, perform local spatial domain energy analysis and frequency domain energy analysis on the feature difference map to generate anomaly feature maps;

[0090] Step 6: In the anomaly scoring generation module, a residual map is generated based on the anomaly feature map, and anomalies are determined in the local images of the instruments based on a preset threshold, and the remote inspection results of the instruments are output.

[0091] In this embodiment, the improved EfficientAD model includes a teacher network, a student network, a directional feature enhancement module, a feature fusion network, a feature difference calculation module, a feature energy anomaly enhancement module, and an anomaly score generation module, wherein the student network includes a detail feature extraction path and a structural feature extraction path.

[0092] In this embodiment, step one specifically includes:

[0093] Camera terminals are installed at predetermined locations in the industrial site. The field of view of the camera terminals covers the instruments and meters. The camera terminals are used to collect on-site images containing the instruments and meters as input images.

[0094] Instruments and meters include pointer-type mechanical instruments and digital display instruments. The pointer-type mechanical instruments include pressure gauges, temperature gauges, differential pressure gauges, vacuum pressure gauges, flow meters, liquid level indicators, ammeters, voltmeters, and power meters. The digital display instruments include LED digital displays, LCD digital displays, digital temperature displays, digital flow displays, digital liquid level displays, and digital electrical measuring instruments.

[0095] The on-site images are images of instruments and their external structures captured by camera terminals. The on-site images include the instrument casing, dial, pointer, scale lines, digital display area, as well as the industrial background and ambient lighting and visual information of the surrounding environment.

[0096] In the remote processing device, the input image is sent to the remote processing device via a wired network;

[0097] The input image is fed into a target detection network with circumferential convolutional kernels in a remote processing device;

[0098] The target detection network with circumferential convolutional kernels includes a backbone feature extraction network, a polar coordinate transformation branch, a circle center localization head, a radius prediction head, and a sector partitioning head;

[0099] Using the input image as input, the backbone feature extraction network performs convolution operations, downsampling operations, and feature mapping to generate the first feature map;

[0100] In the polar coordinate transformation branch, the first feature map is transformed into a polar coordinate feature map with the origin of multiple candidate circle center positions according to the preset polar coordinate sampling rules;

[0101] The preset polar coordinate sampling rule is as follows: taking the center position of each candidate circle in the input image as the origin, the radial direction is divided into [variables] at a preset radius sampling interval. A sampling radius is used to divide the angular direction from 0 to 360 degrees at a preset angular sampling interval. Each discrete angle is used to index the first feature map radially. With angle index Sampling is performed to form a size of The polar coordinate feature map.

[0102] A convolutional layer with a circumferential convolutional kernel is set in the polar coordinate feature map, and convolution operation is performed along the polar coordinate angle direction to obtain a circumferential feature map that characterizes the circular structure of the instrument panel.

[0103] The circumferential feature map is input into the center localization head. Through the prediction structure containing two convolutional layers and one normalization layer, the center confidence map is output for each candidate center position, and the center coordinates of the instrument panel are determined based on the position with the highest confidence.

[0104] The circumferential feature map is input into the radius prediction head, and the radius prediction structure, which includes two convolutional layers, outputs the radius parameter with the center of the circle as a reference.

[0105] The circumferential feature map is input into the sector partitioning head. The classification features in the angular direction are obtained by convolution operation along the angular direction. Based on the preset angular discretization rule, the range from 0 to 360 degrees is divided into equal angles. The output sector category sequence of length N is used as the sector index of the N circumferential sectors obtained by partitioning along the circumferential direction.

[0106] This invention proposes an instrument target detection method based on polar coordinate sampling and circumferential convolution. By constructing a target detection network in a remote processing device, which includes a backbone feature extraction network, a polar coordinate transformation branch, a center positioning head, a radius prediction head, and a sector division head, the method maps image features to polar coordinate space using preset polar coordinate sampling rules and performs circumferential convolution in the angular direction to enhance the perception of circular structures. This allows for high-precision extraction of the center coordinates and radius information of the instrument dial. Based on the angular discretization rules, the method outputs sector indices divided along the circumferential direction, achieving structured and accurate identification of key areas of the instrument.

[0107] In this embodiment, step two specifically includes:

[0108] Based on the center coordinates, radius parameters, and sector index, local images of instruments and meters corresponding to each sector are cropped from the input image, and the cropped local images of instruments and meters are combined into an image data set.

[0109] Simultaneously input local images of each instrument in the image dataset to both the teacher network and the student network;

[0110] In the student network, local images of instruments are input to the orientation feature enhancement module;

[0111] The directional feature enhancement module includes a multi-directional convolution kernel, which performs multi-directional convolution operations on local images of instruments to generate corresponding directional response feature maps.

[0112] The multi-directional convolution kernel is set according to a preset set of directions, which includes 0-degree direction, 45-degree direction, 90-degree direction, 135-degree direction and 180-degree direction;

[0113] The directional response feature maps are concatenated along the channel dimension, and the concatenated feature maps are transformed by a one-dimensional convolutional layer to obtain the input feature map.

[0114] The detail feature extraction path includes a detail feature extraction network consisting of three convolutional layers and two downsampling layers. The number of channels in the three convolutional layers are C1, C2 and C3, respectively, and C2 is greater than C1 and C3 is greater than C2.

[0115] High-frequency detail features are obtained by performing layer-by-layer convolution and downsampling on the input feature map through a detail feature extraction network;

[0116] The structural feature extraction path includes a structural feature extraction network consisting of a convolutional layer and a downsampling layer, wherein the number of channels in the convolutional layer is C4, and C4 is less than C1;

[0117] Low-frequency structural features are obtained by convolution and downsampling local images of instruments and meters using a structural feature extraction network.

[0118] In the teacher network, local images of instruments are used as input, and features are extracted layer by layer through a pre-trained convolutional neural network backbone, outputting corresponding multi-scale teacher features at multiple scales.

[0119] In the teacher network, using local images of instruments as input, feature extraction is performed layer by layer through a pre-trained convolutional neural network backbone. Specifically:

[0120] Convolution operations, nonlinear transformations, and downsampling are performed sequentially according to the hierarchical structure of the backbone network. The continuous convolutional layers in the backbone network are used to extract features from the input image layer by layer.

[0121] As convolutional and downsampling layers are executed sequentially, intermediate feature maps with different spatial dimensions and different numbers of channels are generated at different network depth locations.

[0122] The intermediate feature maps are read from the backbone network according to the multi-scale output nodes set in the network structure, so that each output node corresponds to a feature map of one scale.

[0123] Feature maps at multiple scales are used as multi-scale teacher features for the teacher network.

[0124] This invention addresses local region analysis of instrument images by proposing a student-teacher network structure that combines directional feature enhancement and multi-path feature extraction. Local instrument images are obtained by cropping from the input image based on center coordinates, radius parameters, and sector indices, and then input into both the teacher and student networks. The student network utilizes a directional feature enhancement module with multi-directional convolutional kernels (0°, 45°, 90°, 135°, and 180°) to generate directional response feature maps. These maps are then extracted via detail feature extraction and structural feature extraction paths to obtain high-frequency detail features and low-frequency structural features, respectively. Simultaneously, the teacher network employs a pre-trained convolutional neural network backbone, sequentially performing convolution, nonlinear transformation, and downsampling operations according to a set hierarchical structure. This extracts multi-scale intermediate feature maps with different spatial scales and channel numbers, and outputs corresponding teacher features through multi-scale output nodes. This provides accurate, multi-layered deep feature support for subsequent difference analysis and anomaly detection.

[0125] In this embodiment, the cross-domain joint feature construction mechanism is specifically as follows:

[0126] In the student network, high-frequency detail features and low-frequency structural features are aligned in spatial dimension and then used as the first domain input features and the second domain input features, respectively.

[0127] A spatial domain node set is constructed using the first domain input features, and a frequency domain node set is constructed using the second domain input features;

[0128] Based on the spatial domain node set and the frequency domain node set in spatial location, cross-domain connection edges are established to form a joint graph structure containing spatial domain nodes, frequency domain nodes and cross-domain edges;

[0129] Perform graph convolution update operation on the joint graph structure to obtain the node features after cross-domain update;

[0130] The graph convolution update operation adopts a graph convolution update mechanism based on adjacency propagation, which inputs the node features and adjacent node features in the joint graph structure into the graph convolution layer.

[0131] In the graph convolutional layer, the node features are linearly transformed according to the adjacency relationship of the joint graph structure, and feature aggregation operation is performed in the node dimension to output the cross-domain updated node features.

[0132] A feature reconstruction mechanism based on spatial location mapping is adopted to reconstruct the updated node features in the corresponding spatial locations to generate a cross-domain graph guidance graph;

[0133] The feature reconstruction mechanism based on spatial location mapping writes the updated node features across domains into a two-dimensional feature map matrix of a preset size according to the corresponding spatial coordinates set for spatial domain nodes and frequency domain nodes when establishing the joint graph structure.

[0134] For locations not covered by nodes, local interpolation is performed through a reconstructed network containing convolutional and interpolation layers to generate a complete cross-domain graph guide graph.

[0135] The cross-domain graph guidance graph and the first domain input features are concatenated along the channel dimension to form a query vector. The second domain input features are used as key vectors and value vectors, respectively, and then input into a cross-domain Transformer structure containing multi-head attention computation layers and convolutional layers.

[0136] The similarity calculation between the query vector and the key vector is performed through a multi-head attention computing layer, and the value vector is linearly mapped according to the similarity to output cross-domain mapping features;

[0137] Cross-domain mapped features are transformed through convolutional layers to generate cross-domain Transformer guidance maps;

[0138] The cross-domain graph guidance map and the cross-domain Transformer guidance map are concatenated along the channel dimension and input into a feature fusion network consisting of three convolutional layers, one downsampling layer, and one upsampling layer. Convolution, downsampling, and upsampling operations are performed sequentially to generate fused feature maps with different spatial resolutions. These fused feature maps are then used as multi-scale features for students.

[0139] This invention constructs a cross-domain joint feature mechanism. In the student network, high-frequency detail features and low-frequency structural features are used as input features for the first and second domains, respectively. First, after size alignment in the spatial dimension, node sets in the spatial and frequency domains are constructed, and cross-domain connections are built based on spatial positional correspondences to form a joint graph structure. This joint graph structure updates node features through a graph convolution mechanism based on adjacency propagation, extracting cross-domain updated features that fuse spatial and frequency domain information. Subsequently, a feature reconstruction mechanism using spatial position mapping reconstructs the updated node features into a cross-domain graph guidance graph, and a cross-domain Transformer structure with multi-head attention is used to generate a cross-domain Transformer guidance graph. After concatenating the two types of guidance graphs in the channel dimension, they are input into a feature fusion network composed of multi-layer convolutional structures. Through convolution, downsampling, and upsampling operations, a multi-scale fused feature map is output, which is ultimately used as the student's multi-scale feature input for subsequent difference detection processes, achieving fine-grained cross-semantic domain feature fusion and expression enhancement.

[0140] In this embodiment, step four specifically includes:

[0141] In the feature difference calculation module, a multi-scale feature mapping sub-network is set up for multi-scale teacher features and multi-scale student features, with each scale corresponding to the other.

[0142] The multi-scale feature mapping subnetwork is used to perform spatial dimension alignment and channel transformation on student multi-scale features at different scales, including the sequentially set first Convolutional layer, normalized layer, Convolutional layers and the second Convolutional layer;

[0143] First Convolutional layers are used to compress or expand channels for multi-scale features of students; normalization layers are used to normalize features.

[0144] Convolutional layers are used to preserve spatial structure and extract local features;

[0145] second Convolutional layers are used to map features to BYOL input features of a preset dimension.

[0146] The student multi-scale features at each scale are input into the corresponding multi-scale feature mapping sub-network, and spatial resolution alignment and channel mapping are performed to obtain the student mapping features.

[0147] The student mapping features are input into the online projection subnetwork and the online prediction subnetwork respectively to obtain the student prediction features at the corresponding scale.

[0148] The multi-scale teacher features of the corresponding scale are input into the target projection sub-network to obtain the teacher target features of the corresponding scale.

[0149] The online projection subnetwork, online prediction subnetwork, and target projection subnetwork are derived from the BYOL dual-network structure.

[0150] Based on the differences between student predicted features and teacher target features at each scale, a corresponding multi-scale feature difference map is generated. The multi-scale feature difference map is then fused according to a preset fusion rule to obtain a feature difference map that represents the difference relationship between the teacher network and the student network.

[0151] This invention proposes a teacher-student difference modeling mechanism combining a BYOL feature alignment structure. By setting up a multi-scale feature mapping sub-network, the multi-scale features extracted by the student network are aligned scale-wise in both spatial and channel dimensions. These alignments are then input into an online projection sub-network and an online prediction sub-network to obtain student prediction features for training. Simultaneously, teacher features at the corresponding scale are input into a target projection sub-network to generate teacher target features. These three sub-networks together constitute a BYOL (Bootstrap Your Own Latent) dual-network structure, effectively improving feature representation capabilities in unsupervised scenarios. The differences between student prediction features and teacher target features at each scale are extracted as multi-scale feature difference maps, which are ultimately fused into a unified feature difference map. This map accurately represents the deviations in feature learning between the student and teacher networks, providing a reliable basis for subsequent anomaly enhancement analysis.

[0152] In this embodiment, step five specifically includes:

[0153] In the spatial domain of the feature energy anomaly enhancement module, the feature difference map is divided into M local regions in the spatial dimension according to the preset window size;

[0154] In each local region, the corresponding local spatial domain energy value is calculated based on the amplitude and distribution of the feature values ​​in the current local region, thus obtaining a local spatial domain energy map that corresponds one-to-one with the feature difference map in spatial location.

[0155] The energy value of a local spatial domain is equal to the cumulative amount of the amplitudes of all feature points within that local region, where the contribution of each feature point is the absolute value of its amplitude or the square of its amplitude.

[0156] The local spatial domain energy map is normalized, and a spatial energy weight map is generated according to the preset spatial energy mapping relationship.

[0157] In the frequency domain of the feature energy anomaly enhancement module, a frequency domain transformation is performed on the feature difference map along the channel dimension to convert the feature difference map into the corresponding frequency domain feature map;

[0158] In the frequency domain feature map, the frequency domain features are divided into frequency bands according to the preset frequency band division rules, and the corresponding frequency domain energy value is calculated based on the statistical results of the frequency domain feature amplitude in each frequency band to generate a frequency domain energy distribution map;

[0159] The frequency domain energy value is equal to the cumulative amplitude of all frequency domain characteristic points within a certain frequency band, and is used to characterize the magnitude of the response energy of that frequency band in the frequency domain.

[0160] Normalize the frequency domain energy distribution map and generate a frequency domain energy weight map according to the preset frequency domain energy mapping relationship;

[0161] The spatial energy weight map and the frequency domain energy weight map are fused in spatial location according to a preset fusion rule to obtain a fused energy weight map.

[0162] The fused energy weight map and feature difference map are weighted at their corresponding spatial locations and channel dimensions. The feature responses in the feature difference map are then subjected to energy enhancement processing to generate an anomaly feature map.

[0163] In this embodiment, step six specifically includes:

[0164] In the anomaly scoring generation module, the anomaly feature map is aligned with the local images of each instrument in terms of sector index and spatial location to obtain the anomaly response matrix of each sector.

[0165] The abnormal response values ​​in the abnormal response matrix are normalized, and sector-level residual maps are generated according to the preset residual coding rules.

[0166] Based on the sector-level residual map, statistical calculations are performed on the abnormal response values ​​of each sector within each instrument within a preset spatial range to obtain sector-level anomaly scoring parameters. These parameters are then aggregated according to the instrument dimension to obtain instrument-level anomaly scoring parameters.

[0167] The instrument-level anomaly scoring parameter is compared with a preset threshold. When the instrument-level anomaly scoring parameter is greater than or equal to the preset threshold, the corresponding instrument is marked as an abnormal state. When the instrument-level anomaly scoring parameter is less than the preset threshold, the corresponding instrument is marked as a normal state.

[0168] Based on the status markers of each instrument, generate and output remote inspection result data containing instrument identification information and status information.

[0169] Example 1:

[0170] To verify the feasibility of this invention in practice, it was applied to a concentrated area of ​​equipment with complex inspection needs within a typical industrial park. This area contains a large number of densely distributed and diverse instruments, including pressure gauges, thermometers, ammeters, voltmeters, LED digital displays, and LCD digital displays. For a long time, manual inspection has been relied upon, resulting in low inspection frequency, low anomaly detection rate, and delayed recording. The on-site environment is complex, with some instruments located in high-temperature, vibration, or severely obstructed environments. Traditional remote inspection methods based on general target detection models are prone to detection errors and missed anomaly reports, making it difficult to meet the industrial requirements for high-reliability remote monitoring.

[0171] In the implementation and deployment of this invention, a total of 36 camera terminals were installed at major channel nodes, equipment cabinet areas, and the control center, covering approximately 210 instruments. After acquiring images, the camera terminals transmitted them via a local area network to an edge computing server deployed in the industrial control center. In the target detection module, an improved target detection network with integrated circular convolutional kernels was used to replace the traditional RetinaNet backbone structure to enhance its ability to locate circular structures such as instrument dials. In the polar coordinate transformation branch, the model implemented automatic polar coordinate modeling, center location, and sector division for the instrument dial, resulting in more robust structured target detection results. In the test set of 2600 images annotated in the dataset, the average accuracy (AP) for dial region localization improved by 6.3 percentage points, reaching 91.7%, the center location error decreased from 6.8 pixels in the original model to 3.1 pixels, and the sector recognition accuracy was 96.4%.

[0172] In anomaly detection, the improved EfficientAD model of this invention is applied to image partition detection. The model introduces a multi-directional convolutional directional feature enhancement module to improve the perception of local structural anomalies (such as pointer offset, digit jitter, cracks, and illumination interference), and jointly models high-frequency and low-frequency features through a cross-domain graph convolutional structure and a Transformer fusion mechanism. In actual test data, a total of 19,248 input images were collected over 14 days of operation. After model processing, 456 anomaly candidate images were obtained. After manual confirmation, 389 of these were found to be valid anomalies, resulting in an anomaly detection rate of 85.3%, significantly lower than the traditional Patchcore method (18.1%).

[0173] The BYOL feature alignment module in this invention can effectively learn teacher-student feature differences in unsupervised data scenarios, avoiding the problem of overfitting normal regions by traditional models. Using model stability as a metric, robustness tests were conducted by introducing random occlusion data in an occluded environment. The results show that this model maintains an accurate detection rate of 82.7% for abnormal regions in images with simulated 20% occlusion, which is nearly 9 percentage points higher than the baseline model without cross-domain structure. Specific experimental data are shown in Table 1.

[0174] Table 1. Statistical Table of Anomaly Detection Experiment Results of Remote Inspection System in Actual Deployment

[0175] Instrument type Total number of samples tested Number of abnormal images Number of valid anomaly determinations Abnormal detection rate (%) Average detection time (seconds) Accuracy rate under occlusion (%) pointer pressure gauge 4508 98 85 86.7 1.28 83.2 Pointer thermometer 3216 73 62 84.9 1.30 81.7 Digital voltmeter 2154 46 40 87.0 1.29 84.5 LED digital display temperature gauge 2990 65 56 86.2 1.33 82.6 LCD digital display liquid level gauge 3400 77 68 88.3 1.35 83.9 Digital flow display 3720 97 78 80.4 1.34 81.0 ammeter / voltmeter 3260 81 70 86.4 1.32 82.1 Total / Average 23248 537 459 85.5 1.32 82.8

[0176] As shown in Table 1, from the perspective of response time, when deployed on an edge server, the overall process of this invention, from image acquisition, sector segmentation, feature extraction, anomaly detection to status judgment, takes an average of 1.32 seconds, demonstrating the deployment capability of real-time remote inspection.

[0177] This invention not only achieves comprehensive improvements in anomaly detection rate, robustness, and detection accuracy, but also demonstrates extremely high application value in actual industrial deployments. It can significantly replace traditional manual inspection methods, reduce inspection costs, and improve the safety and operation and maintenance efficiency of industrial sites.

[0178] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A computer vision based method for remote inspection of instruments and meters, characterized in that, The method comprises the following steps: Step one: arranging a camera terminal in an industrial field to collect input images of instruments and meters and send them to a remote processing device, a target detection network with ring convolution is used to detect the input images to obtain a sector index; Step two: according to the sector index, a corresponding local image of the instrument and meter is obtained by cropping, and the local image of the instrument and meter is input into a teacher network and a student network in an improved EfficientAD model; in the student network, the local image of the instrument and meter is input into a directional feature enhancement module to form a directional response feature map through a multi-directional convolution kernel; the directional response feature map is processed through a detail feature extraction path to obtain a high-frequency detail feature; the local image of the instrument and meter is processed through a structure feature extraction path to obtain a low-frequency structure feature; Step three: based on a cross-domain joint feature construction mechanism, a joint graph structure is constructed for the high-frequency detail feature and the low-frequency structure feature, graph convolution update is performed, spatial position reconstruction is completed, a cross-domain graph guide graph and a cross-domain Transformer guide graph are generated, and a student multi-scale feature is output through a feature fusion network; Step four: in a feature difference calculation module, the alignment and mapping of the student multi-scale feature are completed through a multi-scale feature mapping subnetwork, student prediction features and teacher target features are generated through an online projection subnetwork, an online prediction subnetwork and a target projection subnetwork of a BYOL double network structure, and a feature difference graph is constructed; Step five: in a feature energy anomaly enhancement module, local spatial energy analysis and frequency energy analysis are performed on the feature difference graph to generate an abnormal feature graph; Step six: in an abnormal score generation module, a residual graph is generated according to the abnormal feature graph, and the local image of the instrument and meter is abnormally judged based on a preset threshold to output a remote inspection result of the instrument and meter.

2. A computer vision based instrument remote inspection method according to claim 1, wherein, The improved EfficientAD model comprises a teacher network, a student network, a directional feature enhancement module, a feature fusion network, a feature difference calculation module, a feature energy anomaly enhancement module and an abnormal score generation module, wherein the student network comprises a detail feature extraction path and a structure feature extraction path.

3. The computer vision based instrument remote inspection method of claim 1, wherein, The step one is specifically: install a camera terminal at a predetermined position in an industrial field, the field of view of the camera terminal covers the instrument and meter, and the camera terminal is used to collect a field image containing the instrument and meter as an input image; in the remote processing device, the input image is sent to the remote processing device through a wired network; the input image is input into a target detection network with a ring convolution kernel in the remote processing device; the target detection network with the ring convolution kernel comprises a backbone feature extraction network, a polar coordinate transformation branch, a center positioning head, a radius prediction head and a sector division head; the input image is input into the backbone feature extraction network to perform convolution operation, downsampling operation and feature mapping to generate a first feature map; in the polar coordinate transformation branch, the first feature map is converted into a polar coordinate feature map with a plurality of candidate center positions as origins according to a preset polar coordinate sampling rule; The polar coordinate feature map is input into a center positioning head, a radius prediction head and a sector division head. The polar coordinate feature map is input into the center positioning head, the radius prediction head and the sector division head. The polar coordinate feature map is input into the center positioning head, the radius prediction head and the sector division head. The polar coordinate feature map is input into the center positioning head, the radius prediction head and the sector division head. The polar coordinate feature map is input into the center positioning head, the radius prediction head and the sector division head.

4. The computer vision based instrument remote inspection method of claim 1, wherein, The step two is specifically: Based on the center coordinates, the radius parameters and the sector indexes, instrument and meter local images corresponding to the sectors are cropped from the input image, and the instrument and meter local images are grouped into an image data set; Each instrument and meter local image in the image data set is input into a teacher network and a student network; In the student network, the instrument and meter local image is input into a direction feature enhancement module. The direction feature enhancement module includes a multi-directional convolution kernel, which performs multi-directional convolution operation on the instrument and meter local image to generate a corresponding direction response feature map. The multi-directional convolution kernel is set according to a preset direction set, which includes 0-degree direction, 45-degree direction, 90-degree direction, 135-degree direction and 180-degree direction. The direction response feature map is spliced according to the channel dimension, and a one-dimensional convolution layer is used to transform the channel of the spliced feature map to obtain an input feature map. The detail feature extraction path includes a detail feature extraction network composed of three convolution layers and two down-sampling layers. The detail feature extraction network performs layer-by-layer convolution and down-sampling on the input feature map to obtain high-frequency detail features. The structure feature extraction path includes a structure feature extraction network composed of one convolution layer and one down-sampling layer. The structure feature extraction network performs convolution and down-sampling on the instrument and meter local image to obtain low-frequency structure features. In the teacher network, the instrument and meter local image is input into a pre-trained convolutional neural network backbone to perform layer-by-layer feature extraction, and corresponding multi-scale teacher features are output at multiple scales.

5. The computer vision based instrument remote inspection method of claim 1, wherein, The cross-domain joint feature construction mechanism is specifically: In the student network, the high-frequency detail features and the low-frequency structure features are size-aligned in the spatial dimension to obtain first domain input features and second domain input features. The first domain input features are used to construct a spatial domain node set, and the second domain input features are used to construct a frequency domain node set. And according to the corresponding relationship between the spatial domain node set and the frequency domain node set in the spatial position, the cross-domain connection edge is established, and a joint graph structure including the spatial domain node, the frequency domain node and the cross-domain edge is formed; The joint graph structure is executed to perform a graph convolution update operation, and a cross-domain updated node feature is obtained; A feature reconstruction mechanism based on spatial position mapping is adopted to reconstruct the updated node feature at the corresponding spatial position to generate a cross-domain graph guidance graph; The feature reconstruction mechanism based on spatial position mapping writes the cross-domain updated node feature into a two-dimensional feature graph matrix of a preset size according to the spatial corresponding coordinates set for the spatial domain node and the frequency domain node when the joint graph structure is established; A local interpolation operation is performed on the positions not covered by the node through a reconstruction network including a convolution layer and an interpolation layer to generate a complete cross-domain graph guidance graph; The cross-domain graph guidance graph and the first domain input feature are spliced in the channel dimension as a query vector, and the second domain input feature is taken as a key vector and a value vector respectively, which are input into a cross-domain Transformer structure including a multi-head attention calculation layer and a convolution layer; The multi-head attention calculation layer is used to perform similarity calculation between the query vector and the key vector, and the value vector is linearly mapped according to the similarity to output a cross-domain mapping feature; The cross-domain mapping feature is subjected to feature transformation through the convolution layer to generate a cross-domain Transformer guidance graph. The cross-domain graph guidance graph and the cross-domain Transformer guidance graph are spliced in the channel dimension and input into a feature fusion network including three convolution layers, one down-sampling layer and one up-sampling layer, and convolution, down-sampling and up-sampling operations are sequentially performed to generate fusion feature graphs with different spatial resolutions, which are taken as student multi-scale features.

6. The computer vision based instrument remote inspection method of claim 1, wherein, The step four is specifically: In the feature difference calculation module, a multi-scale feature mapping subnetwork is set up one-to-one according to the scale for the multi-scale teacher feature and the student multi-scale feature; The student multi-scale features of each scale are input into the corresponding multi-scale feature mapping subnetwork to perform spatial resolution alignment and channel mapping to obtain student mapping features; The student mapping features are input into an online projection subnetwork and an online prediction subnetwork respectively to obtain student prediction features of the corresponding scale; The multi-scale teacher features of the corresponding scale are input into a target projection subnetwork to obtain teacher target features of the corresponding scale; The online projection subnetwork, the online prediction subnetwork and the target projection subnetwork are derived from the BYOL double network structure; Based on the difference between the student prediction features and the teacher target features of each scale, a corresponding multi-scale feature difference graph is generated, and the multi-scale feature difference graphs are fused according to a preset fusion rule to obtain a feature difference graph for representing the difference relationship between the teacher network and the student network.

7. The computer vision based instrument remote inspection method of claim 1, wherein, The step five is specifically: In the spatial domain of the feature energy anomaly enhancement module, the feature difference graph is divided into M local regions in the spatial dimension according to a preset window size; In each local region, a local spatial energy value corresponding to the feature difference graph is calculated based on the amplitude and distribution of the feature value in the current local region. The local space energy graph is normalized, and a space energy weight graph is generated according to a preset space energy mapping relationship; In the frequency domain of the feature energy anomaly enhancement module, a frequency domain transformation is performed on the feature difference graph in the channel dimension to convert the feature difference graph into a corresponding frequency domain feature graph; In the frequency domain feature graph, the frequency domain features are divided into frequency bands according to a preset frequency band division rule, and the corresponding frequency domain energy values are calculated based on the statistical results of the frequency domain feature amplitudes in each frequency band to generate a frequency domain energy distribution graph; The frequency domain energy distribution graph is normalized, and a frequency domain energy weight graph is generated according to a preset frequency domain energy mapping relationship; The space energy weight graph and the frequency domain energy weight graph are fused in the spatial position according to a preset fusion rule to obtain a fused energy weight graph; The fused energy weight graph and the feature difference graph are weighted in the corresponding spatial position and the channel dimension to perform energy enhancement processing on the feature responses in the feature difference graph to generate an anomaly feature graph.

8. The computer vision based instrument remote inspection method of claim 1, wherein, The step six is specifically: In the anomaly score generation module, the anomaly feature graph and each instrument local image are aligned in the sector index and the spatial position to obtain an anomaly response matrix of each sector; The anomaly response values in the anomaly response matrix are normalized, and a sector-level residual graph is generated according to a preset residual coding rule; According to the sector-level residual graph, the anomaly response values of each sector in each instrument are statistically operated within a preset spatial range to obtain a sector-level anomaly score parameter, which is aggregated in the instrument dimension to obtain an instrument-level anomaly score parameter; The instrument-level anomaly score parameter is compared with a preset threshold value, and when the instrument-level anomaly score parameter is greater than or equal to the preset threshold value, the corresponding instrument is marked as an abnormal state, and when the instrument-level anomaly score parameter is less than the preset threshold value, the corresponding instrument is marked as a normal state; Remote inspection result data containing instrument identification information and state information are generated according to the state marks of each instrument and output.

Images