Intelligent inspection system and method for power distribution network

Abstract

The present application discloses an intelligent inspection system and method for a power distribution network. The intelligent inspection system for the power distribution network comprises: acquiring an inspection analysis model; converting a model parameter into a specific tensor form, to determine whether a network structure of the inspection analysis model matches a network structure supported by a target porting platform; if the two network structures match, on the basis of the network structure supported by the target porting platform, converting the tensor form of the inspection analysis model into a target tensor, and rewriting an inconsistent portion in an operator structure of the inspection analysis model, to complete recompiling of the inspection analysis model; if the two network structures do not match, performing tensor regeneration for a non-matching model parameter type of the inspection analysis model, and performing regeneration for a non-matching operator structure of the inspection analysis model, to complete recompiling of the inspection analysis model; and pushing the recompiled inspection analysis model to the target porting platform, causing the target porting platform to perform inspection analysis on the basis of an inspection task.

Description

Distribution network intelligent inspection system and method

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on September 21, 2023, with application number 202311224968.8, the entire contents of which are incorporated by reference into this application. Technical Field

[0002] The present application relates to the field of intelligent inspection technology for distribution networks, and in particular to an intelligent inspection system and method for distribution networks. Background Art

[0003] Distribution networks are mostly concentrated in townships (towns, districts), and villages (communities). Compared to transmission and transformation scenarios, distribution network environments are more complex and diverse, and inspection information contains more interfering information. To adapt to this complex environment, the same distribution equipment has a variety of shapes and locations, resulting in significant differences in the same defects. Furthermore, due to safety protection requirements, distribution networks are equipped with multiple types of protection devices in different locations. These devices have high similarities between each other and between protection devices and other equipment. However, in transmission and transformation samples, the defect characteristics of different devices vary significantly. This indicates that the characteristics of distribution networks differ significantly from those of transmission and transformation, and the samples are of different orders of magnitude. Transmission and transformation models and related strategies are not suitable for distribution networks and cannot be directly transferred to distribution networks for application.

[0004] The intelligent inspection system for distribution networks applies artificial intelligence recognition models, remote sensing data measurement and other technologies to relevant inspection equipment, effectively improving the inspection efficiency of distribution networks. However, the following problems still exist:

[0005] (1) The data quality required for model construction is low. Due to the uneven quality of distribution network inspection data, there is a large amount of redundant, erroneous or ambiguous data information. During the inspection analysis process, there are problems such as repeated processing of redundant data, failure to identify erroneous information, and low recognition rate of fuzzy data.

[0006] (2) The model construction cycle is long and the update iteration is slow. The construction process of models such as distribution network operation status judgment and abnormal defect location requires the design of large-scale network parameters and continuous iterative optimization of large quantities of data. During model optimization, relevant hyperparameters need to be manually adjusted to achieve model iterative optimization, and autonomous iteration cannot be performed during the use of the model. Therefore, the model construction method of large networks, big data and non-autonomous optimization results in a long construction time and slow update iteration of the distribution network inspection analysis model.

[0007] (3) The model has poor portability across different platforms. During the application process, the model algorithm is difficult to transplant and adapt. The performance of different platforms and equipment varies. The recognition model needs to be individually adapted and developed for different platforms and architectures. The adaptation cycle is long, and the model cannot be quickly upgraded or transplanted.

[0008] (4) Insufficient computing performance of edge devices. The computing resources of edge devices are relatively low, while high-precision models require high-performance computing resources. There is a contradiction between the two. In addition, for model lightweighting methods, models generated under different training architectures lack a unified conversion tool during conversion, making it impossible to uniformly adapt to related edge devices. Moreover, after the model conversion, the model structure is trimmed, resulting in a decrease in model accuracy.

[0009] (5) Poor coordination between multiple equipment in power distribution inspection. Power distribution inspection equipment includes both edge-side equipment and backend analysis platforms. However, due to factors such as insufficient intelligence on the edge side and bandwidth, it is mainly used on the service side with single-node decision-making. Cloud servers and edge-side resources cannot be evaluated in real time, and the interaction time between the service side and the edge side is long, resulting in insufficient coordination.

[0010] Summary of the Invention

[0011] This application proposes a distribution network intelligent inspection system and method, designs a distribution network inspection data intelligent processing method to improve the data quality required for model construction; designs a distribution network intelligent analysis model rapid construction method and parameter fitting self-optimization method with a shared weight mechanism to realize the rapid construction and autonomous optimization of the distribution network intelligent identification model; designs a distribution network to realize the rapid compilation, upgrade and push method of the distribution network inspection analysis model to solve the problem of long adaptation cycle on different platforms; proposes a real-time lossless compression method for edge models to realize real-time quantization in the model reasoning process to solve the contradiction between the high-precision model requiring high-performance computing resources and the low performance of edge-side device computing resources; designs a "cloud-side-edge-side" distribution network intelligent inspection and analysis platform with comprehensive inspection data acquisition, equipment remote control, operation status tracking, data analysis, task hierarchical decision-making and other functions, proposes task hierarchical decision-making and resource adaptive scheduling methods to solve the problem of inability to evaluate cloud servers and edge-side resources in real time and insufficient coordination, and realizes "cloud-side-edge-side" collaborative work division. It solves the key difficulties in distribution network inspection, such as data processing, intelligent recognition model construction, model application, and system interaction, and realizes the three-dimensional inspection of the distribution network with "improved data quality, rapid model construction, simple equipment application, and flexible interaction and collaboration".

[0012] In a first aspect, the present application provides a distribution network intelligent inspection system, comprising:

[0013] The end-side analysis subsystem is configured to obtain multi-channel inspection data according to the inspection task, and use the multi-channel inspection data and the inspection analysis model corresponding to the inspection task to perform inspection analysis on the inspection task, obtain the inspection analysis result, and send the inspection tasks, inspection analysis results and multi-channel inspection data that have not been processed by the end-side analysis subsystem to the edge-side analysis subsystem; wherein, the inspection analysis model is pushed to the end-side analysis subsystem after being processed by the cross-architecture model recompilation method, and when performing the inspection analysis, the weight parameters and network structure are converted into corresponding quantization parameters to perform lossless compression processing on the inspection analysis model; the edge-side analysis subsystem is configured to receive the inspection tasks sent by the end-side analysis subsystem, and Based on multi-channel inspection data, the inspection tasks sent by the end-side analysis subsystem are inspected and analyzed to obtain inspection analysis results, which are sent to the end-side analysis subsystem, and the inspection tasks that have not been completed by the edge-side analysis subsystem, the inspection analysis results of the end-side analysis subsystem, and the inspection analysis results of the edge-side analysis subsystem are sent to the cloud-side decision-making subsystem; the end-side analysis subsystem is also configured to control the execution of actions by the actuator according to the inspection analysis results sent by the edge-side analysis subsystem; the cloud-side decision-making subsystem is configured to summarize the inspection analysis results of the end-side analysis subsystem and the inspection analysis results of the edge-side analysis subsystem, and receive the inspection tasks sent by the edge-side analysis subsystem to make the final inspection task decision.

[0014] As one implementation mode, the end-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is further configured to perform cross-architecture model recompilation; the end-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is configured to perform cross-architecture model recompilation in the following manner: obtaining an inspection analysis model trained with distribution network inspection data samples; converting the model parameters of the inspection analysis model into a specific tensor form; determining whether the network structure of the inspection analysis model matches the network structure supported by the target transplantation platform; if the network structure of the inspection analysis model matches the network structure supported by the target transplantation platform, then, according to the network structure supported by the target transplantation platform, recompilation of the inspection analysis model The tensor form of the inspection analysis model is converted into a target tensor, and the target tensor is used as input to compare the operator structure of the inspection analysis model and the network structure supported by the target transplantation platform, and the inconsistent parts in the operator structure of the inspection analysis model are rewritten, thereby completing the recompilation of the inspection analysis model; if the network structure of the inspection analysis model does not match the network structure supported by the target transplantation platform, tensors are regenerated for the mismatched model parameter types of the inspection analysis model, and the mismatched operator structure of the inspection analysis model is regenerated, thereby completing the recompilation of the inspection analysis model; the recompiled inspection analysis model is pushed to the target transplantation platform, so that the target transplantation platform performs inspection analysis according to the inspection task.

[0015] As one embodiment, the end-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is configured to convert the tensor form of the inspection analysis model into a target tensor according to the network structure supported by the target transplantation platform in the following manner: the model parameters of the inspection analysis model are uniformly converted into a specific tensor form according to the parameter data and position relationship; if the tensor form of the inspection analysis model is inconsistent with the tensor structure expression of the network structure supported by the target transplantation platform, the tensor form of the inspection analysis model is disassembled and combined into a matrix form, or a form consisting of parameter data and direction numbers.

[0016] As one embodiment, the end-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is configured to regenerate tensors for mismatched model parameter types of the inspection analysis model in the following manner: for mismatch between the tensor dimension of the inspection analysis model and the tensor dimension of the target transplantation platform, dimensional decomposition is performed on the high-dimensional tensor of the inspection analysis model; the low-dimensional tensor of the inspection analysis model is dimensionally upgraded and padded with zeros after dimensional upgrade; for mismatch of tensor types, forced type conversion is used to convert the tensor type of the inspection analysis model into a tensor type supported by the target transplantation platform.

[0017] As one embodiment, the end-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is configured to rewrite the inconsistent parts in the operator structure of the inspection analysis model in the following manner: taking the target tensor as the input of the operator structure, and comparing the operator structures of the inspection analysis model and the network structure supported by the target transplantation platform; extracting the basic structure of the operator operation graph of the network structure supported by the target transplantation platform based on the comparison results, and converting the inconsistent branch operation structures in the operator structure of the inspection analysis model into the basic structure based on the basic structure.

[0018] As one embodiment, the end-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is configured to regenerate the mismatched operator structure of the inspection analysis model in the following manner: the mismatched operator structure of the inspection analysis model is minimized and reorganized, and the mismatched operator structure of the minimized decomposition and reorganization inspection analysis model is recompiled according to the operator structure of the network structure supported by the target transplantation platform.

[0019] As one embodiment, the end-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision-making subsystem is further configured to remove redundant information from the inspection images in the distribution network inspection data sample: calculate the MD5 value of the inspection image, filter other inspection images with the same MD5 value, and delete the other inspection images; for two inspection images of the same device and the same inspection location but at different inspection times, determine whether there is redundancy between the two inspection images by extracting image layer features and device area features; for two inspection images with the same content but blurred pixels or compressed, determine whether there is redundancy between the two inspection images by comparing image color features and the structure of foreground and background colors.

[0020] As one embodiment, the end-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision-making subsystem is further configured to use distribution network inspection data samples to perform network structure training on the constructed inspection analysis model; the end-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision-making subsystem is configured to use distribution network inspection data samples to perform network structure training on the constructed inspection analysis model in the following manner: initialize the weight state and the branch structure of the weight connection in the inspection analysis model, and selectively delete or share the corresponding weights and branch structures according to the number of times the weights are used during the training process.

[0021] As one embodiment, the end-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is configured to selectively delete or share corresponding weights and branch structures in the following manner, including: initializing the weight state to zero, and updating the corresponding weight state and branch structure according to the number of times the weight is used during the training process, that is, when the weight is used for the nth time during the training process, the weight state is increased by n; after the training is completed, in response to the existence of a weight with a zero number of uses, the weight with a zero number of uses and the branch structure corresponding to the weight with a zero number of uses are deleted, and in response to the absence of a weight with a zero number of uses, the weight is saved and shared.

[0022] As one embodiment, the terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is also configured such that when the branch structure of the weight changes during the training process, if the changed branch structure is not in the initial branch structure, the branch structure is added to the initial branch structure; if the branch structure of the weight appears repeatedly during the training process, the branch structure is not saved repeatedly, and the saved parameters are directly used for connection.

[0023] As one implementation method, the terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision-making subsystem is also configured to first add the weight, weight status and branch structure corresponding to the weight if the weight is a non-default weight, and then update and share the weight.

[0024] As one embodiment, the end-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision-making subsystem is further configured to use distribution network inspection data samples to perform model parameter training on the inspection analysis model after network structure training; the end-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision-making subsystem is configured to use the distribution network inspection data samples to perform model parameter training on the inspection analysis model after network structure training in the following manner: define a unit vector calculation directed graph, and form an adaptive network architecture with the inspection analysis model after network structure training in the form of a unit vector calculation directed graph, and use a differentiable continuous gradient descent search method on the network architecture to perform model parameter self-optimization, and after the optimized model parameters are secondary connected with the original model parameters, use distribution network inspection data samples to perform model parameter training.

[0025] As one implementation mode, the terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision-making subsystem is further configured to determine the highest dimension of the distribution network inspection data sample, use the highest dimension as a priori dimensional feature, and upgrade the distribution network inspection data sample based on the highest dimension. The model parameters are trained using the upgraded distribution network inspection data sample, and the extraction of continuous distribution features in the data space is achieved through dimensionality upgrade.

[0026] As one embodiment, the end-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is configured to perform self-optimization of model parameters on the network architecture by using a differentiable continuous gradient descent search method in the following manner: arranging the distribution network inspection data samples from high dimension to low dimension, determining an approximate fitting dimensional curve, and fitting a curve or surface to the distribution network inspection data samples based on the dimensional curve, thereby approximating the original data search space of the inspection analysis model from a discrete space to a continuous space, realizing the differentiability of the data search space, and giving a gradient optimization direction for a continuous gradient descent search.

[0027] As one embodiment, the edge-side analysis subsystem is configured to perform inspection analysis on the inspection tasks sent by the end-side analysis subsystem based on the multi-channel inspection data in the following manner: decoding and merging the multi-channel inspection data into a single data stream, and after completing the inspection analysis of the single data stream, encoding and splitting the inspection analysis results of the single data stream into multiple data stream outputs.

[0028] As one embodiment, the lossless compression processing process includes: splitting the weight parameters into minimization parameters, splitting the network structure into the simplest structural units; establishing a quantization dictionary, minimizing the parameters and structural units to achieve weight parameter quantization and network structure simplification through the quantization dictionary.

[0029] As an implementation manner, the weight parameter is quantized by quantizing the floating-point numbers in the minimization parameter into an integer parameter or a binary parameter through a quantization dictionary.

[0030] As an implementation method, the network structure is simplified by reorganizing the network and pruning the structure through a quantized dictionary, and further simplifying the branch structure.

[0031] As one implementation method, the quantization dictionary includes: a dictionary of the number of digits retained for floating-point numbers, which is used to set the number of decimal places retained for floating-point weight values; a dictionary of integer conversions for floating-point numbers, which is used to convert floating-point weight values ​​into integer values; a dictionary of binary parameters for floating-point weight parameters, which is used to convert floating-point weight values ​​into 0 and 1; and a dictionary of network structure simplification, which is used to simplify network branches whose weights correspond to 0.

[0032] As one implementation method, the inspection tasks that are not completed by the end-side analysis subsystem include: inspection tasks that are not supported by the end-side analysis subsystem; inspection tasks that require secondary analysis by the end-side analysis subsystem; inspection tasks that require overall scheduling by the end-side analysis subsystem; and non-emergency inspection tasks that are not supported by the end-side analysis subsystem.

[0033] As one implementation method, the inspection tasks that are not completed by the edge analysis subsystem include: defect types that cannot be determined by both the end-side analysis subsystem and the edge analysis subsystem, task scheduling that cannot be decided, and inspection tasks temporarily sent to the end-side analysis subsystem and the edge analysis subsystem.

[0034] As one implementation method, the cloud-side decision-making subsystem and the edge-side analysis subsystem are also configured to: when performing inspection and analysis, multi-node distributed deployment is performed, the same nodes are deployed simultaneously in different areas, information is shared between different nodes, and all nodes are uniformly managed and scheduled by the master node. When the upload bandwidth, computing power or computing resources of a node reach a critical value, the master node will re-schedule part of the inspection tasks and assign them to other nodes for calculation. At the same time, the master node selects the corresponding node nearby for data upload management based on the location of the end-side device.

[0035] As one embodiment, the distribution network inspection data sample includes the equipment's visible light historical images, equipment's historical sounds, and equipment's historical temperatures. The end-side analysis subsystem, the edge-side analysis subsystem, or the cloud-side decision-making subsystem is further configured to: construct an inspection analysis model for training equipment's visible light images, an inspection analysis model for equipment sounds, and an inspection analysis model for equipment temperatures using the distribution network inspection data sample; perform abnormal diagnosis of the equipment's operating status based on the constructed inspection analysis model; if at least one abnormality exists, determine the equipment defect type, and collect the equipment's visible light image again; perform a secondary analysis on the recollected equipment's visible light image to locate the defect. In a second aspect, the present application provides a distribution network intelligent inspection method, applying the distribution network intelligent inspection system described in the first aspect, including:

[0036] The end-side analysis subsystem obtains multi-channel inspection data, and uses the multi-channel inspection data and the artificial intelligence model corresponding to the inspection task to perform inspection analysis on the inspection task to obtain the inspection analysis result, and sends the inspection tasks, inspection analysis results and multi-channel inspection data that have not been processed by the end-side analysis subsystem to the edge-side analysis subsystem; wherein, the inspection analysis model is pushed to the end-side analysis subsystem after being processed by the cross-architecture model recompilation method, and when performing the inspection analysis, the weight parameters and network structure are converted into corresponding quantization parameters to perform lossless compression processing on the inspection analysis model; the edge-side analysis subsystem receives the inspection task sent by the end-side analysis subsystem, and performs inspection analysis based on the multi-channel inspection data. The inspection tasks sent by the end-side analysis subsystem are subjected to inspection analysis to obtain inspection analysis results, and the inspection analysis results of the edge-side analysis subsystem are sent to the end-side analysis subsystem, and the inspection tasks that have not been completed by the edge-side analysis subsystem, the inspection analysis results of the end-side analysis subsystem, and the inspection analysis results of the edge-side analysis subsystem are sent to the cloud-side decision-making subsystem; the end-side analysis subsystem controls the execution mechanism to perform actions according to the inspection analysis results sent by the edge-side analysis subsystem; the cloud-side decision-making subsystem summarizes the inspection analysis results of the end-side analysis subsystem and the inspection analysis results of the edge-side analysis subsystem, and receives the inspection tasks sent by the edge-side analysis subsystem to make the final inspection task decision. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] FIG1 is a flow chart of a cross-architecture model recompilation method for intelligent inspection of a distribution network provided in Example 1 of the present application;

[0038] FIG2 is a schematic diagram of a weight sharing mechanism provided in Example 1 of the present application;

[0039] FIG3 is a weight connection diagram provided in Example 1 of the present application;

[0040] FIG4 is a schematic diagram of a discrete space provided in Example 1 of the present application;

[0041] FIG5 is a schematic diagram of a continuous space after fitting provided in Example 1 of the present application;

[0042] FIG6 is a schematic diagram of a unit vector calculation directed graph provided in Example 1 of the present application;

[0043] FIG7 is a schematic diagram of a convolutional network architecture provided in Example 1 of the present application;

[0044] FIG8 is a schematic diagram of a cyclic network architecture provided in Example 1 of the present application;

[0045] FIG9 is a schematic diagram of encoding and decoding of multi-channel inspection data provided in Example 2 of the present application;

[0046] Figure 10 is a schematic diagram of a distribution network intelligent inspection system provided in Example 2 of the present application. DETAILED DESCRIPTION

[0047] The present application will be further described below with reference to the accompanying drawings and embodiments.

[0048] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0049] The terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should also be understood that the terms "include" and "have" and any deformation thereof are intended to cover non-exclusive inclusions. For example, the process, method, system, product or equipment comprising a series of steps or units need not be limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or equipment. In the case of no conflict, the embodiments in the application and the features in the embodiments can be combined with each other.

[0050] Example 1

[0051] This embodiment provides a cross-architecture model recompilation method for intelligent inspection of distribution network, as shown in FIG1 As shown, it includes: obtaining an inspection analysis model trained with distribution network inspection data samples; converting the model parameters of the inspection analysis model into a specific tensor form; judging whether the network structure of the inspection analysis model matches the network structure supported by the target transplantation platform; if the network structure of the inspection analysis model matches the network structure supported by the target transplantation platform, then converting the tensor form of the inspection analysis model into a target tensor according to the network structure supported by the target transplantation platform, and using the target tensor as input, comparing the operator structures of the inspection analysis model and the network structure supported by the target transplantation platform, rewriting the inconsistent parts of the operator structure of the inspection analysis model, thereby completing the recompilation of the inspection analysis model; if the network structure of the inspection analysis model does not match the network structure supported by the target transplantation platform, then regenerating tensors for the mismatched model parameter types of the inspection analysis model, and regenerating the mismatched operator structures of the inspection analysis model, thereby completing the recompilation of the inspection analysis model; pushing the recompiled inspection analysis model to the target transplantation platform so that the target transplantation platform performs inspection analysis according to the inspection task.

[0052] In this embodiment, an inspection analysis model is constructed and trained using distribution network inspection data samples; the obtained distribution network inspection data includes equipment inspection images, equipment sounds and equipment temperatures. For equipment inspection images, this embodiment proposes an intelligent processing method for distribution network inspection data, which integrates the deep features and location information of the inspection data and combines a multi-structure comparison method to remove redundant information from the distribution network inspection data to improve data quality.

[0053] The implementation method is: analyze and compare the layer features, color distribution features, location information features and Message-Digest Algorithm 5 (MD5) value features of the inspection image, and obtain the foreground color, background color and pixel distribution structure of the image. Through comprehensive comparison, redundant information is removed to improve data quality.

[0054] The removal of redundant information mainly includes completely repeated inspection images and incompletely repeated inspection images.

[0055] (1) For completely repeated inspection images, calculate the MD5 value of the inspection image. If there are inspection images with the same MD5 value, the two inspection images are considered to be exactly the same and are redundant images, and the redundant images are deleted. The MD5 value is calculated as follows:

[0056] ① Information padding: First, the plaintext information of the inspection image is padded so that its bit length modulo 512 equals 448. Therefore, the bit length of the information is expanded to N*512+448. Then, a 64-bit binary number is appended to this result to represent the information length before padding. The final information byte length is N*512+448+64=(N+1)*512.

[0057] ②Structure initialization: Define a structure that contains the plaintext block (512 bits) that needs to be processed each time and the calculated hash value (128 bits).

[0058] ③ Group files: Group the filled files into groups, with each group consisting of 512 bits, for a total of N groups.

[0059] ④ Use the algorithm to process each set of data: Assign initial values ​​to the four 32-bit chaining variables used in the calculation, namely: A = 0x01234567, B = 0x89abcdef, C = 0xfedcba98, D = 0x76543210; then perform four rounds of the algorithm, the number of rounds being the number of 512-bit information groups in the information.

[0060] The four linked variables are copied to four other variables: A to a, B to b, C to c, and D to d. The main loop has four rounds, each very similar. The first round performs 16 operations. Each operation applies a nonlinear function to three of a, b, c, and d. The result is then added to a fourth variable, a subgroup, and a constant. The result is then shifted right by an indeterminate amount and added to one of a, b, c, or d. Finally, the result replaces one of a, b, c, or d.

[0061] ⑤ Output result: When all information processing is completed, the results of the group processing are processed and the calculation results are output; after all this is completed, A, B, C, and D are added with a, b, c, and d respectively, and then the algorithm is continued with the next group data. The final output is the cascade of A, B, C, and D; when all group processing is completed, the results are cascaded, and the result of MD5 processing is obtained.

[0062] (2) For incompletely repeated inspection images, these are mainly redundant images of the same device, including inspection images of the same device and the same inspection location but at different inspection times, and inspection images with exactly the same content but blurred pixels or compressed original images.

[0063] (2-1) For two inspection images of the same device and the same inspection location but at different inspection times, a multi-structure comparison method of image layer features and device area features is used to determine whether there is information redundancy in the two inspection images; wherein, the acquisition of image layer features includes: obtaining the RGB three-channel layer of the image, taking a clustering method for each layer to achieve cluster segmentation of the adjacent target range, performing masking on the image after RGB three-channel clustering, obtaining the segmentation effect of the final target device in the layer, and statistically analyzing the device area distribution of the RGB three-channel and the pixel distribution of the device area of ​​the masked layer to obtain the overall feature distribution of the image layer. If the similarity of the overall feature distribution of the image layer between different images is greater than 0.8, it is considered that similar information exists between the images;

[0064] The clustering formula is: Among them, C i is the i-th cluster, p is C i The sample points in m i It is C i is the centroid of , and SSE is the clustering error.

[0065] The mask formula is:

[0066] Among them, I(i,j) represents the pixel matrix of the image, and i and j are the corresponding rows and columns respectively.

[0067] The acquisition of device area features includes: statistical clustering and the total number of pixels occupied by different areas after masking, the proportion of pixel areas in the entire image, and the pixel composition of the area boundary.

[0068] Similarity calculation includes: if the similarity of the total number of pixels in the device area between different images is greater than 0.9, and the proportion of the pixel area in the entire image and the overlap of the area boundary pixels are greater than 0.8, then the two images are considered to belong to the same device and there is redundant information.

[0069] (2-1) For two inspection images with exactly the same content but blurred pixels or compressed original images, determine whether there is information redundancy between the two inspection images by comparing the distribution of image color features and the structure of foreground and background colors.

[0070] The acquisition of image color features includes: counting the overall RGB color distribution interval of the image, and normalizing the RGB color interval to obtain the image color feature distribution histogram; the color distribution histogram calculation includes: counting the color value distribution of the three RGB channels respectively, the RGB color digital quantities are all distributed between (0, 255), and merging the color distribution of the three RGB channels to obtain the overall color feature distribution map.

[0071] Acquiring the foreground and background color structures includes: acquiring the color change gradient map of the entire image, extracting the range with obvious gradient changes as the foreground structure of the image, and extracting the range with uniform gradient distribution or changes as the background structure.

[0072] Image gradient calculation includes: Where I(i,j) represents the pixel matrix of the image, i and j correspond to the corresponding rows and columns respectively. When G(x,y) shows a step change from [-1,1], it is considered that the gradient here has changed significantly. The range from this point to the next step change is identified as the foreground structure, and the rest of the range is identified as the background structure.

[0073] The similarity calculation includes: if the similarity of the color feature distribution map is greater than 0.8, and the similarity of the foreground color and the background color pixel range is greater than 0.8, it is considered that there is redundant information in the two images.

[0074] In this embodiment, after the inspection image is processed as described above, the data of the sampling points are scalar normalized during the training sampling process, and then a multi-group weight sharing method mechanism is introduced to realize irrelevant weight replacement and weight projection sharing in the model, thereby realizing the rapid and efficient construction of the inspection analysis model.

[0075] The scalar normalization calculation includes: merging the multidimensional feature data of the distribution network inspection data samples into a multidimensional matrix, using the scalar normalization operator to perform normalized sampling on the multidimensional matrix, converting the vector into a scalar representation, and mapping all real numbers one-to-one to (0, 1). Through mapping, the data is limited to the range of (0, 1), eliminating the adverse effects caused by singular sample data, reducing the order of magnitude difference between different data, ensuring the balance of sample distribution, and while fully retaining the data set information, greatly reducing the representation dimension of the data set information, effectively speeding up the model construction time.

[0076] The scalar normalization operator is expressed as: z∈(-∞,+∞), σ(z)∈(0,1); where z is a real number representing the characteristic data in the multidimensional matrix.

[0077] Therefore, this embodiment uses distribution network inspection data samples to perform network structure training on the constructed inspection analysis model; the network structure training includes: initializing the weight state and weight-connected branch structure in the inspection analysis model, and selectively deleting or sharing the corresponding weights and branch structures according to the number of times the weights are used during the training process.

[0078] As shown in Figure 2-3, the multi-group weight sharing mechanism includes:

[0079] A weight mapping dictionary is designed. During the model training process, the default weights in the model and the branch structures connected to them are selectively deleted, updated, and shared. The weights that meet the training characteristics are retained, and further updated and iterated, new weights are added, weights that did not appear in the training process are deleted, existing weights are repeatedly shared in each layer of the network, and the related network branch structures of weights that are not used in the training process are pruned and deleted. The most concise corresponding dictionary of model structure and weight is established to avoid repeated saving of the same weight parameter and ensure that each weight is only saved once when retaining the parameter.

[0080] The main implementation methods are:

[0081] (1) If the weight is the default weight, the weight operation includes deletion, update and sharing; for example, if the weight number is a, the corresponding connected branch structure is represented as a:{b i},a:{b i} is the weight mapping dictionary, and the initial weight state is a=0. When the weight a is used for the nth time during training, the weight state is increased by n, and now a=n.

[0082] (2) When the corresponding branch structure of a changes, if the changed branch structure is not in b i , then add it to b i In the dictionary.

[0083] (3) During the training process, if the branch structure corresponding to a appears repeatedly, there is no need to save the branch structure repeatedly, and the saved parameters can be directly used for connection.

[0084] (4) After the entire training is completed, if the weight state of a is always 0, it proves that a has not been used, then a and its related branch structures are deleted from the model parameters. If the weight state of a is not always 0, then it can be saved.

[0085] (5) If the weight is a non-default weight and is a randomly generated weight, then add the weight, weight state, and related branch structure, and update and share the weight according to steps (1)-(4).

[0086] The model building process in related technologies is to repeatedly save and read the weights and their corresponding branch structures in different parts of the model. This embodiment adopts a weight comparison and sharing method to design a weight mapping dictionary, which can not only avoid repeated storage of weights, but also realize flexible addition, comparison and call of weights, thereby realizing the simplification of the network structure and the number of weights during the model training process, ensuring the model accuracy while improving the model training efficiency.

[0087] In this embodiment, after completing the construction and training of the model network structure through the above-mentioned network structure training method, a model parameter self-optimization method is proposed, and a heterogeneous architecture self-optimization enhanced search evolutionary algorithm based on the dimensionality upgrade of prior features is designed to realize autonomous updating and rapid optimization of model weights.

[0088] The model parameter self-optimization method includes: defining a unit vector calculation directed graph, forming an adaptive network architecture of the inspection and analysis model after network structure training in the form of a unit vector calculation directed graph, and using a differentiable continuous gradient descent search method to self-optimize the model parameters on the network architecture. After the optimized model parameters are secondary connected with the original model parameters, the distribution network inspection data samples are used to train the model parameters, thereby obtaining the optimal inspection and analysis model.

[0089] The self-optimization enhanced search evolutionary algorithm of the heterogeneous architecture based on the dimensionality increase of prior features improves the ability of model construction and autonomous update. It solves the problems of traditional model construction, such as the inability to differentiate the discrete data space, resulting in a large number of repeated calculations, the random optimization of parameters, which easily leads to random feature discarding or reduction, and the inability to simultaneously optimize and update parameters on heterogeneous architectures such as convolutional network architectures and recurrent architectures, and the inability to match different network architectures.

[0090] Implementation form:

[0091] (1) Perform an overall analysis on the distribution network inspection data samples to determine the highest dimension of the distribution network inspection data samples, use this dimension as a priori dimensional feature, and then upgrade the distribution network inspection data samples based on this dimension to achieve the extraction of continuous distribution features in the data space through dimensionality upgrade.

[0092] (2) The distribution network inspection data samples are arranged from high dimension to low dimension, and the approximate fitting dimensional curve is determined. Based on this, the curve (or surface) is fitted to the data, as shown in Figures 4 and 5. After fitting the data to a curve, a continuous curve or a smooth surface is obtained. Mathematically, this curve or surface is differentiable, thereby approximating the original data search space of the model from a discrete space to a continuous space, realizing the differentiability of the data search space, realizing the gradient calculation in the discrete space of the data, giving the direction of gradient optimization, and converting the calculation mode of the discrete space from a black box trial-and-error search to a continuous gradient descent search, thereby greatly reducing the amount of data calculation.

[0093] (3) Design a unit vector calculation directed graph consisting of an ordered sequence of N nodes, which is a graph structure consisting of data nodes and direction branches, as shown in Figure 6; the ordered sequence is many-to-many or many-to-one node data, including node data and the node pointing direction and the pointed data.

[0094] Then, the inspection and analysis model after network structure training is combined into a corresponding network architecture in the form of a unit vector calculation directed graph. For example, the vector calculation directed graphs are combined with each other to form a convolutional network architecture and a recurrent network architecture, thereby realizing parameter updates on heterogeneous architectures such as the convolutional network architecture and the recurrent network architecture.

[0095] The convolutional network architecture is one in which the vector calculation directed graph is arranged sequentially and has no return direction, as shown in Figure 7; the recurrent network architecture is one in which the vector calculation directed graph has a return direction, as shown in Figure 8; the two out-of-phase architectures are adapted to the network structure used by the model. The composition of the architecture is different, but after decomposition, the different architectures can be decomposed into the same component units, namely the unit vector calculation directed graph.

[0096] Common model self-optimization algorithms can only adapt to specific network architectures. However, this embodiment designs a vector calculation directed graph consisting of an ordered sequence of N nodes, which can flexibly form a variety of structures and perform model parameter self-optimization in different heterogeneous architectures.

[0097] (4) Obtain the optimized model parameters, and then directly migrate the unmodified original model parameters. The original model parameters and the optimized model parameters are reconnected and strengthened to ensure the originality and accuracy of the existing parameters of the model. Therefore, after the model parameters are updated, the new model parameters are used to continue training the data samples after the dimension is upgraded, and the effect analysis is performed. The model parameters are continuously updated according to the model effect, thereby realizing automatic search and optimization update of the model parameters.

[0098] At this point, the construction of the inspection and analysis model is complete. This inspection and analysis model needs to be quickly migrated and applied on different platforms. During the application process, the model algorithm is difficult to transplant and adapt. The performance of different platforms and equipment varies. The model needs to be individually adapted and developed for different platforms and architectures. The adaptation cycle is long, and the model cannot be quickly upgraded and transplanted. To solve the above problems, this embodiment designs a cross-architecture model recompilation method, automatically obtains the relevant model architecture, designs tensor self-generation and operator recompilation methods, and realizes the automatic compilation and upgrade of the model to different frameworks.

[0099] Implementation method: By parsing the trained inspection and analysis model, the network structure and model parameters used in the model training process are obtained, and then compared with the network structure supported by the target transplantation platform. If the network structure and model parameters cannot adapt to the target transplantation platform, the model parameters are parsed and a tensor structure that matches the network structure of the target transplantation platform is automatically generated. At the same time, the operator structure is re-parsed and compiled to finally complete the automatic compilation and upgrade of the model.

[0100] Due to the different architectures of different platforms, the composition of tensor structures also varies. When porting models between different architectures, tensors need to be converted between different architectures. Therefore, this embodiment converts the model parameters of the inspection and analysis model into a specific tensor form; where model parameters include parameter data, data type, and data dimension; tensor form includes tensor dimension, scalar data in the tensor, tensor type, and tensor front and back position; data types include integer data, floating point data, and character data.

[0101] Model parameters are generally expressed in matrix form: Among them, a nn It is parameter data, which can be integer data, floating point data or character data.

[0102] However, different architectures have different representations of the positions before and after the parameters. For example, some architectures use position matrices, while others use position numbers. This embodiment uniformly converts different parameter data and position relationships into a specific tensor form, namely:

[0103] Among them, front-data is the front point of the data, behand-data is the back point of the data, and data-type is the data type.

[0104] In this embodiment, it is determined whether the network structure of the inspection and analysis model matches the network structure supported by the target transplantation platform; if the network structure of the inspection and analysis model matches the network structure supported by the target transplantation platform, it is determined whether the expression forms of the tensor structures of the two are consistent. For example, if the tensor structure of the target transplantation platform is composed of parameter data and direction numbers, and the tensor structure of the inspection and analysis model is directly composed of a matrix, then the operator structure of the inspection and analysis model needs to be adaptively modified at this time; that is, if the tensor structure of the inspection and analysis model is inconsistent with the expression form of the tensor structure of the target transplantation platform, the tensor structure of the inspection and analysis model is automatically converted, that is, according to the tensor form of the target transplantation platform, the tensor form of the inspection and analysis model is converted into the target tensor.

[0105] For example, by disassembling the calculation, becomes: and direction numbering; where direction numbering: other forms of data pointing to front-data and data pointing to behand-data, front-data and behand-data are serialized and arranged in the manner of "row first, column later, from left to right", and correspond one to one with the parameter data; the final form is [a 00 front-data, a 00 behand-data, a01 front-data, a 01 behand-data,......].

[0106] If the network structure of the inspection and analysis model does not match the network structure supported by the target transplantation platform, the tensors of the mismatched model parameter types of the inspection and analysis model are regenerated so that they can adapt to the tensor structure of the network structure supported by the target transplantation platform; for the mismatch between the tensor dimension of the inspection and analysis model and the tensor dimension of the target transplantation platform; if the tensor dimension of the inspection and analysis model is too high, the original high-dimensional tensor of the inspection and analysis model is rewritten into a low-dimensional tensor by dimensional decomposition; if the tensor dimension of the inspection and analysis model is too low, the tensor dimension of the inspection and analysis model is upgraded and filled with 0 after the upgrade; for the mismatch between the tensor type of the inspection and analysis model and the tensor type supported by the target transplantation platform, the tensor data of the inspection and analysis model is converted into the type supported by the target transplantation platform by forced data type conversion.

[0107] This embodiment solves the problem of inconsistent data stream formats during model migration, which requires adjustment of all model parameters. It converts the model parameter tensor into a specific representation and adds a tensor conversion method to achieve real-time conversion of the model tensor during the model reading process, thereby improving the efficiency of model reasoning applications.

[0108] The operation of the model on different platforms mainly relies on operators to combine and connect different tensor structures to form training data streams or inference data streams. In other words, the tensor structure refers to the data composition method, and the operator structure refers to the data operation method. Since the types and composition methods of operators are different in different architectures, in addition to modifying the tensor structure, the operator structure also needs to be converted; therefore, in order to achieve rapid application of the model under different architectures, this embodiment matches and recompiles the model operator structure. For example:

[0109] (1) The basic structure of the operator graph of the network structure supported by the target transplantation platform is extracted, and the tensor form of the inspection and analysis model is converted into the target tensor according to the network structure supported by the target transplantation platform. The target tensor is used as the input of the operator structure, and the operator structure of the inspection and analysis model and the network structure supported by the target transplantation platform are quickly compared.

[0110] (2) Determine the inconsistent branch operation structures based on the comparison results and rewrite them; for example: after obtaining the basic structure of the operator operation graph, convert the operator structure of the inspection analysis model into the basic structure, and realize the rapid compilation and generation of operators under different architectures through adaptive modification.

[0111] Since there may be some structural data with the same form in the operator structure, such as a branch operation structure that is consistent, the operation structure of this part does not need to be rewritten, and only the operation structures with inconsistent output forms need to be rewritten.

[0112] The training architecture of the model is stored in the operator configuration file. By reading the operator configuration file of the network structure supported by the target transplantation platform, the operator operation graph is generated. The minimum structure that constitutes the operator operation graph is the basic structure.

[0113] (3) Recompile and generate the operator structure in the inspection and analysis model that is not supported by the network structure supported by the target transplantation platform, minimize the decomposition and reorganization of the operator structure of the inspection and analysis model, and recompile the operator structure of the inspection and analysis model after the minimized decomposition and reorganization according to the operator structure of the network structure supported by the target transplantation platform.

[0114] Different from the common model adaptation process, when the original model supports operators that are inconsistent with the target architecture operators, the method of completely rewriting the operator structure is adopted. This embodiment recompiles some operators, adds a comprehensive comparison process of tensor graphs and operation graphs, and only modifies the parts of the operator with inconsistent input and output forms. There is no need to modify the entire operator, only partial adaptability modifications are required, which not only improves the compatibility of the operator, but also improves the efficiency of the model adaptation process.

[0115] Example 2

[0116] As shown in FIG10 , this embodiment provides a distribution network intelligent inspection system, including:

[0117] The end-side analysis subsystem 1 is configured to obtain multi-channel inspection data according to the inspection task, and use the multi-channel inspection data and the inspection analysis model corresponding to the inspection task to perform inspection analysis to obtain the inspection analysis result, and send the inspection task, inspection analysis result and multi-channel inspection data that have not been processed by the end-side analysis subsystem 1 to the edge-side analysis subsystem 2; wherein, the inspection analysis model is pushed to the end-side analysis subsystem 1 after being processed by the cross-architecture model recompilation method described in Example 1, and when performing the inspection analysis, the weight parameters and network structure are converted into corresponding quantization parameters to perform lossless compression processing on the inspection analysis model; the edge-side analysis subsystem 2 is configured to receive the inspection task sent by the end-side analysis subsystem 1, and based on the multi-channel inspection The data performs inspection analysis on the inspection tasks sent by the end-side analysis subsystem 1 to obtain the inspection analysis results, sends the inspection analysis results of the edge-side analysis subsystem to the end-side analysis subsystem 1, and sends the inspection tasks that have not been processed by the edge-side analysis subsystem 2, the inspection analysis results of the end-side analysis subsystem 1, and the inspection analysis results of the edge-side analysis subsystem 2 to the cloud-side decision-making subsystem 3; the end-side analysis subsystem 1 is also configured to control the execution mechanism to perform actions according to the inspection analysis results sent by the edge-side analysis subsystem 2; the cloud-side decision-making subsystem 3 is configured to summarize the inspection analysis results of the end-side analysis subsystem 1 and the inspection analysis results of the edge-side analysis subsystem 2, and receive the inspection tasks sent by the edge-side analysis subsystem 2 to make the final inspection task decision.

[0118] The model recompilation process can be deployed in the cloud-based decision-making subsystem, the device-side analysis subsystem, or the edge-side analysis subsystem. Deploying it in the cloud-based decision-making subsystem allows for unified management and distribution of models from the cloud; deploying it in the device-side analysis subsystem or the edge-side analysis subsystem allows for flexible replacement and upgrade of models during application.

[0119] In this embodiment, the device-side analysis subsystem, edge-side analysis subsystem, and cloud-side decision-making subsystem are primarily configured to perform hierarchical judgment and decision-making on inspection tasks, which is distinct from the data collection, management, and model training processes performed by the cloud system in the intelligent perception inspection system. Compared to substation inspections, distribution network inspections are more mobile and subject to significant environmental variations. Therefore, higher requirements are placed on the real-time and reliability of inspection task decisions. Therefore, this embodiment designs hierarchical task decision-making. Through the "cloud-side, edge-side, and device-side" collaborative work division, this allows for rapid task response while reducing data transmission bandwidth and cloud-side service pressure, effectively improving the efficiency of distribution network inspection data analysis and management.

[0120] In this embodiment, the end-side analysis subsystem includes actuators such as drones, robots, handheld devices and vehicle-mounted terminals, which are connected to relevant actuators through external communication devices and control mechanisms to achieve feedback control of multiple types of actuators, thereby realizing real-time tracking, autonomous collection or autonomous operation of distribution network targets.

[0121] For example, when combined with an operating robot arm, operations can be performed on a target that remains stationary. First, the target is detected and located to determine the target position. At the same time, the operating robot arm provides real-time feedback on its position. After determining the operation target, the distance difference between the operating robot arm position and the target position is calculated in real time, and the moving path and moving distance of the operating robot arm are given to control the operating robot arm to operate on the target. For example, when combined with mobile mechanisms such as pan-tilt camera acquisition equipment, real-time tracking and image capture can be performed on targets with constantly changing positions. First, the video stream is acquired and analyzed in real time, and the target position is detected at the same time. The pan-tilt acquisition equipment is controlled not to move abnormally, the target is always within the field of view, and the target can be captured autonomously.

[0122] In this embodiment, the end-side analysis subsystem is configured to perform intelligent analysis on urgent, well-defined inspection tasks, and transmit the inspection analysis results and collected multi-channel inspection data to the edge-side analysis subsystem. Furthermore, in order to simultaneously input and analyze multiple data streams and maximize the performance of edge-side devices while ensuring the real-time nature of multi-channel inspection data analysis, this embodiment designs a high-speed multi-data encoding and decoding method. For example, it acquires multi-channel inspection data such as images and videos, decodes and merges the multi-channel inspection data into a single data stream, completes inspection analysis on the single data stream, and then encodes and splits the inspection analysis results of the single data stream into the original input multi-data stream, achieving real-time detection and output of data streams at the edge of the distribution network, as shown in Figure 9.

[0123] To address the problems of model lightweighting, this embodiment proposes a real-time lossless compression method for edge models, and designs a corresponding conversion method for weight parameters, network structure parameters, and quantization parameters. During the use of the model, instead of adopting the method of quantization first and then deployment, a one-to-one conversion method is directly adopted to dynamically quantize the model during operation, preserve the model structure and parameters to the maximum extent, resolve the contradiction between the computing resources of edge devices and model accuracy, and achieve high-precision operation of the model on edge devices while retaining the accuracy of the model output to the greatest extent. For example, the weight parameters are split into minimization parameters, and a minimization parameter dictionary is established to split the network structure into the simplest structural units, that is, all network structures can be composed of the simplest structural units; a quantization dictionary is established, that is, the minimization parameters after the weight parameters are split and the simplified structural units are used to achieve parameter quantization and structural simplification through the quantization dictionary; wherein, parameter quantization is: if the minimization parameters contain floating-point numbers, the floating-point numbers are directly quantized into integer parameters or binary parameters through the quantization dictionary to reduce the amount of model calculation; structural simplification is: the network structure can achieve network reorganization and structural pruning through the quantization dictionary, and the branch structure in the model is further simplified to reduce the model volume, thereby reducing the resource usage in model inference.

[0124] The quantization dictionary includes: a dictionary of floating-point number retention digits, which sets the number of decimal places retained for floating-point weight values; a dictionary of floating-point number integer conversion, which converts floating-point weight values ​​into integer values; a dictionary of floating-point weight parameter binarization parameters, which converts floating-point weight values ​​directly into 0 and 1; and a dictionary of network structure simplification, which simplifies network branches with weights corresponding to 0, networks with the same structure, and networks with the same parameters.

[0125] After the conversion of the quantized dictionary, the relevant weight parameters of the original model can be retained to the greatest extent possible, and the structural parameters that need to be simplified can also be quantified and simplified, ensuring the accuracy of the model while reducing the resource usage of the model operation.

[0126] In this embodiment, the inspection tasks that are not completed by the end-side analysis subsystem include: inspection tasks that the end-side analysis subsystem does not support analysis, such as models with high precision requirements that the computing power of the end-side device cannot support; inspection tasks that the end-side analysis subsystem needs to perform secondary analysis on, such as the end-side needs to perform secondary confirmation of a defect type; inspection tasks that the end-side analysis subsystem needs to perform overall scheduling on, such as tasks that the end-side needs to perform comprehensive assessment and analysis of multiple defects; non-emergency inspection tasks that are not supported by the end-side analysis subsystem, etc.; the edge-side analysis subsystem provides further inspection analysis for the inspection tasks that are not completed by the end-side analysis subsystem. The edge-side analysis subsystem performs inspection analysis on the above tasks based on the existing model, and then sends the analysis results to the end side.

[0127] In this embodiment, the cloud-side decision-making subsystem is mainly responsible for summarizing and managing the uploaded data, work status, and adjustment of work content on the end side and edge side, and making final decisions on high-level decision-making tasks that the end side and edge side cannot make decisions on; the inspection tasks handled by the cloud-side decision-making subsystem include: defect types that cannot be judged by both the end side and the edge side; task scheduling that cannot be decided on by the end side and the edge side, such as comprehensive analysis of all data information in an area; analysis, collection, and scheduling tasks that need to be temporarily sent to the end side and edge side, etc.

[0128] In this embodiment, through the collaborative work of "cloud side-edge side-end side", inspection tasks can be decomposed according to urgency and analysis difficulty. Urgent and simple-to-analyze inspection tasks can be analyzed directly on the end side; among them, urgent inspection tasks are: inspection tasks that require real-time analysis and decision-making, such as defects in the distribution network that may affect the normal operation of the line; difficulties include: model recognition accuracy, such as defects with an accuracy rate of less than 80%; the end-side equipment cannot meet the computing power required for the analysis and decision-making process, such as an overall analysis of equipment defects in a region; tasks that exceed the decision-making authority of the end-side equipment.

[0129] Non-urgent tasks that require high computing power to analyze can be uploaded by the end-side device to the edge-side intelligent analysis service station for processing and analysis. Important decision-making tasks, batch data, and task management are aggregated from the end-side to the cloud side, which performs unified decision-making and management.

[0130] In this embodiment, an adaptive data scheduling and calculation method for multi-level distributed computing is designed. Through real-time evaluation of upload bandwidth, computing amount and computing resources, the computing is autonomously optimized and scheduled, and the nearest upload and computing nodes are selected to achieve rapid responsive scheduling of the platform.

[0131] Implementation method: The resources of the cloud-side intelligent decision-making platform and the edge-side intelligent analysis service station are deployed in a multi-node distributed manner, that is, the same nodes are deployed simultaneously in different areas, and data, resources and other information are shared between different nodes. At the same time, all nodes are uniformly managed and scheduled by the master node. The master node can obtain the operating status and resource usage information of all nodes. When the upload bandwidth, computing power or computing resources of a node reaches the critical value, the master node can re-schedule part of the tasks and assign them to other nodes for calculation. At the same time, the master node can select the corresponding node nearby for data upload management based on the location of the end-side device, further improving the speed and flexibility of data upload and task operation.

[0132] In this embodiment, when conducting inspection analysis, the determination and positioning of the abnormal type of distribution network equipment is achieved through comprehensive analysis of multi-dimensional hierarchical information feature fusion, including: the multi-dimensional inspection information of the distribution network includes equipment visible light images, equipment sounds and equipment temperatures, thereby generating multi-dimensional feature data of distribution network inspection data, and using the inspection analysis model to perform real-time detection of equipment temperature and equipment sounds to realize the diagnosis of abnormal defects of distribution network equipment, and then combined with the visible light images of the equipment to realize the positioning of the abnormal defect location.

[0133] (1) The multi-dimensional feature data samples include: equipment visible light history images, equipment sound history information and equipment temperature history information, so as to train the constructed inspection and analysis model; wherein, the equipment visible light history images reflecting the appearance shape are obtained, including the equipment normal visible light history images and the equipment abnormal visible light history images, and the inspection and analysis model of the equipment visible light images is constructed; the equipment sound history data are obtained, including the equipment normal sound history data and the equipment abnormal sound history data, and the inspection and analysis model of the equipment sound is constructed; the equipment temperature history data are obtained, including the equipment normal temperature history data and the equipment abnormal temperature history data, and the inspection and analysis model of the equipment temperature is constructed; wherein, in the construction and training process of the above-mentioned inspection and analysis model, the network structure training method and the model parameter self-optimization method are adopted to ensure the model accuracy while improving the model training efficiency.

[0134] (2) Obtaining real-time inspection information of distribution network equipment, including: obtaining equipment sound information, using the trained equipment sound inspection analysis model to determine whether there is any abnormality in the equipment operation sound; obtaining equipment temperature information, using the trained equipment temperature inspection analysis model to determine whether the equipment temperature change trend and change range are within the normal range, and then determining whether the equipment status is normal; obtaining equipment visible light images, obtaining the appearance morphology information of distribution network engineering equipment in the environment, and using the trained equipment visible light image inspection analysis model to determine whether there is any abnormality in the equipment morphology; this embodiment comprehensively analyzes information such as visible light, equipment sound, and equipment temperature, analyzes different operating states of the equipment with different data, and constructs multi-dimensional feature data from dimensions such as appearance, internal temperature, and sound performance; through analysis of the equipment temperature status and operating sound status, it is determined whether the equipment has a fault or defect, and through visible light image analysis, the abnormal equipment morphology and the location of the fault and defect are accurately identified and located.

[0135] In this embodiment, the distribution transformer and distribution cabinet, which are the most commonly used equipment in the distribution network, are taken as an example. The main defect of the distribution transformer is abnormal transformer grounding, and the main defect of the distribution cabinet is poor contact of the distribution cabinet knife switch. Visible light images of normal and abnormal transformer grounding, normal and poor contact of the distribution cabinet knife switch are obtained, historical sound data of the transformer and the distribution cabinet during normal operation and abnormal operation are obtained, and historical temperature data of the transformer and the distribution cabinet during normal operation and abnormal operation are obtained. Then, according to the obtained equipment operation status, the visible light data, operation sound change state data, and temperature change state data of the transformer and the distribution cabinet are scalar-normalized into a unified multi-dimensional feature matrix and sent to the trained inspection analysis model for data detection. When none of the three are abnormal, it indicates that the equipment status is operating normally. If one of the three is abnormal, the visible light image must be collected again and the visible light image must be analyzed again to locate the type and location of the equipment defect.

[0136] Example 3

[0137] The present embodiment provides a distribution network intelligent inspection method, which applies the distribution network intelligent inspection system described in Example 2, including: obtaining multi-channel inspection data by the end-side analysis subsystem, and using the multi-channel inspection data and the artificial intelligence model corresponding to the inspection task to perform inspection analysis on the inspection task to obtain the inspection analysis result, and sending the inspection tasks, inspection analysis results and multi-channel inspection data that have not been processed by the end-side analysis subsystem to the edge-side analysis subsystem; wherein, the inspection analysis model is pushed to the end-side analysis subsystem after being processed by the cross-architecture model recompilation method described in Example 1, and when performing the inspection analysis, the weight parameters and network structure are converted into corresponding quantization parameters to perform lossless compression processing on the inspection analysis model; the edge-side analysis subsystem receives Receive the inspection tasks sent by the end-side analysis subsystem, and perform inspection analysis on the inspection tasks sent by the end-side analysis subsystem based on multi-channel inspection data to obtain the inspection analysis results, send the inspection analysis results of the edge-side analysis subsystem to the end-side analysis subsystem, and send the inspection tasks that have not been completed by the edge-side analysis subsystem, the inspection analysis results of the end-side analysis subsystem, and the inspection analysis results of the edge-side analysis subsystem to the cloud-side decision-making subsystem; the end-side analysis subsystem controls the execution mechanism to perform actions according to the inspection analysis results sent by the edge-side analysis subsystem; the cloud-side decision-making subsystem summarizes the inspection analysis results of the end-side analysis subsystem and the inspection analysis results of the edge-side analysis subsystem, and receives the inspection tasks sent by the edge-side analysis subsystem to make the final inspection task decision.

[0138] This application aims to solve the urgent problems of "low data quality, slow model construction, difficult equipment application and poor collaborative interaction" in the intelligent inspection of distribution networks, and designs an intelligent inspection system and method for distribution networks, realizing a three-dimensional inspection of distribution networks with "improved data quality, rapid model construction, simple equipment application, and flexible interaction and collaboration".

[0139] This application proposes a cross-architecture model recompilation method for intelligent inspection of distribution networks, designs an intelligent inspection system for distribution networks, and removes redundant information and improves the quality of inspection data by fusing deep features and location information and combining multi-structure comparison methods; as well as designs tensor self-generation technology and operator recompilation methods to achieve rapid compilation, upgrading and push of distribution network inspection analysis models, solving the problem of long adaptation cycle on different platforms.

[0140] This application proposes a model construction and network structure training method for sharing multiple sets of weights. Different from the model construction process in which weights and their corresponding branch structures are repeatedly saved and read in different parts of the model, this application introduces a multiple-set weight sharing method mechanism, adopts a weight comparison and sharing method, and designs a weight mapping dictionary, which can avoid repeated storage of weights and realize flexible addition, comparison and call of weights, realize irrelevant weight replacement and weight projection sharing in the model, and thus realize the simplification of network structure and number of weights in the model training process, ensuring model accuracy while improving model training efficiency.

[0141] This application proposes a self-optimization method for model parameter fitting. Different from the common model self-optimization algorithm that uses discrete space data search, this application fits the discrete data space into a continuous data space, designs a differentiable algorithm, and reduces the amount of calculation in the model construction process; different from the common model self-optimization algorithm that can only adapt to specific network architectures, this application designs a vector calculation directed graph composed of an ordered sequence of N nodes, which can be flexibly composed of various structures, and can update parameters in different heterogeneous architecture models, thereby realizing autonomous iterative optimization of the model during use.

[0142] This application proposes a self-optimization method for model parameter fitting, which reconnects and strengthens the optimized model parameters with the unmodified original model parameters to ensure the originality and accuracy of the existing parameters of the model, and avoids the problem of random discarding or reduction of data features caused by random optimization of model parameters; and continuously updates the model parameters according to the model effect, thereby realizing automatic search and optimization update of model parameters.

[0143] This application proposes a task hierarchical decision-making and resource adaptive scheduling method, which is different from the service-side usage form that mainly uses single-node decision-making. It solves the problem of inability to evaluate cloud servers and edge-side resources in real time and insufficient coordination, and realizes three-dimensional inspection with cloud-edge collaboration.

[0144] This application proposes a real-time lossless compression method for edge models, designs corresponding conversion methods for model weight parameters, network structure parameters and quantization parameters, realizes real-time quantization compression during model inference, and solves the contradiction between the high-precision model's requirement for high-performance computing resources and the low performance of edge device computing resources.

[0145] This application designs a "cloud-side-edge-side" distribution network intelligent inspection and analysis platform with comprehensive inspection data acquisition, equipment remote control, operation status tracking, data analysis, task hierarchical decision-making and other functions. Through the "cloud-side-edge-side" collaborative work division, it can not only achieve rapid task response, but also reduce data transmission bandwidth and cloud-side service pressure, and effectively improve the efficiency of distribution network inspection data analysis and management.

Claims

1. A distribution network intelligent inspection system, comprising: The end-side analysis subsystem is configured to obtain multi-channel inspection data according to the inspection task, and use the multi-channel inspection data and the inspection analysis model corresponding to the inspection task to perform inspection analysis on the inspection task to obtain the inspection analysis result, and send the inspection task that has not been processed by the end-side analysis subsystem, the inspection analysis result and the multi-channel inspection data to the edge-side analysis subsystem; The inspection analysis model is pushed to the end-side analysis subsystem after being processed by the cross-architecture model recompilation method, and the inspection analysis model is losslessly compressed by converting the weight parameters and the network structure into corresponding quantization parameters during the inspection analysis; The edge-side analysis subsystem is configured to receive the inspection tasks sent by the end-side analysis subsystem, and perform inspection analysis on the inspection tasks sent by the end-side analysis subsystem based on the multi-channel inspection data to obtain inspection analysis results, send the inspection analysis results of the edge-side analysis subsystem to the end-side analysis subsystem, and send the inspection tasks that have not been processed by the edge-side analysis subsystem, the inspection analysis results of the end-side analysis subsystem, and the inspection analysis results of the edge-side analysis subsystem to the cloud-side decision-making subsystem; The terminal-side analysis subsystem is further configured to control the execution mechanism to perform an action according to the inspection analysis result sent by the edge-side analysis subsystem; The cloud-side decision-making subsystem is configured to summarize the inspection analysis results of the terminal-side analysis subsystem and the inspection analysis results of the edge-side analysis subsystem, and receive the inspection tasks sent by the edge-side analysis subsystem to make a final inspection task decision.

2. The intelligent inspection system for distribution network according to claim 1, wherein: The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is further configured to perform cross-architecture model recompilation; The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is configured to perform cross-architecture model recompilation in the following manner: Obtaining an inspection analysis model trained with distribution network inspection data samples; Converting the model parameters of the inspection analysis model into a specific tensor form; Determining whether the network structure of the inspection analysis model matches the network structure supported by the target transplantation platform; In response to the network structure of the inspection and analysis model matching the network structure supported by the target transplantation platform, the tensor form of the inspection and analysis model is converted into a target tensor according to the network structure supported by the target transplantation platform, and the operator structure of the inspection and analysis model and the network structure supported by the target transplantation platform are compared with the target tensor as input, and the inconsistent parts of the operator structure of the inspection and analysis model are rewritten to complete the recompilation of the inspection and analysis model; In response to the network structure of the inspection and analysis model not matching the network structure supported by the target transplantation platform, regenerating tensors for the unmatched model parameter types of the inspection and analysis model, and regenerating the unmatched operator structure of the inspection and analysis model, thereby completing the recompiling of the inspection and analysis model; The recompiled inspection analysis model is pushed to the target transplantation platform, so that the target transplantation platform performs inspection analysis according to the inspection task.

3. The intelligent inspection system for distribution network according to claim 2, wherein: The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is configured to convert the tensor form of the inspection analysis model into a target tensor according to the network structure supported by the target migration platform in the following manner: The model parameters of the inspection analysis model are uniformly converted into a specific tensor form according to the parameter data and the position relationship; In response to the inconsistency between the tensor form of the inspection and analysis model and the tensor structure representation of the network structure supported by the target transplantation platform, the tensor form of the inspection and analysis model is disassembled and combined into a matrix form, or a form consisting of parameter data and direction numbers.

4. The intelligent inspection system for distribution network according to claim 2, wherein: The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is configured to regenerate tensors for mismatched model parameter types of the inspection analysis model in the following manner: In view of the mismatch between the tensor dimension of the inspection and analysis model and the tensor dimension of the target transplantation platform, the high-dimensional tensor of the inspection and analysis model is dimensionally decomposed; the low-dimensional tensor of the inspection and analysis model is dimensionally upgraded and filled with zeros after the dimension upgrade; in view of the mismatch between the tensor type of the inspection and analysis model and the tensor type of the target transplantation platform, the tensor type of the inspection and analysis model is converted into a tensor type supported by the target transplantation platform by forced type conversion.

5. The intelligent inspection system for distribution network according to claim 2, wherein: The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is configured to rewrite the inconsistent parts in the operator structure of the inspection analysis model in the following manner: Using the target tensor as the input of the operator structure, comparing the operator structure of the inspection analysis model with the network structure supported by the target transplantation platform; The basic structure of the operator operation graph of the network structure supported by the target transplantation platform is extracted according to the comparison result, and the inconsistent branch operation structures in the operator structure of the inspection analysis model are converted into the basic structure according to the basic structure.

6. The intelligent inspection system for distribution network according to claim 2, wherein: The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is configured to regenerate the unmatched operator structure of the inspection analysis model in the following manner: The unmatched operator structure of the inspection and analysis model is decomposed and reorganized in a minimized manner, and the unmatched operator structure of the inspection and analysis model that is decomposed and reorganized in a minimized manner is recompiled according to the operator structure of the network structure supported by the target transplantation platform.

7. The intelligent inspection system for distribution network according to claim 2, wherein: The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision-making subsystem is further configured to remove redundant information from inspection images in the distribution network inspection data samples; The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision-making subsystem is configured to remove redundant information from the inspection image in the distribution network inspection data sample in the following manner: Calculate the information-digest algorithm MD5 value of the inspection image, filter other inspection images with the same MD5 value, and delete the other inspection images; For two inspection images of the same device and the same inspection location but at different inspection times, determine whether there is redundancy between the two inspection images by extracting image layer features and device area features; For two inspection images with the same content but blurred pixels or compressed, it is determined whether there is redundancy between the two inspection images by comparing the image color features and the structure of the foreground color and the background color.

8. The intelligent inspection system for distribution network according to claim 2, wherein: The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision-making subsystem is further configured to use distribution network inspection data samples to perform network structure training on the constructed inspection analysis model; The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision-making subsystem is configured to use the distribution network inspection data samples to perform network structure training on the constructed inspection analysis model in the following manner: Initialize the weight state and the branch structure of the weight connection in the inspection analysis model, and selectively delete or share the corresponding weight and branch structure according to the number of times the weight is used during the training process.

9. The intelligent inspection system for distribution network according to claim 8, wherein: The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is configured to selectively delete or share corresponding weights and branch structures in the following manner: Initialize the weight state to zero, and update the corresponding weight state and branch structure according to the number of times the weight is used during training. When the weight is used for the nth time during training, the weight state increases by n. After the training is completed, in response to the existence of weights with zero usage times, the weights with zero usage times and the branch structure corresponding to the weights with zero usage times are deleted, and in response to the absence of weights with zero usage times, the weights are saved and shared.

10. The intelligent inspection system for distribution network according to claim 9, wherein: The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is further configured to, when a branch structure of the weight changes during the training process, add the branch structure to the initial branch structure in response to the changed branch structure not being in the initial branch structure; In response to the branch structure of the weights repeatedly appearing during the training process, the branch structure is not repeatedly saved, and the saved parameters are directly used for connection.

11. The intelligent inspection system for distribution network according to claim 9, wherein: The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is also configured to, in response to a weight being a non-default weight, first add the weight, the weight state and the branch structure corresponding to the weight, and then update and share the weight.

12. The intelligent inspection system for distribution network according to claim 8, wherein: The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision-making subsystem is further configured to use the distribution network inspection data sample to perform model parameter training on the inspection analysis model after network structure training; The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision-making subsystem is configured to perform model parameter training on the inspection analysis model after network structure training using the distribution network inspection data sample in the following manner: A unit vector calculation directed graph is defined, and the inspection analysis model after network structure training is formed into a corresponding network architecture in the form of a unit vector calculation directed graph. A differentiable continuous gradient descent search method is used on the network architecture to self-optimize model parameters. After the optimized model parameters are secondary connected with the original model parameters, the distribution network inspection data samples are used to train the model parameters.

13. The intelligent inspection system for distribution network according to claim 12, wherein: The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision-making subsystem is also configured to determine the highest dimension of the distribution network inspection data samples, take the highest dimension as a priori dimensional feature, upgrade the distribution network inspection data samples based on the highest dimension, and use the upgraded distribution network inspection data samples to perform model parameter training, so as to realize the extraction of continuous distribution characteristics of data space through dimensional upgrading.

14. The intelligent inspection system for distribution network according to claim 12, wherein: The terminal-side analysis subsystem, the edge-side analysis subsystem or the cloud-side decision subsystem is configured to perform self-optimization of model parameters by using a differentiable continuous gradient descent search on the network architecture in the following manner: The distribution network inspection data samples are arranged from high dimension to low dimension, and an approximate fitting dimensional curve is determined. The distribution network inspection data samples are fitted with a curve or a surface based on the dimensional curve. The original data search space of the inspection analysis model is approximately fitted from a discrete space to a continuous space to achieve the differentiability of the data search space, and a gradient optimization direction is given to perform a continuous gradient descent search.

15. The intelligent inspection system for distribution network according to claim 1, wherein: The edge-side analysis subsystem is configured to perform inspection analysis on the inspection task sent by the terminal-side analysis subsystem based on the multi-channel inspection data in the following manner: Decode and merge the multiple inspection data into a single data stream, and perform inspection analysis on the single data stream Then, the inspection and analysis result of the single data stream is encoded and split into multiple data streams for output.

16. The intelligent inspection system for distribution network according to claim 1, wherein: The lossless compression process includes: splitting weight parameters into minimization parameters and splitting the network structure into the most simplified structural units; establishing a quantization dictionary, and using the quantization dictionary to quantize the weight parameters and simplify the network structure for the minimization parameters and the structural units.

17. The intelligent inspection system for distribution network according to claim 16, wherein: The weight parameter quantization is to quantize the floating point numbers in the minimization parameter into an integer parameter or a binary parameter through the quantization dictionary.

18. The intelligent inspection system for distribution network according to claim 16, wherein: The network structure simplification is to reorganize the network and trim the structure through the quantization dictionary to further simplify the branch structure.

19. The intelligent inspection system for distribution network according to claim 16, wherein: The quantization dictionary includes: a dictionary of floating-point number retention digits, that is, setting the number of decimal places retained by the floating-point weight value; a floating-point number integer conversion dictionary, that is, converting the floating-point weight value into an integer value; a floating-point weight parameter binarization parameter dictionary, that is, converting the floating-point weight value into 0 and 1; a network structure simplification dictionary, that is, simplifying the network branches whose weights correspond to 0.

20. The intelligent inspection system for distribution network according to claim 1, wherein: The inspection tasks that are not processed and completed by the end-side analysis subsystem include: inspection tasks that are not supported by the end-side analysis subsystem; inspection tasks that require secondary analysis by the end-side analysis subsystem; inspection tasks that require overall scheduling by the end-side analysis subsystem; and non-urgent inspection tasks that are not supported by the end-side analysis subsystem.

21. The intelligent inspection system for distribution network according to claim 1, wherein: The inspection tasks that are not completed by the edge analysis subsystem include: defect types that cannot be determined by both the terminal analysis subsystem and the edge analysis subsystem, task scheduling that cannot be decided, and inspection tasks temporarily sent to the terminal analysis subsystem and the edge analysis subsystem.

22. The intelligent inspection system for distribution network according to claim 1, wherein: The cloud-side decision-making subsystem and the edge-side analysis subsystem are further configured as follows: When conducting inspection and analysis, multi-node distributed deployment is performed, and the same nodes are deployed simultaneously in different areas. Information is shared between different nodes, and all nodes are centrally managed and scheduled by the master node. When the upload bandwidth, computing power or computing resources of a node reach a critical value, the master node will re-schedule part of the inspection tasks and allocate them to other nodes for calculation. At the same time, the master node selects the corresponding node nearby for data upload management based on the location of the terminal device.

23. The intelligent inspection system for distribution network according to claim 1, wherein: The distribution network inspection data samples include equipment visible light historical images, equipment historical sounds, and equipment historical temperatures. The terminal-side analysis subsystem, the edge-side analysis subsystem, or the cloud-side decision-making subsystem is further configured as follows: An inspection and analysis model for the visible light image of the training equipment, an inspection and analysis model for the sound of the equipment, and an inspection and analysis model for the temperature of the equipment are constructed using the distribution network inspection data samples. An abnormal diagnosis of the equipment operating status is performed based on the constructed inspection and analysis models. In response to the existence of at least one abnormality, the type of equipment defect is determined, and the visible light image of the equipment is collected again. The re-collected visible light image of the equipment is subjected to a secondary analysis to locate the defect.

24. A distribution network intelligent inspection method, using the distribution network intelligent inspection system according to any one of claims 1 to 23, comprising: The end-side analysis subsystem acquires multi-channel inspection data, and uses the multi-channel inspection data and an artificial intelligence model corresponding to the inspection task to perform inspection analysis on the inspection task to obtain an inspection analysis result, and sends the inspection tasks that have not been processed by the end-side analysis subsystem, the inspection analysis result and the multi-channel inspection data to the edge-side analysis subsystem; The inspection analysis model is pushed to the end-side analysis subsystem after being processed by the cross-architecture model recompilation method, and the inspection analysis model is losslessly compressed by converting the weight parameters and the network structure into corresponding quantization parameters during the inspection analysis; The edge-side analysis subsystem receives the inspection tasks sent by the device-side analysis subsystem, and performs inspection analysis on the inspection tasks sent by the device-side analysis subsystem based on the multi-channel inspection data to obtain inspection analysis results, sends the inspection analysis results of the edge-side analysis subsystem to the device-side analysis subsystem, and sends the inspection tasks that have not been processed by the edge-side analysis subsystem, the inspection analysis results of the device-side analysis subsystem, and the inspection analysis results of the edge-side analysis subsystem to the cloud-side decision-making subsystem; The terminal-side analysis subsystem controls the execution mechanism to perform an action according to the inspection and analysis results sent by the edge-side analysis subsystem; The cloud-side decision subsystem aggregates the inspection analysis results of the terminal-side analysis subsystem and the edge-side analysis subsystem, and receives the inspection tasks sent by the edge-side analysis subsystem to make a final inspection task decision.

Images