Brain inspiration electric power inspection system and method based on incremental learning

Abstract

The invention relates to a brain inspiration electric power inspection system and method based on incremental learning, and the system comprises a plurality of inspection agents carrying edge computing devices. The edge computing device is provided with a site recognition model and a fault detection model and stores a dynamic topological map. Under the condition that a new inspection site is added into the multiple existing inspection sites, the dynamic topological map is updated according to new site data corresponding to the new inspection site; the site recognition model is updated based on an incremental learning algorithm according to the marker image corresponding to the new inspection site; under the condition that a new fault type is added into the multiple existing fault types, the fault detection model is updated based on an incremental learning algorithm according to new equipment fault data corresponding to the new fault type; the electric power inspection system can adapt to the dynamic change of an inspection area and the rapid identification requirement of the fault type of new equipment in a power grid, and has higher intelligence, flexibility and robustness.

Description

Technical Field

[0001] The present application relates to the field of intelligent management of power systems, and in particular to a brain-inspired power inspection system and method based on incremental learning. Background Art

[0002] In the operation monitoring of power grid equipment, regular inspection and maintenance of power equipment is a necessary method and process to ensure the normal operation of the power system. Power inspection refers to a method of regularly inspecting the status of power equipment at each inspection point, discovering faults and handling or reporting them. With the expansion of the scale of the power grid, manual inspection methods are increasingly difficult to cope with complex and dangerous power grid environments. Especially in some dangerous or remote environments, manual inspections are inefficient, and operators need to face safety risks such as high voltage and extreme climate in the power grid. At the same time, the frequency and coverage of manual inspections are limited, making it difficult to discover and handle equipment hazards in a timely manner. With the continuous development of intelligent and automated technologies, automated inspection methods have been widely used in power grids. Automated inspections mainly rely on inspection agents (such as inspection robots, drones, etc.) that perform inspections along fixed paths. Inspection robots can automatically perform inspection tasks in specific areas and have important applications in power, security, emergency rescue and other scenarios.

[0003] At present, in the automated inspection of power grids, inspection robots perform path planning based on the high-precision maps that have been built, and perform inspection site identification and power equipment fault type detection based on the built recognition model and detection model. When inserting a new inspection site, it is necessary to redraw the entire high-precision map, re-plan the global path, and reconstruct the recognition model; and once a new type of power equipment fault appears in the power grid, the existing detection model cannot be effectively detected, and the entire detection model must be rebuilt, which is time-consuming and costly, and cannot adapt to changes in inspection scenarios in a timely manner. Therefore, the current automated inspection methods for power grids lack robustness, flexibility, and intelligence in complex and dynamic environments. Moreover, high-precision maps are not only limited in scalability, but also have high storage costs and large computational workloads, making them difficult to deploy on the edge computing devices of inspection robots, which limits the improvement of the intelligence of inspection robots, resulting in a huge gap between inspection robots and human inspectors. Summary of the invention

[0004] In view of this, the present application proposes a brain-inspired power inspection system and method based on incremental learning, which can quickly and dynamically update the dynamic topology map, site recognition model and fault detection model when the inspection environment changes, adapt to the dynamic changes of the adaptive inspection area and the rapid identification requirements of new equipment fault types in the power grid, thereby improving the intelligence, flexibility and robustness of the power inspection system.

[0005] According to one aspect of the present application, a brain-inspired power inspection system based on incremental learning is provided, the system comprising a plurality of inspection agents equipped with edge computing devices; the inspection agents are used to detect fault conditions of power equipment at inspection sites within an inspection area; the edge computing devices are deployed with a site recognition model and a fault detection model, and store a dynamic topology map; the dynamic topology map is used to reflect the relative position relationship and connectivity relationship between a plurality of existing inspection sites, as well as the markers corresponding to each existing inspection site; the site recognition model is used to identify the plurality of existing inspection sites; the fault detection model is used to determine whether a fault occurs in the power equipment, and in the event of a fault in the power equipment, determine the fault location of the power equipment and determine the fault type of the power equipment from a plurality of existing fault types. fault type; wherein, in the case of adding a new inspection site to the multiple existing inspection sites, the dynamic topology map is updated according to the new site data corresponding to the new inspection site; the new site data includes the relative position information and connectivity information between the new inspection site and the multiple existing inspection sites, and the marker image corresponding to the new inspection site; the site recognition model is updated based on the incremental learning algorithm according to the marker image corresponding to the new inspection site; in the case of adding a new fault type to the multiple existing fault types, the fault detection model is updated based on the incremental learning algorithm according to the new equipment fault data corresponding to the new fault type; the new equipment fault data includes the equipment image corresponding to the fault type of the power equipment when it is the new fault type.

[0006] In a possible implementation, the system further includes a cloud device; each of the inspection agents is connected to the cloud device respectively.

[0007] In one possible implementation, the multiple existing inspection sites include multiple initial inspection sites; the initial state of the dynamic topology map is constructed based on the initial site data corresponding to the multiple initial inspection sites; the initial site data includes the relative position information and connectivity information between the multiple initial inspection sites, and the marker image corresponding to each initial inspection site; the cloud device is used to: pre-train based on the marker image corresponding to each initial inspection site to obtain the initial state of the site recognition model.

[0008] In one possible implementation, the site recognition model includes a site recognition backbone network and a site recognition head network; the cloud device is deployed with a first preset model; the first preset model includes a first backbone network and a first head network; the cloud device is also used to: input the marker image corresponding to each initial inspection site into the first preset model, train the first preset model, and obtain a trained first backbone network and a trained first head network when the first preset training condition is met; use the parameters of the trained first backbone network as the parameters of the site recognition backbone network in the initial state of the site recognition model, and use the parameters of the trained first head network as the parameters of the site recognition head network in the initial state of the site recognition model to obtain the initial state of the site recognition model.

[0009] In a possible implementation, the cloud device is also used to: input the marker image corresponding to the new inspection site into the trained first backbone network to obtain the intermediate features of the new site output by the trained first backbone network; send the new site intermediate features and the existing site intermediate features to the inspection agent; wherein the existing site intermediate features include the features obtained after the marker images corresponding to the existing inspection sites are input into the trained first backbone network; the inspection agent is also used to: input the new site intermediate features and the existing site intermediate features into the site recognition head network of the site recognition model, update the parameters of the site recognition head network of the site recognition model, stop updating when the second preset training condition is met, and obtain an updated site recognition model.

[0010] In one possible implementation, the multiple existing fault types include multiple initial fault types; the cloud device is also used to: pre-train according to initial equipment fault data corresponding to the multiple initial fault types to obtain the initial state of the fault detection model; the initial equipment fault data includes the equipment image corresponding to the fault type of the power equipment when it is each initial fault type.

[0011] In one possible implementation, the fault detection model includes a fault detection backbone network and a fault detection head network; the cloud device is deployed with a second preset model; the second preset model includes a second backbone network and a second head network; the cloud device is also used to: input the initial device fault data into the second preset model, train the second preset model, and obtain a trained second backbone network and a trained second head network when a third preset training condition is met; use the parameters of the trained second backbone network as the parameters of the fault detection backbone network in the initial state of the fault detection model, and use the parameters of the trained second head network as the parameters of the fault detection head network in the initial state of the fault detection model to obtain the initial state of the fault detection model.

[0012] In a possible implementation, the cloud device is also used to: input the new equipment fault data into the trained second backbone network to obtain the intermediate features of the new equipment fault output by the trained second backbone network; send the new equipment fault intermediate features and the existing equipment fault intermediate features to the inspection agent; wherein the existing equipment fault intermediate features include the features obtained after the corresponding equipment image when the fault type of the power equipment is each existing fault type is input into the trained second backbone network; the inspection agent is also used to: input the new equipment fault intermediate features and the existing equipment fault intermediate features into the fault detection head network of the fault detection model, update the parameters of the fault detection head network of the fault detection model, stop updating when the fourth preset training condition is met, and obtain an updated fault detection model.

[0013] In a possible implementation, the dynamic topology map includes multiple local topology maps and multiple activation nodes; the local topology map corresponds to a sub-area in the inspection area; and the activation node is used to connect different local topology maps.

[0014] According to another aspect of the present application, a brain-inspired power inspection method based on incremental learning is provided, which is applied to the inspection intelligent agent in the above-mentioned brain-inspired power inspection system based on incremental learning; the method includes: obtaining an inspection task; the inspection task includes an inspection starting position and a target inspection site; generating an inspection path according to the inspection starting position, the target inspection site and a dynamic topology map; the inspection path is a route from the inspection starting position to the target inspection site; moving according to the inspection path and obtaining environmental data of the surrounding environment; judging whether the target inspection site has been reached based on the environmental data and a site recognition model; when reaching the target inspection site, obtaining an equipment image of the power equipment at the target inspection site, and obtaining the fault condition of the power equipment based on the equipment image and a fault detection model.

[0015] According to another aspect of the present application, an electronic device is provided, comprising: a processor; a memory for storing processor executable instructions; wherein the processor is configured to implement the above-mentioned brain-inspired power inspection method based on incremental learning when executing the instructions stored in the memory.

[0016] According to another aspect of the present application, a non-volatile computer-readable storage medium is provided, on which computer program instructions are stored, wherein the computer program instructions, when executed by a processor, implement the above-mentioned brain-inspired power inspection method based on incremental learning.

[0017] According to another aspect of the present application, a computer program product is provided, comprising a computer-readable code, or a non-volatile computer-readable storage medium carrying a computer-readable code. When the computer-readable code runs in a processor of an electronic device, the processor in the electronic device executes the above-mentioned brain-inspired power inspection method based on incremental learning.

[0018] The power inspection system of the embodiment of the present application can update the dynamic topology map and the site recognition model when a new inspection site is added, and update the fault detection model when a new fault type is added, so as to adapt to the dynamic changes of the inspection scene, making the system scalable in a dynamically changing environment. In addition, the site recognition model and the fault detection model are updated based on the incremental learning algorithm, which can realize the cumulative learning of new knowledge on the basis of retaining the old knowledge, without the need to retrain the entire model, which can greatly improve the learning efficiency of new inspection sites and new fault types, and enhance the recognition ability of new inspection sites and the detection ability of new fault types.

[0019] Other features and aspects of the present application will become apparent from the following detailed description of exemplary embodiments with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate exemplary embodiments, features, and aspects of the present application and, together with the description, serve to explain the principles of the present application.

[0021] Figure 1 A structural schematic diagram of a brain-inspired power inspection system based on incremental learning according to an embodiment of the present application is shown.

[0022] Figure 2 A schematic diagram of a multi-level construction process of a dynamic topology map according to an embodiment of the present application is shown.

[0023] Figure 3A schematic diagram showing training of a site recognition model and a fault detection model based on an incremental learning algorithm according to an embodiment of the present application is shown.

[0024] Figure 4 A schematic diagram showing a dynamic topology map incremental construction process according to an embodiment of the present application.

[0025] Figure 5 A schematic diagram showing the updating of edge devices in a power inspection system based on an incremental learning algorithm according to an embodiment of the present application.

[0026] Figure 6 A flowchart of a brain-inspired power inspection method based on incremental learning according to an embodiment of the present application is shown.

[0027] Figure 7 A block diagram of an electronic device 1900 according to an embodiment of the present application is shown. DETAILED DESCRIPTION

[0028] Various exemplary embodiments, features and aspects of the present application will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0029] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0030] In addition, in order to better illustrate the present application, numerous specific details are provided in the following specific embodiments. It should be understood by those skilled in the art that the present application can also be implemented without certain specific details. In some examples, methods, means, components and circuits well known to those skilled in the art are not described in detail in order to highlight the subject matter of the present application.

[0031] The traditional automated inspection method for power grids has the following problems: 1. It relies on high-precision static maps and global coordinates in inspection route planning, and cannot quickly and dynamically update inspection strategies according to changes in inspection scenarios, resulting in poor robustness and autonomy; 2. It has poor timeliness and adaptability for fault detection of power equipment, and cannot adapt to the rapid detection requirements of new fault types in the power grid. Retraining the entire detection model is costly and time-consuming; 3. The time cost, storage cost, and computing cost of building high-precision maps are high, and the computing platform requirements are high, making it difficult to combine with lightweight unmanned system platforms (such as drone system platforms).

[0032] In view of this, the present application proposes a brain-inspired power inspection system and method based on incremental learning (IL), which draws on the manual inspection method and does not rely on high-precision maps in the path planning process. Instead, it visually observes the surrounding environment and locates its own position in the global context by constantly comparing site landmarks. The power inspection system of the present application can use incremental learning methods to realize the memory of inspection sites, incrementally construct scalable dynamic topological maps, greatly reduce the time cost, storage cost and computing cost of map construction, and can be implemented through a lightweight edge computing platform. It can be combined with a variety of unmanned system platforms (including but not limited to drone system platforms, wheeled robot system platforms, legged robot system platforms, etc.), which can greatly improve the intelligence, flexibility and robustness of the inspection agent.

[0033] Figure 1 A schematic diagram of the structure of a brain-inspired power inspection system based on incremental learning according to an embodiment of the present application is shown. Figure 1 As shown, the system includes multiple inspection agents equipped with edge computing devices; the inspection agents are used to detect fault conditions of power equipment at inspection sites within the inspection area; the edge computing devices are deployed with site recognition models and fault detection models, and store a dynamic topological map (Dynamic Topological Map, DTM); the dynamic topological map is used to reflect the relative position relationship and connectivity relationship between multiple existing inspection sites, as well as the markers corresponding to each existing inspection site; the site recognition model is used to identify the multiple existing inspection sites; the fault detection model is used to determine whether a fault occurs in the power equipment, and in the case of a fault in the power equipment, determine the fault location of the power equipment and determine the fault type of the power equipment from multiple existing fault types.

[0034] Among them, when a new inspection site is added to the multiple existing inspection sites, the dynamic topology map is updated according to the new site data corresponding to the new inspection site; the new site data includes the relative position information and connectivity information between the new inspection site and the multiple existing inspection sites, and the marker image corresponding to the new inspection site; the site recognition model is updated based on the incremental learning algorithm according to the marker image corresponding to the new inspection site; when a new fault type is added to the multiple existing fault types, the fault detection model is updated based on the incremental learning algorithm according to the new equipment fault data corresponding to the new fault type; the new equipment fault data includes the equipment image corresponding to the fault type of the power equipment when it is the new fault type.

[0035] For example, the inspection agent can be an intelligent agent such as a drone, a wheeled robot, or a legged robot that can perform inspection tasks. An edge computing device is a device that can integrate, analyze, and calculate feedback on data at the data collection end or the edge end of the system. An inspection agent equipped with an edge computing device can be called an edge device, which combines the capabilities of the inspection agent and the edge computing capabilities of the edge computing device.

[0036] Exemplarily, the system also includes a cloud device; each inspection intelligent entity equipped with an edge computing device (i.e., each edge device) is connected to the cloud device respectively.

[0037] The power inspection system of the embodiment of the present application can update the dynamic topology map and site recognition model when a new inspection site is added, and update the fault detection model when a new fault type is added, so as to adapt to the dynamic changes of the inspection scene, making the system scalable in a dynamically changing environment. In addition, the site recognition model and the fault detection model are updated based on the incremental learning algorithm, which can realize the cumulative learning of new knowledge without forgetting the old knowledge, and can greatly improve the learning efficiency of new sites and new fault types.

[0038] The following describes the construction process of the initial state of the dynamic topology map, the initial state of the site recognition model, and the initial state of the fault detection model in the power inspection system of the embodiment of the present application.

[0039] In a possible implementation, the multiple existing inspection sites include multiple initial inspection sites; the initial state of the dynamic topology map (ie, the initial dynamic topology map) is constructed based on the initial site data corresponding to the multiple initial inspection sites.

[0040] The technician can set the scope of the initial inspection area and multiple initial inspection sites according to actual needs. After setting the initial inspection area and the initial inspection site, the technician can control one or several intelligent agents (for example, one or a small number of idle inspection intelligent agents in the power inspection system of the embodiment of the present application can be controlled) to explore the initial inspection area and collect the initial site data through sensors and other equipment. The intelligent agent can record the relative position and spatial connectivity between the initial inspection sites, and use the visual sensor to collect images of the markers (such as landmarks, buildings, etc.) around each initial inspection site when arriving at each initial inspection site. The initial site data may include the longitude and latitude information of each initial inspection site, the relative position information and connectivity information between multiple initial inspection sites, and the marker images corresponding to each initial inspection site. The initial position when the intelligent agent explores the initial area can be used as the coordinate origin (0,0), and the relative position information may include the relative coordinates of each initial inspection site relative to the initial position. Connectivity information may include whether the initial inspection sites are connected and the connection direction.

[0041] The collected initial site data can be uploaded to the cloud device. As an example, an initial dynamic topology map can be constructed in the cloud device, and the cloud device then sends the initial dynamic topology map to each edge device in the system. As another example, the cloud device can send the initial site data to each edge device, and each edge device builds the initial dynamic topology map by itself. As another example, the cloud device can send the initial site data to an edge device in the system, and the edge device builds the initial dynamic topology map and uploads the initial dynamic topology map to the cloud device, and the cloud device then sends the initial dynamic topology map to other edge devices in the system.

[0042] Each initial inspection site can be used as a node in the initial dynamic topology map, and the relative coordinates of the initial inspection site can be marked at the node corresponding to each initial inspection site. If the initial inspection site A is connected to the initial inspection site B and the connection direction is from A to B, the node A corresponding to the initial inspection site A and the node B corresponding to the initial inspection site B are connected in the initial dynamic topology map and the connection direction is from node A to node B. The marker corresponding to each initial inspection site can be determined according to the marker image corresponding to each initial inspection site, and the marker corresponding to the initial inspection site can be marked at the node corresponding to each initial inspection site. For example, the marker corresponding to the initial inspection site A is transformer box A, and the marker can be marked as transformer box A at node A of the initial dynamic topology map. In this way, the dynamic topology map of the inspection area can be preliminarily constructed according to the relative position information and connectivity information between multiple initial inspection sites and the marker images corresponding to each initial inspection site. The name of each inspection site, the relative coordinates, the marker, and the connectivity information between the nodes can constitute the core data structure of the dynamic topology map. The initial dynamic topology map is used to reflect the relative position relationship and connectivity relationship between multiple initial inspection sites, as well as the markers corresponding to each initial inspection site.

[0043] Exemplarily, the dynamic topology map may also include auxiliary nodes for indicating turns, U-turns, etc.

[0044] Exemplarily, the dynamic topology map may include multiple local topology maps and multiple activation nodes; the local topology map corresponds to a sub-area in the inspection area; and the activation node is used to connect different local topology maps.

[0045] In order to improve the robustness of map navigation and the efficiency of path planning, in the process of topological map construction, all topological nodes are not a single-layer large network, but analogous to the storage process of the human brain for maps. The scenes are stored in different regions during the storage process, and the key information of the map is stored in the form of scenes. For example, the map of a room or a campus is stored by scene. For the dynamic topological map of the embodiment of the present application, a sub-area in the inspection area can be regarded as a scene, for example, a factory area in the inspection area can be regarded as a scene, and a factory building can also be regarded as a scene. The topological map corresponding to each sub-area in the inspection area is called a local topological map, and the topological map corresponding to the entire inspection area is called a global topological map.

[0046] There are parallel and inclusive relationships between the sub-areas of the inspection area. For example, multiple factory buildings in a factory area in the inspection area are in a parallel relationship, and the factory areas and factory buildings are in an inclusive relationship. Then the local topological maps corresponding to the multiple factory buildings are in a parallel relationship, and the local topological maps corresponding to the factory areas and the local topological maps corresponding to the factory buildings are in an inclusive relationship. The local topological maps corresponding to the multiple factory buildings can be regarded as belonging to the same level, and the local topological maps corresponding to the factory areas can be regarded as belonging to the upper level of the local topological maps corresponding to the factory buildings. In this way, based on the parallel and inclusive relationships between the sub-areas, local topological maps of multiple levels can be constructed. The local topological maps of the same level are in a parallel relationship, and the local topological maps of the high level contain the local topological maps of the low level. The global topological map can be regarded as the highest level. The topological map constructed in this way can be called a "hierarchical topological map."

[0047] Different local topology maps can be connected through activation nodes. For example, activation node A is used to connect local topology map B corresponding to plant B and local topology map C corresponding to plant C. When the inspection agent needs to enter plant C from plant B for inspection, it can jump from local topology map B to local topology map C through activation node A. Connecting all local topology maps through multiple activation nodes can form the entire dynamic topology map.

[0048] For example, different local topology maps can be stored in separate files, and one file can be used to store file name indexes and activation nodes. In this way, when a local topology map needs to be opened, the file corresponding to the local topology map can be found through the file name index, and other local topology maps connected to the local topology map can be found through the activation nodes.

[0049] Figure 2 A schematic diagram of a multi-level construction process of a dynamic topology map according to an embodiment of the present application is shown as follows: Figure 2 As shown, different local topology maps have a parallel relationship or an inclusion relationship. Two different local topology maps are connected by activating nodes, and all local topology maps are connected to form the entire dynamic topology map. The local topology map is stored in separate files. The corresponding local topology map and other local topology maps connected to the local topology map can be found through the file name index and the activation node. When the inspection agent performs the inspection task and uses the dynamic topology map for path planning and navigation, the local topology maps that are not involved can be left unexpanded. For example, when the inspection agent is inspecting a factory area, it only needs to open the local topology map corresponding to the factory area. If you want to enter a factory building from the factory entrance, you can open the corresponding local topology map layer by layer, and irrelevant topology maps will not be expanded, which can improve the efficiency of path planning and reduce costs.

[0050] The dynamic topological map of the embodiment of the present application stores the key information of the map in the form of the relative position of the inspection site, the markers and their connectivity with other inspection sites, which can save storage space and reduce storage costs; it is divided into a global topological map and a local topological map according to the nature of the scene, and the topological maps have a parallel relationship and an inclusion relationship. Different topological maps are stored in separate files for dynamic loading when used. In addition, the dynamic topological map of the embodiment of the present application can cover the global navigation during the inspection process and the local navigation of specific sites, which can greatly reduce the time cost, storage cost and calculation cost brought by the high-precision map during the inspection process, and can provide a more robust navigation method.

[0051] In a possible implementation, the cloud device is used to: perform pre-training based on the marker images corresponding to each initial inspection site to obtain the initial state of the site recognition model.

[0052] Exemplarily, the site recognition model may include a site recognition backbone network and a site recognition head network; the cloud device is deployed with a first preset model; the first preset model includes a first backbone network and a first head network; the process of the cloud device pre-training the initial state of the site recognition model according to the marker images corresponding to each initial inspection site may include:

[0053] (1) Inputting the marker images corresponding to each initial inspection site into a first preset model, training the first preset model, and obtaining a trained first backbone network and a trained first head network when a first preset training condition is met.

[0054] The intelligent agent collects the initial site data and uploads it to the cloud device, which is used to input the marker images corresponding to each initial inspection site into the first preset model for pre-training. The first preset model is used to identify the markers in the marker image to identify the inspection site. The first preset training condition can be that the mean average precision (mAP) of the first preset model is greater than or equal to a preset threshold (e.g., 80%).

[0055] (2) The parameters of the trained first backbone network are used as the parameters of the site recognition backbone network in the initial state of the site recognition model, and the parameters of the trained first head network are used as the parameters of the site recognition head network in the initial state of the site recognition model to obtain the initial state of the site recognition model.

[0056] After the first backbone network and the first head network are trained in the cloud device, the cloud device sends the parameters of the first backbone network and the first head network to each edge device as the parameters of the site recognition backbone network and the site recognition head network in the initial state of the site recognition model, thereby obtaining the initial state of the site recognition model (i.e., the initial site recognition model). The initial site recognition model has the ability to recognize the markers in the marker image and thus recognize the initial inspection site.

[0057] In a possible implementation, the multiple existing fault types include multiple initial fault types; the cloud device is also used to: perform pre-training based on initial device fault data corresponding to the multiple initial fault types to obtain an initial state of the fault detection model.

[0058] The initial equipment failure data includes the equipment images corresponding to the fault types of the power equipment when they are each initial failure type. Equipment images of the power equipment at each initial inspection location can be collected when the intelligent agent explores the initial inspection area, and the technicians can mark the fault location and fault type in the equipment image. These marked equipment images can be called equipment failure data. Since the equipment failure data collected in the initial inspection area is limited, an existing equipment failure data set can also be obtained. The existing equipment failure data set contains a large number of equipment images marked with fault locations and fault types. The equipment failure data in the existing equipment failure data set and the equipment failure data collected in the initial inspection area can together constitute the initial equipment failure data, and the fault type marked in the initial equipment failure data can be called the initial failure type.

[0059] Exemplarily, the fault detection model may include a fault detection backbone network and a fault detection head network; the cloud device is deployed with a second preset model; the second preset model includes a second backbone network and a second head network; the process of the cloud device pre-training according to the initial device fault data to obtain the initial state of the fault detection model may include:

[0060] (1) Inputting initial equipment fault data into a second preset model, training the second preset model, and obtaining a trained second backbone network and a trained second head network when a third preset training condition is met.

[0061] After obtaining the initial equipment fault data, upload it to the cloud device, and the cloud device is used to input the initial equipment fault data into the second preset model for pre-training. The second preset model is used to identify the fault location and fault type of the power equipment, and the third preset training condition can be that the average accuracy of the second preset model is greater than or equal to a preset threshold (e.g., 80%).

[0062] (2) The parameters of the trained second backbone network are used as the parameters of the fault detection backbone network in the initial state of the fault detection model, and the parameters of the trained second head network are used as the parameters of the fault detection head network in the initial state of the fault detection model to obtain the initial state of the fault detection model.

[0063] After the second backbone network and the second head network are trained in the cloud device, the cloud device sends the parameters of the second backbone network and the second head network to each edge device as the parameters of the fault detection backbone network and the fault detection head network in the initial state of the fault detection model, thereby obtaining the initial state of the fault detection model (i.e., the initial fault detection model). The initial fault detection model has the ability to identify the fault type and the fault location from multiple initial fault types.

[0064] The purpose of pre-training the model in the cloud is to obtain general feature extraction capabilities and obtain better test results on known categories. The trained first backbone network and the second backbone network have feature extraction capabilities, that is, the site recognition backbone network and the fault detection backbone network have feature extraction capabilities. The site recognition backbone network and the fault detection backbone network can extract intermediate features from the input image, and the site recognition head network and the fault detection head network can obtain detection results based on the intermediate features extracted by the backbone network. Subsequently, the site recognition model and the fault detection model can be updated based on the incremental learning algorithm. During the update process, the parameters of the site recognition backbone network and the fault detection backbone network are frozen, and only the site recognition head network and the fault detection head network are updated. The incremental learning (also known as continuous learning) algorithm is an artificial intelligence learning method that can gradually absorb new information and maintain old knowledge in an environment where data is constantly changing. It is suitable for dynamic environments.

[0065] Figure 3 A schematic diagram showing training of a site recognition model and a fault detection model based on an incremental learning algorithm according to an embodiment of the present application is shown. Figure 3As shown, after collecting the initial site data and the initial device fault data, upload them to the cloud device and pre-train the first preset model and the second preset model in the cloud device to obtain the trained first backbone network, the first head network, the second backbone network, and the second head network; the parameters of the trained first backbone network, the first head network, the second backbone network, and the second head network are sent to each edge device as the parameters of the site identification backbone network, the site identification head network, the fault detection backbone network, and the fault detection head network, respectively, to obtain the initial site identification model and the initial fault detection model. After obtaining the pre-trained model, the parameters of the site identification backbone network and the fault detection backbone network with feature extraction capabilities are frozen as the feature extraction front end of the model. The intermediate feature data obtained after the initial site data and the initial device fault data pass through the trained cloud backbone network can be used as the continued training data to continue training the site identification head network and the fault detection head network. The site identification head network and the fault detection head network are small trainable networks, which can be called incremental learning network modules. Their scale is limited and online fine-tuning can be achieved. When there is new site data or new equipment fault data that needs to be learned, these new data are passed through the trained cloud backbone network to obtain intermediate feature data, and are mixed with the existing intermediate feature data to obtain mixed feature data. The mixed feature data can be used to continue training the incremental learning network module (i.e., the site recognition head network and the fault detection head network) and update the parameters of the site recognition head network and the fault detection head network. In this way, by using the incremental learning algorithm to continuously optimize the site recognition model and the fault detection model, it is possible to continuously update the model with new data while retaining the old knowledge, and realize the cumulative learning of new knowledge without retraining the entire network. This can greatly improve the learning efficiency of new sites and new fault types, and enhance the ability to recognize new sites and detect new fault types.

[0066] The following is a detailed introduction to the updating process of the dynamic topology map, site identification model and fault detection model of the embodiment of the present application.

[0067] The technician can determine whether it is necessary to add a new inspection site or expand a new inspection area in the current inspection area according to actual needs. Both of these situations belong to the situation of adding a new inspection site to the existing inspection sites. If a new inspection site needs to be added, the technician can control one or several intelligent agents (for example, one or a small number of idle inspection intelligent agents in the power inspection system of the embodiment of the present application can be controlled) to collect new site data corresponding to the new inspection site. The process of collecting new site data can refer to the process of collecting initial site data, which will not be repeated here. The new site data may include the relative position information between the new inspection site and the existing inspection sites (i.e., the relative coordinates of the new inspection site) and connectivity information (i.e., whether the new inspection site is connected to each existing inspection site and the direction of connection), as well as the marker image corresponding to the new inspection site.

[0068] The collected new site data can be uploaded to the cloud device. As an example, the dynamic topology map can be updated in the cloud device, and the cloud device then sends the updated dynamic topology map to each edge device. As another example, the cloud device can send the new site data to each edge device, and each edge device updates the dynamic topology map by itself. As another example, the cloud device can send the new site data to an edge device in the system, and the edge device updates the dynamic topology map and uploads the updated dynamic topology map to the cloud device, and the cloud device then sends the updated dynamic topology map to other edge devices in the system.

[0069] The node corresponding to the new inspection site can be added to the dynamic topology map according to the new site data, the relative coordinates of the new inspection site can be marked at the node corresponding to the new inspection site according to the relative position information corresponding to the new inspection site, the connection relationship between the node corresponding to the new inspection site and other nodes can be determined according to the connectivity information corresponding to the new inspection site, and the marker can be marked at the node corresponding to the new inspection site according to the marker image corresponding to the new inspection site, thereby realizing the update of the dynamic topology map.

[0070] The construction and update process of the dynamic topology map in the embodiment of the present application can be referred to as a dynamic topology map incremental construction process. Figure 4 A schematic diagram showing a dynamic topology map incremental construction process according to an embodiment of the present application is shown as follows: Figure 4As shown, during the initial construction of the dynamic topology map, the intelligent agent is controlled to explore the environment of the initial inspection site, and the location of the initial inspection site is fed back using the intelligent agent's route information to construct the initial dynamic topology map. In the case of a new inspection site, the intelligent agent is controlled to explore the new inspection site, and the location of the new inspection site is topologically recorded through the intelligent agent's motion data, and the dynamic topology map is updated. In this way, in the case of a new inspection site, the power inspection system of the embodiment of the present application can automatically update the dynamic topology map and incrementally add new path information and node information to the topology map.

[0071] Compared with the existing power inspection system, which usually uses drones or robotic devices equipped with depth cameras to collect real-time three-dimensional data of power grid equipment and the surrounding environment to build a three-dimensional model of the environment, the map construction cost is high, the update is complex, and the real-time performance is poor in complex environments. The dynamic topological map of the embodiment of the present application has a low construction cost and can use the knowledge of incremental learning to update the existing topological network. It is suitable for dynamic and complex environment modeling and can be used for brain-like navigation, which effectively improves the efficiency of path planning and the robustness of navigation. The power inspection system of the embodiment of the present application only needs to maintain a dynamically changing topological map when the inspection scene changes, without rebuilding the entire map, which greatly reduces the time cost, storage cost and computing cost of map construction.

[0072] When a new inspection site is added, the power inspection system of the embodiment of the present application will update the site recognition model based on the incremental learning algorithm according to the marker image corresponding to the new inspection site while updating the dynamic topology map, so as to ensure that the system is scalable in a dynamically changing environment. After the dynamic topology map and the site recognition model are updated, the new inspection site becomes an existing inspection site.

[0073] In a possible implementation, the update of the site recognition model is performed collaboratively by the cloud device and the edge device. The cloud device is used to: input the marker image corresponding to the new inspection site into the trained first backbone network to obtain the intermediate features of the new site output by the trained first backbone network; send the intermediate features of the new site and the intermediate features of the existing site to the edge device; the intermediate features of the existing site include the features obtained after the marker images corresponding to each existing inspection site are input into the trained first backbone network; the edge device is used to: input the intermediate features of the new site and the intermediate features of the existing site into the site recognition head network of the site recognition model, update the parameters of the site recognition head network of the site recognition model, stop updating when the second preset training condition is met, and obtain the updated site recognition model.

[0074] During the pre-training process of the initial site recognition model, the intermediate feature data output by the marker image corresponding to the initial inspection site after passing through the trained first backbone network can be stored in the cloud device. In the subsequent update of the site recognition model, the intermediate feature data output by the marker image corresponding to the new inspection site after passing through the trained first backbone network can also be stored in the cloud device. In this way, the cloud device can store the intermediate feature data (i.e., the intermediate features of the existing sites) output by the marker images corresponding to all existing inspection sites after passing through the trained first backbone network.

[0075] When updating the existing site recognition model, the new site intermediate features can be mixed with the existing site intermediate features stored in the cloud device to obtain mixed feature data and send it to the edge device. Since the mixed feature data is large in size, it can be compressed and encoded in the cloud device before being sent to the edge device, which can increase the data transmission speed between the cloud and the edge and reduce the transmission cost. After receiving the compressed mixed feature data, the edge device decompresses it and inputs the decompressed mixed feature data into the site recognition head network to update the parameters of the site recognition head network. The update can be stopped when the average accuracy of the site recognition head network is greater than or equal to the preset threshold, and the updated site recognition head network is obtained. The updated site recognition head network and the original site recognition backbone network constitute an updated site recognition model.

[0076] The above process is a process for updating the dynamic topology map and the site recognition model when the technician determines that a new inspection site needs to be added. In one embodiment, if the inspection agent discovers a new site in the process of performing the inspection task, the site data corresponding to the new site can be collected and the dynamic topology map can be updated. If the inspection agent identifies that the marker corresponding to the new site is a known marker type through the site recognition model, new node information and new path information can be added to the dynamic topology map according to the relative position information, connectivity information and marker type of the new site; if the marker corresponding to the new site is a new marker type (that is, a marker type that cannot be recognized by the existing site recognition model), then while updating the node information and path information in the dynamic topology map, the site recognition model must also be updated so that the site recognition model can recognize this new marker type.

[0077] For the fault detection model, the technician can determine whether it is necessary to update the fault detection model based on actual needs so that the fault detection model can identify new fault types. In the case of needing to identify new fault types, the electric power inspection system of the embodiment of the present application can update the fault detection model based on the incremental learning algorithm according to the new equipment fault data corresponding to the new fault type to ensure that the equipment fault information can also be dynamically expanded. For example, if the equipment fault data set updates a batch of equipment fault data, which contains equipment images marked with new fault types, the equipment images corresponding to these new fault types can be used as new equipment fault data to update the fault detection model. After the fault detection model is updated, the new fault type becomes an existing fault type. After obtaining the new equipment fault data, it can be uploaded to the cloud device.

[0078] In a possible implementation, the update of the fault detection model is carried out collaboratively by the cloud device and the edge device. The cloud device is used to: input the new device fault data into the trained second backbone network to obtain the intermediate features of the new device fault output by the trained second backbone network; send the intermediate features of the new device fault and the intermediate features of the existing device fault to the edge device; the intermediate features of the existing device fault include the features obtained after the corresponding device image when the fault type of the power equipment is each existing fault type is input into the trained second backbone network; the edge device is used to: input the intermediate features of the new device fault and the intermediate features of the existing device fault into the fault detection head network of the fault detection model, update the parameters of the fault detection head network of the fault detection model, stop updating when the fourth preset training condition is met, and obtain an updated fault detection model.

[0079] During the pre-training process of the initial fault detection model, the intermediate feature data output by the initial device fault data after passing through the trained second backbone network can be stored in the cloud device. In the subsequent updating of the fault detection model, the intermediate feature data output by the new device fault data after passing through the trained second backbone network can also be stored in the cloud device. In this way, the cloud device can store the intermediate feature data (i.e., the intermediate features of the existing device faults) output by the device fault data corresponding to all existing fault types after passing through the trained second backbone network.

[0080] When updating the existing fault detection model, the intermediate features of the new device faults and the intermediate features of the existing device faults stored in the cloud device are mixed to obtain mixed feature data and sent to the edge device. The mixed feature data can be compressed and encoded in the cloud device before being sent to the edge, thereby improving the data transmission speed between the cloud and the edge and reducing the transmission cost. After receiving the compressed mixed feature data, the edge device decompresses it and inputs the decompressed mixed feature data into the fault detection head network to update the parameters of the fault detection head network. The update can be stopped when the average accuracy of all categories of the fault detection head network is greater than or equal to the preset threshold, and an updated fault detection head network is obtained. The updated fault detection head network and the original fault detection backbone network constitute an updated fault detection model.

[0081] Existing power inspection systems usually use traditional deep learning methods to train large neural networks for fault detection. The fault detection model trained in this way has poor adaptability to new data and requires frequent retraining of the model, which is time-consuming and costly. There are also studies that use transfer learning methods to apply pre-trained models to power grid fault detection. Although transfer learning can utilize the knowledge of existing models to a certain extent, it still requires a large amount of labeled data for retraining when dealing with new fault types, and the adaptability and flexibility of the model are poor. Compared with the fault detection model trained using deep learning and transfer learning methods, the fault detection model of the embodiment of the present application is trained and updated based on the incremental learning algorithm, which has stronger adaptability and flexibility, and the model's update and iteration efficiency is higher, which can meet the needs of rapid identification of new equipment fault types in the power grid.

[0082] When the site identification model / fault detection model needs to be updated, the cloud device can send the hybrid feature data required for the update to an idle edge device in the system, and the edge device updates the parameters of the site identification head network / fault detection head network and uploads the updated parameters to the cloud device, and the cloud device then sends the updated parameters to each edge device so that the models deployed in all edge devices are updated; or, the cloud device can send the hybrid feature data required for the update to a small number of idle edge devices in the system, and these edge devices update the parameters of the site identification head network / fault detection head network, and compare the average correct rates of all classes of the updated models in these edge devices, and upload the updated parameters of the model with the highest average correct rate of all classes to the cloud device, and the cloud device then sends the updated parameters to each edge device so that the models deployed in all edge devices are updated. In this way, when the inspection scene changes dynamically, the power inspection system of the embodiment of the present application can use a small number of edge devices to update the model, and other edge devices can perform inspection tasks normally during this process without stopping the operation of the entire system for updating, thereby realizing online update of edge devices.

[0083] Figure 5 A schematic diagram showing an edge device in a power inspection system according to an embodiment of the present application being updated based on an incremental learning algorithm is shown, Figure 5 As shown in the figure, the update of the site recognition model and the fault detection model is divided into two parts: the cloud and the edge. The cloud device is responsible for training a large-scale pre-trained model to obtain a pre-trained backbone network with feature extraction capabilities (i.e., the trained first backbone network and the second backbone network), and outputting the input image (including the marker image and the device image) input to the pre-trained backbone network as intermediate feature data, and can store and compress the intermediate feature data. The edge device is responsible for using the site recognition model and the fault detection model to perform real-time reasoning on the current environment. When the model needs to be updated to identify a new category (new inspection site or new fault type), the cloud device mixes the newly obtained intermediate feature data with the existing intermediate feature data to obtain mixed feature data, and compresses it and sends it to the edge device. After decompression, the edge device uses the mixed feature data to quickly update the unfrozen incremental learning network part of the model (i.e., the site recognition head network and the fault detection head network) to adapt the edge device to the new category. The electric power inspection system of the embodiment of the present application adopts a cloud-edge collaborative inspection framework. When the inspection scenario changes dynamically, the edge devices in the system can be quickly updated online through cloud-edge collaboration, and the ability to recognize new categories can be achieved while using a small amount of computing power.

[0084] When updating the dynamic topology map, site identification model or fault detection model, the time for each edge device in the power inspection system to update can be set by the technician according to actual needs. For example, all edge devices can be updated at the same time, or edge devices in idle state can be updated first. For edge devices that are performing inspection tasks, they can be updated after the tasks are completed.

[0085] The brain-inspired power inspection system based on incremental learning in the embodiment of the present application proposes an incremental learning paradigm and model architecture for inspection scenarios, and by constructing a pre-trained model, it has the ability to extract features; through the incremental learning network module, new inspection sites and new equipment fault types can be efficiently learned, and the dynamic changes of the adaptive inspection area and the recognition requirements of new equipment fault types can be avoided, avoiding catastrophic forgetting of old categories during the learning process, avoiding model training from scratch, greatly improving the update and iteration efficiency of the model, and can be deployed on an edge computing platform. The power inspection system of the embodiment of the present application can be applied to power equipment defect inspection scenarios, reducing the inspection process's dependence on manual labor, and at the same time helping to improve the intelligence of existing autonomous inspection systems.

[0086] The present application also proposes a brain-inspired power inspection method based on incremental learning, which can be applied to the inspection agent in the brain-inspired power inspection system based on incremental learning proposed in the present application.

[0087] Figure 6 A flowchart of a brain-inspired power inspection method based on incremental learning according to an embodiment of the present application is shown. Figure 6 As shown, the method may include:

[0088] S601, obtaining an inspection task; the inspection task includes an inspection start position and a target inspection position.

[0089] Technical personnel can determine the inspection tasks and the inspection agents used to perform the inspection tasks according to the needs, and send the inspection tasks to the corresponding inspection agents.

[0090] S602: Generate an inspection path according to the inspection starting position, the target inspection site and a dynamic topology map; the inspection path is a route from the inspection starting position to the target inspection site.

[0091] For example, after obtaining the inspection task, the inspection agent can automatically plan the inspection path based on the path planning algorithm according to the inspection starting position, the target inspection location and the dynamic topological map. The path planning algorithm can be an existing local space path planning algorithm, such as GPS navigation, visual navigation, etc.

[0092] S603: Move according to the inspection path and obtain environmental data of the surrounding environment.

[0093] Exemplarily, the inspection agent includes a control module, which can be designed with a hierarchical controller, including a high-level controller and a low-level controller. The high-level controller is responsible for global path planning and can generate overall navigation instructions for inspection tasks. By using a dynamic topological map for global path planning, the task execution of the inspection agent can be ensured to be stable. The low-level controller is used to convert the instructions of the high-level controller into actual operations and drive the inspection agent to perform inspection actions. The low-level controller can continuously adjust the movement direction and speed of the inspection agent to adapt to the current environmental conditions.

[0094] Exemplarily, the environmental data includes environmental images, and the inspection agent includes a visual sensor, which can collect environmental images through the visual sensor during movement.

[0095] S604: Determine whether the target inspection location has been reached based on the environmental data and the location recognition model.

[0096] During the movement, the patrol agent can input the collected environmental image into the site recognition model deployed by itself, and use the site recognition model to determine whether it has reached the target patrol site. The site recognition model can identify the markers in the environmental image. The patrol agent contains a judgment module, which can determine the current location based on the recognition results of the site recognition model and the markers marked at each patrol site in the dynamic topology map. For example, if the site recognition model recognizes that there is marker A in the currently collected environmental image, and marker A corresponds to site A in the dynamic topology map, then the current location can be determined to be site A; if there are multiple sites in the dynamic topology map and the corresponding markers are all marker A, the judgment module can determine the current location from these multiple sites based on the inspection path and the sites that have been passed.

[0097] During the actual movement of the patrol agent, it may deviate from the planned path due to obstacle avoidance and other reasons. If the patrol agent determines that the current location is not in the planned patrol path, it can re-plan the patrol path based on the current location, the target patrol location and the dynamic topology map. For example, the patrol starting position is location A, the target patrol location is location D, and the planned patrol path is location A-location B-location C-location D. If the patrol agent determines that it is at location E during the movement, and location E is not in the originally planned path, the patrol agent can re-plan the path from location E to location D.

[0098] If the patrol agent determines that the current location is in the planned patrol path, it will continue to move along the patrol path and collect environmental images for the location recognition model to identify, and continue to determine whether the next location to be reached is in the planned patrol path, until the location recognition model identifies the marker corresponding to the target patrol location, and the patrol agent determines that it has reached the target patrol location. In this way, the patrol agent can continuously identify which location it is at based on the environmental image, and can determine its position in the global map based on the location of its location in the map, so as to complete the "road recognition" process by continuously comparing locations.

[0099] S605: When arriving at the target inspection location, obtain a device image of the power equipment at the target inspection location, and obtain a fault condition of the power equipment based on the device image and a fault detection model.

[0100] For example, after the inspection agent arrives at the target inspection site, it can collect device images of the power equipment through visual sensors, input the collected device images into the fault detection model deployed by itself, and call the fault detection model to detect the fault conditions of the power equipment at the target inspection site. The fault detection model can detect whether the power equipment at the location has a fault, and output the fault location and fault type of the power equipment if the power equipment at the location has a fault.

[0101] For example, when the fault detection model detects that a fault has occurred in the power equipment, the inspection agent can take pictures of the faulty part of the power equipment through visual sensors, and identify the faulty part and fault type of the power equipment at that location in the dynamic topology map, so that the fault handling personnel can handle the fault according to the inspection results.

[0102] For example, sites that have been inspected in the inspection area may be marked in the dynamic topology map to avoid repeated inspections.

[0103] The brain-inspired power inspection method based on incremental learning in the embodiment of the present application adopts a brain-inspired inspection framework and can realize brain-like navigation through dynamic topological maps and site recognition models. It does not rely on high-precision maps. The inspection intelligent agent visually observes the surrounding environment and constantly compares site landmarks to locate itself in the global environment. It has higher flexibility and robustness. In addition, by using a fault detection model updated based on an incremental learning algorithm to detect fault conditions of power equipment, new equipment fault types can also be quickly identified.

[0104] The present application also provides a computer-readable storage medium on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the above-mentioned brain-inspired power inspection method based on incremental learning is implemented. The computer-readable storage medium can be a volatile or non-volatile computer-readable storage medium.

[0105] An embodiment of the present application also proposes an electronic device, comprising: a processor; a memory for storing instructions executable by the processor; wherein the processor is configured to implement the above-mentioned brain-inspired power inspection method based on incremental learning when executing the instructions stored in the memory.

[0106] An embodiment of the present application also provides a computer program product, including a computer-readable code, or a non-volatile computer-readable storage medium carrying a computer-readable code. When the computer-readable code runs in a processor of an electronic device, the processor in the electronic device executes the above-mentioned brain-inspired power inspection method based on incremental learning.

[0107] Figure 7 1 shows a block diagram of an electronic device 1900 according to an embodiment of the present application. For example, the electronic device 1900 can be provided as a server or a terminal device. Figure 7 , the electronic device 1900 includes a processing component 1922, which further includes one or more processors, and a memory resource represented by a memory 1932 for storing instructions executable by the processing component 1922, such as an application. The application stored in the memory 1932 may include one or more modules, each corresponding to a set of instructions. In addition, the processing component 1922 is configured to execute instructions to perform the above-mentioned brain-inspired power inspection method based on incremental learning.

[0108] The electronic device 1900 may also include a power supply component 1926 configured to perform power management of the electronic device 1900, a wired or wireless network interface 1950 configured to connect the electronic device 1900 to a network, and an input / output interface 1958 (I / O interface). The electronic device 1900 may operate based on an operating system stored in the memory 1932, such as Windows Server 2003. TM , Mac OS X TM , Unix TM ,Linux TM , FreeBSD TM or similar.

[0109] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions, which can be executed by a processing component 1922 of an electronic device 1900 to complete the above-mentioned brain-inspired power inspection method based on incremental learning.

[0110] The present application may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present application.

[0111] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.

[0112] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0113] The computer program instructions for performing the operation of the present application can be assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine-related instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages, such as Smalltalk, C++, etc., and conventional procedural programming languages, such as "C" language or similar programming languages. Computer-readable program instructions can be executed completely on a user's computer, partially on a user's computer, executed as an independent software package, partially on a user's computer, partially on a remote computer, or completely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, using an Internet service provider to connect through the Internet). In some embodiments, by using the state information of a computer-readable program instruction to personalize an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA) or a programmable logic array (PLA), the electronic circuit can execute a computer-readable program instruction, thereby realizing various aspects of the present application.

[0114] Various aspects of the present application are described herein with reference to the flowcharts and / or block diagrams of the methods, devices (systems) and computer program products according to the embodiments of the present application. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer-readable program instructions.

[0115] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processor of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0116] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operating steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0117] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple embodiments of the present application. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of the module, program segment or instruction includes one or more executable instructions for realizing the logical function of the specification. In some alternative implementations, the function marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can also be executed in the opposite order sometimes, depending on the function involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be realized by a dedicated hardware-based system that performs the function or action of the specification, or can be realized by a combination of special-purpose hardware and computer instructions.

[0118] The embodiments of the present application have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and changes will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A brain-inspired power inspection system based on incremental learning, characterized in that: The system includes a plurality of inspection agents equipped with edge computing devices; The inspection agent is used to detect fault conditions of power equipment at inspection sites within the inspection area; the edge computing device is deployed with a site recognition model and a fault detection model, and stores a dynamic topology map; the dynamic topology map is used to reflect the relative position relationship and connectivity relationship between multiple existing inspection sites, as well as the markers corresponding to each existing inspection site; The site recognition model is used to identify the multiple existing inspection sites; The fault detection model is used to determine whether the power equipment fails, and in the case of a fault in the power equipment, determine the fault location of the power equipment and determine the fault type of the power equipment from a plurality of existing fault types; Among them, when a new inspection site is added to the multiple existing inspection sites, the dynamic topology map is updated according to the new site data corresponding to the new inspection site; the new site data includes the relative position information and connectivity information between the new inspection site and the multiple existing inspection sites, and the marker image corresponding to the new inspection site; the site recognition model is updated based on the incremental learning algorithm according to the marker image corresponding to the new inspection site; when a new fault type is added to the multiple existing fault types, the fault detection model is updated based on the incremental learning algorithm according to the new equipment fault data corresponding to the new fault type; the new equipment fault data includes the equipment image corresponding to the fault type of the power equipment when it is the new fault type.

2. The system according to claim 1, characterized in that The system also includes a cloud device; each of the inspection intelligent bodies is connected to the cloud device respectively.

3. The system according to claim 2, characterized in that The multiple existing inspection sites include multiple initial inspection sites; The initial state of the dynamic topology map is constructed based on the initial site data corresponding to the multiple initial inspection sites; The initial site data includes relative position information and connectivity information between the multiple initial inspection sites, and a marker image corresponding to each initial inspection site; The cloud device is used to: perform pre-training according to the marker images corresponding to each initial inspection site to obtain the initial state of the site recognition model.

4. The system according to claim 3, characterized in that The site recognition model includes a site recognition backbone network and a site recognition head network; the cloud device is deployed with a first preset model; the first preset model includes a first backbone network and a first head network; The cloud device is also used for: Inputting the marker images corresponding to each initial inspection site into the first preset model, training the first preset model, and obtaining a trained first backbone network and a trained first head network when a first preset training condition is met; The parameters of the trained first backbone network are used as the parameters of the site recognition backbone network in the initial state of the site recognition model, and the parameters of the trained first head network are used as the parameters of the site recognition head network in the initial state of the site recognition model to obtain the initial state of the site recognition model.

5. The system according to claim 4, characterized in that The cloud device is also used for: Inputting the marker image corresponding to the new inspection site into the trained first backbone network to obtain the intermediate features of the new site output by the trained first backbone network; Sending the intermediate features of the new site and the intermediate features of the existing site to the inspection agent; wherein the intermediate features of the existing site include features obtained after the marker images corresponding to the existing inspection sites are input into the trained first backbone network; The inspection agent is also used for: The new site intermediate features and the existing site intermediate features are input into the site recognition head network of the site recognition model, the parameters of the site recognition head network of the site recognition model are updated, and the updating is stopped when the second preset training condition is met to obtain an updated site recognition model.

6. The system according to claim 2, characterized in that The multiple existing fault types include multiple initial fault types; the cloud device is also used to: perform pre-training according to the initial device fault data corresponding to the multiple initial fault types to obtain the initial state of the fault detection model; The initial equipment fault data includes equipment images corresponding to each initial fault type of the electric equipment.

7. The system according to claim 6, characterized in that The fault detection model includes a fault detection backbone network and a fault detection head network; the cloud device is deployed with a second preset model; the second preset model includes a second backbone network and a second head network; The cloud device is also used for: Inputting the initial device fault data into the second preset model, training the second preset model, and obtaining a trained second backbone network and a trained second head network when a third preset training condition is met; The parameters of the trained second backbone network are used as the parameters of the fault detection backbone network in the initial state of the fault detection model, and the parameters of the trained second head network are used as the parameters of the fault detection head network in the initial state of the fault detection model to obtain the initial state of the fault detection model.

8. The system according to claim 7, characterized in that The cloud device is also used for: Inputting the new device fault data into the trained second backbone network to obtain the intermediate features of the new device fault output by the trained second backbone network; The intermediate features of new equipment failure and the intermediate features of existing equipment failure are sent to the inspection agent; wherein the intermediate features of existing equipment failure include features obtained after the corresponding equipment images when the fault type of the power equipment is each existing fault type are input into the trained second backbone network; The inspection agent is also used for: The new device fault intermediate features and the existing device fault intermediate features are input into the fault detection head network of the fault detection model, the parameters of the fault detection head network of the fault detection model are updated, and the updating is stopped when the fourth preset training condition is met to obtain an updated fault detection model.

9. The system according to claim 1, characterized in that The dynamic topology map includes a plurality of local topology maps and a plurality of activation nodes; the local topology map corresponds to a sub-area in the inspection area; and the activation node is used to connect different local topology maps.

10. A brain-inspired power inspection method based on incremental learning, characterized in that: An inspection agent applied to a brain-inspired power inspection system based on incremental learning as described in any one of claims 1 to 9; The method comprises: Obtaining an inspection task; the inspection task includes an inspection starting position and a target inspection position; Generate an inspection path according to the inspection starting position, the target inspection location and the dynamic topology map; the inspection path is a route from the inspection starting position to the target inspection location; Move according to the inspection path and obtain environmental data of the surrounding environment; Determining whether the target inspection site has been reached based on the environmental data and the site recognition model; When the target inspection site is reached, an equipment image of the power equipment at the target inspection site is acquired, and the fault condition of the power equipment is obtained based on the equipment image and a fault detection model.

Images