Fault scenario positioning method, apparatus, device, medium, and program product

Abstract

The application relates to the technical field of fault positioning, and provides a fault scene positioning method, device, equipment, medium and program product, the method comprising: obtaining alarm information and resource information of equipment in a target facility at a preset time; based on the alarm information and the resource information, generating a fault fusion feature map at the preset time, the map being used for representing the distribution of equipment state information on a topology structure; inputting the fault fusion feature map into a fault scene positioning model, performing feature extraction and matching processing on the map through a deep context relationship understanding network, and outputting a fault positioning result. The fault scene positioning method provided by the application fuses the alarm information and the resource information to generate a fault fusion feature map, and performs feature extraction and matching processing on the map through a deep context relationship understanding network, so that the model can perceive the distribution law of the equipment state and identify a fault feature area in a global topology perspective, and the adaptability and accuracy of fault positioning are significantly improved.

Description

Technical Field

[0001] This application relates to the field of fault location technology, and in particular to a fault scenario location method, device, equipment, medium and program product. Background Technology

[0002] In the operation and maintenance management of various large-scale facilities, equipment failure presents a huge challenge. Traditional manual fault matching and location methods are inefficient, and there is an urgent need for an intensive and intelligent automatic fault matching, location and repair method to improve the fault handling efficiency of front-line maintenance personnel.

[0003] Currently, existing fault location methods mainly rely on linear traversal analysis of structured text data such as alarms and resources, tree-based classification analysis, and knowledge graph matching analysis. However, the core deficiency of these methods lies in the fact that alarm and resource data exist in the form of structured text, lacking the ability to represent spatial topology, making it difficult for deep learning models to directly extract features and perform generalized matching on this text data. When intermediate layer devices fail to report valid alarms due to reasons such as incorrect alarm configuration, missing alarm generation, or communication anomalies, or when inaccurate resource management data leads to a lack of cross-layer resource relationship information, existing methods struggle to effectively locate faults. Summary of the Invention

[0004] This application provides a fault scene localization method, apparatus, equipment, medium, and program product to solve the problem that the existing technology has insufficient generalization and matching adaptability, which makes it difficult to effectively locate faults.

[0005] Firstly, this application provides a fault scene localization method, including: Obtain alarm and resource information of equipment in the target facility at preset times; Based on the alarm information and the resource information, a fault fusion feature map is generated at the preset time; the fault fusion feature map is used to characterize the distribution of the status information of the equipment in the target facility on the topology. The fault fusion feature map is input into the fault scene localization model to obtain the fault localization result output by the fault scene localization model; the fault localization result is obtained by feature extraction and matching processing of the fault fusion feature map based on the deep contextual understanding network in the fault scene localization model.

[0006] In one embodiment, generating the fault fusion feature map at the preset time based on the alarm information and the resource information includes: Based on the resource information, according to different device levels and attributes, the position level of each device is defined on the base map and the corresponding device icon is drawn to generate an intermediate processing base map; Based on the alarm information, the devices that triggered the alarm are colored and marked on the intermediate processing base map to generate a fault fusion feature map at the preset time.

[0007] In one embodiment, the alarm information includes alarm level, alarm quantity, and alarm type; the step of coloring and marking the devices that triggered alarms on the intermediate processing base map based on the alarm information to generate the fault fusion feature map at the preset time includes: On the intermediate processing base map, the following operations are performed for the devices that trigger alarms to generate a fault fusion feature map at the preset time: Based on the alarm level and the number of alarms, draw node-shaped extension lines within a preset quadrant centered on the device icon; the node-shaped extension lines are used to identify the alarm level and the number of alarms. The devices that have alarmed are colored according to the color corresponding to the highest alarm level, and devices that have no alarms but have resource association with the devices that have alarms are colored with a preset identifier color. Based on the alarm type, rectangular icons corresponding to various alarm types are arranged in chronological order of occurrence, with the device icon as the center point.

[0008] In one embodiment, the fault scene localization model includes a deep context understanding network; the deep context understanding network includes a backbone, a neck section, and a detection end; The main part is used to extract contextual relationships from the fault fusion feature map based on a global self-attention mechanism and output contextual relationship features. The neck portion is used to enhance the spatial understanding of the contextual relationships and output enhanced contextual relationship features; The detection end is used to perform fault scene detection on the enhanced context relationship features and output fault location results; the fault location results include the target box information of the root cause device and the fault scene type.

[0009] In one embodiment, the backbone includes a convolution module, and a context extraction module and a deformable convolution module disposed before and after the convolution branches of the convolution module. The convolution module is used to perform convolution processing on the fault fusion feature map to obtain convolution features; The context relationship extraction module is used to perform global attention pooling and feature transformation on the convolutional features to obtain global self-attention features; The deformable convolution module is used to perform deformable convolution filtering, batch normalization, and nonlinear transformation on the global self-attention features to obtain the context relationship features.

[0010] In one embodiment, the neck portion includes a spatially hollow pyramidal pooling structure; The spatial void pyramid pooling structure is used to extract multi-scale features from the context relationship features using parallel dilated convolutions with multiple dilation sampling rates, resulting in multi-scale spatial features. After upsampling, the multi-scale spatial features are concatenated with the context relationship features output by the corresponding layer of the backbone to obtain the enhanced context relationship features.

[0011] In one embodiment, the fault scene localization model is trained through the following steps: Obtain the fault fusion feature map as a training sample, and obtain the fault feature template map corresponding to the training sample as the annotation result; The training samples are input into the pre-trained model to obtain the prediction results output by the pre-trained model; Based on the difference between the prediction result and the annotation result, the parameters of the pre-trained model are iteratively updated until the preset training termination condition is met.

[0012] In one embodiment, before inputting the training samples into the fault scene localization model to be trained, the method further includes: The coordinates of the left vertex are determined on the training sample. A rectangle is randomly cropped starting from the left vertex. The intersection of the rectangle and the training sample is obtained and a random grayscale transformation is performed to obtain the cropped image. The process involves iteratively determining the coordinates of the left vertex on the training sample, randomly cropping a rectangle starting from the left vertex, obtaining the intersection of the rectangle and the training sample, and performing a random grayscale transformation to obtain the cropped image, until a preset number of steps are reached. Multiple cropped images are merged to obtain training samples after image augmentation. The training samples after image augmentation are subjected to multi-scale transformation to obtain multi-scale transformed training samples.

[0013] In one embodiment, the fault scene localization method further includes: A fault fusion feature graph management structure is constructed with a graph database as its core. The fault fusion feature graph management structure is used to manage the graph model, graph components, graph event responses, and graph layout of the fault fusion feature graph and / or the fault feature template graph.

[0014] In one embodiment, after obtaining the fault location result, the method further includes: Based on the fault location results, fault handling operations are performed on the root cause device to obtain the fault handling results; Based on the fault location results and the fault handling results, a fault case record is generated; Based on the recorded failure cases, an effectiveness evaluation feedback result is generated; Based on the effectiveness evaluation feedback results, the fault scenario localization model is iteratively optimized.

[0015] Secondly, this application also provides a fault scene location device, comprising: The acquisition module is used to acquire alarm information and resource information of the equipment in the target facility at a preset time. The fault fusion feature map generation module is used to generate a fault fusion feature map at the preset time based on the alarm information and the resource information; the fault fusion feature map is used to characterize the distribution of the status information of the equipment in the target facility on the topology. The fault scene localization module is used to input the fault fusion feature map into the fault scene localization model to obtain the fault localization result output by the fault scene localization model; the fault localization result is obtained by feature extraction and matching processing of the fault fusion feature map based on the deep context relationship understanding network in the fault scene localization model.

[0016] Thirdly, this application provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of any of the fault scenario localization methods described above.

[0017] Fourthly, this application also provides a non-transitory computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the fault scenario localization methods described above.

[0018] Fifthly, this application also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by the processor, it implements the steps of any of the fault scenario localization methods described above.

[0019] The fault scene localization method, device, equipment, medium, and program products provided in this application generate a fault fusion feature map with a topological distribution by fusing alarm information and resource information. This transforms the original structured text data into an image representation rich in spatial location relationships and device hierarchical relationships. Then, the deep contextual understanding network is used to extract and match features from the fault fusion feature map, enabling the fault scene localization model to perceive the distribution pattern of device status and identify fault feature areas from a global topological perspective. Thus, even when alarms are missing or resource association information is incomplete, it can still effectively complete the generalized matching and localization of fault scenes, significantly improving the adaptability and accuracy of fault localization. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is one of the flowcharts of the fault scene localization method provided in this application.

[0022] Figure 2 This is a schematic diagram of the device matrix layout structure provided in this application.

[0023] Figure 3 This is a schematic diagram of the fault fusion feature map provided in this application.

[0024] Figure 4 This is a schematic diagram of the structure of the deep contextual understanding network provided in this application.

[0025] Figure 5 This is a schematic diagram of the improved C3-GC3 structure provided in this application.

[0026] Figure 6 This is a schematic diagram of the improved C3-GC3-DCN structure provided in this application.

[0027] Figure 7 This is a schematic diagram of the improved ASPP structure provided in this application.

[0028] Figure 8 This is a schematic diagram of the convergence result of the fault graph generalization matching training process provided in this application.

[0029] Figure 9 This is the second flowchart of the fault scenario localization method provided in this application.

[0030] Figure 10This is a structural schematic diagram of the fault scene location device provided in this application.

[0031] Figure 11 This is a schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] The terms "first," "second," etc., used in this application are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein.

[0034] The fault scene location method provided in this application embodiment can be implemented based on a fault scene location device. Therefore, this application embodiment uses a fault scene location device as the execution subject to describe the fault scene location method.

[0035] The following is combined Figures 1-11 This application describes the fault scenario localization method, apparatus, equipment, medium, and program products provided.

[0036] Figure 1 This is one of the flowcharts of the fault scene localization method provided in this application.

[0037] like Figure 1 As shown, the fault scene localization method includes the following steps: Step 101: Obtain alarm and resource information of the equipment in the target facility at a preset time; Step 102: Based on the alarm information and the resource information, generate a fault fusion feature map at the preset time; the fault fusion feature map is used to characterize the distribution of the status information of the equipment in the target facility on the topology. Step 103: Input the fault fusion feature map into the fault scene localization model to obtain the fault localization result output by the fault scene localization model; the fault localization result is obtained by feature extraction and matching processing of the fault fusion feature map based on the deep contextual understanding network in the fault scene localization model.

[0038] Specifically, the fault scenario localization method provided in this application is mainly applied to cloud computing scenarios, but it should not be construed as being limited to cloud computing scenarios. This technical concept is also applicable to scenarios such as data center operation and maintenance management and communication network fault diagnosis, which require automated localization and identification of equipment faults in complex systems based on multi-source heterogeneous state data. Therefore, the target facility referred to in this application embodiment can be a cloud computing resource pool, a data center, a communication network, or other infrastructure systems containing multi-level devices. The following uses a cloud resource pool in a cloud computing scenario as the target facility to describe the technical solution of this application in detail.

[0039] The system acquires alarm and resource information from the cloud resource pool at a preset time. The preset time can be the current time or a historical time, used to capture a snapshot of the system's operation at that time. The alarm information includes, but is not limited to, data that reflects abnormal device operating status, such as device alarms, performance anomalies, and log errors. The resource information includes structured information such as device hierarchy, type, topology connection relationship, and dependencies between devices.

[0040] After obtaining the alarm and resource information corresponding to the preset time, these two types of structured data are mapped into a two-dimensional image space according to a unified drawing rule to obtain a fault fusion feature map. The fault fusion feature map retains the original topology hierarchy and connection relationship, while embedding information such as alarm type, abnormality level, and resource occupancy status of the corresponding devices at each location as node attributes into the feature map, thus fully presenting the device status distribution of the entire cloud resource pool at that time.

[0041] By aggregating structured text data into unstructured machine vision images, the traditional fault analysis problem based on structured data can be transformed into an image-based fault pattern matching problem. Image-based analysis can more intuitively present the spread trend and associated range of faults, and it also facilitates the direct use of mature computer vision algorithms for subsequent feature extraction and scene matching, reducing the complexity of algorithm design. Furthermore, it preserves the spatial correlation information between multiple data sources, avoiding the problem of losing associated features during structured analysis.

[0042] Furthermore, the fault fusion feature map is input into the fault scene localization model for image matching analysis. The fault scene localization model includes a deep context understanding network, which can deeply perceive and understand the global context relationships in the fault fusion feature map. By extracting multi-scale and multi-level features of the device state distribution in the map and matching them with predefined fault feature patterns, it identifies the root cause device in the current fault fusion feature map and outputs the corresponding fault localization result. The fault localization result can specifically include the target bounding box information of the root cause device and the fault scene type, facilitating maintenance personnel to quickly locate the fault location and clarify the scope of the fault's impact.

[0043] By combining visual fault representation with deep contextual understanding, it is possible to generalize and match fault scenarios from a global perspective when alarm information is missing or resource relationships are incomplete.

[0044] The fault scene localization method provided in this application generates a fault fusion feature map with a topological distribution by fusing alarm information and resource information. This transforms the original structured text data into an image representation rich in spatial location relationships and device hierarchical relationships. Then, it uses a deep contextual understanding network to extract and match features from the fault fusion feature map. This enables the fault scene localization model to perceive the distribution pattern of device status and identify fault feature areas from a global topological perspective. As a result, it can still effectively complete the generalized matching and localization of fault scenes even when alarms are missing or resource association information is incomplete, significantly improving the adaptability and accuracy of fault localization.

[0045] In one embodiment, generating the fault fusion feature map at the preset time based on the alarm information and the resource information includes: Based on the resource information, according to different device levels and attributes, the position level of each device is defined on the base map and the corresponding device icon is drawn to generate an intermediate processing base map; Based on the alarm information, the devices that triggered the alarm are colored and marked on the intermediate processing base map to generate a fault fusion feature map at the preset time.

[0046] Specifically, based on the principle of signal fingerprint features, a general method for generating cloud resource pool base maps is proposed. The fault fusion feature map is formed by superimposing a general resource base map and a general alarm base map, which can characterize the resource, alarm and other conditions of various types of cloud resource pools.

[0047] Pre-defined rules for drawing fault fusion feature maps include general resource base map rules and general alarm base map rules. The general resource base map rules define the location hierarchy of each device on the base map according to different device levels and attributes in the cloud resource pool, and define corresponding device icons based on different device types to complete the drawing of general resource base maps at each level. The general alarm base map rules define the coloring and labeling methods for devices with alarms based on the general resource base map. Specifically, regarding alarm content, node-like extended lines are drawn according to signal fingerprint strength to identify alarm levels and quantities; regarding alarm devices, they are colored according to signal fingerprint status, using the color corresponding to the highest alarm level. Devices without alarms but with resource associations are identified with a color different from alarm devices. Alarm types are represented by various rectangular icons arranged by occurrence time on a coordinate axis centered on the device icon. These two drawing rules allow the alarm base map and resource base map to be merged into a general base map. Both the fault fusion feature map and the fault feature template map are presented using the same drawing rules as the general base map. A fault fusion feature map can be formed by drawing a panoramic view of all devices in a cloud resource pool at a certain time (both alarm and non-alarm devices are drawn); a fault feature template map can be formed by drawing alarm devices in a fault scenario (non-alarm devices are not drawn).

[0048] Based on the pre-defined drawing rules for the fault fusion feature map, when generating the fault fusion feature map at a preset time based on alarm information and resource information, firstly, according to the general resource base map drawing rules, the location hierarchy of each device is defined on the base map and the corresponding device icons are drawn to generate an intermediate processing base map; then, according to the general alarm base map drawing rules, the devices that have alarms are colored and marked on the base map to realize the superposition of resource information and alarm information, and finally obtain the fault fusion feature map at the preset time.

[0049] Figure 2 This is a schematic diagram of the device matrix layout structure provided in this application. Figure 2 (a) shows the location hierarchy of each device on the base map according to different device levels and attributes; Figure 2 (b) illustrates the drawing of corresponding device icons within the device layout. This structure diagram clearly defines the hierarchical position of various devices within the overall architecture and their icon mapping relationships.

[0050] Figure 3 This is a schematic diagram of the fault fusion feature map provided in this application. Figure 3Using the complete topology of the cloud resource pool as a base map, different coloring and labeling methods are used to visualize the fault status information of the devices. The icons of each device are arranged in preset positions on the base map according to their level and type. In this way, alarm information and topological relationships between devices are integrated into a single image.

[0051] This application embodiment integrates scattered multi-source alarm information and resource topology information into a unified and visualized fault fusion feature map. This not only intuitively presents the operating status of all equipment in the facility at a preset time and clearly reflects the location and hierarchical relationship between alarm equipment and associated resources, but also provides a standardized feature basis for subsequent fault scenario matching and positioning, thereby improving the accuracy and efficiency of automatic fault scenario matching and positioning.

[0052] In one embodiment, the alarm information includes alarm level, alarm quantity, and alarm type; the step of coloring and marking the devices that triggered alarms on the intermediate processing base map based on the alarm information to generate the fault fusion feature map at the preset time includes: On the intermediate processing base map, the following operations are performed for the devices that generate alarms to generate a fault fusion feature map at the preset time: Based on the alarm level and the number of alarms, draw node-shaped extension lines within a preset quadrant centered on the device icon; the node-shaped extension lines are used to identify the alarm level and the number of alarms. The devices that have alarmed are colored according to the color corresponding to the highest alarm level, and devices that have no alarms but have resource association with the devices that have alarms are colored with a preset identifier color; Based on the alarm type, rectangular icons corresponding to various alarm types are arranged in chronological order of occurrence, with the device icon as the center point.

[0053] Specifically, based on the base map rich in resource topology information, the devices that trigger alarms are further colored and labeled to form a dynamic fault fusion feature map.

[0054] Regarding alarm content, based on the alarm level and the number of alarms, nodal extension lines are drawn within a preset quadrant (quadrant 1 / 2 / 3 / 4) centered on the device icon, which can be used to identify the alarm level and the number of alarms.

[0055] Regarding alarm devices, devices that have triggered alarms are colored according to the color corresponding to the highest alarm level. Devices without alarms but with resource relationships to these alarm devices are colored differently (e.g., light blue) to visually distinguish potentially affected devices with direct relationships. Furthermore, alarm types are represented by various rectangular icons, arranged chronologically by the X and Y axes centered on a fixed device icon.

[0056] This application embodiment accurately maps multi-dimensional alarm information such as alarm level, alarm quantity, and alarm type to the base map according to a unified drawing rule, so that the fault fusion feature map can not only present the topological layout relationship of the devices, but also intuitively represent the fault status information of each device.

[0057] In one embodiment, the fault scene localization method further includes: A fault fusion feature graph management structure is constructed with a graph database as its core. The fault fusion feature graph management structure is used to manage the graph model, graph components, graph event responses, and graph layout of the fault fusion feature graph and / or the fault feature template graph.

[0058] Specifically, since cloud resource pools involve complex device relationships, and the data sources are scattered and the relationships are inconsistent, it is necessary to build a fault fusion feature map management structure with graph database as the core, so as to realize the unified structured organization and standardized management of fault resources and alarm feature information, and provide reliable underlying data support for the generation and reuse of fault fusion feature maps.

[0059] First, establish a set of fault resources and alarm characteristics. Based on hardware physical devices and software virtual network function (VNF) network element systems, and combined with fault alarm information, summarize the fault resources and alarm characteristic information of four types of technology stacks: computing, storage, network, and virtual layer, as well as each logical level of region, resource pool, availability zone, high availability group, device, and component.

[0060] Based on this, a fault fusion feature graph management system with the Nebula Graph database at its core is constructed. This system has feature graph control capabilities such as graph model management, graph component management, graph event response, graph layout management, graph image drawing, and graph image data management.

[0061] Among them, graph model management is used to define the type attributes and association model of each node in the fault fusion feature graph and / or fault feature template graph; graph component management is used to manage the componentized representation of various device nodes and alarm nodes in the graph; graph event response is used to respond to the event processing logic triggered by the change of node status in the graph; graph layout management is used to control the spatial arrangement and presentation effect of nodes in the graph.

[0062] The fault fusion feature map management system provides unified map management for fault fusion feature maps and / or fault feature template maps. Specifically, the map connection model defines the topological connection relationship between each device node, the map connection component provides a visual configuration of the connection style, the map node association event response realizes the event linkage processing when the node state changes, the map image drawing completes the rendering output from map to image, and the map image data management provides the storage and retrieval capabilities of image data.

[0063] After constructing the fault fusion feature map management system, resource base maps of multiple nodes and resource pools in the target facility are rendered and generated according to the fault fusion feature drawing and calibration rules and map structure. These base maps cover multiple levels, including core network elements, cloud virtual layer, cloud intranet, servers, environmental monitoring, data center, and inter-cloud network. Then, according to the preset alarm labeling rules, various alarm information is accurately mapped to the corresponding base map in real time. Finally, a fault snapshot image is generated for each resource pool at fixed time intervals (e.g., 5 minutes), forming a fault map base dataset ImageNetCloud with fault fusion feature calibration rules. Data annotation tools are then used to assist in labeling the fault feature template map.

[0064] This application embodiment constructs a structured fault fusion feature map management system, providing standardized data support for the generation, management, and reuse of fault fusion feature maps.

[0065] In one embodiment, the fault scene localization model is trained through the following steps: Obtain the fault fusion feature map as a training sample, and obtain the fault feature template map corresponding to the training sample as the annotation result; The training samples are input into the pre-trained model to obtain the prediction results output by the pre-trained model; Based on the difference between the prediction result and the annotation result, the parameters of the pre-trained model are iteratively updated until the preset training termination condition is met.

[0066] Specifically, fault fusion feature maps are obtained as training samples, such as those from ImageNetCloud mentioned above, and fault feature template maps corresponding to the training samples are obtained as annotation results. The annotation content includes fault scene classification labels and target box information of the fault root cause device.

[0067] A pre-trained model based on a deep contextual understanding network is constructed. The deep contextual understanding network adopts a cross-stage network architecture (CSPNet, Cross Stage Partial Network), which consists of an input end, a backbone, a neck part, and a detection end.

[0068] The training samples are fed into the input end, and the backbone extracts contextual features from the input fault fusion feature map. Then, the neck part enhances the extracted features with multi-scale spatial understanding. Finally, the detection end outputs the prediction results containing the fault scene category and the target box of the fault root cause device based on the enhanced features.

[0069] The difference between the predicted and labeled results is calculated using a loss function. The traditional target box loss function, Loss, uses the Intersection-over-Union (IoU) ratio to calculate the area of ​​the predicted box A and the target box B at four points. This embodiment uses a novel α-EIoU (α-Efficient IoU) method, adding a distance loss (for the predicted box). Target box Center point distance ratio), height-width loss (predicted box) / Prediction box / The aspect ratio is calculated and introduced. Power operations enhance the convergence speed and robustness of the loss function.

[0070] Among them, the prediction box Target box ; The coefficients for exponentiation; is the Euclidean distance between the center point of the predicted box and the center point of the target box, and c is the diagonal distance of the smallest box that encloses the predicted box A and the target box B. To predict the difference between the bounding box height and the target bounding box height, The height of the minimum enclosed rectangle; To predict the difference between the bounding box width and the target bounding box width, The width of the minimum enclosed box.

[0071] Based on the difference between the predicted and labeled results, the parameters of the pre-trained model are updated. In the early stages of training, a learning rate with a small update step size is selected to reduce the impact of noise and gradient oscillations. A higher learning rate is restored in the later stages of training to promote stable and efficient convergence of the model. A custom decay factor is introduced during training, and the weights of the new model parameters are updated by weighting the current weights with the previous historical weights and then sliding the update. This mitigates the impact of short-term data fluctuations and reduces the risk of overfitting. These steps are iteratively executed until the preset training termination condition is met, resulting in a successfully trained fault scene localization model.

[0072] The embodiments of this application, through the above training method, enable the fault scene localization model to have the ability to generalize and match fault scenes from the fault fusion feature map and locate the root cause device of the fault, thereby improving the generalization ability and robustness of fault scene localization.

[0073] In one embodiment, before inputting the training samples into the fault scene localization model to be trained, the method further includes: The coordinates of the left vertex are determined on the training sample. A rectangle is randomly cropped starting from the left vertex. The intersection of the rectangle and the training sample is obtained and a random grayscale transformation is performed to obtain the cropped image. The process involves iteratively determining the coordinates of the left vertex on the training sample, randomly cropping a rectangle starting from the left vertex, obtaining the intersection of the rectangle and the training sample, and performing a random grayscale transformation to obtain the cropped image, until a preset number of steps are reached. Multiple cropped images are merged to obtain training samples after image augmentation. The training samples after image augmentation are subjected to multi-scale transformation to obtain multi-scale transformed training samples.

[0074] Specifically, in practical applications of images, there are problems such as insufficient diversity and uneven sample distribution. In addition to conventional image preprocessing methods such as binarization and normalization, this embodiment uses image data augmentation and adaptive anchor box calculation for the dataset to stitch multiple images into a new hybrid image containing multiple categories and scales of fault identification targets, enabling the model to learn to process complex multiple fault features simultaneously during training.

[0075] Find the coordinates of the left vertex on the original training sample, randomly crop a rectangle starting from the left vertex, take the intersection of the rectangle and the original image, and perform a random grayscale transformation to obtain the first cropped image; use the length and width parameters of the randomly obtained cropped rectangle to crop the left vertex a preset number of times (e.g., 4 times), and finally merge all (e.g., 2×2 images) of the rectangle cropping results into a new fault image augmentation result, which is used as the training sample after image augmentation processing.

[0076] Then, according to the preset standard size (e.g., 640×640×3) required by the input end of the fault scene localization model, the fault map augmentation result is subjected to multi-scale image transformation by, for example, 0.5 to 1.5 times, and the black borders appearing in the scaled image are automatically filled with grayscale to obtain the training samples after multi-scale transformation.

[0077] The embodiments of this application effectively improve the diversity of training samples and the model's adaptability to fault features at different scales through image augmentation preprocessing and multi-scale transformation, thereby enhancing the generalization performance and robustness of the fault scene localization model.

[0078] In one embodiment, the fault scene localization model includes a deep context understanding network; the deep context understanding network includes a backbone, a neck region, and a detection end; The main part is used to extract contextual relationships from the fault fusion feature map based on a global self-attention mechanism and output contextual relationship features. The neck portion is used to enhance the spatial understanding of the contextual relationships and output enhanced contextual relationship features; The detection end is used to perform fault scene detection on the enhanced context relationship features and output fault location results; the fault location results include the target box information of the root cause device and the fault scene type.

[0079] Specifically, a deep contextual understanding network consists of an input terminal, a backbone, a neck region, and a detection terminal. For example... Figure 4 As shown, Figure 4 This is a schematic diagram of the structure of the deep contextual understanding network provided in this application.

[0080] In the main body, based on the C3 (ConvBNSiLU×3) neural network of the traditional object detection model, and combined with the principle of global self-attention, a new context relationship extraction module (GlobalContext Block, GC) with global self-attention mechanism is added to extract the context relationship of the input fault fusion feature map and output the context relationship feature.

[0081] In the neck region, to improve the spatial understanding of the context of fault images and alleviate image distortion and insufficient spatial perception caused by local feature extraction, the traditional Spatial Pyramid Pooling (SPP) structure is optimized. A Spatial Dilated Pyramid Pooling structure is adopted, utilizing dilated convolutions with different dilation rates to construct multi-scale receptive fields. Multi-scale feature fusion is performed on the contextual features output from the backbone, expanding the feature receptive range without increasing computational cost. This allows the network to capture global spatial information of fault regions of different sizes and output enhanced contextual features.

[0082] The enhanced contextual features output from the neck region are input to the detection end in the form of feature maps at various scales. The detection end includes multiple parallel detection branches, each corresponding to a feature map at a specific scale, and each feature map at a specific scale corresponds to a preset anchor box size. Each detection branch uses a 2D convolutional detection head to perform target box regression and fault scene classification on the input feature map, outputting predicted candidate boxes and their confidence scores at the corresponding scale. Then, non-maximum suppression is applied to all predicted candidate boxes output by multiple detection branches to filter out redundant candidate boxes with high overlap and low confidence, retaining the target box with the highest confidence as the final detection result, and outputting fault localization results containing target box information of the root cause device and fault scene type.

[0083] This application embodiment designs a deep contextual understanding network with a main structure consisting of a backbone, a neck section, and a detection end. This network can deeply perceive the contextual dependencies in the fault fusion feature map from a global perspective and enhance the spatial understanding of multi-scale fault feature regions. This enables accurate classification of fault scenarios and precise location of the root cause device.

[0084] In one embodiment, the backbone includes a convolution module, and a context extraction module and a deformable convolution module disposed before and after the convolution branches of the convolution module. The convolution module is used to perform convolution processing on the fault fusion feature map to obtain convolution features; The context relationship extraction module is used to perform global attention pooling and feature transformation on the convolutional features to obtain global self-attention features; The deformable convolution module is used to perform deformable convolution filtering, batch normalization, and nonlinear transformation on the global self-attention features to obtain the context relationship features.

[0085] Specifically, based on the C3 (ConvBNSiLU×3) neural network in the backbone of the traditional object detection model, and incorporating the principle of global self-attention, GC modules with a global self-attention mechanism are introduced before and after the three convolutional branches of ConvBNSiLU, forming a new C3-GC3 structure. For example... Figure 5 As shown, Figure 5 This is a schematic diagram of the improved C3-GC3 structure provided in this application.

[0086] The C3 convolutional module uses a ConvBNSiLU structure to perform convolution processing on the fault fusion feature map, namely, sequentially performing two-dimensional convolution, batch normalization, and SiLU nonlinear activation function processing to obtain convolutional features.

[0087] GC3's context relationship extraction module, based on global self-attention, performs global attention pooling and Transform feature transformation on convolutional features to obtain global self-attention features. The calculation formula is as follows: Regarding the first formula: Let w be the feature vector at the j-th spatial location in the input feature map, where w and h are the width and height of the input feature map, respectively, and c is the number of feature channels. This is the key linear transformation matrix, used to map input features to the key space. This represents a linear transformation of the features at the j-th spatial location to obtain the key representation at that location; For attention pooling weights, The total number of spatial locations in the feature map is given by the numerator, which is the attention score of the j-th location after exponential operation, and the denominator is the sum of the attention scores of all locations. After Softmax normalization, the global attention weight of the j-th location in the entire feature map is obtained, which is used to measure the contribution of the location to the global context features. Regarding the second formula: and It is a linear transformation matrix. Used to perform the first linear transformation on the attention-weighted features. This is used to perform a second linear transformation on the features after layer normalization and ReLU activation, together forming a two-layer multilayer perceptron transformation structure. Here, LN is the layer normalization operation used to stabilize the feature distribution during training; ReLU is the linear rectified activation function used to introduce non-linear transformation capability; this transformation function performs further feature mapping and enhancement on the globally attention-weighted features. Regarding the third formula: The final output is the global self-attention feature. This represents a weighted summation of features from all spatial locations based on attention weights, yielding a global context aggregated feature; this aggregated feature is then processed... LN, ReLU After transformation, compared with the original input features By performing residual connections, the output features retain both the local information of the original location and the contextual dependencies of the global scope.

[0088] Based on the improved C3-GC3 structure, deformable convolutional CNNs are further used to replace traditional standard convolutional CNNs. An offset is introduced for each sampling point, allowing the convolutional kernels to dynamically adjust according to the content of the target region as needed. This enhances the adaptability to geometric deformations such as shape and size in different fault scenarios. The ConvBNSiLU convolutional blocks can be replaced with De-ConvBNSiLU variable convolutional blocks, forming a globally variable self-attention extraction structure, C3-GC3-DCN. Figure 6 As shown, Figure 6 This is a schematic diagram of the improved C3-GC3-DCN structure provided in this application.

[0089] De-ConvBNSiLU variable convolutional block extraction results of global self-attention context relations De-Conv2d variable convolution filtering, Batch Normalization (BN), and SiLU nonlinear transformation enhancement are performed. The calculation formula is as follows: in, is the deformable convolution offset of the i-th sampling point; The offset weight coefficient for the i-th sampling point; is a linear transformation matrix for deformable convolution; The result is the calculation result of deformable two-dimensional convolution; The final output of the deformable convolutional block is calculated, where k is the kernel size, s is the stride, and i is the spatial location index. This output is the contextual relationship feature.

[0090] The backbone network based on C3-GC3-DCN fault graph context relationship global variable self-attention extraction has the specific model parameter configurations shown in Table 1 below.

[0091] Table 1. Parameters of the backbone network model based on C3-GC3-DCN

[0092] This application embodiment designs a C3-GC3-DCN globally variable self-attention extraction structure. The backbone can perceive the long-distance dependencies between device nodes in the fault fusion feature map globally. At the same time, the dynamic sampling capability of deformable convolution enhances the adaptability to geometric deformations in complex fault scenarios, thereby outputting contextual relationship features containing rich global contextual information. This provides a high-quality feature representation foundation for subsequent spatial understanding enhancement of the neck part and fault scene localization at the detection end.

[0093] In one embodiment, the neck portion includes a spatially hollow pyramidal pooling structure; The spatial void pyramid pooling structure is used to extract multi-scale features from the context relationship features using parallel dilated convolutions with multiple dilation sampling rates, resulting in multi-scale spatial features. After upsampling, the multi-scale spatial features are concatenated with the context relationship features output by the corresponding layer of the backbone to obtain the enhanced context relationship features.

[0094] Specifically, to improve the spatial understanding of contextual relationships in fault images and alleviate image distortion and insufficient spatial perception caused by local feature extraction, the traditional SPP spatial pyramid pooling structure is optimized. In the backbone, the contextual relationship features extracted by C3-GC3-DCN global self-attention are... The input is fed into a novel ASPP spatially dilated pyramid pooling structure, and multiple receptive field convolutions are constructed using parallel dilated convolutions with various dilation sampling rates r (taking three dilation sampling rates as an example) to enhance the recognition capability of feature regions in multi-scale fault maps.

[0095] In the neck region, ASPP+ upsampling can fuse multi-scale, high- and low-level features of the fault map together, and the ASPP+ upsampling results... and Feature information is then stitched together to enhance the spatial understanding of the fault map. For example... Figure 7 As shown, Figure 7 This is a schematic diagram of the improved ASPP structure provided in this application.

[0096] ASPP dilated convolutional blocks provide contextual features of the input. Receptive field capture was performed at three different hole sampling rates, r. The calculation formula is as follows: in, This is the output of the spatial void pyramid pooling structure at the i-th spatial position, i.e., the enhanced contextual relationship feature; q represents the total number of parallel dilated convolutions, and q is the index of the dilated convolution. The weight parameters are for the q-th parallel dilated convolution branch; The contextual relationship features output at the i-th spatial location of the main body; The sampling rate is the hole sampling rate. The product of the dilated sampling rate r and the dilated convolution index q is used to generate an increasing dilated sampling rate in different parallel branches, thereby constructing a multi-scale receptive field from local details to the global scope.

[0097] The parameter configurations for the neck network part of the ASPP-based model are shown in Table 2 below: Table 2. Neck network hole sampling rate design based on ASPP

[0098] This application utilizes a spatially hollow pyramid pooling structure to effectively expand the receptive field of the convolutional kernel without increasing additional parameters and computational load. This enables the deep contextual understanding network to simultaneously capture global spatial information of fault feature regions of different sizes, achieving multi-scale fusion of high and low layer features. This effectively alleviates the problems of image distortion and insufficient spatial perception caused by local feature extraction, and significantly improves the model's accuracy in identifying and locating the root cause devices in complex fault scenarios.

[0099] Based on the above model structure, under the same training method and environment, an ablation experiment was designed to classify fault scenarios according to fault feature template images and output the average detection accuracy of fault image generalization matching. Experimental results show that adding GC3 improves fault image feature acquisition capability by 1.3%, adding DCN adapts to deformable fault scene recognition by 0.4%, adding ASPP improves the accuracy of distorted images and small target recognition by 0.7%, and adding α-EIoU optimizes model parameter learning capability by 0.5%. Each module has a cumulative gain effect. The fault scenario classification and ablation experiment results are shown in Tables 3 and 4 below, and the fault image generalization matching results are as follows. Figure 8 As shown, Figure 8 This is a schematic diagram of the convergence result of the fault graph generalization matching training process provided in this application.

[0100] Table 3 Fault scenario classification based on fault graph generalization matching

[0101] Table 4 Comparison of ablation experiments using fault graph generalization matching

[0102] In one embodiment, after obtaining the fault location result, the method further includes: Based on the fault location results, fault handling operations are performed on the root cause device to obtain the fault handling results; Based on the fault location results and the fault handling results, a fault case record is generated; Based on the recorded failure cases, an effectiveness evaluation feedback result is generated; Based on the effectiveness evaluation feedback results, the fault scenario localization model is iteratively optimized.

[0103] Specifically, based on the fault location results, fault handling operations are performed on the root cause device to obtain the fault handling result. These operations may include secondary confirmation of the root cause, automatic fault isolation and switching, fault alarm dispatch masking, and automatic fault recovery confirmation. Secondary confirmation of the root cause is used to re-verify the status of the device after locating the root cause to confirm the fault attribution; automatic fault isolation and switching is used to automatically isolate the faulty device from the service link and switch service traffic to the backup device to minimize the impact of the fault on services; fault alarm dispatch masking is used to mask duplicate alarms generated by the faulty device and generate processing work orders to avoid alarm storms interfering with operational decisions; and automatic fault recovery confirmation is used to automatically check the device status after fault handling is completed to confirm whether the fault has been recovered.

[0104] Then, based on the fault location and fault handling results, a fault case record is generated. This case record includes fault scenario information, root cause device information, handling operation information, and handling effect information, which is used to construct a fault handling record set. Further, based on the fault case record, an effectiveness evaluation feedback result is generated to quantitatively assess the matching accuracy of the fault scenario location model and the effectiveness of the handling operations. Finally, the effectiveness evaluation feedback result is fed back into the fault scenario location model to iteratively optimize model parameters, fault graph matching algorithms, and constrained safety control actions for fault scenarios, thereby continuously improving the intelligence level of fault scenario location and automatic handling.

[0105] The embodiments of this application construct a closed-loop process from fault fusion feature map generation, fault scenario localization, automatic handling of fault root causes to case feedback and model iterative optimization, realizing the automation, intelligence and precision of fault handling.

[0106] In summary, combining all the above embodiments, Figure 9 , Figure 9 This is the second flowchart of the fault scenario localization method provided in this application. Figure 9 A specific method for fusion feature map matching based on cloud fault scenarios is provided: (1) Attach fault information such as device alarms, performance abnormalities, and log errors in the cloud resource pool to the cloud topology map, aggregate text structured data into unstructured machine vision images, form a set of fault fusion feature map drawing and calibration rules that combine signal fingerprint features, and use the fault fusion feature map spectrum management method to generate the fault map basic dataset ImageNetCloud. (2) Classify and label fault feature template diagrams according to fault scenarios, input a fault graph deep context relationship understanding algorithm with global self-attention mechanism, generalize matching to locate fault scenarios and fault root cause network element devices, automatically link and execute constrained safety control fault handling actions, and achieve cloud fault graph matching and automatic processing capabilities with higher performance, stronger functions and higher level of digital intelligence.

[0107] This enables the use of fault fusion feature map generalization matching algorithms to effectively improve the adaptability, accuracy, and coverage of fault matching in the cloud resource pool, even when resource and alarm correlations are lacking in the existing network. Based on fault fusion feature map generalization matching, it achieves the correlation and location of faults within the cloud specialty and across business / environmental sectors, effectively reducing work orders and fault location time, facilitating command and dispatch and fault handling, and improving the efficiency of frontline operations and maintenance. A general method for real-time assessment and dynamic updating of the impact of cloud fault events on intra-cloud and inter-cloud services facilitates rapid fault handling and effectively shortens the impact time on inter-cloud services. A general fault fusion feature map matching and automatic fault root cause handling capability cluster supports real-time fault diagnosis and business impact analysis under different fault scenarios, effectively improving horizontal scalability.

[0108] Based on the overall logical architecture, the automatic location and handling model, which is built on multi-dimensional data such as alarms, resources, and indicators, and consists of "fault fusion feature map snapshot generation - fault scenario map matching - automatic handling of faulty devices - fault recovery confirmation", can achieve rapid repair of fault events, intelligently manage cloud network faults from a fully automatic handling perspective, and improve fault handling efficiency.

[0109] Figure 10 This is a structural schematic diagram of the fault scene location device provided in this application.

[0110] like Figure 10 As shown, the fault scene location device includes: The acquisition module 1010 is used to acquire alarm information and resource information of the equipment in the target facility at a preset time. The fault fusion feature map generation module 1020 is used to generate a fault fusion feature map at the preset time based on the alarm information and the resource information; the fault fusion feature map is used to characterize the distribution of the status information of the equipment in the target facility on the topology. The fault scene localization module 1030 is used to input the fault fusion feature map into the fault scene localization model to obtain the fault localization result output by the fault scene localization model; the fault localization result is obtained by feature extraction and matching processing of the fault fusion feature map based on the deep context relationship understanding network in the fault scene localization model.

[0111] The fault scene localization device provided in this application generates a fault fusion feature map with a topological distribution by fusing alarm information and resource information. This transforms the original structured text data into an image representation rich in spatial location relationships and device hierarchical relationships. Then, it uses a deep contextual understanding network to extract and match features from the fault fusion feature map. This enables the fault scene localization model to perceive the distribution pattern of device status and identify fault feature areas from a global topological perspective. Thus, it can still effectively complete the generalized matching and localization of fault scenes even when alarms are missing or resource association information is incomplete, significantly improving the adaptability and accuracy of fault localization.

[0112] In one embodiment, generating the fault fusion feature map at the preset time based on the alarm information and the resource information includes: Based on the resource information, according to different device levels and attributes, the position level of each device is defined on the base map and the corresponding device icon is drawn to generate an intermediate processing base map; Based on the alarm information, the devices that triggered the alarm are colored and marked on the intermediate processing base map to generate a fault fusion feature map at the preset time.

[0113] In one embodiment, the alarm information includes alarm level, alarm quantity, and alarm type; the step of coloring and marking the devices that triggered alarms on the intermediate processing base map based on the alarm information to generate the fault fusion feature map at the preset time includes: On the intermediate processing base map, the following operations are performed for the devices that trigger alarms to generate a fault fusion feature map at the preset time: Based on the alarm level and the number of alarms, draw node-shaped extension lines within a preset quadrant centered on the device icon; the node-shaped extension lines are used to identify the alarm level and the number of alarms. The devices that have alarmed are colored according to the color corresponding to the highest alarm level, and devices that have no alarms but have resource association with the devices that have alarms are colored with a preset identifier color; Based on the alarm type, rectangular icons corresponding to various alarm types are arranged in chronological order of occurrence, with the device icon as the center point.

[0114] In one embodiment, the fault scene localization model includes a deep context understanding network; the deep context understanding network includes a backbone, a neck section, and a detection end; The main part is used to extract contextual relationships from the fault fusion feature map based on a global self-attention mechanism and output contextual relationship features. The neck portion is used to enhance the spatial understanding of the contextual relationships and output enhanced contextual relationship features; The detection end is used to perform fault scene detection on the enhanced context relationship features and output fault location results; the fault location results include the target box information of the root cause device and the fault scene type.

[0115] In one embodiment, the backbone includes a convolution module, and a context extraction module and a deformable convolution module disposed before and after the convolution branches of the convolution module. The convolution module is used to perform convolution processing on the fault fusion feature map to obtain convolution features; The context relationship extraction module is used to perform global attention pooling and feature transformation on the convolutional features to obtain global self-attention features; The deformable convolution module is used to perform deformable convolution filtering, batch normalization, and nonlinear transformation on the global self-attention features to obtain the context relationship features.

[0116] In one embodiment, the neck portion includes a spatially hollow pyramidal pooling structure; The spatial void pyramid pooling structure is used to extract multi-scale features from the context relationship features using parallel dilated convolutions with multiple dilation sampling rates, resulting in multi-scale spatial features. After upsampling, the multi-scale spatial features are concatenated with the context relationship features output by the corresponding layer of the backbone to obtain the enhanced context relationship features.

[0117] In one embodiment, the fault scene localization model is trained through the following steps: Obtain the fault fusion feature map as a training sample, and obtain the fault feature template map corresponding to the training sample as the annotation result; The training samples are input into the pre-trained model to obtain the prediction results output by the pre-trained model; Based on the difference between the prediction result and the annotation result, the parameters of the pre-trained model are iteratively updated until the preset training termination condition is met.

[0118] In one embodiment, before inputting the training samples into the fault scene localization model to be trained, the method further includes: The coordinates of the left vertex are determined on the training sample. A rectangle is randomly cropped starting from the left vertex. The intersection of the rectangle and the training sample is obtained and a random grayscale transformation is performed to obtain the cropped image. The process involves iteratively determining the coordinates of the left vertex on the training sample, randomly cropping a rectangle starting from the left vertex, obtaining the intersection of the rectangle and the training sample, and performing a random grayscale transformation to obtain the cropped image, until a preset number of steps are reached. Multiple cropped images are merged to obtain training samples after image augmentation. The training samples after image augmentation are subjected to multi-scale transformation to obtain multi-scale transformed training samples.

[0119] In one embodiment, the fault scene localization method further includes: A fault fusion feature graph management structure is constructed with a graph database as its core. The fault fusion feature graph management structure is used to manage the graph model, graph components, graph event responses, and graph layout of the fault fusion feature graph and / or the fault feature template graph.

[0120] In one embodiment, after obtaining the fault location result, the method further includes: Based on the fault location results, fault handling operations are performed on the root cause device to obtain the fault handling results; Based on the fault location results and the fault handling results, a fault case record is generated; Based on the recorded failure cases, an effectiveness evaluation feedback result is generated; Based on the effectiveness evaluation feedback results, the fault scenario localization model is iteratively optimized.

[0121] It should be noted that the fault scene location device provided in this application can execute the fault scene location method described in any of the above embodiments during actual operation, which will not be elaborated in this embodiment.

[0122] Figure 11 This is a schematic diagram of the structure of the electronic device provided in this application, such as... Figure 11As shown, the electronic device may include: a processor 1110, a communications interface 1120, a memory 1130, and a communication bus 1140, wherein the processor 1110, the communications interface 1120, and the memory 1130 communicate with each other through the communication bus 1140. The processor 1110 can call logical instructions in the memory 1130 to execute a fault scenario localization method, which includes: acquiring alarm information and resource information of equipment in a target facility at a preset time; generating a fault fusion feature map at the preset time based on the alarm information and the resource information; the fault fusion feature map is used to characterize the distribution of the state information of the equipment in the target facility on the topology; inputting the fault fusion feature map into a fault scenario localization model to obtain a fault localization result output by the fault scenario localization model; the fault localization result is obtained by feature extraction and matching processing of the fault fusion feature map based on a deep contextual understanding network in the fault scenario localization model.

[0123] Furthermore, the logical instructions in the aforementioned memory 1130 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0124] On the other hand, this application also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer can execute the fault scene localization method provided in the above embodiments. The method includes: acquiring alarm information and resource information of equipment in a target facility at a preset time; generating a fault fusion feature map at the preset time based on the alarm information and the resource information; the fault fusion feature map is used to characterize the distribution of the state information of the equipment in the target facility on the topology; inputting the fault fusion feature map into a fault scene localization model to obtain a fault localization result output by the fault scene localization model; the fault localization result is obtained by feature extraction and matching processing of the fault fusion feature map based on a deep contextual understanding network in the fault scene localization model.

[0125] In another aspect, this application also provides a non-transitory computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program is implemented to perform the fault scene localization method provided in the above embodiments. The method includes: acquiring alarm information and resource information of equipment in a target facility at a preset time; generating a fault fusion feature map at the preset time based on the alarm information and the resource information; the fault fusion feature map being used to characterize the distribution of the state information of the equipment in the target facility on the topology; inputting the fault fusion feature map into a fault scene localization model to obtain a fault localization result output by the fault scene localization model; the fault localization result being obtained by feature extraction and matching processing of the fault fusion feature map based on a deep contextual understanding network in the fault scene localization model.

[0126] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0127] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A fault scene localization method, characterized in that, The fault scenario localization method includes: Obtain alarm and resource information of equipment in the target facility at preset times; Based on the alarm information and the resource information, a fault fusion feature map is generated at the preset time; the fault fusion feature map is used to characterize the distribution of the status information of the equipment in the target facility on the topology. The fault fusion feature map is input into the fault scene localization model to obtain the fault localization result output by the fault scene localization model; the fault localization result is obtained by feature extraction and matching processing of the fault fusion feature map based on the deep contextual understanding network in the fault scene localization model.

2. The fault scene localization method according to claim 1, characterized in that, The step of generating the fault fusion feature map at the preset time based on the alarm information and the resource information includes: Based on the resource information, according to different device levels and attributes, the position level of each device is defined on the base map and the corresponding device icon is drawn to generate an intermediate processing base map; Based on the alarm information, the devices that triggered the alarm are colored and marked on the intermediate processing base map to generate a fault fusion feature map at the preset time.

3. The fault scene localization method according to claim 2, characterized in that, The alarm information includes alarm level, alarm quantity, and alarm type; the step of coloring and marking the devices that triggered alarms on the intermediate processing base map based on the alarm information, and generating the fault fusion feature map at the preset time, includes: On the intermediate processing base map, the following operations are performed for the devices that trigger alarms to generate a fault fusion feature map at the preset time: Based on the alarm level and the number of alarms, draw node-shaped extension lines within a preset quadrant centered on the device icon; the node-shaped extension lines are used to identify the alarm level and the number of alarms. The devices that have alarmed are colored according to the color corresponding to the highest alarm level, and devices that have no alarms but have resource association with the devices that have alarms are colored with a preset identifier color. Based on the alarm type, rectangular icons corresponding to various alarm types are arranged in chronological order of occurrence, with the device icon as the center point.

4. The fault scene localization method according to claim 1, characterized in that, The fault scene localization model includes a deep context understanding network; the deep context understanding network includes a backbone, a neck section, and a detection end; The main part is used to extract contextual relationships from the fault fusion feature map based on a global self-attention mechanism and output contextual relationship features. The neck portion is used to enhance the spatial understanding of the contextual relationships and output enhanced contextual relationship features; The detection end is used to perform fault scene detection on the enhanced context relationship features and output fault location results; The fault location results include the target frame information of the root cause device and the fault scenario type.

5. The fault scene localization method according to claim 4, characterized in that, The main body includes a convolution module, as well as a context extraction module and a deformable convolution module set before and after the convolution branches of the convolution module. The convolution module is used to perform convolution processing on the fault fusion feature map to obtain convolution features; The context relationship extraction module is used to perform global attention pooling and feature transformation on the convolutional features to obtain global self-attention features; The deformable convolution module is used to perform deformable convolution filtering, batch normalization, and nonlinear transformation on the global self-attention features to obtain the context relationship features.

6. The fault scene localization method according to claim 4, characterized in that, The neck portion includes a spatially hollow pyramidal pooling structure; The spatial void pyramid pooling structure is used to extract multi-scale features from the context relationship features using parallel dilated convolutions with multiple dilation sampling rates, resulting in multi-scale spatial features. After upsampling, the multi-scale spatial features are concatenated with the context relationship features output by the corresponding layer of the backbone to obtain the enhanced context relationship features.

7. The fault scene localization method according to claim 1, characterized in that, The fault scenario localization model is trained through the following steps: Obtain the fault fusion feature map as a training sample, and obtain the fault feature template map corresponding to the training sample as the annotation result; The training samples are input into the pre-trained model to obtain the prediction results output by the pre-trained model; Based on the difference between the prediction result and the annotation result, the parameters of the pre-trained model are iteratively updated until the preset training termination condition is met.

8. The fault scene localization method according to claim 7, characterized in that, Before inputting the training samples into the fault scene localization model to be trained, the method further includes: The coordinates of the left vertex are determined on the training sample. A rectangle is randomly cropped starting from the left vertex. The intersection of the rectangle and the training sample is obtained and a random grayscale transformation is performed to obtain the cropped image. The process involves iteratively determining the coordinates of the left vertex on the training sample, randomly cropping a rectangle starting from the left vertex, obtaining the intersection of the rectangle and the training sample, and performing a random grayscale transformation to obtain the cropped image, until a preset number of steps are reached. Multiple cropped images are merged to obtain training samples after image augmentation. The training samples after image augmentation are subjected to multi-scale transformation to obtain multi-scale transformed training samples.

9. The fault scene localization method according to claim 7, characterized in that, The fault scenario localization method also includes: A fault fusion feature graph management structure is constructed with a graph database as its core. The fault fusion feature graph management structure is used to manage the graph model, graph components, graph event responses, and graph layout of the fault fusion feature graph and / or the fault feature template graph.

10. The fault scene localization method according to any one of claims 1 to 9, characterized in that, After obtaining the fault location result, the following is also included: Based on the fault location results, fault handling operations are performed on the root cause device to obtain the fault handling results; Based on the fault location results and the fault handling results, a fault case record is generated; Based on the recorded failure cases, an effectiveness evaluation feedback result is generated; Based on the effectiveness evaluation feedback results, the fault scenario localization model is iteratively optimized.

11. A fault scene location device, characterized in that, The fault scene location device includes: The acquisition module is used to acquire alarm information and resource information of the equipment in the target facility at a preset time. The fault fusion feature map generation module is used to generate a fault fusion feature map at the preset time based on the alarm information and the resource information; the fault fusion feature map is used to characterize the distribution of the status information of the equipment in the target facility on the topology. The fault scene localization module is used to input the fault fusion feature map into the fault scene localization model to obtain the fault localization result output by the fault scene localization model; the fault localization result is obtained by feature extraction and matching processing of the fault fusion feature map based on the deep context relationship understanding network in the fault scene localization model.

12. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the fault scene localization method as described in any one of claims 1 to 10.

13. A non-transitory computer-readable storage medium, wherein a computer program is stored on the non-transitory computer-readable storage medium, characterized in that, When the computer program is executed by the processor, it implements the steps of the fault scenario localization method as described in any one of claims 1 to 10.

14. A computer program product, the computer program product comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the fault scenario localization method as described in any one of claims 1 to 10.

Images