Disaster damage monitoring method and identification device for electric power facilities

Abstract

The invention relates to an electric power facility disaster damage monitoring method and an identification device, and belongs to the technical field of electric power facility disaster damage monitoring. The method comprises the steps of constructing a multi-dimensional feature portrait of a disaster damage identification device by collecting multi-source data, and dynamically scheduling and matching a pre-layering model for concurrent tasks based on the multi-dimensional feature portrait; performing decoupling analysis on the multi-dimensional feature portrait through the pre-layering model, and completing classification and positioning in parallel; data uploaded by all disaster damage identification devices are aggregated to train and generate a new model, a multi-dimensional feature weight is set based on a release target to calculate a comprehensive score, and a gray release target is screened. According to the method, dynamic scheduling of model resources is realized through the multi-dimensional feature portraits, and the utilization efficiency and response real-time performance of edge-side resources are improved; classification and positioning tasks are processed in parallel by adopting a decoupling architecture, so that the recognition precision is improved; a closed-loop self-evolution system is constructed by means of an intelligent gray release mechanism based on cloud edge cooperation, and the monitoring accuracy, predictability and continuous optimization capability are comprehensively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of disaster damage monitoring of power facilities, and particularly relates to a disaster damage monitoring method and identification device of power facilities. BACKGROUND

[0002] As a key asset related to the national economy and people's livelihood, power infrastructure is facing severe challenges from sudden natural disasters such as rainstorms, floods, earthquakes, landslides, and mudslides. These disasters can easily trigger a chain reaction from mechanical damage to electrical damage of power facilities, leading to large-scale power outages.

[0003] However, the existing disaster damage monitoring technology has obvious limitations: in the perception layer, most systems rely on a single type of mechanical or electrical quantity monitoring, and fail to effectively integrate multi-source heterogeneous data, resulting in unclear revelation of the disaster damage evolution mechanism of "mechanical-electrical" coupling, and incomplete and inaccurate disaster damage state perception; in the analysis and decision-making layer, traditional methods mostly use fixed models and static threshold criteria, which are difficult to adapt to the heterogeneous hardware resources of field-deployed terminals and dynamically changing disaster scenes, resulting in low utilization efficiency of computing resources, poor response real-time performance, and inability to provide early warning based on changing trends in the early stages of disasters; in the system evolution layer, the updating and verification process of algorithm models is rigid, lacking a risk-controllable, gradual intelligent release and closed-loop optimization mechanism in a real operating environment, which limits the overall self-adaptation and continuous improvement capabilities of the system. SUMMARY

[0004] To solve the above problems in the prior art, the application provides a disaster damage monitoring method and identification device of power facilities, which realizes dynamic scheduling of model resources through multi-dimensional feature profiling, improves the utilization efficiency of edge-side resources and response real-time performance, uses a decoupling architecture to parallelly process classification and positioning tasks, significantly improves the identification accuracy, integrates static threshold and dynamic trend analysis to realize early warning, and builds a closed-loop self-evolution system relying on the intelligent gray release mechanism of cloud-edge collaboration, to comprehensively improve the monitoring accuracy, predictability, and continuous optimization capabilities.

[0005] The object of the application can be achieved by the following technical solutions: The first aspect of the disclosure provides a disaster damage monitoring method of power facilities, comprising the following steps: S1, data perception and dynamic scheduling: constructing a multi-dimensional feature profile of the disaster damage identification device through multi-source data acquisition, and dynamically scheduling a matching pre-layered model for concurrent tasks based on the multi-dimensional feature profile; S2, decoupling identification and local early warning: decoupling analysis of the multi-dimensional feature profile is performed by the pre-layered model, and classification and positioning are completed in parallel, and local early warning and data uploading are triggered based on the analysis results according to the preset threshold; S3, optimization and intelligent gray release: train a new model by aggregating data uploaded by all disaster identification devices, calculate a comprehensive score based on the release target and set multi-dimensional feature weight, and screen gray release targets; The decoupling identification and local early warning includes the following steps: S21, intelligent identification: based on the dispatching instruction, load the selected pre-layered model, and perform decoupling analysis on the incoming sensor data to complete the classification and positioning tasks; S22, local early warning: based on the results of decoupling analysis, perform early warning according to the preset threshold rules, and encapsulate the identification results, key feature data and applied pre-layered model ID, and upload them through the optimal communication link.

[0006] Further, the data perception and dynamic scheduling includes the following steps: S11, multi-dimensional feature portrait construction: construct a multi-dimensional digital archive containing hardware performance, environmental risk and business criticality for each disaster identification device; S12, model pre-layering and dynamic scheduling: according to the multi-dimensional feature portrait, select pre-layered models for concurrent multiple identification tasks from the pre-processed pre-layered model library with different calculation depths, and generate dispatching instructions.

[0007] Further, the pre-layered model library includes pre-layered processing based on chip-level disaster identification devices: Select multiple split points for each disaster identification DNN model for pre-layering; Generate a model lineage from fast response to fine analysis, record the calculation amount, memory amount and output data size of each pre-layered model; Construct a pre-layered combination solution to prepare multiple calculation depth optional models for each AI task.

[0008] Further, the decoupling analysis includes the following steps: Standardize the original data of the multi-dimensional feature portrait, extract features from the standardized data based on the selected pre-layered model, and generate a set of shared feature maps containing different levels of information; Send the shared feature maps to two independent decoding branches for parallel processing, including: Classification branch: use a step-by-step semantic enhancement module to aggregate features at different semantic levels to determine the disaster type; Positioning branch: use a spatial information refinement module to combine high and low resolution features to accurately predict the location, shape or extent of the disaster.

[0009] Further, the early warning according to the preset threshold rules includes performing: The time series data of the monitoring index is constructed from the state quantity of the decoupling analysis result, the first order derivative and the second order derivative of each index are calculated, and a trend analysis model of a sliding time window is established; A double criterion is used for early warning level determination: Threshold criterion: Emergency threshold: monitoring value > 90% of the preset safety threshold; Attention threshold: monitoring value > 70% and ≤ 90% of the preset safety threshold; Acceleration rate criterion: Emergency acceleration rate: change acceleration > 80% of the acceleration threshold; Attention acceleration rate: change acceleration > 50% and ≤ 80% of the acceleration threshold.

[0010] Further, the early warning according to the preset threshold rule further includes performing: Early warning level decision: When the monitoring value enters the emergency threshold interval, and the change acceleration enters the emergency acceleration rate state, a red early warning is triggered immediately; When the monitoring value enters the emergency threshold interval, and the change acceleration is in the attention acceleration rate state, an orange early warning is triggered; When the monitoring value is in the attention threshold interval, and the change acceleration enters the emergency acceleration rate state, a yellow early warning is triggered; When the monitoring value is in the attention threshold interval, and the change acceleration is also in the attention acceleration rate state, a blue early warning is triggered.

[0011] Further, the optimization and intelligent gray release include the following steps: S31, data aggregation and model optimization: the cloud computing center aggregates the identification results and original data reported from all disaster damage identification devices, starts a distributed training task, optimizes specific features by combining pre-training and fine-tuning strategies, and introduces machine learning methods, executes the same pre-layer processing as the old model for the trained new model, and generates a series of sub-models with different calculation amounts and stores them in the pre-layer model library; S32, intelligent gray release decision: after the new model is verified, a release task is automatically created on the gray release platform, and a preset weight configuration scheme is called according to the release target; According to the current weight configuration scheme, the comprehensive scores of all disaster damage identification devices are calculated, and according to the preset release ratio, the gray target list is automatically selected from the comprehensive score ranking; S33, shunting and deployment: according to the gray target list, the container management application precisely and automatically deploys the new model algorithm container to the target disaster damage identification device through a secure channel; S34, double-track monitoring and closed-loop decision: run the new model in parallel with the old model, monitor the performance indicators of the two, and use statistical methods to determine whether the new model meets the expectations.

[0012] Further, the preset weight configuration scheme comprises: In the stability verification phase, the weights are set as hardware performance: 0.6, environmental risk: 0.2, and business criticality: 0.2; in the performance verification phase, the weights are set as hardware performance: 0.3, environmental risk: 0.5, and business criticality: 0.2; and in the effectiveness verification phase, the weights are set as hardware performance: 0.2, environmental risk: 0.4, and business criticality: 0.4.

[0013] The second aspect of the present disclosure provides a disaster damage identification device of a power facility, which executes a disaster damage monitoring method of a power facility as described above, and comprises a multi-dimensional feature perception module, an edge computing module, a decoupling identification analysis module, an in-situ early warning module, and a containerized operation and maintenance module. The multi-dimensional feature perception module is used for the collection, convergence and standardization of multi-source heterogeneous data, and provides energy management and communication link adaptation functions for the disaster damage identification device. The edge computing module is used for performing resource-aware model loading and computing adaptation, calling and instantiating a specified AI model in a local pre-hierarchical model library according to a scheduling instruction, and completing an inference task. The decoupling identification analysis module is used for performing feature extraction and parallel decoding based on a deep neural network, respectively completing classification and regression tasks on input data, and outputting type confidence and spatial positioning information of disaster damage.

[0014] As a preferred technical solution of the present disclosure, the in-situ early warning module is used for performing rule-based state evaluation and early warning decision on the identification result, and managing a northbound communication protocol stack to realize the encapsulation and uploading of alarm information and key data. The containerized operation and maintenance module is used for being responsible for the containerized life cycle management, resource isolation of upper-layer application software of the disaster damage identification device, and the receiving, deployment and A / B test verification of a new version of an algorithm model.

[0015] The present disclosure has the following beneficial effects: The application realizes dynamic and accurate scheduling and adaptation of model resources by constructing a multi-dimensional feature image of the disaster damage identification device, effectively improving the edge side computing efficiency and task real-time performance; then, the decoupling identification architecture is used to process classification and positioning tasks in parallel, which significantly enhances the disaster damage identification accuracy; and the static threshold and dynamic trend analysis are combined to realize the leap from lagging alarm to early warning, greatly extending the emergency response window; finally, through data aggregation and intelligent gray release mechanism, a closed-loop self-evolution system of cloud iterative optimization and edge risk controllability is formed, which comprehensively improves the accuracy, predictability and continuous optimization capability of the monitoring system. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to facilitate the understanding of those skilled in the art, the present application will be further described below with reference to the accompanying drawings.

[0017] Figure 1 A step schematic diagram of a disaster damage monitoring method of a power facility provided by an embodiment of the present application is shown in Figure 2 A step schematic diagram of data sensing and dynamic scheduling provided by an embodiment of the present application is shown in Figure 3 A step schematic diagram of decoupling identification and local early warning provided by an embodiment of the present application is shown in Figure 4 A step schematic diagram of optimization and intelligent gray release provided by an embodiment of the present application is shown in Figure 5 A structure schematic diagram of a disaster damage identification device of a power facility provided by an embodiment of the present application is shown in DETAILED DESCRIPTION

[0018] In order to further illustrate the technical means and effects adopted by the present application to achieve the predetermined invention purpose, the specific embodiments, structures, features and effects according to the present application are described in detail below with reference to the accompanying drawings and preferred embodiments.

[0019] The present embodiment provides a disaster damage monitoring method of a power facility, as shown in Figure 1 The method comprises the following steps: S1, data sensing and dynamic scheduling: constructing a multi-dimensional feature image of the disaster damage identification device by collecting multi-source data, and dynamically scheduling and matching a pre-layered model for concurrent tasks based on the multi-dimensional feature image, as shown in Figure 2 The method comprises the following steps: S11, multi-dimensional feature image construction: constructing a multi-dimensional digital archive containing hardware performance, environmental risk and business criticality for each disaster damage identification device.

[0020] It should be noted that the hardware performance dimension includes state data from the operating system and hardware monitoring unit of the device itself, such as CPU / GPU usage, memory occupation, battery power, chip temperature, etc.; the environmental risk dimension includes historical disaster data, real-time meteorological / geological warning level and heterogeneous data collected by MEMS sensors (three-axis acceleration, inclination, settlement, displacement, etc.); and the business criticality dimension is derived from the asset management system and dispatching system of the power grid company, and is pre-configured, relatively static metadata, such as load importance, network structure location and recovery difficulty, etc.

[0021] In this embodiment, the disaster loss identification device collects data of all sensors through RS485 bus (southbound communication) and Modbus protocol; and encapsulates multi-source heterogeneous data into a unified internal data format, realizing decoupling with the underlying hardware sensors. At the same time, the device automatically selects the optimal communication path according to the deployment scene, such as preferentially using RJ45 wired network, and secondarily using 4G wireless communication (supporting APN private network) or Beidou short message as backup. The power supply uses seamless switching: in the easy power supply scene (factory station, power distribution), wideband AC / DC alternating current power supply is used. In the no city power supply scene (power transmission tower), solar energy + iron phosphate lithium battery power supply scheme is used.

[0022] S12, model pre-layering and dynamic scheduling: according to the multi-dimensional feature portrait, the most suitable model combination is selected from the pre-layering model library with different calculation depths for concurrent multiple identification tasks (such as simultaneous earthquake and landslide analysis), and a scheduling instruction is generated.

[0023] Specifically, first, the state of the disaster loss identification device is obtained through the multi-dimensional feature portrait, and then the metadata of all selectable models in the cloud computing center pre-layering model library are queried, including calculation amount, memory occupation and identification accuracy; based on the current disaster loss identification device state (resource constraints and task priority), the most suitable model layering is selected for each concurrent task through a multi-objective optimization algorithm, such as selecting a deep model for fine identification when the computing power is sufficient, and selecting a shallow model to ensure rapid response when the resources are scarce; the scheduling instruction containing the specific model version and resource allocation is generated and sent to the edge computing module of the device; the edge computing module loads the specified model after receiving the instruction, and completes the dynamic adaptation of computing resources, thereby maximizing the resource utilization efficiency while ensuring real-time performance.

[0024] The pre-layering model library includes pre-layering processing based on the chip-level disaster loss identification device: A plurality of split points are selected for each disaster identification DNN model (earthquake identification, landslide identification, flood identification, etc.) for pre-layering; A model spectrum from rapid response to fine analysis is generated, recording the calculation amount, memory access amount and output data size of each pre-layering model; Construct a pre-hierarchical combined solution set to prepare optional models with multiple computational depths for each AI task.

[0025] It should be noted that pre-layering processing enables the system to dynamically balance computational efficiency and recognition accuracy based on real-time resource conditions, providing the optimal model configuration scheme for disaster loss monitoring tasks in different scenarios.

[0026] S2. Decoupled Identification and On-Site Early Warning: A pre-layered model is used to decouple and analyze multi-dimensional feature profiles, performing classification and localization in parallel. Based on the results, on-site early warnings and data uploads are triggered according to preset thresholds. Figure 3 As shown, it includes the following steps: S21. Intelligent identification: Based on scheduling instructions, load the selected pre-hierarchical model and perform decoupling analysis on the incoming sensor data to complete the classification and localization tasks.

[0027] The decoupling analysis includes the following steps: The raw data of the multidimensional feature profile is standardized, and features are extracted from the standardized data based on the selected pre-hierarchical model to generate a set of shared feature maps containing information at different levels. The shared feature maps are fed into two independent decoding branches for parallel processing, including: Classification branch: Adopting a structure such as the hierarchical semantic enhancement module (HSEM), it focuses on learning the high-level semantic information of the target by aggregating features from different semantic levels, judging the type of disaster damage, and outputting a classification confidence score to answer the question "what is it" (e.g., "This is a seismic wave, confidence score 97%"). The localization branch employs a structure similar to the Spatial Information Refinement Module (SIRM). Combining high- and low-resolution features, it focuses on the spatial details and texture information of the features to accurately predict the location, shape, or extent of damage. Within this branch, an IoU-aware prediction sub-branch is introduced, specifically to evaluate the overlap between the predicted bounding box and the actual location, thereby strengthening the correlation between classification results and localization quality. This branch ultimately outputs localization information, answering the questions "Where is it, and how severe is it?" (e.g., "Tower deviated 15 degrees east, settlement 3 cm").

[0028] It should be noted that classification tasks focus on high-level semantic features, while localization tasks require fine texture details. Coupled with one module, the two will interfere with each other. Therefore, they are decoupled and each is assigned to the most suitable network structure for processing.

[0029] S22. Local Early Warning: Based on the results of decoupling analysis, early warning is issued according to preset threshold rules, and the identification results, key feature data and the applied pre-layered model ID are encapsulated and uploaded through the optimal communication link.

[0030] Issue an alert based on preset threshold rules and execute the following: From the state variables of the decoupling analysis results, construct time series data of monitoring indicators (such as tilt angle, settlement rate, vibration frequency, etc.), calculate the first derivative (rate of change) and second derivative (acceleration of change) of each indicator, and establish a trend analysis model of the sliding time window (such as the most recent 10 sampling periods); A dual-criteria approach is used to determine the warning level: Threshold criterion: Emergency threshold: Monitored value > 90% of preset safety threshold; Note the threshold: The monitored value is greater than 70% of the preset safety threshold and less than or equal to 90%. Acceleration rate criterion: Emergency acceleration rate: Variable acceleration > 80% of the acceleration threshold; Note the acceleration rate: the change in acceleration is greater than 50% of the acceleration threshold and less than or equal to 80%; Early warning level decision: Red Alert (Highest Level): Triggered immediately when the monitored value enters the emergency threshold range and the change in acceleration enters the emergency acceleration rate state; Orange alert: Triggered when the monitored value enters the emergency threshold range, but the change in acceleration is only in the attention acceleration state; Yellow alert: Triggered when the monitored value is within the attention threshold range, but the change in acceleration enters the emergency acceleration state; Blue alert (lowest level): Triggered when the monitored value is within the attention threshold range and the change acceleration is also in the attention acceleration state.

[0031] In this embodiment, the on-site early warning decision-making mechanism has evolved from an alarm mode based on static threshold criteria to a dynamic trend early warning mode that integrates real-time change rate and acceleration analysis. It can not only issue alarms when monitored parameters exceed static safety thresholds, but also issue early warning signals before the absolute value of the parameters reaches the critical threshold by identifying a significant increase in the rate of change of the monitored parameters. This trend prediction-based early warning strategy significantly extends the response time window for emergency response and effectively suppresses the probability of emergency response failure caused by the sudden acceleration of the disaster damage evolution process.

[0032] S3. Optimization and Intelligent Gray-Scale Release: A new model is generated by aggregating data uploaded from all disaster damage identification devices. Based on the release target, multi-dimensional feature weights are set to calculate a comprehensive score and filter gray-scale targets, such as... Figure 4 As shown, it includes the following steps: S31. Data Aggregation and Model Optimization: The cloud computing center gathers the identification results and raw data reported from all disaster damage identification devices, starts a distributed training task, adopts a strategy that combines pre-training and fine-tuning, and introduces machine learning methods to optimize specific features. The trained new model is subjected to the same pre-layering process as the old model, generating a series of sub-models with different computational loads and storing them in the pre-layered model library.

[0033] S32. Intelligent Gray-Scale Release Decision: After the new model has passed basic verification, a release task is automatically created on the gray-scale release platform, and the preset weight configuration scheme is called according to the release goal (such as verifying stability).

[0034] The preset weight configuration scheme includes: During the stability verification phase, the weights were set as follows: hardware performance: 0.6, environmental risk: 0.2, and business criticality: 0.2. Devices with strong computing power, sufficient memory, and low load were prioritized, while high-risk disaster areas and business-critical equipment were avoided to keep the risks to a minimum.

[0035] During the performance verification phase, the weights were set as follows: hardware performance: 0.3, environmental risk: 0.5, and business criticality: 0.2. Devices in medium- to high-risk areas were gradually included to verify the performance of the new algorithm in real disaster scenarios. This included medium-critical equipment to test the algorithm's adaptability in important but not core scenarios.

[0036] During the effective verification phase, the weights were set as follows: hardware performance: 0.2, environmental risk: 0.4, and business criticality: 0.4. Ultimately, the algorithm was covered to verify the maximum value of the algorithm in core business scenarios and ensure that critical infrastructure received optimal protection.

[0037] Calculate the comprehensive score of all disaster damage identification devices based on the current weight configuration scheme, and automatically filter out the gray target list from the comprehensive score ranking according to the preset release ratio (e.g., 5%).

[0038] S33. Offloading and Deployment: Based on the grayscale target list, the container management application deploys the new model algorithm container to the target disaster assessment device accurately and automatically through a secure channel; S34. Dual-track monitoring and closed-loop decision-making: Run the new model and the old model in parallel, monitor and compare their performance indicators, and use statistical methods to determine whether the new model has achieved the expected results.

[0039] The performance metrics include business metrics (such as recognition accuracy, false alarm rate, and false negative rate) and system metrics (such as model inference time, CPU / memory usage, and energy consumption).

[0040] Automated decision-making based on confidence intervals: Success criteria: When a 95% confidence level is reached to confirm that the new version is significantly better than the old version in core metrics, the weighting strategy is automatically adjusted and the release scope is gradually expanded (50%→100%). Circuit breaker mechanism: When a key indicator deteriorates to the threshold (such as a false positive rate increase of >5%), the release will be automatically paused immediately, upgraded devices will be rolled back to the old version, and the root cause analysis will be recorded.

[0041] This embodiment also provides a disaster damage identification device for power facilities, such as... Figure 5 As shown, it includes a multi-dimensional feature perception module, an edge computing module, a decoupled identification and analysis module, an on-site early warning module, and a containerized operation and maintenance module.

[0042] The multi-dimensional feature perception module is used for the collection, aggregation and standardization of multi-source heterogeneous data, and provides energy management and communication link self-adaptation functions for the disaster damage identification device.

[0043] Specifically, the multi-dimensional feature sensing module periodically polls various connected MEMS sensors via a southbound RS485 bus, following the Modbus protocol, to acquire their state measurement data. Simultaneously, the power management unit within the module monitors the input power status and seamlessly switches between a wideband AC / DC power supply and a solar photovoltaic + lithium iron phosphate battery energy storage system according to a preset strategy. The communication unit dynamically selects the optimal path among multi-mode communication interfaces (RJ45, 4G, BeiDou) based on link status and strategies, and standardizes and encapsulates all collected heterogeneous data before transmitting it to the edge computing core.

[0044] The edge computing module is used to perform resource-aware model loading and computation adaptation, and to call and instantiate the specified AI model in the local pre-layered model library according to the scheduling instructions to complete the inference task.

[0045] Specifically, the edge computing module receives and parses scheduling instructions from the cloud, which include the identifier of the target model and resource allocation requirements. Subsequently, the module accesses the local pre-layered model library, loads the corresponding model files into memory, and allocates computing units (CPU / GPU cores) and memory resources to the model instances according to the instructions, completing the construction and initialization of the computing graph and preparing for subsequent data inference.

[0046] The decoupled identification and analysis module is used to perform feature extraction and parallel decoding based on deep neural networks, respectively complete the classification and regression tasks of the input data, and output the disaster type confidence and spatial location information.

[0047] Specifically, the decoupled identification and analysis module receives preprocessed and standardized sensor data streams and inputs them into an instantiated pre-layered model. The model's feedforward network first performs multi-layer convolutions and nonlinear transformations on the input data to generate a set of shared feature maps. These feature maps are then fed into two independent decoder branches: the classification branch outputs a probability distribution vector representing different damage types through global pooling and fully connected layers; the localization branch outputs a tensor containing the target bounding box coordinates or keypoint locations through deconvolution or upsampling operations, combined with spatial information preserved by skip connections.

[0048] The local early warning module is used to perform rule-based status assessment and early warning decisions on the identification results, and manage the northbound communication protocol stack to encapsulate and upload alarm information and key data.

[0049] Specifically, the on-site early warning module continuously receives structured results output by the decoupled identification and analysis module. The early warning decision unit performs sliding window analysis on the time-series data of the monitoring indicators based on a preset threshold rule base, calculates their first derivative (rate of change) and second derivative (acceleration of change), and integrates static thresholds and dynamic trend criteria to generate corresponding early warning level signals. Simultaneously, the communication protocol stack (e.g., supporting the IEC 104 protocol) encapsulates the early warning information, key data features, and metadata into application service data units of a specific protocol, which are then sent to the main station system via an established physical link session.

[0050] The containerized operation and maintenance module is responsible for the containerized lifecycle management and resource isolation of the upper-layer application software of the disaster loss identification device, as well as the reception, deployment and A / B testing of new version algorithm models.

[0051] Specifically, the containerized operations and maintenance module runs a container management engine. When it receives a deployment command and a new version model image from the canary release platform, the engine creates a new containerized isolated environment locally and loads the image. Subsequently, the engine starts the new container according to a policy, allowing it to run in parallel with the old version container (dual-track operation). The monitoring agent continuously collects performance metrics of the two container instances and reports the data. The decision-making unit compares the differences in metrics based on a preset statistical verification method, and ultimately performs version switching, scaling up production, or rollback operations based on the results.

[0052] This invention constructs a multi-dimensional feature profile of the disaster damage identification device, enabling dynamic and precise scheduling and adaptation of model resources, effectively improving edge computing efficiency and task real-time performance. Furthermore, it employs a decoupled identification architecture to process classification and location tasks in parallel, significantly enhancing disaster damage identification accuracy. By integrating static thresholds and dynamic trend analysis, it achieves a leap from delayed alarms to early warnings, greatly extending the emergency response window. Finally, through data aggregation and intelligent gray-scale release mechanisms, it forms a closed-loop self-evolving system with cloud-based iterative optimization and controllable edge risks, comprehensively improving the accuracy, predictability, and continuous optimization capabilities of the monitoring system.

[0053] This invention first constructs a multi-dimensional feature profile of the disaster damage identification device through multi-source data acquisition, and then dynamically schedules the optimal model combination in the pre-layered model library based on the profile; subsequently, decoupled identification is performed at the edge, and classification and location results are output in parallel; hierarchical early warning is triggered according to the dual-criteria trend model; finally, the on-site data is aggregated to train a new model, and grayscale targets are screened through a weighted scoring mechanism to achieve closed-loop optimization, forming a complete monitoring chain from data perception to system self-evolution.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for monitoring disaster damage to power facilities, characterized in that: Includes the following steps: Data perception and dynamic scheduling: By collecting multi-source data, a multi-dimensional feature profile of the disaster damage identification device is constructed, and a pre-layered model is matched for dynamic scheduling of concurrent tasks based on the multi-dimensional feature profile; Decoupled identification and on-site early warning: Decoupled analysis of multi-dimensional feature profiles is performed through a pre-layered model, and classification and localization are completed in parallel. On-site early warning and data upload are triggered based on the analysis results according to preset thresholds. Optimization and Intelligent Gray-Scale Release: A new model is generated by aggregating data uploaded by all disaster damage identification devices, and a comprehensive score is calculated based on the release target by setting multi-dimensional feature weights to select gray-scale release targets; The decoupling identification and on-site early warning include the following steps: Intelligent identification: Based on scheduling instructions, the selected pre-hierarchical model is loaded, and decoupling analysis is performed on the incoming sensor data to complete the classification and localization tasks; Local early warning: Based on the results of decoupling analysis, early warning is issued according to preset threshold rules, and the identification results, key feature data and the applied pre-layered model ID are encapsulated and uploaded through the optimal communication link.

2. The method for monitoring disaster damage to power facilities according to claim 1, characterized in that: The aforementioned data perception and dynamic scheduling. Includes the following steps: Multi-dimensional feature profile construction: Construct a multi-dimensional digital profile for each disaster damage assessment device, including hardware performance, environmental risks, and business criticality; Model pre-layering and dynamic scheduling: Based on multi-dimensional feature profiles, pre-layered models are selected from a pre-processed library of pre-layered models with different computational depths for multiple concurrent recognition tasks, and scheduling instructions are generated.

3. The method for monitoring disaster damage to power facilities according to claim 2, characterized in that: The pre-layered model library includes pre-layered processing based on chip-level disaster detection devices: For each disaster damage identification DNN model, multiple splitting points are selected for pre-stratification; Generate a model hierarchy from fast response to fine analysis, and record the computational cost, memory access cost, and output data size for each pre-hierarchical model; Construct a pre-hierarchical combined solution set to prepare optional models with multiple computational depths for each AI task.

4. The method for monitoring disaster damage to power facilities according to claim 1, characterized in that: The decoupling analysis includes the following steps: The raw data of the multidimensional feature profile is standardized, and features are extracted from the standardized data based on the selected pre-hierarchical model to generate a set of shared feature maps containing information at different levels. The shared feature maps are fed into two independent decoding branches for parallel processing, including: Classification branch: A hierarchical semantic enhancement module is used to determine the type of disaster damage by aggregating features from different semantic levels; Location Branch: Using a spatial information refinement module, combined with high and low resolution features, it accurately predicts the location, shape, or extent of disaster damage.

5. The method for monitoring disaster damage to power facilities according to claim 4, characterized in that: The method of issuing warnings based on preset threshold rules includes executing: Time series data of monitoring indicators are constructed from the state variables of the decoupling analysis results. The first and second derivatives of each indicator are calculated, and a trend analysis model with a sliding time window is established. A dual-criteria approach is used to determine the warning level: Threshold criterion: Emergency threshold: Monitored value > 90% of preset safety threshold; Note the threshold: The monitored value is greater than 70% of the preset safety threshold and less than or equal to 90%. Acceleration rate criterion: Emergency acceleration rate: Variable acceleration > 80% of the acceleration threshold; Note the acceleration rate: the change in acceleration is greater than 50% of the acceleration threshold and less than or equal to 80%.

6. The method for monitoring disaster damage to power facilities according to claim 5, characterized in that: The method of issuing warnings based on preset threshold rules also includes executing: Early warning level decision: When the monitored value enters the emergency threshold range and the change in acceleration enters the emergency acceleration state, a red alert is immediately triggered. An orange alert is triggered when the monitored value enters the emergency threshold range and the change in acceleration is in the state of attention acceleration. A yellow alert is triggered when the monitored value is within the attention threshold range and the change in acceleration enters an emergency acceleration state. A blue alert is triggered when the monitored value is within the attention threshold range and the acceleration change is also within the attention acceleration state.

7. The method for monitoring disaster damage to power facilities according to claim 1, characterized in that: The optimization and intelligent canary release include the following steps: Data aggregation and model optimization: The cloud computing center gathers the identification results and raw data reported from all disaster damage identification devices, starts a distributed training task, adopts a strategy that combines pre-training and fine-tuning, and introduces machine learning methods to optimize specific features. The trained new model is subjected to the same pre-layering process as the old model, generating a series of sub-models with different computational loads and storing them in the pre-layered model library. Intelligent canary release decision: After the new model has passed basic validation, a release task is automatically created on the canary release platform, and the preset weight configuration scheme is called according to the release target; Calculate the comprehensive score of all disaster damage identification devices based on the current weight configuration scheme, and automatically filter out the gray target list from the comprehensive score ranking according to the preset release ratio. Offloading and Deployment: Based on the grayscale target list, the container management application deploys the new model algorithm containers to the target disaster assessment device in a precise and automated manner through a secure channel; Dual-track monitoring and closed-loop decision-making: The new model and the old model are run in parallel, their performance indicators are monitored and compared, and statistical methods are used to determine whether the new model has achieved the expected results.

8. A method for monitoring disaster damage to power facilities according to claim 7, characterized in that: The preset weight configuration scheme includes: During the stability verification phase, the weights are set as follows: hardware performance: 0.6, environmental risk: 0.2, and business criticality: 0.

2. During the performance verification phase, the weights are set as follows: hardware performance: 0.3, environmental risk: 0.5, and business criticality: 0.

2. During the effectiveness verification phase, the weights are set as follows: hardware performance: 0.2, environmental risk: 0.4, and business criticality: 0.

4.

9. A disaster damage identification device for power facilities, executing a disaster damage monitoring method for power facilities as described in any one of claims 1-8, characterized in that: It includes a multi-dimensional feature perception module, an edge computing module, a decoupled identification and analysis module, an on-site early warning module, and a containerized operation and maintenance module; The multi-dimensional feature perception module is used for the collection, aggregation and standardization of multi-source heterogeneous data, and provides energy management and communication link self-adaptation functions for the disaster damage identification device. The edge computing module is used to perform resource-aware model loading and computation adaptation, and to call and instantiate the specified AI model in the local pre-layered model library according to the scheduling instructions to complete the inference task; The decoupled identification and analysis module is used to perform feature extraction and parallel decoding based on deep neural networks, respectively complete the classification and regression tasks of the input data, and output the disaster type confidence and spatial location information.

10. A disaster damage identification device for power facilities according to claim 9, characterized in that: The local early warning module is used to perform rule-based status assessment and early warning decision-making on the identification results, and manage the northbound communication protocol stack to realize the encapsulation and uploading of alarm information and key data. The containerized operation and maintenance module is responsible for the containerized lifecycle management and resource isolation of the upper-layer application software of the disaster loss identification device, as well as the reception, deployment and A / B testing of new version algorithm models.

Images