Substation hard strap state monitoring system based on visual identification

Abstract

The invention discloses a transformer substation hard pressing plate state monitoring system based on visual identification, belongs to the field of hard pressing plate data analysis and mining, and aims to solve the problems that in the prior art, the efficiency of a monitoring mode is low, and the identification precision cannot be guaranteed. Hard strap image information can be collected in real time through a sensing unit, the state and abnormity of the hard strap can be accurately judged by means of a visual recognition algorithm, the recognition effect is optimized by fusing environmental data, efficient capture of the state and abnormity of the hard strap is achieved, then a Prognomics engine constructs and trains hard strap digital twinborn bodies through a fractional order physical information neural network, and the hard strap digital twinborn bodies are obtained. According to the method, a neural differential equation solver is used for simulating a performance attenuation track under mechanical stress and electric stress, and a potential fault root cause is positioned through an abnormality traceability device based on maximum likelihood estimation, so that precise simulation of performance attenuation of the hard pressing plate and effective tracing of the fault root cause are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of hardboard data analysis and mining, and particularly relates to a substation hardboard state monitoring system based on visual recognition. BACKGROUND

[0002] As a key control component of the protection device of the power system, the state of the substation hardboard directly relates to the correct action of the relay protection system and plays a crucial role in the safe and stable operation of the power grid.

[0003] At present, the existing monitoring methods mostly rely on manual inspection or single visual recognition technology. The manual inspection is inefficient and labor-intensive, and is easily affected by factors such as personnel experience, environmental light, and stain coverage, resulting in misjudgment and missed judgment. The single visual recognition technology can only adapt to standard hardboards, and has poor adaptability to non-standard hardboards, position deviation, damage, foreign matter shielding and other complex scenes, and does not effectively fuse environmental temperature and humidity data, so that the recognition accuracy cannot be guaranteed. At the same time, the existing technology lacks monitoring means for the physical state of the hardboard mechanical deformation and micro-cracks, and cannot comprehensively capture potential equipment fault hazards.

[0004] Therefore, the present application provides a substation hardboard state monitoring system based on visual recognition. SUMMARY

[0005] The present application aims to provide a substation hardboard state monitoring system based on visual recognition, which solves the problem of low efficiency and unguaranteed recognition accuracy of the monitoring method in the background art.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a substation hardboard state monitoring system based on visual recognition, comprising: a perception unit, configured to: collect hardboard image information in real time, accurately judge the state and abnormalities thereof through a visual recognition algorithm, fuse environmental data to optimize the recognition effect and preprocess the uploaded data; a transmission unit, configured to: adopt multiple communication methods to realize reliable data transmission between the perception unit and the platform unit, and guarantee efficient, safe and stable data transmission through compression encryption and link backup; a platform unit, configured to: be responsible for data storage, state analysis, electronic management of the whole process of hardboard operation, alarm linkage and visual display, and support data sharing with the existing system; Meanwhile, a Prognostics engine is introduced, which comprises: a fractional order physical information neural network, configured to construct and train a digital twin of the hardboard; A neural differential equation solver for simulating the performance degradation trajectory of a hard pad under mechanical and electrical stress through a multi-physics simulation model of the digital twin; An anomaly tracer based on maximum likelihood estimation for locating the root cause of potential failures through an anomaly tracing algorithm; An application unit for: Providing hard pad state query, operation processing, and log tracing functions for users through multiple terminals, and intuitively displaying information to assist operation and decision-making.

[0007] Further, the perception unit includes an industrial camera, an image processing module, a state recognition module, and an auxiliary sensor module. The industrial camera has an IP65 level waterproof and dustproof shell, an embedded electromagnetic shielding layer, and is suitable for outdoor or indoor installation environments in substations. It integrates an RJ45 Ethernet interface, a 5G module, and an RS485 interface, supports wired and wireless dual backup communication, and ensures data transmission stability. The image processing module uses grayscale and image segmentation algorithms to improve image quality. The state recognition module introduces a template matching algorithm, pre-stores input and exit state templates for standard hard pads, and quickly matches and identifies them. Based on the YOLOv8 deep learning algorithm, it trains models by labeling different types of hard pad samples in substations to improve state classification accuracy. It identifies abnormal states such as hard pad position deviation, damage, and foreign object obstruction by comparing consecutive frame image changes, and triggers secondary alarms. The auxiliary sensor module integrates temperature and humidity sensors to fuse environmental data and visual recognition results. When the temperature and humidity exceed the threshold, it automatically adjusts the image acquisition parameters to compensate for environmental influences.

[0008] Further, the transmission unit includes a communication transmission module and a data optimization module. The communication transmission module selects a communication method. When the screen cabinet is centrally deployed indoors, it uses industrial Ethernet. When the screen cabinet is centrally deployed in remote outdoor areas, it uses 5G wireless communication and supports breakpoint resume. The data optimization module uses data compression algorithms to reduce image data volume, with a transmission delay of ≤500ms. Key state data, including hard pad state changes and alarm information, is transmitted using encryption, while non-key data is transmitted using plaintext, balancing security and transmission efficiency.

[0009] Further, the platform unit includes a data reception and storage module, a state analysis and judgment module, an electronic management module, an alarm module, and a visual display module. The data receiving and storing module receives the hard pad state data, image data and environment data uploaded by the sensing unit, adopts mixed storage of MySQL+MongoDB, stores structured data such as state information and operation log in MySQL, stores unstructured data such as images in MongoDB, supports data life cycle management, and automatically backs up regularly; The state analysis and judgment module analyzes the hard pad state data in real time, compares with the preset normal running state library, judges whether there is abnormality such as misfeeding, miswithdrawing and state mutation, and supports custom state rules; The electronic management module is used for grading operation permissions, setting three-level permissions of administrator, operator and viewer, only the operator can initiate hard pad operation, and the administrator is responsible for approval and permission configuration; The alarm module is divided into three alarm states, namely state abnormality alarm, operation abnormality alarm and equipment failure alarm, the alarm mode is selected from platform pop-up window, sound prompt and APP push, and the alarm levels are first-level alarm, second-level alarm and third-level alarm; The visual display module supports the panoramic electronic map of the transformer substation, the real-time state of the corresponding hard pad is viewed by clicking the screen cabinet icon, the historical state change curve of the single hard pad, the image comparison before and after operation, and the environment data trend chart are displayed, and data statistical analysis is also supported to generate visual reports.

[0010] Further, the application unit supports PC browser access, mobile terminal APP or small program access, and monitoring center large screen access.

[0011] Further, the micro strain gauge is introduced in the auxiliary sensor module, attached to the stress point of the hard pad, used for measuring the mechanical deformation, and the acoustic emission sensor is also introduced, used for capturing the micro crack propagation signal in the material.

[0012] Further, the platform unit also integrates an intelligent decision center, which comprises: The dynamic confidence fusion module is used for receiving the original recognition result and its confidence output by the state recognition module, the environment data of the auxiliary sensor module, and the predictive state trajectory output by the Prognostics engine, calculating the global state confidence which integrates real-time sensing and long-term prediction based on Bayesian fusion algorithm; The strategy generation and scheduling module is used for dynamically generating and issuing collaborative strategy instructions according to the global state confidence, the historical record of the alarm module and the operation and maintenance plan of the electronic management module; the collaborative strategy instruction comprises: issuing sensing enhancement strategy to the sensing unit, instructing it to adjust image acquisition frequency, switch recognition algorithm or start redundant sensor; A transmission optimization strategy is issued to the transmission unit, instructing it to dynamically switch communication links, adjust data compression ratios, or change encryption levels, etc. A human-computer interaction strategy is issued to the application unit, instructing it to change the dimension and manner of information presentation.

[0013] Further, the Bayesian fusion algorithm executed by the dynamic credibility fusion module is as follows:

[0014] E represents the image recognition result and environmental data provided by the perception unit, P represents the prognostic evidence provided by the Prognostics engine, is the output global state credibility, representing the probability of the state being S under the condition of integrating all evidence, are the conditional probabilities of perception and prediction evidence, respectively, which are learned through historical data, is the prior probability.

[0015] Further, the strategy generation and scheduling module is implemented through a reinforcement learning agent, whose state space is the set of the global state credibility, device health index, and network condition, and whose action space is the various cooperative strategy instructions that can be issued, and its reward function R is designed as:

[0016] wherein is a weight coefficient, and the agent autonomously learns the optimal cooperative strategy under different system states by maximizing the cumulative reward.

[0017] Further, the application unit responds to the human-computer interaction strategy issued by the platform, and its display mode is: When displaying the state of the hard press plate, the global state credibility thereof is rendered synchronously to intuitively prompt the reliability of the result in a visual manner; When receiving a predictive maintenance suggestion, instead of directly popping up an alarm, it is integrated into the operation and maintenance calendar, and a visual report containing the decision basis and recommended operation is generated for the user to review; A strategy simulator interface is provided, allowing the user to replay the system state at any historical moment and try to issue different strategy instructions to observe the virtual deduction results, for the purpose of evaluating the effectiveness of the strategy.

[0018] Compared with the prior art, the present application has the following advantages: The application provides a substation hard pressboard state monitoring system based on visual recognition. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The application provides a substation hard pressboard state monitoring system based on visual recognition. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the application.

[0021] To solve the technical problems of low monitoring efficiency and low recognition accuracy in the prior art, the application provides a substation hard pressboard state monitoring system based on visual recognition. Figure 1 As shown in the figure, the following preferred technical solutions are provided: A substation hard pressboard state monitoring system based on visual recognition comprises: A perception unit is configured to: collect hard pressboard image information in real time, accurately determine the state and abnormality of the hard pressboard by using a visual recognition algorithm, fuse environmental data to optimize the recognition effect, and preprocess uploaded data; A transmission unit is configured to: adopt multiple communication modes to realize reliable data transmission between the perception unit and the platform unit, and guarantee efficient, safe and stable data transmission by compression encryption and link backup; A platform unit is configured to: be responsible for data storage, state analysis, electronic management of the whole process of hard pressboard operation, alarm linkage and visual display, and support data sharing with an existing system; Meanwhile, a Prognostics engine is introduced, which comprises: a fractional order physical information neural network configured to construct and train a digital twin of the hard pressboard; and a neural differential equation solver for simulating the performance degradation trajectory of the hard pad under mechanical and electrical stress through the multi-physics simulation model of the digital twin; an abnormality tracer based on maximum likelihood estimation for locating the root cause of potential faults through an abnormality tracing algorithm; an application unit for: providing hard pad state query, operation processing, and log tracing functions for users through multiple terminals, and intuitively displaying information to assist operation and maintenance decisions.

[0022] The perception unit includes an industrial camera, an image processing module, a state recognition module, and an auxiliary sensor module. The industrial camera has an IP65 level waterproof and dustproof shell, an embedded electromagnetic shielding layer, and is suitable for outdoor or indoor installation environments in substations. It integrates an RJ45 Ethernet interface, a 5G module, and an RS485 interface, supports wired and wireless dual backup communication, and ensures data transmission stability. The image processing module uses grayscale and image segmentation algorithms to improve image quality. The state recognition module introduces a template matching algorithm, pre-stores input and exit state templates for standard hard pads, and quickly matches and identifies them. Based on the YOLOv8 deep learning algorithm, for non-standard hard pads, stained covers, and light changes, the model is trained by labeling different types of hard pad samples in substations, improving state classification accuracy. By comparing the changes in consecutive frames of images, the system identifies abnormal states such as pad position deviation, damage, and foreign object obstruction, triggering secondary alarms. The auxiliary sensor module integrates temperature and humidity sensors to fuse environmental data with visual recognition results. When the temperature and humidity exceed the threshold, the system automatically adjusts the image acquisition parameters to compensate for environmental influences.

[0023] The transmission unit includes a communication transmission module and a data optimization module. The communication transmission module selects a communication method. When the screen cabinet is centrally deployed in an indoor environment, it uses industrial Ethernet. When the screen cabinet is centrally deployed in a remote outdoor area, it uses 5G wireless communication and supports breakpoint resume. The data optimization module uses data compression algorithms to reduce image data volume, with a transmission delay of ≤500ms. Key state data, including hard pad state changes and alarm information, is transmitted using encryption, while non-critical data is transmitted using plaintext, balancing security and transmission efficiency.

[0024] The platform unit includes a data reception and storage module, a state analysis and judgment module, an electronic management module, an alarm module, and a visual display module. Data receiving and storage module, receiving the hard pressure plate state data, image data, environment data uploaded by the sensing unit, using MySQL+MongoDB hybrid storage, structured data such as state information, operation log is stored in MySQL, unstructured data such as image is stored in MongoDB, while supporting data life cycle management, regular automatic backup; State analysis and judgment module, real-time analysis of hard pressure plate state data, compared with the preset normal running state library, judge whether there is misplacement, misplacement, state mutation, in addition, support custom state rules; (such as specific protection device corresponding to the pressure plate combination logic, when the combination logic does not meet the alarm is triggered) Electronic management module, used for hierarchical operation permission, setting administrator, operator, viewer three-level permission, only the operator can initiate the hard pressure plate operation, the administrator is responsible for approval and permission configuration, specifically: first, the operator initiates the operation application through the platform, needs to fill in the operation reason and operation range, then the administrator online approval, then operation execution, after on-site operation, the system automatically compares the state before and after operation to confirm the operation effectiveness, finally generate electronic operation ticket, including operator, approver, operation time, state change record, support PDF export and print; Alarm module, divided into three kinds of alarm state, respectively state abnormal alarm, operation abnormal alarm and equipment fault alarm, alarm mode selected platform pop-up, sound prompt and APP push, alarm level is divided into first alarm (most serious), second alarm, third alarm (lightest); Visual display module, support substation panoramic electronic map, click on the screen cabinet icon to view the real-time state of the corresponding hard pressure plate, at the same time show the single hard pressure plate historical state change curve, operation before and after image comparison, environment data trend chart, in addition, also support data statistical analysis (monthly operation times, abnormal alarm times), generate visual report.

[0025] Application unit, support PC browser access, mobile APP or applet access, monitoring center large screen access.

[0026] Auxiliary sensor module introduces micro strain gauge, attached to the stress point of hard pressure plate, used for measuring its mechanical deformation, at the same time, also introduces acoustic emission sensor, used for capturing the micro crack propagation signal inside the material.

[0027] Application unit can show health index curve, residual useful life probability distribution diagram and fault mode confidence.

[0028] The performance degradation of hard pressure plate is a process with long memory effect and nonlinearity, fractional differential equation is used to describe the relationship between its state, stress and time, as follows: Define the health state of the system as a hidden variable wherein represents a brand new, represents a complete failure; The multi-physics simulation model of the digital twin is a dynamics model based on fractional differential equation, and the evolution of the health state is described by the following fractional neuro-differential equation: ; wherein, is the Caputo fractional differential operator, wherein , when , it is a conventional integer order differential, is a learnable parameter, which is used to capture the memory and genetic effect of the degradation process (i.e. the influence of the historical state on the current rate of change); is a physical information neural network, and its parameters are The inputs of the network include: is the current health state; is the real-time mechanical stress calculated from the micro-strain gauge measurement value; is the environmental temperature; is an external excitation vector, specifically the operation number cumulative amount and vibration energy; The digital twin is trained by a fractional physical information neural network, and the training process is achieved by minimizing the following loss function: A deep neural network is used to approximate the real health state , which takes time t as input and outputs the predicted value of the health state; A physical constraint loss function is introduced, specifically: ; ; ; wherein, is the data loss, is the health state observation value estimated based on the contact resistance conversion at time The health state observation value estimated based on the contact resistance conversion can be estimated by the following fusion formula: , is the estimated contact resistance at time t, is the initial value of the contact resistance and the defined failure threshold, is the energy release rate measured by the acoustic emission sensor at time t, is the failure threshold of the energy release rate, For the weighting coefficients, satisfying ; For physical loss, the neural network is forced to... The output must satisfy a fractional differential equation. It involves sampling a large number of points in the time domain; These are hyperparameters used to balance the weights of the two losses; The number of time points with health status observations; For the first Each observation time point; For neural networks in time Predictions of health status; Represents the time domain The internal configuration points sampled for calculating physical loss are typically far more numerous than the data points and are uniformly or randomly distributed in the time domain, without needing to be compared with... coincide; Output of the neural network In time Caputo fractional derivative at point; When an anomaly is detected, it is necessary to trace its source, which can be conceived as a hypothesis testing problem, as follows: Construct a set of failure hypotheses, assuming there are K failure modes, denoted as Kf. ; Calculate the likelihood function for each fault mode. Assuming it will affect the parameters in the state equation, the likelihood of this failure mode occurring can be calculated based on the current observation data D: ; in, State equations The parameter set in; Given parameters and failure modes, the probability of observing the current data; Fault mode The corresponding prior distribution of parameters; Using Bayes' theorem, calculate the posterior probability of each failure mode:

[0029] in, Fault mode The prior probability of occurrence can be obtained according to historical statistics, and the system finally outputs the fault mode with the highest posterior probability, for example: , Quantitative traceability is realized.

[0030] By introducing a fractional differential equation to characterize the long memory effect and genetic characteristics of the performance degradation of the hard press plate, the health state evolution trajectory of the device throughout its life cycle, especially at the end of its life, is predicted. Compared with the integer order model and the simple exponential model, the fractional differential equation can more accurately describe the degradation process with memory and genetic characteristics, and can reduce the prediction error.

[0031] The platform unit is also integrated with an intelligent decision hub, which includes: A dynamic credibility fusion module is configured to receive the original recognition result and its confidence level output by the state recognition module, the environmental data of the auxiliary sensor module, and the predictive state trajectory output by the Prognostics engine, and calculate the global state credibility that integrates real-time perception and long-term prediction based on a Bayesian fusion algorithm; A strategy generation and scheduling module is configured to dynamically generate and issue a collaborative strategy instruction according to the global state credibility, the historical record of the alarm module, and the operation and maintenance plan of the electronic management module; the collaborative strategy instruction includes: issuing a perception enhancement strategy to the perception unit, instructing it to adjust the image acquisition frequency, switch the recognition algorithm, or start the redundant sensor; issuing a transmission optimization strategy to the transmission unit, instructing it to dynamically switch the communication link, adjust the data compression ratio, or change the encryption level; issuing a human-computer interaction strategy to the application unit, instructing it to change the dimension and way of information presentation.

[0032] The Bayesian fusion algorithm executed by the dynamic credibility fusion module is as follows:

[0033] S represents the true state of the hard press plate, E represents the image recognition result and environmental data provided by the perception unit, and P represents the predictive evidence provided by the Prognostics engine, is the output global state credibility, and represents the probability of the state being S under the condition of integrating all evidence, are the conditional probabilities of perception and prediction evidence, respectively, which are learned from historical data, is the prior probability.

[0034] The policy generation and scheduling module is realized by a reinforcement learning agent, a state space of which is a set of the global state credibility, the device health index and the network condition, an action space of which is a variety of cooperative policy instructions that can be issued, and a reward function R of which is designed as:

[0035] wherein is a weight coefficient, the agent autonomously learns the optimal cooperative policy under different system states by maximizing the cumulative reward.

[0036] The application unit responds to the human-computer interaction policy issued by the platform, and the display mode is: When the state of the hard press plate is displayed, the global state credibility is rendered synchronously to intuitively prompt the reliability of the result in a visual manner; When the predictive maintenance suggestion is received, instead of directly popping up a window alarm, it is integrated into the operation and maintenance calendar, and a visual report containing the basis for decision-making and recommended operations is generated for the user to review; A policy simulator interface is provided to allow the user to replay the system state at any historical moment and try to issue different policy instructions to observe the virtual deduction results, so as to evaluate the effectiveness of the policy.

[0037] The visual recognition results are de-pretended and stored and weighted fused by the dynamic credibility fusion module and the Bayesian algorithm, and a quantitative and reliable global credibility is output.

[0038] Through the policy generation and scheduling module and the reinforcement learning agent, the system can intelligently schedule computing resources, communication resources and human resources according to the current state, such as the network condition, whether the result is reliable, and the overall utility of the system is maximized.

[0039] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any skilled person in the art can make equivalent replacement or change according to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A substation hard plate condition monitoring system based on visual recognition, characterized in that, include: Sensing unit, used for: Real-time acquisition of hard plate image information, accurate judgment of its status and abnormalities through visual recognition algorithm, integration of environmental data to optimize recognition effect and preprocessing of uploaded data; Transmission unit, used for: Multiple communication methods are used to achieve reliable data transmission between the sensing unit and the platform unit, and compression encryption and link backup are used to ensure efficient, secure and stable data transmission. Platform unit, used for: Responsible for data storage, status analysis, full-process electronic management of hard platen operation, alarm linkage and visualization, and also supports data sharing with existing systems; The Prognostics engine is also introduced, which includes: A fractional-order physical information neural network is used to construct and train a digital twin of a hardened pressure plate. A neural differential equation solver is used to simulate the performance degradation trajectory of a hardened platen under mechanical and electrical stress through the multiphysics simulation model of the digital twin. An anomaly source tracer based on maximum likelihood estimation is used to locate the root cause of potential failures through anomaly source tracing algorithms. Application unit, used for: By providing users with functions such as hard board status query, operation processing, and log tracing through multiple terminals, information can be displayed intuitively and operation and maintenance decisions can be assisted.

2. The substation hard plate status monitoring system based on visual recognition as described in claim 1, characterized in that: The sensing unit includes an industrial camera, an image processing module, a status recognition module, and an auxiliary sensor module; The industrial camera features an IP65-rated waterproof and dustproof housing with an internal electromagnetic shielding layer, making it suitable for outdoor or indoor installation in substations. It integrates an RJ45 Ethernet interface, a 5G module, and an RS485 interface, supporting wired and wireless dual backup communication to ensure stable data transmission. The image processing module employs grayscale and image segmentation algorithms to improve image quality; The status recognition module introduces a template matching algorithm. For standard model hard pressure plates, it pre-stores input and output status templates for fast matching and recognition. Based on the YOLOv8 deep learning algorithm, it trains the model by labeling samples of different types of hard pressure plates in substations to improve the accuracy of status classification. By comparing changes in continuous frame images, it identifies abnormal states such as pressure plate position displacement, damage, and foreign object obstruction, triggering a secondary alarm. The auxiliary sensor module integrates temperature and humidity sensors, fusing environmental data with visual recognition results. When the temperature and humidity exceed the threshold, it automatically adjusts the image acquisition parameters to compensate for environmental influences.

3. The substation hard plate status monitoring system based on visual recognition as described in claim 2, characterized in that: The transmission unit includes a communication transmission module and a data optimization module; The communication transmission module is used to select the communication method. When the centralized deployment scenario of the cabinet is indoors, industrial Ethernet is used, and when the centralized deployment scenario of the cabinet is in a remote outdoor area, 5G wireless communication is used, and it supports breakpoint resume. The data optimization module is used to reduce the amount of image data by employing data compression algorithms, with a transmission latency of ≤500ms. Critical status data, including hard plate status changes and alarm information, are transmitted in encrypted form, while non-critical data is transmitted in plaintext, balancing security and transmission efficiency.

4. The substation hard plate condition monitoring system based on visual recognition as described in claim 3, characterized in that: The platform unit includes a data receiving and storage module, a status analysis and judgment module, an electronic management module, an alarm module, and a visualization display module; The data receiving and storage module receives hard plate status data, image data, and environmental data uploaded by the sensing unit. It adopts a hybrid storage of MySQL and MongoDB. Structured data such as status information and operation logs are stored in MySQL, while unstructured data such as images are stored in MongoDB. It also supports data lifecycle management and regular automatic backups. The status analysis and judgment module parses the hard plate status data in real time, compares it with the preset normal operation status database, and judges whether there are any abnormalities such as accidental activation, accidental deactivation, or sudden status changes. In addition, it supports custom status rules. The electronic management module is used to classify operation permissions, setting three levels of permissions: administrator, operator, and viewer. Only operators can initiate hard-panel operations, while administrators are responsible for approval and permission configuration. The alarm module is divided into three alarm states: abnormal status alarm, abnormal operation alarm, and equipment failure alarm. The alarm methods include platform pop-up, sound reminder and APP push. The alarm levels are divided into level 1 alarm, level 2 alarm and level 3 alarm. The visualization module supports a panoramic electronic map of the substation. Clicking on the cabinet icon allows you to view the real-time status of the corresponding hard plate. It also displays the historical status change curve of a single hard plate, image comparison before and after operation, and environmental data trend chart. In addition, it supports data statistical analysis and generates visualization reports.

5. The substation hard plate condition monitoring system based on visual recognition as described in claim 4, characterized in that: The application unit supports access via PC browser, mobile APP or mini-program, and monitoring center large screen.

6. The substation hard plate condition monitoring system based on visual recognition as described in claim 5, characterized in that: The auxiliary sensor module incorporates micro-strain gauges, which are attached to the stress points of the hard plate to measure its mechanical deformation. It also incorporates acoustic emission sensors to capture signals of microcrack propagation within the material.

7. The substation hard plate status monitoring system based on visual recognition as described in claim 6, characterized in that: The platform unit also integrates an intelligent decision-making center, which includes: The dynamic confidence fusion module is used to receive the original recognition result and its confidence level output by the state recognition module, the environmental data of the auxiliary sensor module, and the predictive state trajectory output by the Prognostics engine. Based on the Bayesian fusion algorithm, it calculates and outputs a global state confidence level that integrates real-time perception and long-term prediction. The strategy generation and scheduling module is used to dynamically generate and issue collaborative strategy instructions based on the global state reliability, the historical records of the alarm module, and the operation and maintenance plan of the electronic management module; the collaborative strategy instructions include: The sensing unit is instructed to adjust the image acquisition frequency, switch the recognition algorithm, or activate redundant sensors. The transmission unit is given a transmission optimization strategy, which instructs it to dynamically switch communication links, adjust the data compression ratio, or change the encryption level. The human-computer interaction strategy is issued to the application unit, instructing it to change the dimensions and methods of information presentation.

8. The substation hard plate status monitoring system based on visual recognition as described in claim 7, characterized in that: The Bayesian fusion algorithm executed by the dynamic credibility fusion module is as follows: ； E represents the actual state of the hard platen, E represents the image recognition results and environmental data provided by the sensing unit, and P represents the predictive evidence provided by the Prognostics engine. Let S be the global state confidence score, representing the probability that state S is obtained by considering all the evidence. These are the conditional probabilities of perceived and predicted evidence, respectively, learned from historical data. This represents the prior probability.

9. The substation hard plate condition monitoring system based on visual recognition as described in claim 8, characterized in that: The policy generation and scheduling module is implemented through a reinforcement learning agent. Its state space is a set of global state credibility, device health index, and network status. Its action space consists of various identifiable cooperative policy instructions. Its reward function R is designed as follows: ； in The weights are used as coefficients, and the agent learns the optimal cooperative strategy under different system states by maximizing the cumulative reward.

10. The substation hard plate condition monitoring system based on visual recognition as described in claim 9, characterized in that: The application unit responds to the human-computer interaction strategy issued by the platform, and its display method is as follows: When displaying the status of the hard platen, its global status credibility is rendered simultaneously to visually and intuitively indicate the reliability of the result. When a predictive maintenance suggestion is received, instead of directly popping up an alarm, it is integrated into the operation and maintenance calendar and a visual report containing the decision basis and recommended actions is generated for the user to review. It provides a policy simulator interface, allowing users to replay the system state at any historical moment, try issuing different policy commands, observe the virtual simulation results, and evaluate the effectiveness of the policy.

Images