An intelligent transformer abnormal state identification system

Abstract

The present application relates to the technical field of intelligent inspection and state diagnosis of power equipment, in particular to a transformer abnormal state intelligent identification system; comprising data acquisition, environment parameter estimation, optical decoupling, state diagnosis, interference quantification and coupling diagnosis units; the system captures high dynamic pixel flow and estimates environment parameters; the core is to adopt reverse optical transmission decoupling, physically decompose the pixel flow into device inherent reflectivity graph and environment light and interference graph; the system performs defect diagnosis on the reflectivity graph, quantifies the interference graph to evaluate the reliability, and finally generates an interpretable diagnosis report; the present application solves the problem that traditional models cannot distinguish between device state and environmental interference, and improves the robustness of diagnosis.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection and condition diagnosis technology for power equipment, specifically an intelligent identification system for abnormal transformer conditions. Background Technology

[0002] In the visual inspection and maintenance of power equipment such as transformers, the system needs to analyze images to identify early defects such as oil stains and cracks. Existing data-driven recognition models, when processing high-dynamic images acquired on-site, highly obfuscate the actual material state of the equipment with dynamic environmental optical interference, such as fog scattering, rain reflection, or strong light overexposure. Traditional models cannot decouple these two signals at the physical level, resulting in an irreconcilable contradiction between recognition sensitivity and environmental robustness. To resist strong noise, the model has to reduce its sensitivity to minor defects, which can easily lead to missed detections; conversely, it may generate a large number of false alarms because it cannot distinguish between interference and defects. In addition, these models are usually black boxes, unable to quantify the degree of interference encountered during diagnosis, and unable to provide physically interpretable diagnostic evidence. Therefore, how to solve the physical obfuscation between equipment status signals and environmental interference signals, and achieve sensitive and reliable identification of early minor defects in high-interference environments, has become an urgent technical problem to be solved. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an intelligent identification system for abnormal transformer conditions. Specifically, the technical solution of this invention includes:

[0004] The scene perception and data acquisition unit is used to capture the high dynamic range visual image stream of the transformer target area and obtain the original high dynamic range pixel stream.

[0005] The physical environment parameter estimation unit is used to receive the original high dynamic range pixel stream, extract environmental parameters characterizing physical interference, and generate an environmental state vector.

[0006] The reverse optical transmission decoupling unit is used to receive the original high dynamic range pixel stream and the environmental state vector, and decompose the original high dynamic range pixel stream into a device intrinsic reflectivity map and an ambient illumination and interference map;

[0007] The equipment material condition diagnostic unit is used to perform defect detection on the inherent reflectivity map of the equipment, identify defect areas, and generate defect feature signals.

[0008] The interference quantification and credibility generation unit is used to receive the ambient light and interference map, analyze the interference signal, and generate an environmental interference index.

[0009] The coupled diagnostic result generation unit is used to receive the defect feature signal and the environmental interference index, perform physical interpretability coupling between the two, and generate an interpretable diagnostic report.

[0010] Preferably, the physical environment parameter estimation unit generates the environmental state vector by:

[0011] By analyzing the global contrast reduction and color shift of the original high dynamic range pixel stream, the atmospheric scattering coefficient used to characterize the concentration of fog or smoke is determined.

[0012] By detecting high-frequency, flowing instantaneous highlights or low-frequency jitter in the original high dynamic range pixel stream, dynamic interference patterns used to characterize rain or strong winds are determined.

[0013] The environmental state vector is generated based on the atmospheric scattering coefficient and the dynamic disturbance mode.

[0014] Preferably, the reverse optical transmission decoupling unit is used for:

[0015] Using the environmental state vector as control input, the physical optics model is dynamically selected and configured;

[0016] When the environmental state vector indicates a high scattering coefficient, an inverse decoupling algorithm based on an atmospheric scattering physics model is initiated to perform the decomposition.

[0017] Preferably, the equipment material condition diagnostic unit is used for:

[0018] A highly sensitive defect detection algorithm is used to analyze the inherent reflectivity map of the equipment in order to accurately identify areas in the inherent reflectivity map that match the characteristics of early oil stains or hairline cracks.

[0019] Preferably, the interference quantization and credibility generation unit is used for:

[0020] By analyzing the energy of the ambient light and interference map, including the statistical characteristics of the intensity, range and frequency of the interference signal, a comprehensive evaluation value is generated as a confidence adjustment factor, which serves as the environmental interference index. The environmental interference index is used to quantitatively reflect the degree of meteorological noise pollution on the original image.

[0021] Preferably, the coupling diagnostic result generation unit is used for:

[0022] The environmental interference index is used as a diagnostic confidence regulator to actively regulate the output logic of the defect feature signal.

[0023] When the defect feature signal indicates the presence of a defect, the environmental interference index and the credible interference threshold are used for determination.

[0024] Preferably, the process of the coupling diagnosis result generation unit performing the determination includes:

[0025] If the environmental interference index is lower than the credible interference threshold, then a high-credibility defect is output.

[0026] If the environmental interference index is higher than the credible interference threshold, the confidence level of the diagnostic result is actively reduced, or it is marked as a suspected defect under high interference, or a request for manual review is triggered.

[0027] Preferably, the credible interference threshold is a benchmark set by statistical analysis of the decoupling quality and final diagnostic accuracy under different interference intensities on a validation dataset.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] 1. This system decomposes the original high dynamic pixel stream into a device-inherent reflectivity map and an ambient illumination and interference map through physical optics decoupling. It then diagnoses and quantifies both maps separately, thus resolving the contradiction between recognition sensitivity and environmental robustness caused by the inability of traditional models to distinguish between device status signals and environmental interference signals.

[0030] 2. This system performs defect detection on the inherent reflectivity map of the equipment after physical separation, which has been stripped of environmental noise. This makes it possible to use a highly sensitive defect detection algorithm, thereby accurately identifying small defects such as early oil stains or hairline cracks that traditional models would inevitably ignore due to strong denoising or robust training. This significantly improves the early warning capability of inspection.

[0031] 3. This system analyzes the statistical characteristics of the separated ambient light and interference maps, such as energy, intensity, range, and frequency, by setting up interference quantification and credibility generation units. For the first time, it has achieved a quantitative assessment of environmental interference. The system no longer operates as a black box, but generates objective environmental interference indexes, realizing self-awareness in diagnosing the severity of the environment.

[0032] 4. This system uses the quantified environmental interference index as a regulator of diagnostic confidence by coupling the diagnostic result generation unit. When the interference is determined to be severe, the system actively reduces the confidence of the output result or triggers manual review. When the interference is determined to be minor, it outputs a high-confidence defect. This design greatly suppresses false alarms in high-interference environments and realizes a physically interpretable intelligent diagnosis that combines high sensitivity and high reliability. Attached Figure Description

[0033] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0034] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0036] Example 1: Please refer to Figure 1 A transformer abnormality state intelligent identification system, comprising:

[0037] The scene perception and data acquisition unit is used to capture the high dynamic range visual image stream of the transformer target area and obtain the original high dynamic range pixel stream.

[0038] The physical environment parameter estimation unit is used to receive the original high dynamic range pixel stream, extract environmental parameters that characterize physical disturbances, and generate an environmental state vector.

[0039] The reverse optical transmission decoupling unit is used to receive the original high dynamic range pixel stream and the environmental state vector, and decompose the original high dynamic range pixel stream into a device intrinsic reflectivity map and an ambient illumination and interference map.

[0040] The equipment material condition diagnostic unit is used to detect defects in the inherent reflectivity map of the equipment, identify defect areas, and generate defect feature signals.

[0041] The interference quantization and credibility generation unit is used to receive ambient light and interference maps, analyze interference signals, and generate environmental interference indexes.

[0042] The coupled diagnostic result generation unit is used to receive defect feature signals and environmental interference level indicators, perform physical interpretability coupling between the two, and generate an interpretable diagnostic report.

[0043] In one specific embodiment of the present invention, the overall workflow of the system exhibits an innovative logic based on physical perception that differs from the traditional data-driven model;

[0044] The system's starting point is the scene perception and data acquisition unit; the core purpose of this unit is to acquire high-fidelity raw visual data for subsequent precise physical analysis; in this embodiment, it is configured with industrial-grade high dynamic range... Image sensors, for example, employ multi-exposure fusion or logarithmic response CMOS sensors, configured to capture a high dynamic range visual image stream of the transformer target area; the output of this step is a raw high dynamic range pixel stream; this raw high dynamic range pixel stream is an uncompressed or nonlinear tone-mapped raw sensor data stream, whose key characteristic is its ability to fully preserve details in both bright and dark areas under extreme lighting conditions such as strong midday reflections or shadows with high bit depth; this high fidelity is a necessary and indispensable raw input for subsequent physical-optical decoupling, because standard 8-bit image streams will permanently lose critical physical information due to overexposure or underexposure; this pixel stream is concurrently transmitted to a physical environment parameter estimation unit and a reverse optical transmission decoupling unit;

[0045] The purpose of the physical environment parameter estimation unit is to independently evaluate the optical state of the current environmental medium and provide key prior physical constraints for the subsequent decoupling unit. It receives the raw high dynamic range pixel stream and analyzes it in real time to extract environmental parameters characterizing physical disturbances. The output of this unit is an environmental state vector. This environmental state vector is a structured control signal used to physically describe the current optical state of the environment in real time. It is immediately transmitted to the reverse optical transmission decoupling unit to dynamically configure the algorithm of this unit.

[0046] The reverse optical transmission decoupling unit, as the core technology of this invention, aims to decouple reverse rendering from the physical scene, physically separating the highly confused device state signals and environmental interference signals in the pixel space. It receives the original high-dynamic pixel stream and the environmental state vector from the physical environment parameter estimation unit. The innovation of this unit lies in its use of the environmental state vector as a dynamic control input, rather than a static algorithm, to dynamically select and activate the corresponding physical optical model. For example, when the vector indicates high scattering, the atmospheric scattering model is activated. This unit performs the decoupling action, physically decomposing the original high-dynamic pixel stream into two independent, orthogonal data layers.

[0047] Inherent reflectivity map of the device: This layer represents the true material state of the device surface, that is, it theoretically restores the optical reflectivity characteristics that the device surface should have under ideal lighting conditions without weather interference, fog, or rain.

[0048] Ambient lighting and interference map: This layer captures all the isolated ambient interference signals, such as scattered light caused by fog, and highlight reflections and refractions caused by rain or water stains;

[0049] This unit transmits the device's inherent reflectivity map to the device material status diagnosis unit, and transmits the ambient light and interference map to the interference quantification and credibility generation unit;

[0050] The system executes two key downstream branches in parallel:

[0051] The purpose of the equipment material condition diagnostic unit is to achieve highly sensitive identification of fine-grained defects on clean data. It receives the equipment's inherent reflectivity map and performs defect detection only on this layer. Since the expected input equipment inherent reflectivity map has physically stripped away high-frequency environmental noise such as rain reflection and low-frequency obstructions such as fog, this unit can use a detection algorithm that is highly sensitive to noise but extremely sensitive to defects, without worrying about the massive false alarms in traditional applications. Its output is a defect feature signal, a data structure indicating the defect type, location, and extent.

[0052] The purpose of the interference quantification and credibility generation unit is to achieve self-awareness of the system, that is, to quantify environmental interference. It receives ambient lighting and interference map, and generates an environmental interference index by analyzing the statistical characteristics of the interference signal in this layer, such as intensity, spatial range, and frequency distribution. This environmental interference index is a comprehensive evaluation value used to quantitatively and objectively reflect the extent to which the original image is contaminated by meteorological noise.

[0053] The purpose of the coupled diagnostic result generation unit is to achieve physical interpretability and output highly reliable, non-black-box diagnostic conclusions. It simultaneously receives defect feature signals to answer what was diagnosed and environmental interference indicators to answer under what conditions the diagnosis was made. The core logic of this unit is to physically couple these two aspects for interpretability. It uses the environmental interference indicator as a confidence modifier for the diagnostic results to determine the final reliability of the defect feature signals. Its output is an interpretable diagnostic report; for example, this report not only includes the discovery of oil stains but also the physical evidence: under high fog interference, the indicator... Under quantization, the system successfully decoupled the surface reflectance to match oil stains rather than water stains. The diagnosis was highly reliable oil contamination.

[0054] Through the precise collaborative work of the aforementioned units, this invention constructs an innovative physical perception-based identification architecture. The innovation of this architecture lies in its information flow: it is not a simple denoising-recognition serial pipeline, but rather establishes a parallel dual-branch logic of perception-decoupling-diagnosis and interference quantification-reliability coupling. Through real-time physical-optical decoupling, it solves the problem of recognition failure caused by the confusion between the material state of the device and the optical state of the environment in existing image recognition technologies. This system no longer needs to painfully balance defect recognition sensitivity and noise-ignoring robustness; instead, through physical separation, it simultaneously achieves fine-grained sensitivity to micron-level defects and high robustness to dynamic meteorological interference. Compared with existing technologies, this invention, for the first time, achieves physical interpretability of diagnostic conclusions—knowing why it is a defect—and quantified self-awareness of interference—knowing under what conditions it is identified—greatly improving the reliability of diagnostic results and their credibility in actual operation and maintenance scenarios.

[0055] Example 2:

[0056] The physical environment parameter estimation unit generates an environment state vector including:

[0057] By analyzing the global contrast reduction and color shift of the original high dynamic range pixel stream, the atmospheric scattering coefficient used to characterize the concentration of fog or smoke is determined.

[0058] By detecting high-frequency, flowing instantaneous highlights or low-frequency jitter in the original high dynamic range pixel stream, dynamic interference patterns used to characterize rain or strong winds are determined.

[0059] An environmental state vector is generated based on the atmospheric scattering coefficient and dynamic disturbance pattern.

[0060] In a specific embodiment of the present invention, the method by which the physical environment parameter estimation unit generates the environmental state vector is described in detail. The purpose of this unit is to perform physical quantization, rather than simple weather phenomenon classification. For example, it distinguishes between optical scattering and foggy weather, and between instantaneous highlights and rainy weather.

[0061] This unit analyzes the raw high dynamic range pixel stream; based on physical optics principles such as Rayleigh scattering and Mie scattering, fog or smoke causes light to be severely scattered in the air; this unit quantifies this scattering effect through algorithms, specifically as follows:

[0062] By calculating the dynamic range of the image histogram, it can be observed that in dense fog, the histogram tends to concentrate towards the intermediate gray levels, resulting in a significant compression of the dynamic range.

[0063] By analyzing the R / G / B channel mean values ​​of distant areas or known neutral color areas such as white insulators in the image; fog usually causes an overlay of bluish or grayish atmospheric light on the image, resulting in color cast;

[0064] The atmospheric scattering coefficient is a quantitative indicator that integrates the above analysis results. Its value, for example, is a floating-point number normalized to 0 to 1, used to characterize the concentration of fog or smoke, and is directly related to the key parameters of subsequent atmospheric scattering models.

[0065] This unit simultaneously detects dynamic changes in the pixel stream in the time dimension; the dynamic disturbance mode is a feature descriptor used to characterize transient disturbances such as rain or strong winds, such as an enumeration value or feature vector.

[0066] When rain is present, the unit detects a large number of high-frequency, flowing instantaneous highlights in the image, namely the reflections of raindrops on the lens or device surface, by means of methods such as inter-frame difference or optical flow.

[0067] When there is strong wind, the unit may detect low-frequency jitter or periodic motion blur caused by device or camera bracket shaking by, for example, tracking feature points such as SIFT or ORB features on consecutive frames.

[0068] This unit combines the atmospheric scattering coefficients, which may be floating-point numbers, with dynamic disturbance patterns, which may be enumerated values ​​or vectors, to generate a structured multidimensional environmental state vector, for example: ;

[0069] With this design, the present invention no longer relies on external independent weather station data, which often does not match the micro-environmental optical state of the camera. Instead, it directly infers the environmental state from the pixel stream. This makes the environmental state vector a real-time, accurate, and local physical description of the current optical interference rather than the weather itself, providing an extremely accurate prior physical constraint for the dynamic model selection of the subsequent reverse optical transmission decoupling unit.

[0070] Example 3:

[0071] The reverse optical transmission decoupling unit is used for:

[0072] Using the environmental state vector as the control input, the physical optics model is dynamically selected and configured;

[0073] When the environmental state vector indicates a high scattering coefficient, an inverse decoupling algorithm based on an atmospheric scattering physics model is initiated to perform decomposition.

[0074] In specific embodiments of the present invention, the adaptive working mechanism of the reverse optical transmission decoupling unit is explained in detail;

[0075] Its core processing logic lies in its dynamic and physical characteristics, which is fundamentally different from traditional static or data-driven black box models; it receives environmental state vectors from the physical environment parameter estimation unit and uses them as a key control input, rather than just additional data.

[0076] This unit has a pre-built or dynamically constructed library of physical optics models, including atmospheric scattering models based on Mie scattering, scattering and absorption models based on physical optics, and water stain / raindrop models based on motion and reflection. The unit's processing logic is based on strict condition judgment and configuration.

[0077] The system parses the contents of the environment state vector in real time;

[0078] When the environment state vector indicates a high scattering coefficient, for example... The system determines that the current dominant interference is fog or smoke;

[0079] The decoupling unit dynamically selects and configures a specific physical optics model; specifically, it initiates an inverse decoupling algorithm based on an atmospheric scattering physical model, such as a model known in the art or a physical model that estimates atmospheric light using dark channel priors; this algorithm is then used to perform decomposition, i.e., from the original image... Solving in the middle ;in, It is the inherent reflectivity of the equipment. It is the original high dynamic range pixel stream. It is atmospheric light. It refers to transmittance; in this field, A can be estimated from the original image I(x) through, for example, the dark channel prior principle, while the transmittance t(x) can be further calculated from the estimated dark channel map and atmospheric light A to ensure the physical accuracy of decoupling; while the ambient lighting and interference map can be obtained from the atmospheric light With transmittance Export the raw high dynamic range pixel stream into a device-intrinsic reflectance map and an ambient illumination and interference map.

[0080] It should be understood that if the environmental state vector indicates high-frequency instantaneous specular highlights... This unit will dynamically select and configure another decoupling model, such as an algorithm based on time series median filtering or robust principal component analysis (RPCA), to achieve optimal separation of different physical disturbances.

[0081] The non-obviousness of this design lies in its implementation of specific physical problems and specific physical analysis; the system no longer relies on data-driven black-box dehazing or deraining algorithms, which are prone to failure when faced with interference not seen in the training set, or inevitably erase real minute defects during denoising; this invention solves the current physical problem by dynamically selecting the correct physical tools, making the decoupling process highly targeted, accurate and physically interpretable, thereby preserving the authenticity and integrity of the device's inherent reflectivity map to the greatest extent.

[0082] Example 4:

[0083] The equipment material condition diagnostic unit is used for:

[0084] A highly sensitive defect detection algorithm is used to analyze the inherent reflectivity map of the equipment in order to accurately identify areas in the inherent reflectivity map that match the characteristics of early oil stains or hairline cracks.

[0085] In a specific embodiment of the present invention, the working method of the equipment material condition diagnosis unit is described in detail;

[0086] The core purpose of this unit is to achieve highly sensitive identification of fine-grained defects; the input it receives is not the original image, but the device's inherent reflectivity map generated by the reverse optical transmission decoupling unit after careful processing.

[0087] The core working logic of this unit is that since its input layer, the device inherent reflectivity map, has physically stripped away most of the environmental noise such as rain reflection and fog obstruction, this unit can and should adopt a highly sensitive defect detection algorithm without compromising sensitivity for robustness at the algorithm level, as is the case with traditional CNNs.

[0088] In this invention, a high-sensitivity defect detection algorithm refers to a class of algorithms specifically designed to detect minute variations in texture, color, or reflectivity; in practice, this may include:

[0089] Texture analysis algorithms based on Gabor filter banks or wavelet transforms are employed. These algorithms process the inherent reflectance map of the device. Since hairline cracks introduce high-frequency texture discontinuities, while early oil stains cause low-frequency smooth changes in local reflectance, these algorithms can efficiently distinguish it from the normal texture of background materials such as ceramics or metals on a clean reflectance map.

[0090] An anomaly detection method based on machine learning is adopted, such as using an autoencoder or a single-class support vector machine. In this method, the model is trained using only the device’s inherent reflectivity map under normal conditions. During inference, the unit inputs the decoupled real-time reflectivity map into the model. Any region that cannot be reconstructed by the model with low error, i.e., a region with high reconstruction error, will be marked as an anomaly, i.e., identified as early oil stains or hairline cracks.

[0091] This design is impossible to implement in traditional systems because training such models on noisy raw images would cause the models to learn raindrops or reflections as anomalous, resulting in a massive number of false alarms.

[0092] This design resolves the core conflict between sensitivity and robustness mentioned in the background art. The non-obviousness of this invention lies in creating conditions that, through a process of decoupling before diagnosis, enable, for the first time, high-sensitivity algorithms that are only effective under laboratory conditions, such as the methods described above (method one or two), to be deployed and applied in real-world environments filled with interference. The ultimate effect is that the system can reliably detect early defects such as hairline cracks that traditional systems would inevitably ignore or erase due to strong denoising or robustness training, greatly improving the early warning capability and maintenance value of inspections.

[0093] Example 5:

[0094] The interference quantization and credibility generation unit is used for:

[0095] By analyzing the energy of ambient light and interference maps, including the statistical characteristics of the intensity, range, and frequency of interference signals, a comprehensive evaluation value is generated as a confidence adjustment factor, serving as an environmental interference index. This index quantitatively reflects the degree to which the original image is contaminated by meteorological noise.

[0096] In a specific embodiment of the present invention, the working method of the interference quantization and credibility generation unit is described in detail.

[0097] The purpose of this unit is to achieve self-awareness of the system, that is, to objectively quantify the severity of the current environmental interference; the input it receives is the ambient light and interference map, i.e., the noise layer, separated by the reverse optical transmission decoupling unit.

[0098] This unit performs quantization by analyzing the energy of ambient lighting and the interference map; this energy analysis is a multi-dimensional process, not a simple summation; it extracts statistical features of the interference layer using image processing algorithms, including but not limited to:

[0099] The strength of the interference signal: for example, by calculating the root mean square of the pixel values ​​of the interference layer. Or quantify by average amplitude; dense fog corresponds to high-intensity low-frequency signals, and heavy rain corresponds to high-intensity instantaneous high-frequency signals;

[0100] Range of interference signals: For example, by thresholding the interference layer, the percentage of non-zero pixels that detect interference can be calculated, and the area affected by the interference can be quantified.

[0101] The frequency of the interference signal: for example, by performing a Fast Fourier Transform on the interference layer. Alternatively, wavelet analysis can be used to determine whether the interference energy is mainly concentrated in low frequencies, such as fog, or in high frequencies, such as raindrops.

[0102] This unit combines these statistical characteristics into a comprehensive evaluation value using a pre-defined evaluation model, such as a simple weighted algorithm or a small regression neural network; for example, in a preferred embodiment, the weighted algorithm can be defined as: Environmental disturbance index = ;in, These are the quantified characteristic values ​​of interference intensity, range, and frequency, calculated based on the aforementioned statistical characteristic analysis. It is a preset weight obtained by optimizing on the validation dataset according to the calibration method, used to characterize the importance of different interference features; this value is defined as the environmental interference index.

[0103] In this invention, the environmental interference index has a specific technical meaning: it is a reliability adjustment factor; its function is to quantitatively reflect the degree of meteorological noise pollution of the original image; the higher the index, for example, a value of 0 to 100, 85 points, the more serious the interference of the original image, the greater the difficulty of the decoupling unit's reverse optical transmission decoupling unit, and the higher the uncertainty of its output device inherent reflectivity map.

[0104] This system design is the first to possess the ability to quantify environmental interference. Traditional existing image recognition technology models are black boxes, unable to inform users that the recognition was performed in dense fog, resulting in low reliability. However, this invention achieves this self-awareness through this unit. This environmental interference index is a key input for subsequent physical interpretability diagnosis and risk control, and is an important component of the inventive logic flow of this invention.

[0105] Example 6:

[0106] The coupled diagnostic result generation unit is used for:

[0107] The environmental interference index is used as a diagnostic confidence regulator to actively control the output logic of the defect feature signal.

[0108] When a defect feature signal indicates the presence of a defect, the environmental interference level index and the credible interference threshold are used for determination.

[0109] In a specific embodiment of the present invention, the coupling mechanism of the coupling diagnosis result generation unit is explained; this unit is the core of realizing the identification-interference bidirectional coupling mechanism of the present invention and is the ultimate manifestation of system intelligence;

[0110] This unit simultaneously receives two independent input streams: the defect feature signal comes from the equipment material condition diagnosis unit, which answers what was diagnosed, and the environmental interference index comes from the interference quantification and credibility generation unit, which answers under what conditions the diagnosis was made.

[0111] The core logic of this unit lies in active control; the environmental interference index is not just an additional value for human reference, but is used by the system internally as a diagnostic confidence regulator, which directly controls the output logic of defect characteristic signals.

[0112] The specific control logic, i.e., the physical interpretability coupling, is as follows:

[0113] When a defect feature signal indicates the presence of a defect, for example, when the equipment material condition diagnostic unit reports in coordinates... An area matching the texture characteristics of hairline cracks was discovered.

[0114] Initiate a decision checkpoint: This unit will not immediately report the defect as a high confidence alarm; instead, it will execute an internal decision procedure: compare the current value of the environmental interference index from the interference quantification unit with a system-preset confidence interference threshold.

[0115] The creativity of this design lies in establishing a metacognitive or self-reflective checkpoint; it prevents the system from making potentially erroneous judgments under uncertain, highly disruptive conditions; by forcibly coupling defect detection with the quantification of environmental interference through logical coupling, the system begins to possess physical interpretability, and its diagnostic conclusions are no longer the black box of traditional image recognition technology models. Instead of outputting data, it transforms into a conditional, interpretable, and risk-controlled intelligent output.

[0116] Example 7:

[0117] The process of the coupled diagnostic result generation unit making a judgment includes:

[0118] If the environmental interference index is lower than the credible interference threshold, then output a high-credibility defect.

[0119] If the environmental interference index is higher than the credible interference threshold, the confidence level of the diagnostic result will be actively reduced, or it will be marked as a suspected defect under high interference, or a request for manual review will be triggered.

[0120] In a specific embodiment of the present invention, the specific output logic of the coupling diagnosis result generation unit after executing the determination of embodiment 6 is described in detail;

[0121] Its processing logic is based on decision conditions:

[0122] Condition 1: Low interference, high reliability: If the environmental interference index is lower than the reliable interference threshold; the physical meaning of this situation is: the original image is clear, or the interference is slight, or the reverse optical transmission decoupling unit has an excellent separation effect on the interference and is confident that the ambient light and interference image energy are very low.

[0123] Action 1: Under these high-confidence conditions, the system determines that defect characteristic signals, such as the detection of cracks, are reliable; therefore, the system outputs a high-confidence defect; for example, it generates a structured report.

[0124] ;

[0125] Condition 2: High Interference, Low Confidence: If the environmental interference index exceeds the confidence interference threshold, the physical meaning of this situation is that the original image is severely contaminated, such as by heavy rain or dense fog, resulting in extremely high energy in the ambient light and interference map. This indicates that even if the decoupling unit tries its best to separate the images, the output device inherent reflectivity map still has high uncertainty, or the defect feature signal itself may be caused by strong interference artifacts that have not been completely separated. Caused by;

[0126] Action Two: At this point, the system proactively intervenes, no longer blindly trusting defect characteristic signals; it will perform one or more of the following risk avoidance operations: proactively reducing the confidence level of the diagnostic results. For example, even if the equipment material condition diagnostic unit outputs a 99% confidence level, this unit will reduce it to 60%, or mark it as a suspected defect under high interference. For example, clearly noting in the report... and This could trigger a request for manual review, for example, by sending a request to the operations and maintenance system. The request, or direct control of the pan-tilt unit after the interference is reduced, for example... The area will be reshot in minutes;

[0127] This design achieves a balance between robustness and interpretability in diagnostics. Under condition one, the system ensures a high detection rate and low false negatives. Under condition two, the system actively reduces confidence or requests manual review, greatly suppressing false alarms caused by strong environmental noise. This dual logic ensures that the final interpretable diagnostic report has both high sensitivity and high reliability, perfectly solving the pain points of inaccurate reporting and false alarms in actual operation and maintenance.

[0128] Example 8:

[0129] The credible interference threshold is a benchmark set by statistically analyzing the decoupling quality and final diagnostic accuracy under different interference intensities on a validation dataset.

[0130] In specific embodiments of the present invention, the source and setting method of the credible interference threshold used as the key judgment criterion in embodiments 6 and 7 are described in detail to ensure that those skilled in the art can implement it;

[0131] This threshold is a configurable benchmark used to determine whether the environmental interference level is too high, i.e., beyond the system's reliable operating range;

[0132] It should be understood that the value of this threshold is not a fixed or guessed value; it is set by those skilled in the art, such as algorithm engineers or system deployment personnel, through statistical analysis on a validation dataset during the system deployment or calibration phase.

[0133] Specific setup method and implementation steps:

[0134] The complete system of this invention processes the verification dataset;

[0135] For each image in the validation set, the system generates an environmental interference index and a final diagnostic accuracy rate, i.e., whether the system's diagnostic results are consistent with the gold standard manual annotation.

[0136] Perform statistical analysis; for example, you can plot a curve. The axis represents the environmental disturbance index from low to high. The axis represents the final diagnostic accuracy, for example... or ;

[0137] Those skilled in the art will observe that, with the increase in environmental disturbance index The increase in axis accuracy ultimately affects diagnostic accuracy. The axis usually remains stable, but after a certain point it will start to drop sharply because excessive disturbance causes decoupling failure;

[0138] The credible interference threshold is set near this inflection point where accuracy drops sharply;

[0139] More precisely, this can be achieved by plotting receiver operational features on the validation dataset. The curves are used to compare the decoupling quality under different interference intensities and environmental interference levels. For example, the signal-to-noise ratio of the decoupled image is compared with the final true positive rate. and the false positive rate To find a balance; for example, the threshold could be set to maintain diagnostic accuracy, for example... The maximum environmental disturbance level that the system can tolerate when it exceeds 95%;

[0140] Furthermore, the credible interference threshold can be dynamically adjusted based on the importance level or risk appetite of different devices. For example, for critical 500kV transformers in the backbone network, the operation and maintenance strategy may require zero false alarms, so the threshold can be appropriately increased to tolerate higher interference and accept more manual review requests. For non-critical areas, the threshold can be reduced to pursue a lower false alarm rate.

[0141] This design ensures the scientific validity, implementability, and adjustability of the core judgment logic embodiment 7 of this invention; by providing a method based on statistical analysis and operational requirements, such as... The threshold setting method of this invention ensures that the system can be optimized through data-driven approaches in different deployment scenarios to achieve the best diagnostic performance that conforms to actual operation and maintenance risk preferences.

[0142] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A transformer abnormality state intelligent identification system, characterized in that, include: The scene perception and data acquisition unit is used to capture the high dynamic range visual image stream of the transformer target area and obtain the original high dynamic range pixel stream. The physical environment parameter estimation unit is used to receive the original high dynamic range pixel stream, extract environmental parameters characterizing physical interference, and generate an environmental state vector. The reverse optical transmission decoupling unit is used to receive the original high dynamic range pixel stream and the environmental state vector, and decompose the original high dynamic range pixel stream into a device intrinsic reflectivity map and an ambient illumination and interference map; The equipment material condition diagnostic unit is used to perform defect detection on the inherent reflectivity map of the equipment, identify defect areas, and generate defect feature signals. The interference quantification and credibility generation unit is used to receive the ambient light and interference map, analyze the interference signal, and generate an environmental interference index. The coupled diagnostic result generation unit is used to receive the defect feature signal and the environmental interference index, perform physical interpretability coupling between the two, and generate an interpretable diagnostic report. The physical environment parameter estimation unit generates the environmental state vector by including: By analyzing the global contrast reduction and color shift of the original high dynamic range pixel stream, the atmospheric scattering coefficient used to characterize the concentration of fog or smoke is determined. By detecting high-frequency, flowing instantaneous highlights or low-frequency jitter in the original high dynamic range pixel stream, dynamic interference patterns used to characterize rain or strong winds are determined. Based on the atmospheric scattering coefficient and the dynamic disturbance pattern, the environmental state vector is generated; The interference quantization and credibility generation unit is used for: By analyzing the energy of the ambient light and interference map, including the statistical characteristics of the intensity, range and frequency of the interference signal, a comprehensive evaluation value is generated as a confidence adjustment factor, which serves as the environmental interference index. The environmental interference index is used to quantitatively reflect the degree of meteorological noise pollution on the original image. The coupling diagnostic result generation unit is used for: The environmental interference index is used as a diagnostic confidence regulator to actively regulate the output logic of the defect feature signal. When the defect feature signal indicates the presence of a defect, the environmental interference index and the credible interference threshold are used for determination. The process by which the coupling diagnostic result generation unit makes the determination includes: If the environmental interference index is lower than the credible interference threshold, then a high-credibility defect is output. If the environmental interference index is higher than the credible interference threshold, the confidence level of the diagnostic result will be actively reduced, or it will be marked as a suspected defect under high interference, or a request for manual review will be triggered. Environmental disturbance index = ;in These are the quantified characteristic values ​​of interference intensity, range, and frequency, respectively, calculated based on the aforementioned statistical characteristic analysis. It is a preset weight obtained by optimizing the calibration method on the validation dataset to characterize the importance of different interference features; this value is defined as the environmental interference index.

2. The intelligent identification system for abnormal transformer states according to claim 1, characterized in that, The reverse optical transmission decoupling unit is used for: Using the environmental state vector as control input, the physical optics model is dynamically selected and configured; When the environmental state vector indicates a high scattering coefficient, an inverse decoupling algorithm based on an atmospheric scattering physics model is initiated to perform the decomposition.

3. The intelligent identification system for abnormal transformer states according to claim 1, characterized in that, The equipment material condition diagnostic unit is used for: A highly sensitive defect detection algorithm is used to analyze the inherent reflectivity map of the equipment in order to accurately identify areas in the inherent reflectivity map that match the characteristics of early oil stains or hairline cracks.

4. The intelligent identification system for abnormal transformer states according to claim 3, characterized in that, The credible interference threshold is a benchmark set by statistically analyzing the decoupling quality and final diagnostic accuracy under different interference intensities on a validation dataset.

Images