Intelligent operation and maintenance and fault early warning system for photovoltaic power station

Abstract

The invention relates to the technical field of photovoltaic power station operation and maintenance, and discloses a photovoltaic power station intelligent operation and maintenance and fault early warning system. A multi-band image fusion algorithm and a temperature gradient enhancement algorithm are applied to identify composite defects; establishing a defect interaction influence model to analyze the synergistic effect of different defects; constructing a defect evolution prediction algorithm to predict a defect development trend; and developing a defect-performance mapping model and a processing priority algorithm to optimize a maintenance decision. According to the method, the defect detection sensitivity and accuracy are improved, accurate identification and evolution prediction of composite defects are realized, a prospective maintenance decision is effectively supported, the operation and maintenance efficiency of a photovoltaic power station is remarkably improved, the service life of the power station is prolonged, and the safety risk is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and more specifically, it relates to a method for intelligent operation and maintenance and fault warning of a photovoltaic power station, as well as a related system. Background Art

[0002] In recent years, with the rapid development of photovoltaic power generation technology and the continuous expansion of the application scale, large-scale photovoltaic power stations have increased rapidly worldwide. However, as outdoor power equipment operating for a long time, photovoltaic modules face various types of fault risks, including hot spots, PID (Potential Induced Degradation), hidden cracks, string soldering fractures, etc. These faults not only reduce the power generation efficiency but may also cause safety accidents. Therefore, the effective operation and maintenance and fault warning of photovoltaic power stations have become key links to ensure the safe and stable operation of the power stations.

[0003] Traditional operation and maintenance methods for photovoltaic power stations mainly rely on manual inspection with infrared thermal imagers. This method has obvious limitations: on the one hand, manual inspection is inefficient, requires a large amount of human resources, and it is difficult to cover all areas of large-scale power stations; on the other hand, traditional infrared thermal imagers have insufficient detection sensitivity in low-temperature difference environments (such as early morning or cloudy days), and it is difficult to identify early thermal anomalies with a temperature difference less than 2°C, resulting in early defects not being detected in time.

[0004] In the prior art, although drones equipped with infrared cameras have begun to be used for automated inspection, these systems generally use a single band (mainly a single infrared band) for fault detection, and have limited ability to identify complex defects. In actual operation, photovoltaic modules often have multiple types of defects at the same time, such as the coexistence of hot spots and PID, the co-occurrence of hidden cracks and string soldering fractures, etc. These different types of defects will affect each other and jointly act on the component performance. However, the prior art lacks the ability to model the mutual influence of different types of defects and cannot accurately evaluate the comprehensive impact of complex defects on the system.

[0005] In addition, the prior art generally lacks a scientific defect grading mechanism and priority evaluation method. Defects in different positions, different types, and different severity levels have different impacts on the output power of components and different risks to system safety. However, the prior art mainly relies on empirical judgment to determine the maintenance priority order, lacks a quantitative evaluation standard, resulting in unreasonable allocation of operation and maintenance resources and inability to achieve precise maintenance.

[0006] Most critically, the prior art lacks the ability to predict the evolution trend of defects. Defects in photovoltaic modules often have the characteristics of progressive development, such as the gradual expansion of hidden cracks and the gradual aggravation of PID. If the development trend of defects can be accurately predicted, proactive maintenance can be achieved, and intervention can be carried out before the defects cause serious losses. However, the prior art mainly focuses on the static characteristics of defects and lacks a model-based analysis of their dynamic evolution process, and cannot support preventive maintenance decisions.

[0007] Therefore, there is an urgent need for an intelligent operation and maintenance and fault warning method for photovoltaic power stations that can comprehensively utilize multi-spectral information, process composite defects, predict defect evolution, and guide precise maintenance, so as to improve the operation and maintenance efficiency and reliability of photovoltaic power stations, extend the service life of the power stations, and reduce the operation and maintenance costs. Summary of the Invention

[0008] The present invention provides an intelligent operation and maintenance and fault warning method and system for a photovoltaic power station, which solves the technical problem of the lack of an effective composite defect identification and warning mechanism in the related art.

[0009] The present invention discloses an intelligent operation and maintenance and fault warning method for a photovoltaic power station, including the following steps:

[0010] Use a drone system equipped with a multi-spectral camera to inspect the photovoltaic power station and collect multi-spectral image data including visible light, near-infrared, and mid-infrared bands;

[0011] Based on the multi-spectral image data, apply a multi-band image fusion algorithm and a temperature gradient enhancement algorithm to identify composite defects in photovoltaic modules;

[0012] Establish a defect interaction influence model, quantitatively analyze the synergistic effect when different types of defects coexist, and evaluate the comprehensive influence of composite defects on the performance of photovoltaic modules;

[0013] Based on the time-series multi-spectral data, construct a defect evolution prediction algorithm, establish a growth model of defect severity, and predict the development trend of defects;

[0014] Develop a defect-performance mapping model and a risk assessment and treatment priority algorithm. Based on features such as defect type, location, and area, quantitatively evaluate the actual impact of defects on the output power of the module, and determine the maintenance priority according to the degree of influence and safety risk, so as to realize the optimal allocation of operation and maintenance resources.

[0015] Further, the acquisition of the multi-spectral image data includes: configuring an integrated multi-spectral camera system, including a visible light camera, a near-infrared camera, and a mid-infrared thermal imager; based on the layout information of the photovoltaic power station and meteorological conditions, use an adaptive path planning algorithm to generate the optimal flight path of the drone; execute an automatic inspection task according to the planned flight path, collect images in three bands and record metadata such as acquisition time, ambient temperature, and irradiance; preprocess the collected multi-band images, including image correction, multi-spectral image registration, image enhancement, and image mosaicing.

[0016] Furthermore, the multi-band image fusion algorithm includes: extracting the feature representations of each band using dedicated feature extraction networks according to the characteristics of different band images; adjusting the features of each band to the same dimension and spatial size through feature channel alignment and feature space alignment; calculating the fusion weights of different feature maps at each position and performing adaptive weighting based on feature importance; and performing weighted fusion on the features of the three bands according to the calculated weights to generate a fusion feature map containing multi-band complementary information.

[0017] Furthermore, the temperature gradient enhancement algorithm includes: calculating the two-dimensional temperature gradient and gradient amplitude of the thermal image; calculating an adaptive threshold based on the temperature statistical characteristics of the component region; and generating enhanced thermal image features according to the calculated temperature gradient and adaptive threshold to improve the detection sensitivity of thermal anomalies.

[0018] Furthermore, the establishment of the defect interaction influence model includes: quantifying the features of various identified defects and extracting key parameters related to performance influence; establishing a single influence model of each defect type on the component performance parameters; defining a co-influence function between defects to represent the additional influence generated when two defects coexist; and calculating the comprehensive influence of multiple coexisting defects on the key performance parameters of the component based on the single defect influence and co-influence function.

[0019] Furthermore, the defect evolution prediction algorithm includes: constructing a time-series defect database to record the same defect feature data collected at different time points; constructing growth models for different types of defects respectively to achieve the mathematical expression of the defect evolution process; predicting the future evolution trajectory of the defect based on the constructed growth model to form a defect evolution curve; and predicting the key time points in the defect development process to provide a time reference for maintenance decisions.

[0020] Furthermore, the defect growth model includes: an exponential growth model of hot spot area and temperature difference; a Sigmoid growth model of the PID influence area; a piecewise linear or power-law growth model of the crack length; and adapting to environmental factors through adaptive estimation of model parameters.

[0021] Furthermore, the defect-performance mapping model and the risk assessment and treatment priority algorithm include: integrating the key features of the defect into a feature vector, including parameters such as defect type, severity, location, area, and development stage; training a defect-performance mapping model based on historical defect data and corresponding measured performance parameters to predict the impact of the defect on the component performance; calculating the annual power generation loss and economic loss caused by the defect; establishing a safety risk assessment model to evaluate the potential risk of the defect to system safety; and calculating the maintenance priority score by comprehensively considering the economic loss, safety risk level, and treatment difficulty to determine the priority order of defect treatment.

[0022] Furthermore, the maintenance priority algorithm further includes: classifying defects into four processing priority levels of emergency handling, high priority, medium priority, and low priority according to the priority score; clustering and optimizing the scheduling of maintenance tasks based on defect priority, geographical location, and maintenance cost to generate an optimal maintenance plan; recommending appropriate processing methods according to defect type, severity, and location, including component replacement, isolation processing, or local repair; calculating the input-output ratio of maintenance processing, including maintenance cost, expected revenue, and payback period.

[0023] The present invention also discloses a photovoltaic power station intelligent operation and maintenance and fault warning system, including: a multi-spectral data acquisition module for inspecting a photovoltaic power station by using an unmanned aerial vehicle system equipped with a multi-spectral camera to acquire multi-band image data of components; a composite defect identification module for identifying various defects in photovoltaic components based on the preprocessed multi-spectral image data by applying a multi-band image fusion algorithm and a deep learning model; a defect interaction influence analysis module for establishing a defect interaction influence model, quantitatively analyzing the synergistic effect when multiple defects coexist, and evaluating the comprehensive influence of composite defects on the performance of photovoltaic components; a defect evolution prediction module for developing a defect evolution prediction algorithm based on time-series multi-spectral data, establishing a growth model of defect size and severity, and predicting the development trend of defects; a loss assessment and maintenance decision-making module for developing a defect-performance mapping model, evaluating the actual influence of defects on the output power of components, and establishing a risk assessment and processing priority algorithm to guide accurate maintenance decisions.

[0024] The photovoltaic power station intelligent operation and maintenance and fault warning method provided by the present invention solves the technical problem of the lack of an effective composite defect identification and warning mechanism in the prior art and achieves the following beneficial effects:

[0025] Improve the sensitivity and accuracy of fault detection: Through multi-spectral fusion technology and temperature gradient enhancement algorithm, early hot spot defects with a temperature difference as small as 1.2 °C can be detected, and the automatic detection accuracy of hot spots is increased to 97.5%, significantly reducing misjudgment and missed judgment situations. For defects such as early PID and micro-cracks that are difficult to detect by traditional single-band detection, the detection sensitivity is increased by about 40%, making the hidden danger investigation of photovoltaic power stations more comprehensive and reliable.

[0026] Realize the accurate identification and evaluation of composite defects: By establishing a defect interaction influence model, the synergistic effect of different types of defects is analyzed from a quantitative perspective for the first time, breaking through the limitation of traditional technologies that analyze various defects in isolation, and being able to accurately evaluate the comprehensive influence of composite defects on component performance, providing an important basis for scientific maintenance decisions.

[0027] Support forward-looking maintenance decisions: The defect evolution prediction algorithm based on time-series multi-spectral data can accurately predict the development trend of defects, identify potential risks in advance, realize the transformation from "repair after failure" to "predictive maintenance", and significantly reduce maintenance costs and power generation losses. Field tests show that this function can detect potential serious defects 15 - 30 days in advance on average, reducing power generation losses caused by defect deterioration by about 35%.

[0028] Improve operation and maintenance efficiency: Through the defect-performance mapping model and the risk assessment and treatment priority algorithm, optimize the allocation of maintenance resources, improve the operation and maintenance efficiency by about 3 times, and reduce the operation and maintenance costs by more than 20% at the same time. Especially for large-scale photovoltaic power plants, this method can avoid unnecessary component replacements, reduce the mis-dismantling rate by about 50%, and significantly reduce the operation and maintenance costs.

[0029] Prolong the service life of the power plant: By accurately identifying early defects and intervening in advance, effectively prevent the expansion and evolution of defects, and can extend the average service life of components by 2 - 3 years, improving the overall return on investment of photovoltaic power plants.

[0030] Reduce safety risks: Discover and handle high-risk hot spots in a timely manner, effectively avoid safety accidents such as component fires. Field tests show that this method can reduce the risk of safety accidents related to hot spots by about 80%, significantly improving the operation safety of photovoltaic power plants.

[0031] In summary, the intelligent operation and maintenance and fault warning method for photovoltaic power plants provided by the present invention, through the combination of multi-spectral imaging technology and deep learning algorithms, realizes the accurate identification, evolution prediction and maintenance decision optimization of composite defects of photovoltaic components, significantly improves the operation and maintenance efficiency of photovoltaic power plants, prolongs the service life of the power plant, reduces the operation and maintenance costs and safety risks, and has significant economic and social benefits. Brief Description of the Drawings

[0032] Figure 1 is the overall flow chart of the intelligent operation and maintenance and fault warning method for photovoltaic power plants provided by the present invention;

[0033] Figure 2 is the detailed flow chart of the multi-spectral data acquisition step provided by the present invention;

[0034] Figure 3 is the detailed flow chart of the composite defect identification step provided by the present invention;

[0035] Figure 4 is the detailed flow chart of the defect interaction impact analysis step provided by the present invention;

[0036] Figure 5 is the detailed flow chart of the defect evolution prediction step provided by the present invention;

[0037] Figure 6It is a detailed flowchart of the loss assessment and maintenance decision-making steps provided by the present invention. Detailed implementation manners

[0038] Before describing the present application in detail, in order to help understand the technical solution of the present application, the terms to be used in the present application are first explained below:

[0039] Photovoltaic module: A functional unit composed of multiple solar cells connected in series or parallel, usually composed of tempered glass, encapsulation material, backplane and aluminum alloy frame, and is the core component of a photovoltaic power generation system.

[0040] Multispectral imaging: An imaging technique that acquires the reflection or radiation characteristics of a target in different spectral regions by simultaneously or sequentially collecting images of the same target in different wavelength bands (such as visible light, near infrared, mid infrared, etc.).

[0041] Hot spot: An abnormal temperature rise phenomenon that occurs in a local area of a photovoltaic module, usually caused by problems such as broken solar cells, shading, and poor connection. In severe cases, it may cause component damage or fire.

[0042] PID (Potential Induced Degradation): A degradation phenomenon that affects the performance of photovoltaic modules, caused by charge accumulation under high voltage in humid conditions, and will cause a decrease in the output power of the modules.

[0043] Hidden crack: A tiny crack that is invisible to the naked eye but exists inside a photovoltaic cell. Over time, it may expand into a complete fracture, resulting in a decrease in the performance of the module.

[0044] In the prior art, the following main problems exist in the fault detection and operation and maintenance of photovoltaic power station components:

[0045] First of all, the traditional fault detection method for photovoltaic power stations mainly relies on manual inspection with an infrared thermal imager. This method not only has low efficiency and requires a large amount of human resources, but also has limited detection accuracy. In practical applications, it is often difficult for manual inspections to cover all areas of large-scale photovoltaic power stations, and it is easy to overlook some fault points. More importantly, traditional infrared thermal imagers have insufficient detection sensitivity in low temperature difference environments (such as early morning or cloudy days), and it is difficult to identify early thermal anomalies with a temperature difference less than 2°C, resulting in the inability to detect early defects (such as microcracks and PID in the initial stage) in a timely manner.

[0046] Secondly, existing drone imaging analysis systems generally use a single band (mainly a single infrared band) for fault detection, which has limited ability to identify composite defects. In actual operation, photovoltaic modules often have multiple types of defects at the same time, such as hot spots and PID coexisting, hidden cracks and string welding fractures coexisting, etc. These different types of defects will affect each other and act together on the performance of the module, but the existing technology lacks the ability to model the mutual influence of different types of defects, and cannot accurately evaluate the comprehensive impact of composite defects on the system.

[0047] Third, existing technologies generally lack scientific defect classification mechanisms and priority assessment methods. Defects of different locations, types, and severity have different impacts on component output power and different risks to system safety. However, existing technologies mainly rely on experience to determine maintenance priorities and lack quantitative assessment standards, resulting in unreasonable allocation of operation and maintenance resources and the inability to achieve precise maintenance.

[0048] Finally, existing technologies lack the ability to predict the evolution trend of defects. Defects in photovoltaic modules often have a progressive development characteristic, such as the gradual expansion of hidden cracks and the gradual aggravation of PID. If the development trend of defects can be accurately predicted, proactive maintenance can be achieved and intervention can be carried out before the defects cause serious losses. However, existing technologies mainly focus on the static characteristics of defects and lack modeling analysis of their dynamic evolution process, which cannot support preventive maintenance decisions.

[0049] The present application provides a method for intelligent operation and maintenance and fault warning of a photovoltaic power station, which combines multispectral imaging technology with a deep learning algorithm to solve the problems existing in the above-mentioned prior art.

[0050] According to the embodiments of the present application, the technical solution mainly includes the following aspects:

[0051] First, this application uses multi-spectral imaging technology to simultaneously obtain image data of photovoltaic modules in the visible light, near infrared and mid-infrared bands. Through the complementarity of multi-band information, the detection sensitivity of various defects is improved, especially the ability to identify early defects. Compared with traditional single infrared band detection, multi-spectral fusion technology can effectively detect tiny thermal anomalies in low temperature difference environments and can distinguish different types of defect characteristics.

[0052] Secondly, this application proposes an innovative multi-band image fusion algorithm and temperature gradient enhancement algorithm. By extracting the complementary features of different bands and effectively fusing them, a thermal anomaly pattern library that takes into account interference factors such as shadows and reflections is constructed, and accurate identification of multiple defect types such as hot spots, PID, hidden cracks, and string welding fractures is achieved. This algorithm greatly improves the recognition accuracy of composite defects and reduces the misjudgment and missed judgment rates.

[0053] Third, this application establishes a defect interaction impact model, and for the first time analyzes the synergy effects of different types of defects from a quantitative perspective. This model breaks through the limitation of traditional technologies that analyze various defects in isolation, can accurately evaluate the comprehensive impact of compound defects on component performance, and provides an important basis for scientific maintenance decisions.

[0054] Fourth, this application constructs a defect evolution prediction algorithm based on time-series multi-spectral data. By analyzing the characteristic changes of defects at different time points, a growth model for defect size and severity is established, achieving accurate prediction of defect development trends and supporting proactive maintenance decisions.

[0055] Finally, this application proposes a defect-performance mapping model and a risk assessment and treatment priority algorithm. Based on characteristics such as defect type, location, and area, it quantitatively evaluates the actual impact of defects on the output power of components, and determines the maintenance priority order according to the degree of impact and safety risks, realizing the optimal allocation of operation and maintenance resources.

[0056] It should be understood that the intelligent operation and maintenance and fault warning method of this application is mainly applied to the operation and maintenance management of large-scale ground photovoltaic power stations and distributed photovoltaic systems. Its typical application scenarios include:

[0057] Large-scale ground photovoltaic power stations: Suitable for centralized ground photovoltaic power stations with an installed capacity greater than 10 MW. Such power stations usually have a large number of components (usually more than 100,000 components), a wide distribution area, and low efficiency of manual inspection. The method of this application can quickly obtain multi-band images of all components through a drone system equipped with a multi-spectral camera, and perform defect identification, evaluation, and prediction through a backend intelligent analysis system, realizing efficient and accurate fault diagnosis and warning.

[0058] Industrial and commercial rooftop distributed photovoltaic systems: Suitable for small and medium-sized photovoltaic systems installed on the rooftops of industrial factories and commercial buildings. Since such systems are installed on rooftops, there are safety risks in manual inspection. The method of this application can perform automatic inspection by drones, avoiding the risk of personnel working at heights, and at the same time realizing comprehensive detection of the system.

[0059] Photovoltaic power stations in remote areas: Suitable for photovoltaic power stations located in areas with inconvenient transportation such as deserts, wastelands, and mountains. It is difficult for maintenance personnel to reach such power stations. The method of this application can reduce the need for on-site maintenance personnel and provide accurate fault information and treatment suggestions through remote data analysis.

[0060] Photovoltaic systems in complex environments: Suitable for photovoltaic systems operating under complex environmental conditions (such as high temperature, high humidity, high salt mist, high sand and dust, etc.). The frequency of defects in such systems is high and the evolution speed is fast. The method of this application can track the defect evolution process by regularly collecting multi-spectral data and timely discover potential risks.

[0061] The implementation steps of the intelligent operation and maintenance and fault warning method for photovoltaic power stations proposed in this application are described in detail below:

[0062] Step S01: Multi-spectral data acquisition;

[0063] According to an embodiment of the present application, in this step, a drone system equipped with a multi-spectral camera is used to inspect the photovoltaic power station, and multi-band image data of the components is collected as the basis for subsequent analysis. The specific acquisition process includes the following sub-steps:

[0064] Step S01-1: Configuration of the multi-spectral camera system;

[0065] According to the implementation manner of the present application, an integrated multi-spectral camera system is adopted. The system includes a visible light camera (wavelength range 400 - 700nm), a near-infrared camera (wavelength range 750 - 1400nm), and a mid-infrared thermal imager (wavelength range 3000 - 5000nm). The three cameras are coaxially and fixedly installed on the drone mounting platform to ensure good spatial correspondence of different band images of the same target. It should be noted that the system is also equipped with a GPS / IMU module to record the precise position and attitude information of each image for geolocation and image stitching during subsequent data processing.

[0066] Step S01-2: Drone flight path planning;

[0067] In the embodiment of the present application, based on the photovoltaic power station layout information (including the arrangement mode, inclination angle, orientation, etc. of the components) and meteorological conditions, an adaptive path planning algorithm is used to generate the optimal flight path. The factors considered by this algorithm include: lighting conditions, photovoltaic module layout, drone performance parameters (such as endurance time, flight speed), and image overlap rate requirements, etc. The path planning results include parameters such as flight height, speed, flight route, and shooting point position, ensuring that the collected images cover all areas of the power station and have sufficient spatial resolution.

[0068] It should be understood that the determination of the flight height needs to meet the following conditions:

[0069]

[0070] Among them, H is the flight height (meters), GSD is the ground sampling distance (centimeters / pixel), W sensor is the sensor width (millimeters), f is the lens focal length (millimeters), W FOV is the field of view width (pixels).

[0071] Step S01-3: Multi-spectral image acquisition and storage;

[0072] According to an embodiment of the present application, an automatic inspection task is performed according to the planned flight path to maximize the guarantee of image quality and acquisition efficiency. Under good lighting conditions (irradiance greater than 600 W / m 2 ), images of three bands are collected: visible light images are used to identify component appearance defects (such as broken glass, deformed frames, etc.); near-infrared images are used to detect defects such as PID and hidden cracks that are not easily found under visible light; mid-infrared thermal images are used to detect thermal anomalies. In addition, metadata such as acquisition time, ambient temperature, and irradiance are recorded simultaneously to provide reference information for subsequent analysis.

[0073] Step S01-4: Image preprocessing;

[0074] In the embodiment of the present application, the multi-band images collected are preprocessed, including operations such as image correction, registration, and enhancement:

[0075] Image correction: Correct for problems such as lens distortion and uneven illumination, including geometric correction and radiometric correction.

[0076] Multi-spectral image registration: Since there may be differences in the field of view angle and resolution of cameras in different bands, the images of the three bands need to be accurately registered so that pixels at the same position correspond to the same physical target. The registration uses the phase cross-correlation algorithm, and the registration accuracy is better than 1 pixel.

[0077] Image enhancement: According to the characteristics of images in different bands, different enhancement algorithms are used. For thermal images, the temperature gradient enhancement algorithm is applied to highlight the temperature anomaly area; for visible light and near-infrared images, contrast enhancement and noise suppression algorithms are used to improve the image quality.

[0078] Image mosaicking: The single images are stitched according to the position information recorded by GPS / IMU to generate an orthophoto map covering the entire power station, providing a basis for subsequent global analysis.

[0079] Step S02: Composite defect identification;

[0080] According to an embodiment of the present application, in this step, based on the preprocessed multi-spectral image data, a multi-band image fusion algorithm and a deep learning model are applied to identify various defects in photovoltaic components, especially composite defects. The specific process includes the following sub-steps:

[0081] Step S02-1: Multi-band feature extraction;

[0082] In the embodiment of the present application, according to the characteristics of images in different bands, dedicated feature extraction networks are respectively used to extract the feature representations of each band:

[0083] Visible light feature extraction: An improved ResNet-50 network structure is adopted to extract the appearance features of components. On the basis of the standard structure, this network adds an attention mechanism module to enhance the feature extraction ability for subtle appearance defects (such as cell cracks and hidden cracks). The visible light feature can be expressed as:

[0084] f vis =φ vis (I vis )

[0085] Among them, f vis is the extracted visible light feature map, φ vis is the visible light feature extraction network, and I vis is the input visible light image.

[0086] Near-infrared feature extraction: A densely connected convolutional network (Dense-CNN) designed specifically for the characteristics of near-infrared images is adopted. This network has strong edge and texture feature extraction capabilities and is suitable for detecting defects such as PID and cell hidden cracks in photovoltaic modules. The near-infrared feature can be expressed as:

[0087] f nir =φ nir (I nir )

[0088] Among them, f nir is the extracted near-infrared feature map, φ nir is the near-infrared feature extraction network, and I nir is the input near-infrared image.

[0089] Mid-infrared thermal image feature extraction: An improved FPN (Feature Pyramid Network) structure is adopted. This structure has good detection capabilities for thermal anomalies of different scales and can simultaneously identify large-area hot spots and small-area hot spot defects. The mid-infrared feature can be expressed as:

[0090] f mir =φ mir (I mir )

[0091] Among them, f mir is the extracted mid-infrared feature map, φ mir is the mid-infrared feature extraction network, and I mir is the input mid-infrared image.

[0092] Step S02-2: Multi-band feature fusion;

[0093] According to the embodiments of the present application, the features extracted from three bands are fused to make full use of the complementarity of image information in different bands. The present application proposes an innovative multi-scale selective fusion (MSF) algorithm, which can adaptively select and fuse key features in different bands. The fusion process can be expressed as:

[0094] F = MSF(f vis , f nir , f mir )

[0095] The specific implementation of the MSF algorithm includes the following steps:

[0096] Feature channel alignment: Since the feature dimensions extracted by different networks may be different, first, the number of channels of each feature map is adjusted to the same value C through 1×1 convolution.

[0097] Feature space alignment: The spatial dimensions of all feature maps are adjusted to the same size H×W through bilinear interpolation.

[0098] Attention weight calculation: Calculate the fusion weights of different feature maps at each position (i, j):

[0099]

[0100] where g m is a fully connected layer for calculating feature importance, and w nir and w mir are calculated in a similar manner.

[0101] Weighted fusion: According to the calculated weights, the features of the three bands are weighted and fused:

[0102] F(i, j) = w vis (i, j)·f vis (i, j) + w nir (i, j)·f nir (i, j) + w mir (i, j)·f mir (i, j)

[0103] It should be noted that the fused feature map F contains the complementary information of the three-band images, providing a more comprehensive feature representation for subsequent defect recognition.

[0104] In some embodiments, optionally, a simplified feature fusion method is adopted, such as directly concatenating channels and then performing 1×1 convolution for feature fusion:

[0105] F = Conv 1×1 (Concat[f vis , f nir , fmir )

[0106] This simplified method has higher computational efficiency and is applicable to edge computing scenarios with limited computing resources.

[0107] In another alternative implementation, a feature fusion method based on channel attention can be adopted. This method first calculates the attention weights in the channel dimension and then performs weighted fusion on the features of different bands:

[0108]

[0109] Among them, Attention(·) is a channel attention calculation function that can adaptively adjust the importance weights of features in different bands.

[0110] Step S02-3: Temperature Gradient Enhancement Algorithm;

[0111] For the temperature anomaly region in the thermal image, this application proposes a Temperature Gradient Enhancement (TGE) algorithm to improve the detection sensitivity of thermal anomalies, especially for small thermal anomalies in a low temperature difference environment. The steps of the TGE algorithm are as follows:

[0112] Calculate the temperature field gradient: For the thermal image I mir Calculate the two-dimensional temperature gradient:

[0113]

[0114] Gradient magnitude calculation:

[0115]

[0116] Adaptive threshold calculation: Based on the temperature statistical characteristics of the component region, calculate the adaptive threshold:

[0117] T th = μ + α·σ

[0118] Among them, μ is the average temperature of the component region, σ is the temperature standard deviation, and α is an adaptive coefficient that is dynamically adjusted according to the environmental temperature and light conditions, and usually ranges from 1.5 to 3.0.

[0119] Calculation of enhanced thermal image features:

[0120]

[0121] Among them, β is the enhancement coefficient, and I(·) is an indicator function that takes the value 1 when the condition is met and 0 otherwise.

[0122] Optionally, for application scenarios under different light intensity conditions, the parameters of the temperature gradient enhancement algorithm can be adaptively adjusted. For example, under weak light conditions (irradiance < 400 W / m 2 ), a lower threshold coefficient α (usually taken as 1.2 to 1.8) can be adopted to improve the detection sensitivity to small temperature differences; while under strong light conditions (irradiance > 800 W / m 2 ), a higher threshold coefficient α (usually taken as 2.0 to 3.0) can be adopted to filter out environmental noise.

[0123] In a specific application example, for the hot spot detection scenario in the early morning low temperature environment, this algorithm adopts a low threshold coefficient of α = 1.5 and successfully identifies an early hot spot defect with a temperature difference of only 1.2 °C that cannot be detected by conventional methods. Such early hot spots are often ignored in traditional infrared detection, but after being processed by the temperature gradient enhancement algorithm, their features are significantly amplified, and the detection sensitivity of the system is increased by about 40%, providing an effective means for early detection of potential risks.

[0124] In another alternative implementation, time-domain enhancement can be combined with a multi-frame thermal image sequence to further improve the reliability of anomaly detection by calculating the temperature change rate:

[0125]

[0126] where I mir (i, j, t) represents the temperature value at position (i, j) at time t, and Δt is the sampling time interval. By combining spatial gradient and time gradient information, the system can more effectively distinguish real thermal anomalies from environmental interference.

[0127] Step S02-4: Composite defect recognition and classification;

[0128] According to the embodiments of the present application, based on the fused features and enhanced thermal image features, a composite defect recognition model is constructed. This model adopts a multi-task learning framework to simultaneously implement defect detection, segmentation, and classification tasks.

[0129] Defect detection and localization: An improved Faster R-CNN structure is adopted to detect the positions and types of defects in the image. The improvements include: introducing a feature pyramid structure to handle defects of different scales; adopting an attention mechanism to enhance the recognition ability of key defect regions; introducing a context information enhancement module to consider the features of the regions around the defects.

[0130] Defect fine segmentation: The DeepLab v3+ semantic segmentation structure is adopted to precisely segment the detected defect regions, generating a pixel-level defect mask for subsequent defect area and shape analysis.

[0131] Defect Classification: Based on the segmentation results and fusion features, defect feature vectors are extracted, and a multi-layer perceptron classifier is used to classify the defects into the following categories: hot spots, PID, hidden cracks, string welding fractures, glass breakage, backplane delamination, etc. The classification model outputs the probability distribution of each defect type.

[0132] Composite Defect Recognition: For multiple defects that may exist in the same area, a multi-label classification strategy is adopted, allowing multiple defect types to be assigned to the same area simultaneously, thereby realizing the recognition of composite defects. The recognition result of the composite defect can be expressed as:

[0133]

[0134] where Y is the set of labels assigned to the defect area, K is the number of detected defect types, and C is the total number of predefined defect categories.

[0135] Step S02-5: Construction of the Thermal Anomaly Pattern Library;

[0136] To address the interference of environmental factors such as shadows and reflections on the interpretation of thermal images, this application constructs a thermal anomaly pattern library to distinguish the thermal anomalies caused by real defects from the false hot spots caused by environmental factors.

[0137] Thermal Anomaly Feature Extraction: For each thermal anomaly area, the following features are extracted: temperature difference (relative to the surrounding normal area), area, shape features (roundness, eccentricity, etc.), edge gradient features, time stability features, etc.

[0138] Thermal Anomaly Pattern Clustering: An unsupervised learning method (such as the improved DBSCAN algorithm) is used to cluster the thermal anomaly features to form different thermal anomaly pattern categories.

[0139] Pattern Library Establishment and Update: Based on the clustering results and expert annotations, a thermal anomaly pattern library is established, and a defect type label or an environmental interference label is assigned to each pattern. The pattern library is continuously updated through incremental learning to adapt to the thermal anomaly patterns under different environmental conditions.

[0140] It should be understood that through the thermal anomaly pattern library, the system can effectively filter out the false hot spots caused by environmental factors such as shadows and reflections, and improve the accuracy of thermal anomaly detection.

[0141] Step S03: Analysis of Defect Interaction Effects;

[0142] According to an embodiment of the present application, in this step, a defect interaction effect model is established to quantitatively analyze the synergistic effect when multiple defects coexist, and evaluate the comprehensive impact of the composite defect on the performance of the photovoltaic module. The specific process includes the following sub-steps:

[0143] Step S03-1: Quantification of Defect Features;

[0144] In the embodiments of the present application, various identified defects are characterized and quantified, and key parameters related to performance impact are extracted:

[0145] Quantification of hot spot defect characteristics: Extract the maximum temperature T of the hot spot area max , temperature gradient area A hs , relative position P hs (relative coordinates on the component), and other parameters. The severity of the hot spot can be expressed as:

[0146]

[0147] where T avg is the average temperature of the component, T ref is the reference temperature difference (usually taken as 5°C), A module is the total area of the component, is the reference temperature gradient, and λ1, λ2, and λ3 are weight coefficients determined according to the characteristics of different types of hot spots.

[0148] Quantification of PID defect characteristics: Based on the PID characteristics in the near-infrared image, extract the area A of the PID-affected region pid , severity coefficient α pid (based on the change in gray value), and other parameters. The severity of PID can be expressed as:

[0149]

[0150] Quantification of crack defect characteristics: Extract the crack length L cr , width W cr , position P cr , direction θ cr , and other parameters. The severity of the crack can be expressed as:

[0151]

[0152] where L cell is the characteristic length of the cell, f(P cr , θ cr ) is the influence function considering the position and direction of the crack, and β1, β2, and β3 are weight coefficients.

[0153] Quantification of string soldering fracture characteristics: Extract the fracture position P sb , fracture degree γ sb , and other parameters. The severity of the fracture can be expressed as:

[0154] S sb = γ sb ·g(P sb )

[0155] Among them, g(P sb ) is a function considering the influence of the fracture position.

[0156] Step S03-2: Single defect performance influence model;

[0157] According to an embodiment of the present application, for each type of defect, an influence model of its impact on component performance parameters (such as output power, open-circuit voltage, short-circuit current, etc.) is established:

[0158] Hot spot influence model: Hot spots mainly affect the component fill factor FF and the parallel resistance R sh , and their relationship can be expressed as:

[0159] ΔFF hs =-k1·S hs

[0160] ΔR sh,hs =-k2·S hs ·R sh,0

[0161] Among them, ΔFF hs is the change in fill factor caused by hot spots, ΔR sh,hs is the change in parallel resistance caused by hot spots, R sh,0 is the parallel resistance of a normal component, and k1 and k2 are empirical coefficients.

[0162] PID influence model: PID mainly affects the component open-circuit voltage V oc and short-circuit current I sc , and their relationship can be expressed as:

[0163] ΔV oc,pid =-k3·S pid ·V oc,0

[0164] ΔI sc,pid =-k4·S pid ·I sc,0

[0165] Among them, ΔV oc,pid and ΔI sc,pid are the changes in open-circuit voltage and short-circuit current caused by PID respectively, V oc,0 and I sc,0 are the open-circuit voltage and short-circuit current of a normal component respectively, and k3 and k4 are empirical coefficients.

[0166] Hidden crack influence model: Hidden cracks mainly affect the short-circuit current I sc , and their relationship can be expressed as:

[0167] ΔI sc,cr =-k5·Scr ·I sc,0

[0168] Among them, ΔI sc,cr is the short - circuit current change caused by the hidden crack, and k5 is an empirical coefficient.

[0169] String soldering fracture influence model: String soldering fracture mainly affects the series resistance R s , and its relationship can be expressed as:

[0170] ΔR s,sb = k6·S sb ·R s,0

[0171] Among them, ΔR s,sb is the series resistance change caused by the string soldering fracture, R s,0 is the series resistance of the normal component, and k6 is an empirical coefficient.

[0172] Step S03 - 3: Defect interaction influence model construction;

[0173] In the embodiments of the present application, when different types of defects co - exist, their influence on the component performance is not a simple superposition, but there are complex interactions. The present application proposes a defect interaction influence model, considering the synergistic effect between defects:

[0174] Synergistic influence function definition: Define the synergistic influence function Φ(D i , D j ) between defect types i and j, representing the additional influence generated when two defects co - exist.

[0175] Dual - defect synergistic model: For the case where hot spot and PID co - exist, their synergistic influence can be expressed as:

[0176] Φ(D hs , D pid ) = δ hs,pid ·S hs ·S pid

[0177] Among them, δ hs,pid is the synergistic coefficient of hot spot and PID, obtained by regression of experimental data.

[0178] Multi - defect synergistic influence calculation: For the case where n kinds of defects co - exist, their comprehensive influence can be expressed as:

[0179]

[0180] Among them, Impact(D i ) is the individual influence of defect i, Φ(D i , D j) is the co - influence of double defects i and j, ξ(D1, D2,..., D n ) is the high - order co - effect term, representing the additional influence when three or more defects co - exist.

[0181] Step S03 - 4: Comprehensive calculation of performance parameters;

[0182] According to the embodiments of the present application, based on the defect interaction influence model, calculate the comprehensive influence of composite defects on the key performance parameters of the component:

[0183] Calculate the comprehensive influence of composite defects on the fill factor FF:

[0184]

[0185] Among them, FF0 is the fill factor of the defect - free component, and ΔFF i is the change in the fill factor caused by defect i, and ΔFF Φ(i,j) is the additional change in the fill factor caused by the co - effect of defects i and j.

[0186] Similarly, calculate the comprehensive influence of composite defects on the open - circuit voltage V oc , short - circuit current I sc , series resistance R s and shunt resistance R sh .

[0187] Based on the corrected performance parameters, calculate the output power of the component:

[0188] P = FF·V oc ·I sc

[0189] Step S03 - 5: Visualization of defect influence;

[0190] To facilitate the understanding of the interaction influence of composite defects, the present application constructs a defect influence visualization module:

[0191] Component performance heat map: Use pseudo - colors to represent the degree of performance impairment in each area of the component, and intuitively display the impairment conditions in different areas.

[0192] Defect network diagram: Show the interaction relationship between different defects in the form of a network diagram, where nodes represent defects and connection lines represent the intensity of interaction influence.

[0193] Defect co - influence matrix: Construct an n×n matrix to represent the co - influence intensity between n defects. The diagonal elements represent the influence of single defects, and the non - diagonal elements represent the co - influence of two defects.

[0194] It should be understood that through defect interaction impact analysis, the system can accurately evaluate the comprehensive impact of compound defects on component performance, providing a scientific basis for subsequent maintenance decisions.

[0195] Step S04: Defect evolution prediction;

[0196] According to an embodiment of the present application, in this step, based on time-series multi-spectral data, a defect evolution prediction algorithm is constructed to establish a growth model for defect size and severity, and predict the defect development trend. The specific process includes the following sub-steps:

[0197] Step S04-1: Construction of time-series defect database;

[0198] In the implementation manner of the present application, a time-series defect database is constructed to record the same defect feature data collected at different time points, providing a basis for defect evolution analysis:

[0199] Defect time-series tracking: Based on geographical coordinates and defect features, the same defect detected at different time points is matched to establish a time series of the defect. The matching uses a comprehensive score of feature similarity and spatial position similarity:

[0200] S match = α·S feature +(1 - α)·S position

[0201] Among them, S feature is the defect feature similarity, S position is the spatial position similarity, and α is the weight coefficient (usually taken as 0.6 - 0.8). The matching threshold is set to 0.85, and observations at different time points exceeding this threshold are regarded as those of the same defect.

[0202] Optionally, in some implementation manners, defect time-series tracking can adopt deep learning methods. By training a defect feature matching network, the matching relationship of defects at different time points can be directly learned. This method can still maintain a high tracking accuracy even when the appearance of the defect changes greatly. The structure of the matching network can adopt the Siamese Network architecture, and it judges whether it is the same defect by comparing the similarity of the defect feature vectors at two time points:

[0203] M(D1, D2)=σ(F θ (D1)·F θ (D2))

[0204] Among them, F θ is the feature extraction network, σ is the sigmoid activation function, and M(D1, D2) represents the matching probability of defects D1 and D2.

[0205] Defect Feature Time Series Record: For each traced defect, record its characteristic parameters at each time point, including information such as defect type, area, severity, location, etc., as well as environmental parameters during acquisition (such as environmental temperature, irradiance, etc.).

[0206] Preliminary Analysis of Defect Evolution: Conduct a preliminary analysis on the time series data, calculate the change rates of defect characteristic parameters, such as area growth rate, severity change rate, etc., to provide input for the subsequent evolution model.

[0207] Step S04-2: Construction of Defect Growth Model;

[0208] According to the embodiments of the present application, based on the defect time series data, construct growth models for different types of defects to achieve a mathematical expression of the defect evolution process:

[0209] Hot Spot Growth Model: The area A hs (t) and temperature difference ΔT hs (t) of the hot spot usually follow an exponential model:

[0210]

[0211] where A hs (t0) and ΔT hs (t0) are the hot spot area and temperature difference at the initial observation, r A and r T are the growth rate parameters of the area and temperature difference, and t and t0 are the predicted time point and the initial observation time point respectively.

[0212] PID Evolution Model: The growth of the area A pid (t) of the PID influence region usually follows a Sigmoid model:

[0213]

[0214] where A max is the maximum possible influence area of the PID (usually the total area of the component), k is the growth rate parameter, and t m is the growth midpoint time.

[0215] In some embodiments, optionally, a piecewise growth model is adopted according to the environmental conditions in different seasons to more accurately describe the evolution law of PID under different environmental conditions. For example, in high-temperature and high-humidity seasons (such as summer), the PID expansion rate is usually faster, and a higher k value can be adopted; while in low-temperature and low-humidity seasons (such as winter), the PID expansion rate slows down, and a lower k value can be adopted. Specifically, the piecewise PID evolution model can be expressed as:

[0216] When t is in the i-th season;

[0217] Among them, k i is the growth rate parameter of the i-th season, which is determined according to the typical temperature and humidity conditions of this season.

[0218] Hidden crack evolution model: The growth of the hidden crack length L cr (t) generally follows a piecewise linear or power-law model:

[0219] L cr (t) = L cr (t0) + β · (t - t0) γ

[0220] Among them, L cr (t0) is the hidden crack length at the initial observation, β and γ are model parameters. When γ = 1, it is linear growth; when γ > 1, the growth accelerates; when 0 < γ < 1, the growth slows down.

[0221] In a specific application example, for the monitoring of the hidden crack evolution of components in a large-scale photovoltaic power station in a desert area, through comparative analysis, it is found that the hidden crack expansion rate in the area with severe temperature cycling is significantly higher than that in the area with stable temperature. Specifically, in the area where the daily temperature difference exceeds 25°C, the growth of the hidden crack length conforms to the acceleration model (γ ≈ 1.4), while in the area where the daily temperature difference is less than 15°C, the hidden crack growth is close to the linear model (γ ≈ 1.1). Based on this observation, the system sets different warning thresholds for hidden cracks in different temperature environment areas, significantly improving the accuracy of early warning.

[0222] Adaptive estimation of model parameters: By fitting historical time series data, the least squares method is used to estimate each model parameter. To improve the adaptability of the model, an environmental factor adjustment term is introduced:

[0223]

[0224] Among them, r is the adjusted growth rate parameter, r0 is the benchmark growth rate, E i is the environmental factor (such as temperature, humidity, irradiance, etc.), f i is the environmental factor influence function, η i is the weight coefficient.

[0225] Optionally, in another implementation, a deep learning method based on a recurrent neural network (RNN) or a long short-term memory network (LSTM) can be used to construct a defect evolution model. This method can automatically learn the complex patterns in the defect time series data and is applicable to defect types with non-linear development laws. The deep learning model can be expressed as:

[0226] X t+1 = f θ (X t , X t-1 ,..., Xt-n , E t )

[0227] Among them, X t represents the defect feature vector at time t, and E t is the environmental parameter vector, and f θ is a deep learning model with parameter θ, and n is the number of historical time steps considered.

[0228] Step S04-3: Evolution trajectory prediction;

[0229] In the embodiments of the present application, based on the constructed defect growth model, the future evolution trajectory of the defect is predicted to form a defect evolution curve:

[0230] Short-term prediction: Based on the latest observation data and current environmental conditions, predict the defect evolution trend in the next 7-30 days, and generate a time series of defect characteristic parameters (such as area, severity).

[0231] Medium- and long-term prediction: Combine the seasonal environmental change model to predict the defect evolution trend in the next 3-12 months, mainly focusing on the change in defect severity.

[0232] Prediction uncertainty assessment: Use the Monte Carlo simulation method to generate multiple possible evolution trajectories by adding random perturbations to the model parameters, and calculate the confidence interval of the prediction result:

[0233] [F(t) - z α / 2 ·σ(t), F(t) + z α / 2 ·σ(t)]

[0234] Among them, F(t) is the average trajectory of the predicted value, σ(t) is the standard deviation of the predicted value, and z α / 2 is the quantile of the standard normal distribution corresponding to the confidence level (1.96 is taken for a 95% confidence interval).

[0235] Step S04-4: Critical time point prediction;

[0236] According to the embodiments of the present application, based on the defect evolution trajectory, the critical time points in the defect development process are predicted to provide a time reference for maintenance decisions:

[0237] Threshold crossing time prediction: Predict the time point when the defect characteristic parameter reaches a preset threshold. For example, the hot spot area reaches 5% of the component area, the temperature difference reaches 20°C, etc.

[0238] Performance degradation threshold prediction: Predict the time point when the defect causes the component performance to decline to a specific threshold, such as the output power drops by 10%, 20%, etc.

[0239] Safety risk threshold prediction: Predict the time point when a defect develops into a safety risk, such as when the hot spot temperature difference reaches the dangerous temperature threshold (usually 50°C).

[0240] It should be noted that the key time point prediction uses the interpolation method. Based on the predicted evolution curve, calculate the time point t when the parameter reaches the threshold threshold :

[0241]

[0242] where V threshold is the threshold, V(t i ) and V(t i+1 ) are the parameter values at adjacent predicted time points, and t i and t i+1 are the corresponding time points.

[0243] Step S04-5: Visualization of defect evolution;

[0244] In the embodiments of the present application, a defect evolution visualization module is constructed to intuitively display the historical development and future trends of defects:

[0245] Time series change curve: Plot the change curves of defect characteristic parameters (such as area, severity) over time, including historical observation data points and predicted curves.

[0246] Spatial expansion animation: For spatially expanding defects (such as PID, hot spots, etc.), generate dynamic visualizations of the defect area expanding over time to intuitively display the defect expansion process.

[0247] Component performance impact curve: Plot the time change curves of the impact of defects on component performance, such as the output power decline rate, fill factor change rate, etc., to facilitate the evaluation of the long-term impact of defects on power generation.

[0248] It should be understood that through defect evolution prediction, the system can identify potential risks in advance, support forward-looking maintenance decisions, and avoid major losses caused by defects developing to a severe stage.

[0249] Step S05: Loss assessment and maintenance decision-making;

[0250] In this step, a defect-performance mapping model is developed to evaluate the actual impact of defects on the output power of components based on defect types, locations, areas, etc.; a risk assessment and treatment priority algorithm is established to guide precise maintenance decisions and achieve optimal allocation of maintenance resources. The specific process includes the following sub-steps:

[0251] Step S05-1: Construction of the defect-performance mapping model;

[0252] Establish the mapping relationship between defect features and component performance parameters, and quantify the impact of defects on the output power of components:

[0253] Defect feature vector construction: Integrate the key features of defects into the feature vector X, including parameters such as defect type T, severity S, location P, area A, development stage E, etc.:

[0254] X = [T, S, P, A, E,...]

[0255] Among them, the defect type T is represented by one-hot encoding, and other features are appropriately normalized.

[0256] Performance impact regression model: Based on historical defect data and corresponding measured performance parameters, train a defect-performance mapping model to predict the impact of defects on component performance. The model uses an ensemble learning method, combining multiple regression algorithms (such as gradient boosting trees, random forests, neural networks, etc.) to improve prediction accuracy:

[0257]

[0258] Among them, ΔP is the power loss ratio, F i is the i-th basic regression model, w i is the corresponding weight coefficient, and m is the number of basic models.

[0259] Calculation of annual power generation loss: Based on the power loss ratio, combined with the historical irradiance data of the photovoltaic power station and the typical meteorological year data of the power station location, calculate the annual power generation loss caused by defects:

[0260]

[0261] Among them, ΔE is the annual power generation loss (kWh), P rated is the rated power of the component (kW), η is the system efficiency, I i,j is the irradiance at the i-th day and j-th hour (kW / m 2 ), and Δt is the time interval (usually 1 hour).

[0262] Economic loss assessment: Based on the annual power generation loss and electricity price information, calculate the economic loss caused by defects:

[0263] L = ΔE·C electricity

[0264] Among them, L is the annual economic loss (yuan), C electricity is the electricity price (yuan / kWh).

[0265] Step S05-2: Construction of a risk assessment model;

[0266] In addition to economic losses, establish a security risk assessment model to evaluate the potential risks of defects to system security:

[0267] Definition of security risk levels: According to the defect type and severity, define the security risk levels (levels 1-5, the higher the value, the greater the risk):

[0268] Level 1: Minor defects, no security risks, only slightly affecting performance;

[0269] Level 2: Slight defects, no security risks in the short term, performance impact observable;

[0270] Level 3: Medium defects, may develop into security risks in the medium to long term, significant performance degradation;

[0271] Level 4: Severe defects, security hazards exist in the short term, need to be dealt with promptly;

[0272] Level 5: Critical defects, with urgent security risks, need to be dealt with immediately;

[0273] Calculation of security risk score: Based on defect characteristics and evolution prediction results, calculate the security risk score:

[0274] R safety = σ(w1·f1(T) + w2·f2(S) + w3·f3(E) + w4·f4(P) + w5·f5(G))

[0275] Where σ is the Sigmoid function, used to map the result to the interval [0,1]; f1 to f5 are the risk contribution functions of defect type T, severity S, development stage E, location P, and growth trend G respectively; w1 to w5 are the corresponding weight coefficients.

[0276] Determination of security risk level: According to the security risk score, determine the security risk level L of the defect safety :

[0277]

[0278] Where represents rounding up x.

[0279] Step S05-3: Maintenance priority algorithm;

[0280] Based on the economic loss assessment and security risk assessment results, develop a maintenance priority algorithm to determine the priority order of defect handling:

[0281] Calculation of priority comprehensive score: Considering the economic loss L, security risk level L safety and handling difficulty D, calculate the maintenance priority score:

[0282]

[0283] Among them, L max is the maximum possible economic loss, D max is the maximum processing difficulty, and α, β, and γ are weight coefficients, satisfying α + β + γ = 1. Usually, α = 0.3, β = 0.6, and γ = 0.1 are taken to reflect the dominant position of safety risks in decision-making.

[0284] Priority classification: According to the priority score P, the defects are divided into four treatment priority levels:

[0285] Emergency treatment: P ≥ 0.8, and repair needs to be arranged immediately;

[0286] High priority: 0.6 ≤ P < 0.8, and it needs to be treated in the near future (within 1 - 2 weeks);

[0287] Medium priority: 0.4 ≤ P < 0.6, and it needs to be treated in the medium term (within 1 month);

[0288] Low priority: P < 0.4, and it can be treated during routine maintenance;

[0289] Optimal allocation of maintenance resources: Based on defect priority, geographical location, and maintenance cost, cluster and optimize the scheduling of maintenance tasks to generate an optimal maintenance plan:

[0290]

[0291] Meet the constraint conditions:

[0292]

[0293] Among them, c i is the cost of the i-th maintenance plan, x i is the corresponding decision variable (1 means selected, 0 means not selected), U j is the set of all possible maintenance plans for the j-th defect, n is the total number of plans, and m is the number of defects.

[0294] Step S05 - 4: Generation of maintenance suggestions;

[0295] Based on maintenance priority and defect characteristics, generate specific maintenance suggestions:

[0296] Recommendation of treatment methods: According to defect type, severity, and location, recommend appropriate treatment methods, including component replacement, isolation treatment, local repair, etc. Specifically, a rule-based expert system is recommended for the recommendation, and rule examples are as follows:

[0297] IF the hot spot area > 10% AND the temperature difference > 30°C THEN recommend component replacement;

[0298] IF the PID area > 30% AND the power decline > 15%, then it is recommended to isolate and perform PID recovery processing;

[0299] IF the length of the hidden crack < 5 cm AND the power decline < 5%, then it is recommended to continue observing;

[0300] Suggestions on the processing timing: Combine the defect evolution prediction results and maintenance priorities to give suggestions on the best processing timing, considering factors such as weather conditions and power station operation plans.

[0301] Economic benefit analysis: Calculate the input-output ratio of the maintenance process, including maintenance costs, expected benefits (avoided losses), and payback period:

[0302]

[0303] Among them, ROI is the return on investment, L is the annual economic loss, T remaining is the remaining life of the component (years), C repair is the maintenance cost.

[0304] Step S05-5: Maintenance decision support system;

[0305] Develop a maintenance decision support system, integrate the above analysis results, and provide an intuitive decision support interface:

[0306] Defect map: Intuitively display the defect location, type, and priority on the power station geographic information system, and use different colors to mark defects with different priorities.

[0307] Decision dashboard: Display key decision-making indicators, such as overall health status, the number of high-risk defects, expected economic losses, etc.

[0308] Maintenance plan generation: Automatically generate a maintenance plan based on the priority algorithm and the result of resource optimization allocation, including the processing sequence, required resources, and expected completion time.

[0309] Decision adjustment interface: Provide a human-computer interaction interface that allows operation and maintenance personnel to adjust maintenance decisions according to the actual situation on site. The system records the reasons for the adjustment and updates the decision model.

[0310] Through the loss assessment and maintenance decision-making steps, the system realizes the closed-loop management from defect identification to maintenance decision-making, ensures the optimal allocation of photovoltaic power station operation and maintenance resources, maximizes the maintenance benefits, and extends the service life of the power station.

[0311] To implement the above method, the present application also proposes a photovoltaic power station intelligent operation and maintenance and fault warning device. The device includes a plurality of function modules connected to each other, corresponding one by one to the above method steps. The device includes:

[0312] The multi - spectral data acquisition module is used to inspect a photovoltaic power station by using an unmanned aerial vehicle (UAV) system equipped with a multi - spectral camera, and collect multi - band image data of components; the multi - spectral data acquisition module includes a multi - spectral camera system configuration unit, a UAV flight path planning unit, a multi - spectral image acquisition and storage unit, and an image pre - processing unit;

[0313] The composite defect identification module is used to identify various defects in photovoltaic components based on the pre - processed multi - spectral image data by applying a multi - band image fusion algorithm and a deep learning model; the composite defect identification module includes a multi - band feature extraction unit, a multi - band feature fusion unit, a temperature gradient enhancement algorithm unit, a composite defect identification and classification unit, and a thermal anomaly pattern library construction unit;

[0314] The defect interaction impact analysis module is used to establish a defect interaction impact model, quantitatively analyze the synergistic effect when multiple defects co - exist, and evaluate the comprehensive impact of composite defects on the performance of photovoltaic components; the defect interaction impact analysis module includes a defect feature quantification unit, a single - defect performance impact model unit, a defect interaction impact model construction unit, a performance parameter comprehensive calculation unit, and a defect impact visualization unit;

[0315] The defect evolution prediction module is used to develop a defect evolution prediction algorithm based on time - series multi - spectral data, establish a growth model for the size and severity of defects, and predict the development trend of defects; the defect evolution prediction module includes a time - series defect database construction unit, a defect growth model construction unit, an evolution trajectory prediction unit, a key time - point prediction unit, and a defect evolution visualization unit;

[0316] The loss assessment and maintenance decision - making module is used to develop a defect - performance mapping model, evaluate the actual impact of defects on the output power of components, and establish a risk assessment and treatment priority algorithm to guide accurate maintenance decisions; the loss assessment and maintenance decision - making module includes a defect - performance mapping model construction unit, a risk assessment model construction unit, a maintenance priority algorithm unit, a maintenance recommendation generation unit, and a maintenance decision - making support system unit.

[0317] In a possible implementation manner, the multi - band feature fusion unit adopts a multi - scale selective fusion algorithm, which can adaptively select and fuse key features in different bands, and the fusion process includes steps of feature channel alignment, feature space alignment, attention weight calculation, and weighted fusion.

[0318] In a possible implementation manner, the temperature gradient enhancement algorithm unit includes functions of temperature field gradient calculation, gradient amplitude calculation, adaptive threshold calculation, and enhanced thermal image feature calculation, and is used to improve the detection sensitivity of thermal anomalies, especially for small thermal anomalies in a low - temperature difference environment.

[0319] In a possible implementation manner, the defect interaction impact model construction unit defines a collaborative impact function, representing the additional impact generated when two defects coexist, and evaluates the comprehensive impact when n defects coexist through a multi-defect collaborative impact calculation formula.

[0320] In a possible implementation manner, the defect growth model construction unit constructs a hot spot growth model, a PID evolution model, and a hidden crack evolution model for different types of defects respectively, and optimizes the model parameters through an adaptive parameter estimation method.

[0321] In a possible implementation manner, the maintenance priority algorithm unit determines the priority order of defect handling through three steps: priority comprehensive scoring calculation, priority grading, and optimized allocation of maintenance resources, so as to realize the optimized allocation of maintenance resources.

[0322] This application also proposes a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the above-mentioned various method steps are implemented.

[0323] This computer device includes:

[0324] A processor, used to execute various functional applications and data processing, such as running an instruction set stored in the memory to implement each step of the above method;

[0325] A memory, used to store various types of data, such as application programs and data required to implement the above method, and provide a working space for the operation of the processor;

[0326] A communication interface, used to communicate with external devices (such as an unmanned aerial vehicle multispectral camera system, a power station monitoring system);

[0327] A bus, used to connect the processor, the memory, and the communication interface, enabling communication between components.

[0328] The computer device of this application is configured with a dedicated graphics processing unit (GPU) and an artificial intelligence processing unit (AI accelerator), used to accelerate the operation of deep learning models and improve the processing efficiency of core algorithms such as defect recognition, interaction impact analysis, and evolution prediction.

[0329] The memory in the computer device of this application stores multiple functional modules, including: a multispectral data processing module, a composite defect recognition module, a defect interaction impact analysis module, a defect evolution prediction module, and a loss assessment and maintenance decision module. These modules are all called and executed by the processor through a program to implement each step of the above method.

[0330] It should be noted that those skilled in the art can understand that the methods and systems of the above embodiments can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solutions of the above embodiments, in essence, or the parts that contribute to the prior art can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, or optical disc, etc., and includes several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of various embodiments or certain parts of the embodiments of the present invention.

Claims

1. A photovoltaic power station intelligent operation and maintenance and fault warning method, characterized in that: The following steps are involved: Use a drone system equipped with a multispectral camera to inspect photovoltaic power plants and collect multispectral image data including visible light, near infrared and mid-infrared bands; Based on the multispectral image data, a multi-band image fusion algorithm and a temperature gradient enhancement algorithm are applied to identify composite defects in photovoltaic modules; Establish a defect interaction model, quantify and analyze the synergistic effect of different types of defects when they coexist, and evaluate the comprehensive impact of compound defects on the performance of photovoltaic modules; Based on time-series multispectral data, a defect evolution prediction algorithm is constructed, a defect severity growth model is established, and defect development trends are predicted; Develop a defect-performance mapping model and a risk assessment and processing priority algorithm to quantitatively evaluate the actual impact of defects on component output power based on defect type, location and area characteristics, determine maintenance priorities based on the degree of impact and safety risks, and achieve optimal allocation of operation and maintenance resources.

2. The photovoltaic power station intelligent operation and maintenance and fault warning method according to claim 1 is characterized in that: The acquisition of the multispectral image data includes: Configure an integrated multispectral camera system, including a visible light camera, a near infrared camera, and a mid-infrared thermal imager; Based on the photovoltaic power plant layout information and meteorological conditions, an adaptive path planning algorithm is used to generate the optimal flight path for the UAV; Perform automatic inspection tasks according to the planned flight path, collect images in three bands and record the acquisition time, ambient temperature and irradiance metadata; The collected multi-band images are preprocessed, including image correction, multispectral image registration, image enhancement and image mosaicking.

3. The photovoltaic power station intelligent operation and maintenance and fault warning method according to claim 1, characterized in that: The multi-band image fusion algorithm includes: According to the characteristics of images in different bands, a dedicated feature extraction network is used to extract the feature representation of each band; Through feature channel alignment and feature space alignment, the features of each band are adjusted to the same dimension and spatial size; Calculate the fusion weights of different feature maps at each position and perform adaptive weighting based on feature importance; According to the calculated weights, the features of the three bands are weighted fused to generate a fused feature map containing multi-band complementary information.

4. The photovoltaic power station intelligent operation and maintenance and fault warning method according to claim 1, characterized in that: The temperature gradient enhancement algorithm includes: Calculate the two-dimensional temperature gradient and gradient amplitude for the thermal image; Calculate the adaptive threshold based on the statistical characteristics of the component area temperature; Based on the calculated temperature gradient and adaptive threshold, enhanced thermal image features are generated to improve the detection sensitivity of thermal anomalies.

5. The photovoltaic power station intelligent operation and maintenance and fault warning method according to claim 1, characterized in that: The establishment of the defect interaction impact model includes: Quantify the characteristics of various identified defects and extract key parameters related to performance impact; For each defect type, a single impact model on component performance parameters is established; Define the synergistic impact function between defects, which represents the additional impact when two defects coexist; Based on the single defect impact and synergistic impact function, the comprehensive impact of multiple defects on the key performance parameters of the component is calculated.

6. The photovoltaic power station intelligent operation and maintenance and fault warning method according to claim 1, characterized in that: The defect evolution prediction algorithm includes: Build a time series defect database to record the same defect feature data collected at different time points; Construct growth models for different types of defects to achieve mathematical expression of the defect evolution process; Based on the constructed growth model, the future evolution trajectory of defects is predicted to form a defect evolution curve; Predict critical time points in the defect development process and provide time reference for maintenance decisions.

7. The photovoltaic power station intelligent operation and maintenance and fault warning method according to claim 6, characterized in that: The defect growth model includes: Exponential growth model of hot spot area and temperature difference; Sigmoid growth model of the area affected by PID; Piecewise linear or power-law growth models of crack length; Adaptation to environmental factors is achieved through adaptive estimation of model parameters.

8. The photovoltaic power station intelligent operation and maintenance and fault warning method according to claim 1, characterized in that: The defect-performance mapping model and risk assessment and processing priority algorithm include: Integrate the key features of defects into feature vectors, including defect type, severity, location, area and development stage parameters; Based on historical defect data and corresponding measured performance parameters, a defect-performance mapping model is trained to predict the impact of defects on component performance; Annual power generation loss and economic losses caused by calculation defects; Establish a security risk assessment model to evaluate the potential risks of defects to system security; Taking into account the economic losses, safety risk level and handling difficulty, the maintenance priority score is calculated to determine the priority of defect handling.

9. The photovoltaic power station intelligent operation and maintenance and fault warning method according to claim 8, characterized in that: The maintenance priority algorithm also includes: According to the priority score, defects are divided into four processing priority levels: emergency, high priority, medium priority and low priority; Cluster and optimize the scheduling of maintenance tasks based on defect priority, geographic location and maintenance cost to generate the optimal maintenance plan; Recommend appropriate treatment methods based on defect type, severity and location, including component replacement, isolation treatment or local repair; Calculate the cost-effectiveness of maintenance treatments, including repair costs, expected benefits, and payback period.

10. A photovoltaic power station intelligent operation and maintenance and fault warning system, characterized in that: include: Multispectral data acquisition module, used to inspect photovoltaic power plants using a drone system equipped with a multispectral camera to collect multi-band image data of components; Composite defect recognition module, which is used to identify various defects in photovoltaic modules based on pre-processed multispectral image data and apply multi-band image fusion algorithm and deep learning model; Defect interaction impact analysis module, which is used to establish a defect interaction impact model, quantify and analyze the synergistic effect of multiple defects coexisting, and evaluate the comprehensive impact of compound defects on the performance of photovoltaic modules; Defect evolution prediction module, which is used to develop defect evolution prediction algorithms based on time-series multispectral data, establish growth models for defect size and severity, and predict defect development trends; The loss assessment and maintenance decision module is used to develop a defect-performance mapping model, evaluate the actual impact of defects on component output power, and establish a risk assessment and processing priority algorithm to guide accurate maintenance decisions.

Images