Deep learning based distributed photovoltaic module hot spot fault detection method

Abstract

The application discloses a distributed photovoltaic module hot spot fault detection method based on deep learning, and relates to the technical field of photovoltaic module fault detection, and comprises the following steps: acquiring infrared thermal imaging frames collected in multiple inspection cycles of the same photovoltaic module, and deducting the temperature component to obtain the residual temperature rise field of the infrared thermal imaging frames in each inspection cycle; identifying precursor lines composed of linearly extended temperature rises, and determining the extension direction, centroid and micro-temperature rise intensity thereof; establishing cross-cycle correlation for the precursor lines whose centroid and extension direction meet a preset correlation condition in different inspection cycles, and forming an evolution causal chain diagram; determining true precursors according to the proportion of positive difference value directed edges and the consistency degree of the direction of the centroid migration vector; determining a precursor degradation image based on a temperature rise rate sequence and an expansion rate sequence, and marking the early warning trend diagram of the photovoltaic module output; and the application can reliably distinguish true hidden crack precursors and give graded early warnings in the early stage of weak temperature rise and numerous interferences.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic module fault detection technology, specifically a method for detecting hot spot faults in distributed photovoltaic modules based on deep learning. Background Technology

[0002] During long-term outdoor operation, photovoltaic modules may experience localized overheating at current mismatch points due to factors such as microcracks in the cells and localized aging. These hot spots can lead to a decrease in the power generation efficiency of the modules and, in severe cases, cause burnout or even fire. Timely detection of hot spot hazards is of great significance for the safe operation and maintenance of photovoltaic power plants.

[0003] Existing hot spot detection methods mostly rely on manual inspection of each hot spot using infrared thermal imagers, or on the deployment of drones or fixed cameras to collect infrared thermal images and then interpret them. Based on this, image segmentation or deep learning methods are further used to identify and locate areas in a single infrared image where the temperature is significantly higher than the surrounding area. Areas with temperatures exceeding a set level are identified as hot spots and alarms are triggered.

[0004] However, defects such as microcracks often take a long time to develop from their inception into hot spots with significantly higher temperatures. In the early stages of this process, the temperature rise at the defect is very small, and the temperature difference with the normal area is not obvious. At the same time, in infrared images taken outdoors, factors such as cloud shadows, bird droppings, and dust accumulation can also cause local temperatures to be higher, and their appearance in the image is similar to that of real defects.

[0005] Therefore, in the early stages when defects are still in the stage of slight temperature rise and there are multiple similar interferences in the image, how to distinguish the real defects from the background and interference, and thus detect them before they develop into obvious hot spots, has become an urgent technical problem to be solved. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a deep learning-based method for detecting hot spot faults in distributed photovoltaic modules. This method can distinguish real defects from background and interference in the early stages when the defect is still in the early stages and there are multiple similar interferences in the image.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] In a first aspect, the present invention discloses a method for detecting hot spot faults in distributed photovoltaic modules based on deep learning, comprising the following steps:

[0009] Infrared thermal imaging frames of the same photovoltaic module were acquired in multiple inspection cycles. The temperature component determined by the acquisition environment was subtracted from the infrared thermal imaging frames of each inspection cycle to obtain the residual temperature rise field.

[0010] For the remaining temperature rise field of each inspection cycle, identify the precursors formed by linear temperature rises, and determine the extension direction, centroid and micro-temperature rise intensity of each precursor.

[0011] For precursor patterns whose centroid and extension direction meet the preset correlation conditions in different inspection cycles, they are determined to belong to the same precursor and cross-cycle correlation is established.

[0012] For the same precursor, the intensity of its micro-temperature rise in each inspection cycle is set as a node. Directed edges carrying the difference in micro-temperature rise intensity between adjacent nodes are established from early to late according to the inspection cycle to form the evolutionary causal chain diagram of the precursor. The difference between the centroids of the precursor in adjacent inspection cycles is taken as the centroid migration vector.

[0013] Calculate the proportion of directed edges with positive differences in the evolutionary causal chain graph, as well as the degree of consistency between the direction of each centroid migration vector and the extension direction. Precursors with a proportion higher than a preset cumulative threshold and a degree of consistency higher than a preset consistency threshold are judged as true precursors.

[0014] For true precursors, the heating rate sequence and the expansion rate sequence are determined based on the difference in micro-temperature rise intensity of their directed edges and the centroid migration vector, respectively.

[0015] Based on the heating rate sequence and the expansion rate sequence, a precursor degradation profile of true precursors is determined. The precursor degradation profile includes degradation mode labels, degradation urgency, and self-aggravation markers.

[0016] The centroid, extension direction, and degradation profile of the precursors are labeled onto the photovoltaic modules, and an early warning trend chart is output.

[0017] Secondly, this invention discloses a deep learning-based distributed photovoltaic module hot spot fault detection system, comprising:

[0018] The data acquisition module is used to acquire infrared thermal imaging frames of the same photovoltaic module in multiple inspection cycles. The temperature component determined by the acquisition environment factors is subtracted from the infrared thermal imaging frames of each inspection cycle to obtain the remaining temperature rise field.

[0019] The feature extraction module is used to identify the precursors of temperature rise consisting of linearly extending temperature rises in the remaining temperature rise field of each inspection cycle, and to determine the extension direction, centroid, and micro-temperature rise intensity of each precursor.

[0020] The cross-cycle correlation module is used to determine that the precursor patterns whose centroid and extension direction meet the preset correlation conditions in different inspection cycles belong to the same precursor and establish cross-cycle correlation.

[0021] The causal graph construction module is used to set the micro-temperature rise intensity of the same precursor in each inspection cycle as a node, and establish directed edges between adjacent nodes from early to late according to the inspection cycle to carry the difference in micro-temperature rise intensity between two adjacent nodes, thus forming the evolutionary causal chain graph of the precursor, and taking the difference of the centroid of the precursor in adjacent inspection cycles as the centroid migration vector.

[0022] The true precursor determination module is used to calculate the proportion of directed edges with positive difference values ​​in the evolutionary causal chain graph, as well as the degree of consistency between the direction of each centroid migration vector and the extension direction. Precursors with a proportion higher than a preset cumulative threshold and a degree of consistency higher than a preset consistency threshold are determined as true precursors.

[0023] The degradation profile generation module is used to determine the heating rate sequence and the expansion rate sequence based on the micro-temperature rise intensity difference of the directed edges and the centroid migration vector of true precursors. Based on the heating rate sequence and the expansion rate sequence, the precursor degradation profile of the true precursor is determined. The precursor degradation profile includes degradation mode label, degradation urgency and self-intensification mark.

[0024] The early warning output module is used to annotate the centroid, extension direction, and early warning degradation profile of the precursor to the photovoltaic module and output an early warning trend map.

[0025] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0026] 1. This invention first subtracts the temperature component determined by the acquisition environment from the infrared thermal imaging frames of each inspection cycle to obtain the remaining temperature rise field, and removes environmental fluctuations to highlight weak precursors; then it identifies precursor patterns composed of linearly extending temperature rises and determines their extension direction, centroid, and micro-temperature rise intensity, using linear shapes to eliminate point interference; furthermore, it establishes cross-cycle correlations and constructs evolutionary causal chain diagrams for precursor patterns in different inspection cycles that meet preset correlation conditions for centroid and extension direction, and determines true precursors by the proportion of directed edges with positive differences exceeding a preset cumulative threshold and the degree of direction consistency exceeding a preset consistency threshold. Thus, in the early stage when the temperature rise is weak and there is a lot of interference, the true precursors of microcracks can be reliably distinguished from the background and interference, and detected before they develop into obvious hot spots.

[0027] 2. This invention constructs a structured representation of an evolutionary causal chain diagram by setting the intensity of micro-temperature rise of the same precursor in each inspection cycle as nodes and establishing directed edges between adjacent nodes from early to late to carry the difference in micro-temperature rise intensity. This explicitly encodes the degradation process of the precursor over time into a chain structure that can be read along the causal direction. Based on this, the proportion of directed edges with positive differences measures the monotonic accumulation of micro-temperature rise over time, and the consistency between the centroid migration vector and the extension direction measures the unidirectional expansion of the hot zone along the crack direction. The two criteria respectively characterize the degradation law of the real hidden crack from the temperature dimension and the spatial dimension and are orthogonal to each other. Fixed hot spots that continuously heat up but do not expand spatially and shadows that drift laterally with sunlight only satisfy one of these criteria and are therefore excluded. This significantly improves the reliability and interpretability of distinguishing between true and false precursors under early weak conditions. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is an overall block diagram of the method in Embodiment 1 of the present invention;

[0030] Figure 2 This is a schematic diagram illustrating the principle of linear directional precursor recognition in this invention.

[0031] Figure 3 This is a schematic diagram of the phase plane in the deterioration state of the present invention;

[0032] Figure 4 This is an overall block diagram of the system in Embodiment 2 of the present invention. Detailed Implementation

[0033] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0034] Example 1:

[0035] like Figures 1-3 As shown, the deep learning-based distributed photovoltaic module hot spot fault detection method includes the following steps:

[0036] For ease of understanding, this embodiment uses a polycrystalline silicon photovoltaic module, numbered A-07, in a ground-mounted photovoltaic power station as an example. The module's position in the array is given by the layout metadata. One of the cells developed a microcrack defect during long-term operation. The operation and maintenance system arranges for drones to conduct periodic infrared inspections of the module, collecting one frame of infrared thermal image for the module in each inspection cycle. The inspection cycle is approximately once every seven days. This embodiment selects infrared thermal image frames collected from five consecutive inspection cycles (denoted as cycle 1 to cycle 5, where cycle 1 is the normal time point before the defect has developed) for a complete explanation. In the early stages, the microcrack only causes a slight local temperature rise, with a temperature increase of less than 0.6°C in the infrared image. It is difficult to directly distinguish from normal cell areas and interference such as cloud shadows and dust accumulation. This is precisely the early identification problem that this embodiment aims to solve.

[0037] Step S1: Obtain infrared thermal imaging frames of the same photovoltaic module in multiple inspection cycles, and subtract the temperature component determined by the acquisition environment factors from the infrared thermal imaging frames of each inspection cycle to obtain the remaining temperature rise field.

[0038] In this step, an infrared thermal imaging frame refers to a single-frame infrared temperature image captured by a drone or fixed camera of the same photovoltaic module. It is a two-dimensional pixel matrix, with each pixel representing the temperature reading at its corresponding physical location. Infrared thermal imaging frames from multiple inspection cycles originate from repeated acquisitions of the same module by the acquisition device at different times, with one frame acquired per inspection cycle.

[0039] During the data collection phase, infrared thermal imaging frames acquired from each session are received. Each frame is assigned an inspection cycle index in ascending order of its acquisition timestamp, and the inspection cycle interval is obtained from the difference between adjacent acquisition timestamps. Simultaneously, each frame is bound to its corresponding acquisition environment vector, component location identifier, and solar azimuth angle at the time of acquisition. The acquisition environment vector contains the irradiance and ambient temperature at the time of acquisition, sourced from the power plant's environmental sensors; the component location identifier is the inspected component's number and coordinates within the array, sourced from the power plant's layout metadata; and the solar azimuth angle is calculated from the geographic coordinates corresponding to the acquisition timestamp and component location identifier. Each frame is then associated with the aforementioned binding information, forming an infrared thermal imaging frame sequence ordered by the inspection cycle index and oriented towards the same component location identifier.

[0040] Using the example of component A-07, five inspection cycles are collected sequentially at similar times in five consecutive weeks. Each frame is bound to the irradiance, ambient temperature and solar azimuth angle collected at that time, and uniformly points to the component A-07 as the component location identifier.

[0041] Because the viewing angle and position of each drone acquisition are slightly different, the same physical point may fall on different pixel coordinates in frames of different cycles. To ensure that cross-cycle temperature comparisons are conducted under the same spatial reference, infrared thermal imaging frames of each cycle under the same component location are spatially registered with the component outline corner points and boundaries as the reference, so that the same physical point in each cycle falls on the same pixel coordinates. Frames whose cross-cycle coordinate offset exceeds the tolerance of the component outline boundary after registration are marked as registration failure frames and are not included in subsequent cross-cycle sequences, thereby avoiding false migration caused by registration deviation. When a frame of a certain inspection cycle is removed due to registration failure, or a certain cycle is not acquired for some reason, the two adjacent valid cycles are still processed according to their actual inspection cycle index in subsequent mapping. The corresponding inspection cycle interval is the difference between the acquisition timestamps of the two valid cycles, so that subsequent measurements can still be correctly calculated along the actual time scale even when there are missing frames. In the example of component A-07, the frames of all five cycles fall into a uniform pixel grid after the four corners of the component are aligned with the border, and no registration failure frames appear.

[0042] Furthermore, considering that changes in the solar azimuth angle between different cycles can cause shifts in shadow positions, it is necessary to determine whether the solar azimuth angle between cycles has changed sufficiently to cause a discernible shadow drift. Therefore, the cross-cycle variation amplitude of the solar azimuth angle at the time of data collection in each cycle is calculated, and any variation amplitude exceeding the discernible azimuth angle threshold is considered. The azimuth angle is set to true if it is within a certain range, and false if it is not higher. The azimuth angle resolution threshold is... The solar azimuth angle is used to determine whether the inter-period variation is large enough to produce a discernible shadow position shift on the image. Its value is calibrated based on the latitude of the power station, the season, and the imaging resolution, and is typically chosen such that the corresponding shadow displacement reaches several pixels. This flag determines whether to introduce an exclusion criterion based on solar drift in subsequent steps. In the example of component A-07, the acquisition times of each period are close but span several weeks, and the inter-period variation of the solar azimuth angle is higher than... Therefore, the azimuth angle can be identified as true.

[0043] After completing the above collection and registration, the temperature component determined by the acquisition environment factors is subtracted from the infrared thermal imaging frames of each inspection cycle to obtain the residual temperature rise field. The principle of this subtraction process is as follows: the temperature on the infrared image consists of two parts. One part is the temperature component caused by acquisition environment factors such as irradiance and ambient temperature, which changes with the environment. The other part is the temperature component caused by defect precursors, which does not change with the environmental metadata. Since the environment-related component can be interpreted from the environmental metadata through a mapping relationship, while the defect-related component cannot, the temperature component remaining after subtracting the environment-related component is the residual temperature rise field related to the defect. This achieves the separation of defect signals from environmental interference, making the weak precursor signals that were originally submerged in environmental fluctuations stand out.

[0044] Specifically, the process of determining the residual temperature rise field includes: reconstructing the infrared thermal imaging frames of each inspection cycle to obtain the reconstructed temperature field, and binding each frame with an acquisition environment vector containing irradiance and ambient temperature. Here, reconstruction refers to reading the temperature reading pixel by pixel of the infrared thermal imaging frames indexed in each inspection cycle to restore the temperature distribution at each pixel location in that cycle.

[0045] Subsequently, the inspection cycle of the photovoltaic module after it has been verified to be defect-free is taken as the normal baseline cycle. A regression relationship is fitted pixel-by-pixel using the environmental vector acquired during the normal baseline cycle as the independent variable and the reconstructed temperature field during the normal baseline cycle as the dependent variable. Substituting the environmental vector acquired during each inspection cycle into this regression relationship, the temperature component determined by the environmental factors is obtained pixel-by-pixel. The remaining temperature rise field is obtained by subtracting this temperature component pixel-by-pixel from the reconstructed temperature field. This regression relationship is used to characterize the components of the temperature distribution that can be explained by irradiance and ambient temperature.

[0046] It should be noted that since only one frame is acquired per inspection cycle, a single cycle only provides one set of "environmental vector to temperature" observations at each pixel, which is insufficient to fit a pixel-by-pixel regression relationship containing two independent variables: irradiance and ambient temperature. Therefore, the normal baseline cycle should be taken from multiple inspection cycles of the photovoltaic module that have passed inspection and are defect-free. These normal baseline cycles are located in different weather conditions and at different times, with different irradiance and ambient temperature values, thus providing multiple sets of observation samples at each pixel. In practical implementation, the above pixel-by-pixel regression relationship can be implemented as an environmental interpretation mapping: for each pixel coordinate... Using irradiance and ambient temperature collected in multiple normal baseline cycles as inputs and the reconstructed temperature of the pixel under the corresponding acquisition as output, a function from the environment vector to temperature is fitted. The simplest form can be linear, where the pixel temperature is assigned a regression coefficient to both irradiance and ambient temperature, with a constant term appended. The regression coefficients maintain their sign (e.g., a positive coefficient corresponds to an increase in temperature due to increased irradiance) to conform to the physical principle that "increased irradiance leads to increased temperature." Since this mapping is independently calibrated pixel-by-pixel, it can characterize the differences in the speed and intensity of the component's response to the environment at different locations (e.g., faster heat dissipation near the edge, slower heating in the center). Substituting the acquired environment vector from each cycle into the corresponding environment interpretation mapping for each pixel yields the environment interpretation component for that cycle, which is the temperature component determined by the acquired environmental factors.

[0047] Therefore, the normalized residual temperature rise field for each period can be expressed as:

[0048] ;

[0049] In the formula, For pixels In the cycle The remaining temperature rise field, For the reconstructed temperature field of the same location and the same period, The acquisition environment vector (including irradiance and ambient temperature) is bound to this cycle. The output is a pixel-by-pixel calibrated environment interpretation map. This refers to the temperature component (environmental interpretation component) determined by environmental factors for that pixel and that period. The remaining component after subtracting the environmental interpretation component from the reconstructed temperature field pixel by pixel is the residual temperature rise field on the left side of the above equation. The physical significance of the residual temperature rise field is that it strips away the contribution of external environmental factors such as irradiation and air temperature to the surface temperature of the module, retaining only the abnormal temperature rise caused by defects in the module itself, which does not change with the environment. Therefore, it is a direct carrier of defect precursors.

[0050] For component A-07, the defect-free period 1, along with several earlier defect-free inspection periods under varying weather conditions, is taken as the normal baseline period. An environmental interpretation map is then fitted pixel-by-pixel. Taking a pixel at the center of a defective cell as an example, its reconstructed temperature in period 1 can be almost completely explained by the current irradiance and ambient temperature through this mapping, with the residual temperature rise close to zero. However, in periods 2 to 5, where defects have already developed, even though the current irradiance and ambient temperature are different, after subtracting the environmental component through the environmental interpretation map, the pixel still retains residual temperature rises of approximately 0.18℃, 0.27℃, 0.38℃, and 0.52℃ respectively. This slight temperature rise, unaffected by environmental changes, is the precursor signal of microcrack defects. Through this normalization process, environmental fluctuations are effectively filtered out, and the defect-related residual temperature rise field is extracted from the images of each period, providing clean input for subsequent linear precursor identification. In addition, the normalized temperature field sequence (i.e., the temperature field period by period after deducting the environmental component) can also be used in subsequent steps to verify whether each period is at the same registration reference.

[0051] Step S2: For the remaining temperature rise field of each inspection cycle, identify the precursor patterns formed by linear temperature rises, and determine the extension direction, centroid, and micro-temperature rise intensity of each precursor pattern.

[0052] The rationale behind this step is that microcracks in photovoltaic cells physically manifest as cracks extending in a certain direction. Due to current mismatch, the crack generates heat, resulting in a weak, linear temperature rise band extending along the crack's direction in the residual temperature rise field, rather than an isolated circular hot spot. Therefore, "linear extension" is used as a shape criterion to distinguish true microcrack precursors from point-like interference (such as bird droppings or localized dust accumulation). This shape criterion allows point-like interference to be eliminated during the identification stage.

[0053] In practice, the process of identifying precursor lines and determining their extension direction, centroid, and micro-temperature rise intensity includes the following steps.

[0054] First, the directional responses of each pixel in the remaining temperature rise field are extracted in multiple candidate directions. The ratio of the maximum directional response to the average directional response of the same pixel in multiple candidate directions is recorded as the linear saliency of the pixel. The directional response is obtained by directional matching of the remaining temperature rise field through a directional filter.

[0055] To further suppress random noise in infrared imaging and enhance the weak linear temperature rise structure, before orientation matching, a convolutional neural network can be optionally used to preprocess the remaining temperature rise field. This involves layer-by-layer convolution and nonlinear activation, outputting an enhanced temperature rise response map pixel by pixel, and then calculating the orientation response on this enhanced temperature rise response map. Without this preprocessing, the orientation response is calculated directly on the remaining temperature rise field. One executable process for this convolutional enhancement is as follows: a preset convolutional kernel slides across the remaining temperature rise field with a stride, calculating the weighted sum of the kernel weight and the local temperature neighborhood at each position. This is then transformed element-by-element by a nonlinear activation function to obtain the feature map for that layer. These layers are stacked and mapped from the last layer to a single-channel enhanced temperature rise response map. This process suppresses random noise, enhances the locally coherent temperature rise structure, and highlights the weak linear temperature rise.

[0056] In its implementation, this convolutional neural network can employ stacked convolutional layers, such as three to five layers. Each convolutional layer consists of a two-dimensional convolution, batch normalization, and a non-linear activation function (such as a linear rectified function, i.e., taking the larger of the input and zero) connected in sequence. The kernel size can be 3×3, and the number of channels can be 16, 32, or 64 from shallow to deep layers. The final layer uses another convolution to map the multi-channel features into a single-channel enhanced temperature rise response map. Each convolutional layer uses a stride of 1 and fills the boundaries with equal width to ensure that the spatial size of the output feature map of each layer is consistent with the input residual temperature rise field. This aligns the output enhanced temperature rise response map with the residual temperature rise field pixel by pixel, facilitating the calculation of the directional response pixel by pixel. The training of this convolutional neural network can be carried out as follows: Pixels in areas of battery cells confirmed to have hidden cracks during historical inspections are labeled as positive, and pixels in areas confirmed to be defect-free are labeled as negative, thus constructing a pixel-by-pixel label. The remaining temperature rise field is input into the network, and the network outputs the response value of each pixel belonging to a linear temperature rise structure. The cross-entropy loss between the pixel-by-pixel response value and the pixel-by-pixel label is used as the training loss. Through forward computation, loss calculation, and backpropagation, the kernel weights and batch normalization parameters of each convolutional layer are iteratively updated using a gradient descent-type optimization method until the loss converges or the preset number of training rounds is reached. After training, the network parameters are fixed and used to output an enhanced temperature rise response map for the remaining temperature rise field in each inspection cycle.

[0057] Based on this, the directional response field is calculated: a set of discrete candidate directions for each orientation of the coverage plane is defined (e.g., a direction is taken at certain angles from 0° to 180°). For each pixel, directional matching is performed along each candidate direction using a directionally adjustable kernel (i.e., the aforementioned directional filter) to obtain the linear extension response intensity of the pixel along that direction. The response for each candidate direction constitutes the directional response field of that pixel. The directionally adjustable kernel has a longer support length in its matching direction and narrows perpendicular to the matching direction, thus providing a strong response to "slender temperature rise bands extending along that direction" and a weak response to point-like or other oriented structures. The support length of this kernel in the matching direction is determined by the directional kernel scale parameter. control, The value corresponds to the spatial scale of the linear structure to be matched and is related to the crack width order. It can be set according to the typical width of the microcrack in the solar cell and the imaging resolution. If the value is too small, it will be easily interfered with by noise, and if the value is too large, fine cracks will be missed.

[0058] Subsequently, for each pixel, the ratio of its maximum directional response to the mean of all directional responses is taken as the linear saliency, thus constructing a linear saliency field: ;

[0059] In the formula, For pixels linear significance, For this pixel along the candidate direction directional response, For the candidate direction set, The number of candidate directions. The ratio represents the maximum response of a pixel across all candidate directions, with the denominator being the average response of that pixel across all candidate directions. The physical meaning of this ratio is as follows: if a pixel exhibits a distinct linear structure, its response along the crack direction will be significantly higher than in other directions, with the maximum value much larger than the average, resulting in a larger ratio; conversely, if a pixel only exhibits isotropic point hotspots or noise, the responses in all directions will be similar, and the ratio will approach 1. Therefore, linear saliency, measured by "direction selectivity," naturally distinguishes between linear and point temperature rises. Furthermore, because it uses a ratio, it is insensitive to the absolute magnitude of the temperature rise. Even extremely weak linear temperature rises can achieve high saliency as long as they possess direction selectivity. This characteristic is crucial for this step to capture early signs of temperature rises at very small intervals.

[0060] Next, linear significance levels higher than a preset significance threshold are set. The pixels are designated as candidate precursor pixels, and the connected regions formed by these candidate pixels are designated as candidate precursor regions. The direction in which the pixels contained in the candidate precursor region obtain the maximum directional response is taken as the extension direction of the candidate precursor region, and the length along the extension direction is greater than a preset length threshold. The candidate precursor region is determined as a precursor ripple. Among them, the salience threshold... Used to filter out point-like and noisy pixels with insufficient directional selectivity; length threshold. Used to filter out fragmented areas that are directional but too short to constitute a crack, both ensure that the identified precursor cracks are both linear and have sufficient extension span; both can be determined by using the detection rate of known precursor cracks and the false alarm rate of point interference on labeled historical inspection data as evaluation indicators, and by using grid search to find the values ​​that achieve the expected compromise between the two.

[0061] Finally, for each identified precursor ripple, the centroid of the pixels contained in the precursor ripple is taken as its centroid, i.e., the centroid coordinates are obtained by taking the arithmetic mean of all pixel coordinates in the region; the regional peak value of the residual temperature rise of the pixels contained in the precursor ripple in the residual temperature rise field is taken as its micro-temperature rise intensity, i.e., the maximum value of the residual temperature rise in that region is taken. Considering that there is random noise in infrared imaging, its noise equivalent temperature difference is often on the same order of magnitude as the weak temperature rise of the early precursor. If the strict maximum value is directly taken as the micro-temperature rise intensity, a single thermal noise pixel may raise this value. Therefore, when determining the above-mentioned regional peak value, a small-scale spatial smoothing (such as small window Gaussian filtering) can be performed on the residual temperature rise field first, or the high quantile value (such as the 95th quantile) of the residual temperature rise in the precursor ripple region can be taken instead of the strict maximum value, thereby reducing the interference of single pixel noise while preserving the weak precursor signal. Thus, each precursor crack in each inspection cycle is characterized by three descriptions: the precursor extension direction (represented by a unit vector) representing the crack's orientation, the precursor centroid coordinates representing the crack's location, and the micro-temperature rise intensity representing the crack's heating intensity.

[0062] Using component A-07 as an example, in the remaining temperature rise field of period 1, because defects have not yet occurred, the linear saliency of the pixelless lines exceeds [a certain threshold]. No precursor lines were detected; starting from cycle 2, a weak linear temperature rise extending at approximately 30° appeared on the defective cell, with a pixel line saliency higher than that of the previous one. The length of the connected region along this direction is also higher than Therefore, it was identified as a precursor crack, with its precursor extending as a unit vector at approximately 30°, its centroid located at a pixel in the middle of the cell, and its micro-temperature rise intensity (the noise-resistant peak value of the remaining temperature rise in the area) approximately 0.18℃. Precursor cracks with similar directions were identified at similar locations in periods 3, 4, and 5, with micro-temperature rise intensities of approximately 0.27℃, 0.38℃, and 0.52℃, respectively. Furthermore, the span of the precursor crack along its extension direction increased slightly with each period, reflecting the slow propagation of the crack. As a control, a point-like temperature rise caused by bird droppings at another location on the same component in a certain period showed similar responses in all directions and a linear significance close to 1, failing to exceed [a certain threshold]. Therefore, it was not identified as a precursor pattern and was thus excluded by the shape criterion in this step.

[0063] Step S3: For precursor patterns whose centroid and extension direction meet the preset correlation conditions in different inspection cycles, determine that they belong to the same precursor and establish cross-cycle correlation.

[0064] The purpose of this step is to connect isolated precursor cracks identified in each cycle across cycles based on whether they represent the same crack at different points in time. This transforms the static identification of a single cycle into a time-series object whose evolution can be tracked. This connection is the prerequisite for all subsequent time-series analysis. Determining whether precursor cracks in two different cycles represent the same crack is based on the physical fact that cracks are spatially stable in location and orientation.

[0065] Accordingly, the aforementioned preset correlation condition is: the coordinate difference of the centroid of two precursor patterns in different inspection cycles does not exceed the preset position tolerance. Furthermore, the directional deviation corresponding to the vector inner product of the two extension directions does not exceed the preset directional tolerance. This condition can be expressed as:

[0066] ;

[0067] In the formula, Two signs , The association indicator has a value of 1 indicating that the association condition is met and a value of 0 indicating that it is not met. The coordinates of the centroids of the two precursor fringes are given. The difference in coordinates between the two centroids; The unit vector representing the direction of the precursor extension of the two precursor fringes. It is the dot product of two unit vectors in two directions. It is used to characterize the deviation between two directions (the closer the directions are, the closer the inner product is to 1, and the closer the deviation is to 0). To preset the position tolerance, To preset the directional tolerance, For logical AND, Indicative function. Position tolerance. Orientation tolerance is used to accommodate centroidal micro-movements caused by registration residuals and slow crack propagation. The values ​​of the jitter used to accommodate the orientation estimation need to be balanced between "accommodating reasonable fluctuations of the same crack" and "not misinterpreting different structures in the vicinity". They can be set according to the typical amplitude of the registration residual and the typical jitter range of the orientation estimation.

[0068] In implementation, a correlation model is used to describe the precursor patterns across inspection cycles. For precursor patterns in adjacent and nearby cycles, the proximity of the precursor centroid coordinates and the consistency of the precursor extension direction are jointly measured. For precursor patterns whose centroid coordinate differences do not exceed a certain threshold... And the deviation in the direction of extension does not exceed If a precursor pattern is not identified during a given cycle, it is considered to have entered the linear persistence condition. This condition requires that each precursor pattern entering the correlation must be continuously identified as a valid precursor pattern in each of the relevant inspection cycles, thereby eliminating spurious responses that only appear occasionally and lack temporal persistence. If a precursor pattern is not identified in an intermediate cycle due to interference, a correlation can still be established between the two adjacent cycles that were successfully identified before and after it, and it can be regarded as a continuous evolution of the same precursor. The missing cycle will not have a node in the subsequent evolutionary causal chain diagram, and will be directly connected by the valid cycles before and after it.

[0069] Furthermore, considering that cloud shadows and component frame shadows in outdoor infrared imaging will move laterally on the image as the solar azimuth changes, the resulting pseudo-temperature rise areas may happen to meet the correlation conditions between position and direction during a certain two-week period. To distinguish this interference that drifts with sunlight from real cracks with stable positions, this embodiment introduces a centroid anchoring criterion when a discernible change in the solar azimuth occurs. Specifically, the infrared thermal imaging frame is also bound to the solar azimuth corresponding to the acquisition time; when the change in the solar azimuth in each inspection cycle exceeds a preset discernible azimuth threshold... When the azimuth identifiable marker is true, the conditions for determining that they belong to the same precursor and establishing a cross-cycle correlation are as follows: the precursor pattern involved in the correlation is identified in each inspection cycle involved, and the centroid of the precursor does not drift with the solar azimuth.

[0070] The determination that the centroid does not drift with the solar azimuth angle includes: taking the maximum value of the ratio of the centroid displacement to the corresponding change in solar azimuth angle in each adjacent inspection cycle of the precursor as the anchoring measure, and ensuring that the anchoring measure does not exceed a preset anchoring threshold. At that time, it is determined that the centroid of the precursor does not drift with the solar azimuth angle. This anchoring measure can be expressed as:

[0071] ;

[0072] In the formula, For anchoring measurement, For this precursor in the cycle The precursor centroid coordinates, The centroid displacement is between two adjacent periods. For period The corresponding solar azimuth angle, This represents the change in solar azimuth angle between two adjacent cycles. This represents the maximum value of the above ratio across all adjacent cycles throughout the entire process. The physical meaning of this measure is as follows: if a "precursor" is actually a shadow moving with the sun, its centroid will move laterally synchronously with the sun's azimuth angle, and the ratio of centroid displacement to azimuth angle change is large; if it is a real crack with a fixed location, its centroid will remain almost stationary regardless of changes in the sun's azimuth angle, and this ratio is very small. Therefore, anchoring measures and anchoring thresholds are used... By comparison, it can be determined whether the centroid drifts with the sunlight. The value corresponds to the maximum allowable ratio of the centroid's micro-motion relative to the azimuth angle change caused by the registration residual. It should be noted that the azimuth angle discernibility indicator being true only ensures that the total variation in the solar azimuth angle over the entire process is sufficient, not that the azimuth angle change in each adjacent period is not close to zero. To avoid the ratio becoming meaningless when the denominator of the above formula approaches zero, when calculating the anchoring measure, this ratio is only calculated for adjacent period segments where the absolute value of the solar azimuth angle change is not less than a preset minimum azimuth angle change, and the maximum value is taken. Adjacent period segments with excessively small azimuth angle changes are not included in the anchoring measure calculation. Here, we examine the lateral drift of the centroid relative to the solar azimuth angle change, which is a displacement in a different direction than the longitudinal migration of the centroid along the crack extension direction examined in subsequent steps. Their physical meanings are independent and do not substitute for each other.

[0073] It should be noted that when the change in solar azimuth angle during each inspection cycle does not exceed the azimuth angle discernibility threshold... When the azimuth is identifiable as a false sign, the shadow drift caused by the solar azimuth on the image is indistinguishable, and the anchoring measure does not have a stable denominator that can be calculated. In this case, the anchoring measure is not calculated, and the linear persistence condition only requires that each precursor fringe maintains linear directionality in each of its involved cycles. The elimination of the common cause of drift with sunlight is instead undertaken by the monotonic accumulation along the extension direction in subsequent steps and the linear directionality criterion in this step.

[0074] After establishing cross-cycle correlations, for each cross-cycle precursor ripple identified as the same precursor, two temporal processing steps are performed: First, the micro-temperature rise intensity of the precursor in each cycle is arranged in ascending order according to the inspection cycle index, forming a temporal cumulative trajectory; second, the difference in the centroid coordinates of the precursor in adjacent inspection cycles is taken to form the centroid migration vector for each adjacent cycle. Considering that the micro-temperature rise intensity in a single cycle may contain jitter due to the influence of infrared noise, a slight temporal smoothing can be performed when processing the temporal cumulative trajectory to ensure that subsequent judgments based on the differences between adjacent cycles are not distorted by noise jitter in individual cycles. In addition, the normalized temperature field sequence can be used to confirm that the precursor ripples in each cycle are at the same registration reference. The aforementioned temporal cumulative trajectory and centroid migration vector, along with the precursor extension direction, precursor centroid coordinates, anchoring measure (left blank when azimuth is indistinguishable), and azimuth distinguishable marker, are used as the temporal description of the precursor and passed to subsequent steps. By introducing the centroid anchoring criterion, this step eliminates shadow-like interference that drifts laterally with the solar azimuth angle during the association stage, making the precursors for subsequent causal mapping and authenticity determination purer and preventing shadow artifacts from contaminating the entire evolutionary causal chain.

[0075] Using the example of component A-07, the coordinate differences between any two centroids of the four precursor fringes identified in periods 2 to 5 do not exceed [a certain value]. The deviation in the extension direction does not exceed Furthermore, each cycle was consistently identified as a valid precursor streak, satisfying the linear continuity condition. Since the azimuth indicator in this example is genuine, its anchoring measure was further calculated—despite the change in solar azimuth in each cycle, the precursor centroid remained almost stationary, and the maximum ratio of the centroid displacement to the azimuth change in each adjacent segment was much smaller than... The centroid was determined not to drift with the solar azimuth angle. Based on this, the four precursor lines in periods 2 to 5 were identified as belonging to the same precursor and a cross-period correlation was established. The temporal cumulative trajectory of this precursor was found to be a micro-temperature rise intensity sequence [0.18, 0.27, 0.38, 0.52]℃, and the centroid migration vector was the difference between the adjacent centroids in periods 2→3, 3→4, and 4→5. The vector direction of each segment was approximately along an extension direction of about 30°, with an amplitude of sub-pixel to several pixel level, reflecting the slow expansion of the crack along its direction. As a contrast, the aforementioned point-like temperature rise caused by bird droppings did not continuously appear in multiple periods and did not meet the linear persistence condition. Even if there was an occasional response, it would not be established as a cross-period precursor.

[0076] Step S4: For the same precursor, set its micro-temperature rise intensity in each inspection cycle as a node, and establish directed edges between adjacent nodes from early to late according to the inspection cycle to carry the difference in micro-temperature rise intensity between adjacent nodes, thus forming the evolutionary causal chain graph of the precursor, and take the difference of the centroid of the precursor in adjacent inspection cycles as the centroid migration vector.

[0077] This step organizes the temporal evolution of the precursor into a directed graph structure to explicitly carry the causal direction of "the earlier state advancing to the later state" and the temperature increase increment at each step, laying a structured foundation for subsequent authenticity discrimination and rate estimation using graph measurements; it is this structured representation that makes the temporal causal order of degradation and the temperature change at each step explicitly encoded and directly readable.

[0078] Specifically, for the same precursor, its precursor state at each inspection cycle index is set as an evolutionary causal node. This node carries the micro-temperature rise intensity of that cycle, as well as the precursor extension direction and centroid coordinates of that cycle. Directed evolutionary causal edges are established between adjacent evolutionary causal nodes according to the inspection cycle index from early to late, with the edge direction pointing from the earlier cycle to the later cycle to reflect the causal direction of degradation progressing over time. Each evolutionary causal edge carries the difference in micro-temperature rise intensity between its two adjacent nodes, i.e., the cumulative increment obtained by subtracting the micro-temperature rise intensity of the earlier cycle from the micro-temperature rise intensity of the later cycle. Connecting the evolutionary causal nodes of the same precursor sequentially along the evolutionary causal edges constitutes the evolutionary causal chain graph of that precursor. When there is an intermediate cycle missing due to registration failure or misidentification, the evolutionary causal nodes corresponding to the two valid cycles before and after it are directly connected by a directed edge, and the cumulative increment of this edge is the difference in micro-temperature rise intensity between the two valid cycles. The cumulative increment carried by the above directed edge and the edge direction can be expressed as:

[0079] ;

[0080] In the formula, For period The intensity of micro-temperature rise carried by evolutionary causal nodes, The intensity of the slight temperature rise borne by the causal node of the subsequent adjacent cycle evolution. The difference between the two is the cumulative increment of that segment; , For two adjacent evolutionary causal nodes, For the reason point to and based on cumulative increments The evolutionary causal edges are weighted. The physical meaning of this evolutionary causal chain graph is that the nodes depict the heating intensity state of the precursor at each time point, and the directed edges encode the temporal causal order of the degradation with their direction and the net change in heating intensity between two adjacent time points with their weight (cumulative increment), thus completely representing the temperature evolution process of a crack as a chain structure that can be read along the causal direction.

[0081] Simultaneously, the difference between the centroids of adjacent inspection cycles of the precursor is taken as the centroid migration vector. That is, the centroid coordinates of the precursor in the later cycle are subtracted from the centroid coordinates of the earlier cycle to obtain the vector characterizing the spatial displacement of the crack hot zone within that segment. The centroid migration vectors of each adjacent cycle are arranged periodically, corresponding to the evolutionary causal chain. Figure 1 This constitutes a complete temporal representation of the precursor in both the temperature and spatial dimensions.

[0082] Using the example of component A-07, this precursor has four evolutionary causal nodes in periods 2 to 5, carrying micro-temperature increases of 0.18℃, 0.27℃, 0.38℃, and 0.52℃ respectively. Three directed evolutionary causal edges are established between adjacent nodes, with directions corresponding to periods 2→3, 3→4, and 4→5, respectively. The cumulative increments they carry are 0.27−0.18=0.09℃, 0.38−0.27=0.11℃, and 0.52−0.38=0.14℃, respectively. Connecting the four nodes along these three edges yields the evolutionary causal chain diagram of this microcrack precursor. Correspondingly, its centroid migration vector is the difference between the three adjacent centroids in periods 2→3, 3→4, and 4→5, all pointing approximately in an extension direction of about 30°.

[0083] Step S5: Calculate the proportion of directed edges with positive differences in the evolutionary causal chain graph, and the degree of consistency between the direction of each centroid migration vector and the extension direction. Precursors with a proportion higher than a preset cumulative threshold and a degree of consistency higher than a preset consistency threshold are determined as true precursors.

[0084] This step is crucial for ultimately identifying genuine microcrack precursors from various interferences. Its ability to distinguish them stems from a dual characterization of the physical laws governing genuine microcracks. The underlying physical laws are: the slight temperature rise of a genuine microcrack precursor monotonically increases along the causal direction (heat accumulation), and its hot zone expands along the crack's direction, causing the centroid to migrate unidirectionally along the extension direction; while fixed hot spots that continuously heat up but do not expand spatially, and shadows that drift with the sun, do not simultaneously possess these two characteristics. Therefore, this step establishes two mutually orthogonal criteria in the temperature and spatial dimensions.

[0085] The first criterion is the monotonically cumulative criterion for the temperature dimension: calculating the proportion of directed edges with positive differences in the evolutionary causal chain graph. That is, reading the cumulative increment carried by each edge along the direction of the evolutionary causal edge, and using the proportion of evolutionary causal edges with positive cumulative increments among all evolutionary causal edges as the measure of this criterion, denoted as the monotonically cumulative measure: ;

[0086] In the formula, It is a monotonically cumulative measure. Let be the set of all evolutionary causal edges of this precursor. To represent the total number of causal edges in evolution, To evolve causal edges The cumulative increment it carries The value is 1 when the cumulative increment is positive and 0 otherwise. The physical meaning of this measure lies in characterizing the degree to which the precursory micro-temperature rise continues to increase along the causal direction of time: the actual precursory heat of degradation accumulates monotonically, and the cumulative increment of the vast majority of directed edges is positive. The value is close to 1; while the pseudo-precursors of random fluctuations rise and fall unpredictably. Significantly low.

[0087] The second criterion is the spatial dimension expansion consistency criterion: calculating the degree of directional consistency between each centroid migration vector and its extension direction. That is, for each centroid migration vector of the precursor, calculate the dot product between its unit vector and the unit vector of the precursor's extension direction, and take the minimum value among the corresponding dot products of each centroid migration vector as the degree of directional consistency, denoted as the expansion consistency measure.

[0088] ;

[0089] In the formula, To extend the consistency measure (i.e., the degree of directional consistency). For period arrive The centroid migration vector is determined by the centroid coordinates of the later period. Reduce the centroid coordinates of the earlier period get, Its unit vector, The precursor extension direction is the unit vector. The inner product of the two, This indicates that the minimum value of the inner product is taken for the centroid migration vector of each segment throughout the entire process. Taking the minimum value, rather than the average, is intended to require that the precursor extends along the extension direction in each migration segment. If any segment shows a reverse direction or a significant deviation, the value is lowered, thus more strictly eliminating false precursors with chaotic migration directions. The physical meaning of this measure is to characterize whether the centroid of the hot zone always migrates unidirectionally along the crack direction: In a true crack propagation, the centroid moves forward along the extension direction, and the inner product of each segment is close to 1. Too high; fixed hotspots have no spatial expansion, while shadows drift laterally, and their centroid migration and extension directions are inconsistent. Too low.

[0090] It should be noted that the consistency criterion is based on the core principle of "migration direction along the extension direction," without setting a minimum migration amplitude threshold. This design addresses two real-world engineering problems: First, early microcrack propagation is extremely slow, and the centroid migration amplitude of a certain segment may be lower than the centroid resolvable value. Its centroid migration vector is approximately zero, and its inner product with the extension direction approaches zero rather than being negative, thus still being accommodated by the criterion. Second, cross-period spatial registration has residuals, and the centroid migration vector may contain pseudo-migration components caused by registration errors, with amplitudes comparable to true micro-migrations but random directions. The inner product of such pseudo-migrations fluctuates near zero and does not remain stable at positive. Therefore, taking the minimum inner product throughout the entire process without setting an amplitude threshold ensures that fluctuations with amplitudes near zero and random directions will not have a stable positive inner product, thus avoiding misjudgment as extension along the direction. Simultaneously, it does not mistakenly exclude early precursors that are slow-growing but have the correct direction. Only when the centroid continuously and stably migrates along the same extension direction across multiple periods will the inner products of each segment be simultaneously positive, and only then will the minimum inner product exceed the consistency threshold. This is precisely the key point of this criterion: it can both ensure that early, weak precursors are not missed and eliminate interference from registration residuals and fixed hotspots. In addition, when the number of centroid migration vectors is less than two (i.e., the number of cycles spanned by the precursor is insufficient), the extended consistency measure degenerates into being determined solely by the inner product of the extended directions of that single centroid migration vector.

[0091] In distinguishing between true and false signals, this embodiment employs two combinations of judgments based on the azimuth-identifiable marker. When the azimuth-identifiable marker is true, the monotonic cumulative measure is applied. Higher than the preset cumulative threshold Extended consistency measure Higher than the preset consistency threshold And anchored measurement Not exceeding the preset anchoring threshold If all three conditions are met simultaneously, the precursor is considered a true precursor; otherwise, it is considered a non-precursor. When the azimuth marker is identifiable as a false precursor, the anchoring measure is not calculated; instead, the monotonic cumulative measure is used. Higher than Extended consistency measure Higher than Furthermore, a precursor is considered a true precursor if all three corresponding precursor patterns are present in all inspection cycles it is identified; otherwise, it is considered a non-precursor. Among these, the cumulative threshold... Define the minimum acceptable monotonic increase rate, the consistency threshold. Define the minimum acceptable forward expansion consistency; the above cumulative threshold Consistency threshold Both can be evaluated using the detection rate of true precursors and the false alarm rate of non-precursors on labeled historical inspection data. The evaluation indicators can be determined by grid search or by a compromise between the detection rate and the false alarm rate. For example, the threshold that minimizes false alarms can be selected while ensuring that all known true microcrack samples are detected.

[0092] The two criteria mentioned above are physically orthogonal and complementary: a fixed hot spot that continuously heats up but does not expand spatially may meet the monotonic accumulation measure, but fails to meet the expansion consistency measure because its centroid does not migrate along the extension direction, and is therefore excluded by the second criterion; a shadow that drifts with the sun is excluded by the corresponding criterion either because the anchoring measure exceeds the threshold due to the lateral drift of the centroid (when the azimuth is discernible), or because it does not have continuous linear directionality (when the azimuth is not discernible). The combination of these two criteria ensures that only objects that simultaneously satisfy "monotonically accumulating temperature over time" and "unidirectional expansion of the centroid along space," and whose centroid does not drift with sunlight, are determined to be true precursors. This allows for reliable identification of true cryptorchid fracture precursors even under weak and highly interfering early conditions. After this separation, true precursors proceed to subsequent steps, while non-precursors are excluded.

[0093] Using the example of component A-07, the evolutionary causal chain diagram of this precursor to the microcrack has three evolutionary causal edges, and its cumulative increments of 0.09℃, 0.11℃, and 0.14℃ are all positive, therefore the monotonic cumulative measure is... Above the cumulative threshold (In this example, we take approximately 0.6, meaning that if more than half of the directed edges are positive, it is considered to have monotonicity); the centroid migration vectors of its three segments are all roughly along an extension direction of approximately 30°, and the inner product of each segment's unit vector and the unit vector in the extension direction is close to 1. We take the minimum value to extend the consistency measure. The value is approximately 0.98, which is higher than the consistency threshold. (In this example, we take approximately 0.7, corresponding to an angle of less than approximately 45° between the migration and extension directions); In this example, the azimuth can be identified as true, and its anchoring measurement is... Much smaller than the anchoring threshold If all three conditions are met simultaneously, the precursor is determined to be a true precursor. In contrast, if a component has a fixed hot spot caused by a localized poor solder joint, its temperature rise will gradually increase over time (…). (Meets the standard), but the hot zone never expands, the centroid remains almost stationary, the centroid migration vectors of each segment are close to zero and have no direction, and the dot product of its unit vector with any "extension direction" is not consistently positive along the direction. If the condition does not meet the standard, it will be determined as a non-precursor condition and terminated, and will not be falsely reported as a hidden crack.

[0094] Step S6: For true precursors, determine the heating rate sequence and the expansion rate sequence based on the micro-temperature rise intensity difference of their directed edges and the centroid migration vector, respectively; determine the precursor degradation profile of the true precursor based on the heating rate sequence and the expansion rate sequence, the precursor degradation profile includes degradation mode label, degradation urgency and self-aggravation mark.

[0095] After obtaining the true precursors, this step first converts their evolution in the temperature and spatial dimensions into a rate sequence over time. Then, based on the two rates, it describes the pattern, degree, and trend of their deterioration, forming a precursor deterioration profile. This transforms the qualitative judgment obtained in the previous step into a quantitative characterization of the deterioration state.

[0096] First, let's look at the determination of the rate sequence. The basis for setting this step is that the cumulative increment of the evolutionary causal edge and the centroid migration vector themselves only reflect the absolute change between two adjacent cycles, while the intervals of each inspection cycle may not be completely equal (especially when there are missing frames or when the effective cycles are directly connected, the interval is multiple inspection cycles). Only by dividing by the corresponding inspection cycle interval and converting it into the rate of change per unit time can we compare the speed of each segment's progress on a unified time scale and further determine its acceleration trend.

[0097] The determination of the heating rate sequence includes: for each directed edge of the true precursor, dividing its micro-temperature rise intensity difference by the corresponding inspection cycle interval to obtain the segmented heating rate of the directed edge, and arranging the segmented heating rates according to the inspection cycle to form a heating rate sequence, i.e.: ;

[0098] In the formula, For period arrive The stepwise heating rate, This represents the cumulative increment (i.e., the difference in micro-temperature rise intensity) carried by the causal edge of this evolutionary segment. This represents the inspection cycle interval between two adjacent cycles. The physical meaning of this step-by-step heating rate is to characterize the rate of increase in precursor heat intensity per unit time, i.e., the propagation speed of crack heat generation.

[0099] The determination of the spread rate sequence includes: for each centroid migration vector of the true precursor, dividing its projection in the precursor extension direction by the corresponding inspection cycle interval to obtain the segmented spread rate. The segmented spread rates are arranged according to the inspection cycle to form the spread rate sequence, i.e.: ;

[0100] In the formula, For period arrive The segmented expansion rate, This is the centroid migration vector for this segment. The precursor extension direction is the unit vector. Let be the projection magnitude of the centroid migration vector in the extension direction. This corresponds to the inspection cycle interval. The reason for using the projection of the centroid migration vector along the extension direction instead of its modulus is to eliminate the centroidal jitter component perpendicular to the crack direction, retaining only the effective migration along the crack propagation direction. The physical meaning of this segmental propagation rate is to characterize the speed at which the crack hot zone propagates along the crack direction per unit time, i.e., the speed of crack geometric propagation. The heating rate and propagation rate characterize the degradation progression from the two dimensions of temperature and space, respectively. Each constitutes a sequence, allowing for the subsequent determination of both the current degradation state and its second-order acceleration trend.

[0101] Using the example of component A-07, each inspection cycle interval is one inspection cycle. The cumulative increments of the three evolutionary causal edges of the true precursor, 0.09℃, 0.11℃, and 0.14℃, divided by the corresponding cycle intervals, yield a heating rate sequence of approximately [0.09, 0.11, 0.14] (temperature increment per inspection cycle); the projections of the three centroid migration vectors in the approximately 30° extension direction, divided by the corresponding cycle intervals, yield approximately [0.20, 0.30, 0.45] (migration amplitude per inspection cycle). It can be seen that regardless of heating or expansion, the segmental rate increases with the cycle, indicating that the microcrack is deteriorating at an accelerated rate. It should be noted that if the intervals between each inspection cycle are not equal, for example, if two cycles are separated by two inspection cycles due to missing frames, then the cumulative increment and centroid migration projection of that segment need to be divided by the actual interval to obtain a comparable segment rate. At this time, the value of the rate sequence will no longer be equal to the cumulative increment sequence. The step of dividing by the interval reflects the role of normalizing the unequal interval sampling to a unified time scale.

[0102] After obtaining the two rate sequences, a precursor degradation profile is determined, which includes degradation pattern labels, degradation urgency and self-aggravation markers, in three layers.

[0103] The first layer involves locating the deterioration state phase plane. The heating rate sequence and the expansion rate sequence are weighted and aggregated using weights that increase from early to late with the inspection cycle, and then normalized to obtain the normalized heating rate and expansion rate. A two-dimensional plane is established with the normalized heating rate as the first coordinate axis and the normalized expansion rate as the second coordinate axis. Points with the normalized heating rate and normalized expansion rate as their two coordinate components are designated as deterioration state points. The weighted aggregation using periodically increasing weights aims to make the aggregation results more reflective of recent deterioration trends and less reflective of earlier, outdated rates. The resulting aggregated values ​​are the current heating rate and the current expansion rate, respectively. These periodically increasing weights can be linearly increasing (e.g., the first...). Segment weight is proportional to ), or exponential memory weight (the first Segment weight is proportional to ,in The attenuation factor has a value between 0 and 1. The current rate sequence consists of a total number of segments (with the weight of segments closer to the current one increasing exponentially). The rates of each segment are normalized according to this weight and then summed to obtain the current heating rate and the current expansion rate. Since the heating rate and expansion rate are located in two different dimensions (temperature and space) and have different numerical dimensions, they are each normalized according to their respective dimensional references so that they can be compared on the same plane. The normalized heating rate and expansion rate are obtained after normalization. The degradation mode label is determined by the orientation of the degradation state point relative to the origin, and the degradation urgency is determined by the distance of the degradation state point relative to the origin. This orientation is the angle from the first coordinate axis to the degradation state point, and the distance is the Euclidean distance from the degradation state point to the origin, which can be expressed as:

[0104] ;

[0105] In the formula, This refers to the orientation of the deteriorated state point relative to the origin (i.e., the orientation of the deterioration mode). This represents the distance between the deteriorated state point and the origin (i.e., the urgency of deterioration). Given the current heating rate, For the current expansion rate, As a dimensional standard for heating rate, As a dimensional reference for the expansion rate, For the normalized heating rate, To extend the normalized rate. Wherein, the dimensional standard for the heating rate is... Dimensional reference for expansion rate To eliminate the difference in dimensions between two rates and make them comparable on the same plane, their values ​​can be taken from typical scales of the heating rate and expansion rate in historical true precursor samples (such as their respective medians or quantiles), thus ensuring that the normalized heating rate and the normalized expansion rate are of similar orders of magnitude, and that the orientation corresponding to their ratio is... It can accurately reflect the relative strength of the two propulsion methods, rather than being dominated by dimensional differences. Since for a true precursor, the expansion rate and heating rate are both positive along the extension direction, and both normalized rates are non-negative, It falls between 0° and 90°. The physical meaning of this azimuth and distance is as follows: the azimuth angle reflects the relative strength of the two propulsion modes of heating and expansion, which can be used to determine which mode of degradation is advancing; the distance to the origin, i.e. the combined magnitude of the two uniform rates, reflects the intensity of the overall degradation, which can be used to measure the urgency of degradation.

[0106] The second layer is degradation mode identification. This involves determining the orientation of the degradation state point relative to the origin. According to the preset pattern orientation threshold set The sector into which the degradation mode is identified is labeled. During the determination, the distance from the degradation state point to the origin is first calculated. If the distance is below the threshold for determining stagnation, it is directly classified as stagnant and no longer differentiated by direction. This indicates that the precursor's current two rates are both low, the deterioration is weak, and there is no significant development yet. Otherwise, it is classified by direction. The sector that falls into the degradation mode is identified by the label. Since the first axis is the temperature rise rate axis and the second axis is the expansion rate axis, the orientation... The closer to 90° (the degradation state point is closer to the normalized rate axis of expansion), the more dominant the expansion is relative to the temperature rise; the closer to 0° (the closer to the normalized rate axis of temperature rise), the more dominant the temperature rise is relative to the expansion. Therefore, A sector approaching 90° is identified as an extension-dominant type, used to characterize rapid crack propagation without a significant increase in heat generation per unit area; Sectors approaching 0° are classified as temperature-dominant, indicating that the current situation is characterized by an increase in heat intensity with relatively insignificant lateral expansion. The specific mechanism needs to be comprehensively judged in conjunction with the degree of deterioration urgency. High urgency often indicates that the crack is stabilizing and local heat generation is rapidly intensifying, approaching the critical point of manifestation. Low urgency may be an early precursor with insignificant expansion. Sectors approaching 45° and with a significant distance from the origin are classified as co-accelerated, characterizing a simultaneous surge in crack propagation and heat generation. Among these, the mode orientation threshold set... A set of azimuth boundaries for dividing the above sectors can be calibrated according to the needs of operation and maintenance for handling different degradation mechanisms.

[0107] The third layer involves determining the self-intensifying trend. The heating rate sequence is analyzed to determine its periodic variation along the inspection cycle, yielding the heating acceleration. Similarly, the expansion rate sequence is analyzed to determine its periodic variation along the inspection cycle, yielding the expansion acceleration. If either the heating acceleration or the expansion acceleration is positive and increases along the inspection cycle, the self-intensifying flag is set to true, indicating that the precursor degradation is self-intensifying, and the corresponding warning lead time needs to be shortened. Here, acceleration refers to the second-order change of the rate sequence, its physical meaning being to reflect whether the degradation process is accelerating or slowing down: if a rate is not only positive but also increases periodically, it indicates that the degradation is self-intensifying uncontrollably. It should be noted that when the rate sequence length is less than two cycles (i.e., the precursor spans less than three cycles), its periodic variation trend cannot be obtained; in this case, acceleration is not calculated, and the self-intensifying flag is set to empty.

[0108] Finally, the degradation pattern labels and degradation urgency are determined according to a preset urgency grading threshold set. The resulting urgency rating, along with the self-aggravating markers, together constitute the prodromal deterioration profile of this true precursor. Among them, the urgency rating threshold set... To increase the urgency of degradation A set of distance boundaries, divided into several alarm levels, can be set according to the response time limits for different levels of urgency by operations and maintenance. Thus, the early warning degradation profile simultaneously provides three aspects of information about the hidden crack: "what mode it is progressing in (degradation mode label), what level of urgency it has reached (degradation urgency and its classification), and whether it is escalating out of control (self-escalation marker)". This provides operations and maintenance decision-making with a basis that can distinguish the degradation mechanism and determine the priority of handling.

[0109] Using the example of component A-07, the heating rate sequence [0.09, 0.11, 0.14] and the expansion rate sequence [0.20, 0.30, 0.45] are weighted and aggregated with increasing weights over time. Taking an exponential memory weight and a decay factor of 0.5 as an example, the normalized weights of the three segments are approximately 0.143, 0.286, and 0.571, respectively. Therefore, the current heating rate is approximately 0.09×0.143+0.11×0.286+0.14×0.571≈0.124, and the current expansion rate is approximately 0.20×0.143+0.30×0.286+0.45×0.571≈0.371. Each component is then normalized to its dimensional reference. In this example, the temperature rise dimensional reference is approximately 0.10, and the expansion dimensional reference is approximately 0.20. Both are taken from the median of the rates of the historical microcrack samples of this power plant and are independent of the specific values ​​of this precursor. Based on this, the normalized rate of temperature rise is approximately 1.24, and the normalized rate of expansion is approximately 1.86. The degradation state point of this precursor is determined on the degradation state plane. The distance of this state point from the origin is significantly higher than the distance corresponding to the stagnation judgment, so it does not belong to the stagnation type. Therefore, its orientation relative to the origin is calculated. Because the current propagation rate of this precursor is more prominent relative to its historical scale, the normalized propagation rate (1.86) is higher than the normalized heating rate (1.24). This orientation is greater than 45°, biased towards the propagation rate axis, and falls into the sector corresponding to the propagation-dominant type. Therefore, the degradation mode label is judged as propagation-dominant, indicating that the current microcrack is mainly characterized by crack propagation and the heat generation per unit area has not yet increased dramatically. Calculate its distance from the origin. Threshold set based on urgency Falling into the corresponding (higher) level, the urgency level of this precursor is obtained. Since both its heating rate sequence and propagation rate sequence increase periodically, and the corresponding heating acceleration and propagation acceleration are positive and increase periodically, the self-aggrandizing flag is set to true, indicating that the microcrack deterioration is self-aggrandizing and its warning lead time should be shortened and it should be addressed as a priority. The above-mentioned deterioration mode label (propagation-dominant), urgency level, and self-aggrandizing flag (true) constitute the precursor deterioration profile of component A-07 for this microcrack precursor.

[0110] Step S7: Mark the centroid, extension direction, and precursor degradation profile of the precursor onto the photovoltaic module, and output the early warning trend chart.

[0111] After obtaining the precursor degradation profiles of each true precursor, the precursor centroid coordinates, precursor extension direction, and precursor degradation profile of each true precursor are marked to the corresponding location on the inspected photovoltaic module according to its module location identifier, and aggregated into an early warning trend map covering the module. The form of this early warning trend map is a module-level two-dimensional labeled map, which is marked with the precursor centroid coordinates (indicating the location of the crack), the precursor extension direction (indicating the direction of the crack), and the precursor degradation profile (including degradation mode label, urgency level, and self-aggravation mark) of each true precursor.

[0112] The significance of setting this output is that: the ratio of the heating rate to the expansion rate represents the mode of degradation, the combined magnitude of the two represents the urgency of degradation, and the second-order change of the two represents whether the degradation is out of control. The three together depict the mode, degree and trend of precursor degradation. The early warning trend map formed by this can issue graded warnings on the location, direction, degradation mechanism and urgency level of hidden dangers before the appearance of visible hot spots. This provides a direct basis for operation and maintenance personnel to distinguish the degradation mechanism and prioritize the handling of self-intensifying precursors, thereby supporting predictive operation and maintenance decisions for photovoltaic power plants.

[0113] Using module A-07 as an example, the precursor centroid of the microcrack is marked in the center of the cell on the 2D diagram of module A-07, and the crack direction is marked along an approximately 30° angle. A precursor degradation profile is then added, indicating "dominantly expanding type, corresponding urgency level, and self-intensifying." Maintenance personnel can then determine that module A-07 has an early microcrack that is spreading along an approximately 30° angle and is self-intensifying. Although its current temperature rise is only about 0.5°C and it has not yet formed a visible hot spot, it should still be given higher priority and intervened for early intervention to eliminate it before it develops into a visible hot spot or even causes burnout. Compared to traditional methods that only alert for visible hot spots after the temperature significantly exceeds the limit, this embodiment removes environmental components in the temperature dimension to eliminate interference, captures weak linear precursors with linear saliency, distinguishes between true and false precursors using an evolutionary causal chain diagram and a biorthogonal criterion, and then characterizes the degradation mode, urgency, and self-aggravation trend with a degradation state phase plane. Thus, in the early stage when the defect temperature rise is weak and there are multiple similar interferences such as cloud shadows, bird droppings, and dust accumulation, it can reliably distinguish the true precursors of hidden cracks from the background and interference and provide graded early warnings, thereby realizing early detection and predictive maintenance of hot spot faults.

[0114] Example 2:

[0115] like Figure 4 As shown, a deep learning-based distributed photovoltaic module hot spot fault detection system includes:

[0116] The data acquisition module is used to acquire infrared thermal imaging frames of the same photovoltaic module in multiple inspection cycles. The temperature component determined by the acquisition environment factors is subtracted from the infrared thermal imaging frames of each inspection cycle to obtain the remaining temperature rise field.

[0117] The feature extraction module is used to identify the precursors of temperature rise consisting of linearly extending temperature rises in the remaining temperature rise field of each inspection cycle, and to determine the extension direction, centroid, and micro-temperature rise intensity of each precursor.

[0118] The cross-cycle correlation module is used to determine that the precursor patterns whose centroid and extension direction meet the preset correlation conditions in different inspection cycles belong to the same precursor and establish cross-cycle correlation.

[0119] The causal graph construction module is used to set the micro-temperature rise intensity of the same precursor in each inspection cycle as a node, and establish directed edges between adjacent nodes from early to late according to the inspection cycle to carry the difference in micro-temperature rise intensity between two adjacent nodes, thus forming the evolutionary causal chain graph of the precursor, and taking the difference of the centroid of the precursor in adjacent inspection cycles as the centroid migration vector.

[0120] The true precursor determination module is used to calculate the proportion of directed edges with positive difference values ​​in the evolutionary causal chain graph, as well as the degree of consistency between the direction of each centroid migration vector and the extension direction. Precursors with a proportion higher than a preset cumulative threshold and a degree of consistency higher than a preset consistency threshold are determined as true precursors.

[0121] The degradation profile generation module is used to determine the heating rate sequence and the expansion rate sequence based on the micro-temperature rise intensity difference of the directed edges and the centroid migration vector of true precursors. Based on the heating rate sequence and the expansion rate sequence, the precursor degradation profile of the true precursor is determined. The precursor degradation profile includes degradation mode label, degradation urgency and self-intensification mark.

[0122] The early warning output module is used to annotate the centroid, extension direction, and early warning degradation profile of the precursor to the photovoltaic module and output an early warning trend map.

[0123] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

[0124] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0125] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A deep learning-based method for detecting hot spot faults in distributed photovoltaic modules, characterized in that, Includes the following steps: Infrared thermal imaging frames of the same photovoltaic module were acquired in multiple inspection cycles. The temperature component determined by the acquisition environment was subtracted from the infrared thermal imaging frames of each inspection cycle to obtain the residual temperature rise field. For the remaining temperature rise field of each inspection cycle, identify the precursors formed by linear temperature rises, and determine the extension direction, centroid and micro-temperature rise intensity of each precursor. For precursor patterns whose centroid and extension direction meet the preset correlation conditions in different inspection cycles, they are determined to belong to the same precursor and cross-cycle correlation is established. For the same precursor, the intensity of its micro-temperature rise in each inspection cycle is set as a node. Directed edges carrying the difference in micro-temperature rise intensity between adjacent nodes are established from early to late according to the inspection cycle to form the evolutionary causal chain diagram of the precursor. The difference between the centroids of the precursor in adjacent inspection cycles is taken as the centroid migration vector. Calculate the proportion of directed edges with positive difference values ​​in the evolutionary causal chain graph, and the degree of consistency between the direction of each centroid migration vector and the extension direction. Determine a precursor whose proportion is higher than a preset cumulative threshold and whose degree of consistency is higher than a preset consistency threshold as a true precursor. For the true precursor, the heating rate sequence and the expansion rate sequence are determined based on the micro-temperature rise intensity difference of its directed edges and the centroid migration vector, respectively. Based on the heating rate sequence and the expansion rate sequence, a precursor degradation profile of the true precursor is determined, the precursor degradation profile including degradation mode label, degradation urgency and self-aggravation marker; The centroid, extension direction, and degradation profile of the precursor are labeled onto the photovoltaic module, and an early warning trend chart is output.

2. The deep learning-based distributed photovoltaic module hot spot fault detection method according to claim 1, characterized in that: The process of determining the residual temperature rise field includes: The reconstructed temperature field is obtained by reconstructing the infrared thermal imaging frames of each inspection cycle, and each frame is bound to the acquisition environment vector containing irradiance and ambient temperature. The inspection cycle of the photovoltaic module after it has been verified to be defect-free is taken as the normal baseline cycle. The acquisition environment vector of the normal baseline cycle is used as the independent variable and the reconstructed temperature field of the normal baseline cycle is used as the dependent variable to fit a regression relationship pixel by pixel. The acquisition environment vector of each inspection cycle is substituted into the regression relationship to obtain the temperature component determined by the acquisition environment factors pixel by pixel. The reconstructed temperature field is subtracted from the temperature component pixel by pixel to obtain the remaining temperature rise field. The regression relationship is used to characterize the components of the temperature distribution explained by irradiance and ambient temperature.

3. The deep learning-based distributed photovoltaic module hot spot fault detection method according to claim 1, characterized in that: The process of identifying precursor streaks composed of linearly extending temperature rises and determining the extension direction, centroid, and intensity of each precursor streak includes: The directional responses of each pixel in the remaining temperature rise field are extracted in multiple candidate directions. The ratio of the maximum directional response to the average directional response of the same pixel in multiple candidate directions is recorded as the linear saliency of the pixel. The directional responses are obtained by directional matching of the remaining temperature rise field through a directional filter. Connected regions composed of pixels whose linear salience is higher than a preset salience threshold are taken as candidate precursor regions. The direction in which the pixels contained in the candidate precursor region obtain the maximum directional response is taken as the extension direction of the candidate precursor region. Candidate precursor regions whose length along the extension direction is higher than a preset length threshold are determined as a precursor ridge. The centroid of the pixel coordinates contained in the precursor ripple is taken as its centroid, and the peak value of the region of residual temperature rise of the pixel contained in the precursor ripple in the residual temperature rise field is taken as its micro temperature rise intensity.

4. The deep learning-based distributed photovoltaic module hot spot fault detection method according to claim 1, characterized in that: The preset association condition is that the coordinate difference of the centroid of two precursor patterns with different inspection cycles does not exceed the preset position tolerance, and the directional deviation corresponding to the vector inner product of their extension directions does not exceed the preset directional tolerance.

5. The deep learning-based distributed photovoltaic module hot spot fault detection method according to claim 1, characterized in that: The infrared thermal imaging frame is also bound to the solar azimuth angle corresponding to the time of acquisition; When the change in solar azimuth angle in each inspection cycle exceeds the preset azimuth angle distinguishable threshold, the conditions for determining that they belong to the same precursor and establishing cross-cycle correlation must also be met: the precursor pattern involved in the correlation is identified in each inspection cycle involved, and the centroid of the precursor does not drift with the solar azimuth angle. The determination that the centroid does not drift with the solar azimuth angle includes: taking the maximum value of the ratio of the centroid displacement to the corresponding change in solar azimuth angle in each adjacent inspection cycle of the precursor as the anchoring measure; when the anchoring measure does not exceed the preset anchoring threshold, it is determined that the centroid of the precursor does not drift with the solar azimuth angle.

6. The deep learning-based distributed photovoltaic module hot spot fault detection method according to claim 5, characterized in that: When the change in solar azimuth angle during each inspection cycle is not higher than the azimuth angle discernibility threshold, the anchoring measurement is not calculated, and precursors that meet the following conditions are determined to be true precursors: The proportion of directed edges with positive difference values ​​in the evolutionary causal chain graph is higher than a preset cumulative threshold, the degree of directional consistency is higher than a preset consistency threshold, and the precursor is identified as a corresponding precursor pattern in each inspection cycle involved. The determination of the degree of directional consistency includes: for each centroid migration vector of the true precursor, calculating the inner product of its unit vector and the unit vector of the extension direction, and taking the minimum value of the inner product of each centroid migration vector as the degree of directional consistency.

7. The deep learning-based distributed photovoltaic module hot spot fault detection method according to claim 1, characterized in that: The determination of the heating rate sequence includes: for each directed edge of the true precursor, the segmented heating rate of the directed edge is obtained by dividing its micro-temperature rise intensity difference by the corresponding inspection cycle interval, and the segmented heating rates are arranged according to the inspection cycle to form the heating rate sequence. The determination of the expansion rate sequence includes: for each centroid migration vector of the true precursor, dividing its projection in the extension direction by the corresponding inspection cycle interval to obtain the segmental expansion rate, and arranging the segmental expansion rates according to the inspection cycle to form the expansion rate sequence.

8. The deep learning-based distributed photovoltaic module hot spot fault detection method according to claim 1, characterized in that: Determining the precursor degradation profile of the true precursor based on the heating rate sequence and the expansion rate sequence includes: The heating rate sequence and the expansion rate sequence are weighted, aggregated, and normalized with weights that increase from early to late with the inspection cycle to obtain the normalized heating rate and the normalized expansion rate. A two-dimensional plane is established with the heating rate as the first coordinate axis and the expansion rate as the second coordinate axis. The point with the heating rate and the expansion rate as the two coordinate components is taken as the deterioration state point. The degradation mode label is determined by the orientation of the degradation state point relative to the origin, and the degradation urgency is determined by the distance of the degradation state point relative to the origin. The orientation is the angle from the first coordinate axis to the point of degradation, and the distance is the Euclidean distance from the point of degradation to the origin.

9. The deep learning-based distributed photovoltaic module hot spot fault detection method according to claim 8, characterized in that: The determination of the degradation mode label includes: determining the degradation mode label by the sector into which the orientation of the degradation state point relative to the origin falls according to a preset mode orientation threshold set. The determination of the self-intensification flag includes: obtaining the heating acceleration by calculating the cycle-by-cycle change trend of the heating rate sequence along the inspection cycle, obtaining the expansion acceleration by calculating the cycle-by-cycle change trend of the expansion rate sequence along the inspection cycle, and setting the self-intensification flag to true when there is a positive one between the heating acceleration and the expansion acceleration that increases along the inspection cycle. The self-aggravating marker is used to characterize the worsening of the true precursor.

10. A distributed photovoltaic module hot spot fault detection system based on deep learning, characterized in that, include: The data acquisition module is used to acquire infrared thermal imaging frames of the same photovoltaic module in multiple inspection cycles. The temperature component determined by the acquisition environment factors is subtracted from the infrared thermal imaging frames of each inspection cycle to obtain the remaining temperature rise field. The feature extraction module is used to identify the precursors of temperature rise consisting of linearly extending temperature rises in the remaining temperature rise field of each inspection cycle, and to determine the extension direction, centroid, and micro-temperature rise intensity of each precursor. The cross-cycle correlation module is used to determine that the precursor patterns whose centroid and extension direction meet the preset correlation conditions in different inspection cycles belong to the same precursor and establish cross-cycle correlation. The causal graph construction module is used to set the micro-temperature rise intensity of the same precursor in each inspection cycle as a node, and establish directed edges between adjacent nodes from early to late according to the inspection cycle to carry the difference in micro-temperature rise intensity between two adjacent nodes, thus forming the evolutionary causal chain graph of the precursor, and taking the difference of the centroid of the precursor in adjacent inspection cycles as the centroid migration vector. The true precursor determination module is used to calculate the proportion of directed edges with positive difference values ​​in the evolutionary causal chain graph, and the degree of consistency between the direction of each centroid migration vector and the extension direction. Precursors with the proportion higher than a preset cumulative threshold and the degree of consistency in direction higher than a preset consistency threshold are determined as true precursors. The degradation profile generation module is used to determine the heating rate sequence and the expansion rate sequence for the true precursor based on the micro-temperature rise intensity difference of its directed edges and the centroid migration vector, respectively. Based on the heating rate sequence and the expansion rate sequence, the module determines the precursor degradation profile of the true precursor. The precursor degradation profile includes a degradation mode label, degradation urgency, and self-aggravation marker. The early warning output module is used to annotate the centroid, extension direction and the degradation profile of the precursor to the photovoltaic module and output an early warning trend map.

Images