A method and system for measuring solar radiation based on window state recognition

Abstract

The application discloses a kind of solar radiation measurement method and system based on window state identification, belong to solar radiation measurement and sensor state diagnosis technical field.The method is based on public clock synchronization to obtain radiation, dome window, sky and attitude data, recognize pollution state by background suppression and partition mapping, and generate window influence and attitude influence according to incident direction, sky radiance and detection response;Accordingly, output limited correction, quality record and disposal instruction, when static determination is uncertain, implement diagnostic excitation by heating or air flow pulse, disable automatic correction during recovery period and take verification again.The scheme can reduce the influence of window pollution and attitude deviation on measurement results, and improve effective data rate.

Description

Technical Field

[0001] This invention relates to the field of solar radiation measurement and sensor condition diagnosis technology, and in particular to a solar radiation measurement method and system based on window condition recognition. Background Technology

[0002] A solar radiation sensor with a hemispherical transparent dome receives incident radiation through the dome and the radiation detection unit generates measurement results. Dust accumulation, fixed obstructions, condensation, frost, ice, raindrops, water films, and snow accumulation on the dome surface can cause light transmission attenuation, scattering, or absorption, but it is usually difficult to distinguish these factors and their coverage locations based solely on radiation output.

[0003] Even if existing visual aids can capture images of the dome or sky, there may be problems such as asynchronous radiation, image, environmental and attitude data, interference from the sky background on the dome surface, failure to convert local pollution conditions into measurement effects according to the incident direction and detection response, and failure to jointly process cloud conditions, solar shading and installation attitude. Therefore, it is easy to cause insufficient basis for quality marking or correction.

[0004] Existing heating, defrosting, or cleaning controls are often triggered based on fixed cycles or single temperature and humidity conditions. When the confidence level of static visual classification is insufficient or when visual judgment is inconsistent with radiation changes, direct treatment may fail to verify the pollution category, and the temperature difference between the dome and the shell caused by heating, as well as the zero-point drift of radiation, may further affect the correction results.

[0005] Therefore, there is a need for a solar radiation measurement method and system based on window state recognition that can overcome the shortcomings of the existing technology. Summary of the Invention

[0006] One objective of this invention is to propose a solar radiation measurement method and system based on window state recognition. Addressing the limitations of existing solar radiation sensors in simultaneously identifying dome contamination, cloud conditions, and attitude, and in quantifying the impact of each factor on radiation measurements, this invention proposes a technical solution involving multi-source synchronous acquisition, window background suppression and partitioning mapping, joint sky and attitude recognition, construction of weighted influence quantities based on directional response, and constrained correction and closed-loop processing. When static visual confidence is insufficient or the visual influence quantity is inconsistent with the radiation residual, constrained heating pulses and airflow pulses are used as diagnostic stimuli, and automatic correction is disabled during thermal recovery. This solution improves the interpretability of contamination classification and measurement quality assessment, reduces the risk of erroneous corrections and thermal zero bias, and enhances the accuracy and effectiveness of radiation data.

[0007] This invention provides a solar radiation measurement method based on window status recognition, comprising:

[0008] S1. Simultaneously acquire raw solar radiation sampling values, window images of the hemispherical transparent dome, sky images, and attitude detection data to form a synchronous sampling data set;

[0009] S2. Perform sky background suppression and dome surface partitioning mapping on the window image to generate window state information including pollution category, estimated transmittance and confidence level, and generate sky state information and attitude state information based on the sky image and the attitude detection data.

[0010] S3. Map the window partitions to the direction of the incident celestial sphere. Based on the solar radiation detection direction response, weight the transmittance loss of each window partition and the relative radiance of the corresponding sky partition to generate the window influence quantity. Generate the attitude influence quantity based on the attitude state information and the sky state information.

[0011] S4. Determine the quality of the original solar radiation sampling value based on the window influence, the attitude influence, and the confidence level, and generate a radiation sampling quality record and processing instructions.

[0012] S5. Perform the treatment according to the treatment instructions and simultaneously re-collect the window image and the original solar radiation sampling value; when the pollution category cannot be determined according to the confidence level, or when the change of the window influence amount is inconsistent with the change of the original solar radiation sampling value, use a heating pulse or airflow pulse as a diagnostic excitation, review the pollution category according to the diagnostic response information before and after the excitation, and verify the treatment effect based on the re-collection results.

[0013] Optionally, S1 includes: sending a trigger signal with the same sampling sequence number to the radiation detection unit, the window imaging channel, the sky imaging channel, the active illumination unit, and the environment and attitude detection unit via a common clock;

[0014] The window imaging channel acquires the first window frame and the second window frame at adjacent exposure times when the active illumination unit is turned off and on, respectively.

[0015] The synchronous sampling data set also records each acquisition time, exposure parameters, lighting status, ambient temperature and humidity, dome surface temperature, shell temperature, and inertial measurement data.

[0016] Optionally, S2 includes: generating the window state information includes: registering the first window frame and the second window frame, and generating a background suppression map based on the registered pixel difference or pixel ratio;

[0017] Based on the pre-defined mapping relationship between image pixels and dome surface positions, the background suppression map is divided into multiple window partitions;

[0018] Extract at least one of brightness variation, texture, edge morphology and polarization features from each window partition to identify pollution categories in dust accumulation, fixed occlusion, condensation, frost, ice, raindrops, water film and snow accumulation. Determine the coverage rate based on the partition ratio of pixels of each category, determine the estimated transmittance based on the transmittance response ratio of the clean state and the current state, and determine the confidence level based on the degree of matching between the classification output and the distribution of the calibrated samples.

[0019] Furthermore, generating the sky state information and the attitude state information includes: extracting the actual solar image points, the confidence level of the actual solar image points, the solar occlusion status, cloud masking, cloud cover, cloud type, and relative radiance of sky zones from the sky image;

[0020] The theoretical solar vector is calculated based on the data acquisition time and the geographical location of the sensor.

[0021] When the sun is not obscured and the confidence level of the actual image point of the sun reaches the image point determination threshold determined by the calibrated sky image, the theoretical solar vector is compared with the observed solar vector calculated from the actual image point of the sun based on the camera calibration parameters to obtain the solar vector consistency residual.

[0022] When the sun is obscured or the confidence level of the actual image point of the sun does not reach the image point determination threshold, the solar vector consistency residual is marked as unusable.

[0023] The attitude deviation is obtained by comparing the gravity vector in the inertial measurement data with the reference gravity vector in the sensor calibration coordinate system.

[0024] The direct scattering state is determined based on the solar shading state, the cloud cover, and the relative radiance of the sky zones.

[0025] Optionally, S3 includes: determining the set of incident directions corresponding to each window partition based on the calibration mapping between the window partition and the incident celestial sphere direction;

[0026] The response weights of each incident direction set are determined by the directional response function or cosine response calibration table of the radiation detection unit;

[0027] The window influence amount is obtained by summing the product of the transmittance loss obtained by subtracting the estimated transmittance from the transmittance of each window partition, the relative radiance of the corresponding sky partition, and the response weight.

[0028] The incident angle of the direction response function is transformed according to the attitude deviation, and the influence of the direction deviation of the direct component and the influence of the cosine response deviation of the scattering component are calculated respectively in combination with the direct scattering state to obtain the attitude influence quantity.

[0029] Optionally, S4 includes: a quality state point composed of the window influence, the attitude influence, and the confidence level;

[0030] Based on the radiation deviation distribution corresponding to different quality state points in the cleanroom calibration data, the automatic correction allowable area and the effective allowable area are determined, wherein the automatic correction allowable area is located within the effective allowable area;

[0031] The category determination threshold is determined based on the cleanliness status and the confidence distribution of the calibrated samples for each pollution category;

[0032] When the quality status point is located within the automatic correction allowable area and its confidence level reaches the category determination threshold, the original solar radiation sample value is corrected according to the window influence amount and the attitude influence amount, and a limited optional correction value is generated.

[0033] When the quality status point is outside the valid allowable area, the corresponding radiation sample is marked as invalid and automatic correction is disabled;

[0034] In other cases, the corresponding radiation sample is marked as suspicious and automatic correction is disabled;

[0035] The radiation sampling quality record includes the original solar radiation sampling value, the restricted optional correction value, the quality level, the cause code, and the disposal instruction, and writes the pollution category, area status, and attitude deviation that triggered the quality determination into the cause code.

[0036] Optionally, S5 includes: heating treatment instructions for condensation, frost, ice or snow formation, which are constrained by maximum power, dome allowable temperature rise and duration.

[0037] Generate cleaning alerts for dust accumulation or fixed obstructions;

[0038] Automatically correct any delays caused by raindrops or water films generating precipitation.

[0039] A sampling identifier is generated for each of the synchronous sampling data groups by a common clock used for synchronous acquisition;

[0040] After each processing, the window state information, the window influence amount, and the attitude influence amount are regenerated, and the corresponding results before and after the processing are written into the anomaly recovery verification record associated with the same sampling identifier;

[0041] Furthermore, using heating pulses or airflow pulses as diagnostic excitation includes: determining a clean baseline based on the original solar radiation sampling value during the clean period, and determining a consistency criterion based on the joint calibration data of the clean period and the known contamination period. The joint calibration data includes the window effect and the radiation residual of the original solar radiation sampling value relative to the clean baseline. The consistency criterion is used to determine whether the change of the window effect is consistent with the change of the original solar radiation sampling value.

[0042] During the period when freezing radiation output is allowed, the clean baseline and the pre-excitation baseline are latched. A heating pulse constrained by maximum power, dome allowable temperature rise and duration is applied. Window images, temperature data and electrical parameters are collected before, during and after the pulse. The area shrinkage rate, edge migration speed, texture entropy change, optical flow, estimated transmittance recovery rate and optical recovery amount corresponding to unit input energy are calculated according to window partitions to form a thermal response vector.

[0043] An airflow pulse is applied to the window partition where the thermal response vector still does not pass the category distinction criterion determined by the calibration sample, and the displacement, deformation, shedding rate and re-coverage rate of the polluted area are calculated to form the aerodynamic response vector;

[0044] By combining the thermal response vector, the aerodynamic response vector, ambient temperature and humidity, precipitation duration, and dome surface temperature, the pollution category is verified, and disposal instructions such as continued heating, airflow cleaning, delayed disposal, or artificial early warning are generated.

[0045] Furthermore, from the start of the diagnostic excitation until the difference between the dome surface temperature and the shell temperature returns to the allowable range corresponding to the baseline before excitation and the radiation zero-point drift returns to the cleanroom zero-point fluctuation range, the radiation sampling quality record remains in an automatic correction disabled state and is written with the thermal zero-bias risk cause code.

[0046] The treatment is considered successful when the influence of the window decreases after treatment compared to before treatment, and the original solar radiation sample value after re-collection is consistent with the direction of radiation change determined based on the direct scattering state at the same time.

[0047] After confirming successful treatment, the clean baseline is updated only when the public clock sampling cycles continuously reach a preset number and all meet the allowable range, the radiation change direction is consistent, and the window influence amount remains within the verification boundary determined by the treatment verification sample. The preset number is determined by the clean period radiation zero-point stability calibration.

[0048] On the other hand, the present invention also provides a solar radiation measurement system based on window state recognition, comprising:

[0049] The synchronous acquisition module includes a radiation detection unit, a window imaging channel, a sky imaging channel, an active illumination unit, an environment and attitude detection unit, and a common clock that sends synchronous trigger signals to each unit, for forming the synchronous sampling data set; a state recognition module is used to generate the window state information, the sky state information, and the attitude state information; an influence quantity construction module is used to generate the window influence quantity and the attitude influence quantity; a quality control module is used to generate the radiation sampling quality record and, when the correction constraint is met, generate a limited optional correction value based on the window influence quantity and the attitude influence quantity; and a closed-loop treatment module includes a heating device and an airflow device, for executing the treatment command, applying the diagnostic excitation, verifying the contamination category, verifying the treatment effect based on the resampling results, and maintaining a clean baseline for comparing radiation changes.

[0050] The beneficial effects of this invention are:

[0051] 1. By synchronizing multi-source data through a common clock and using active illumination off frames and on frames to form a background suppression map and dome surface calibration mapping, the interference of sky background changes on window pollution identification can be reduced, and the pollution category, coverage, estimated transmittance and confidence level of each zone can be obtained.

[0052] 2. By mapping the viewport partitions to the direction of the incident celestial sphere, and combining the relative radiance of the sky partitions, the directional or cosine response of the radiation detection unit, as well as the attitude deviation and direct scattering state to generate influence quantities, traceable quantitative basis can be provided for restricted automatic correction, quality level, cause code and disposal instructions.

[0053] 3. When static judgment is uncertain or the visual impact quantity is inconsistent with the radiation residual, heating pulses and airflow pulses subject to safety constraints are used as diagnostic stimuli. Automatic correction, post-treatment re-verification, and updating of the clean baseline are disabled during thermal recovery. This can reduce contamination misjudgment, thermal zero bias correction, and ineffective treatment, forming a closed loop of diagnosis, quantification, treatment, and verification. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0055] Figure 1 This is a flowchart of the solar radiation measurement method based on window state recognition according to the present invention.

[0056] Figure 2 This is a flowchart of the S2 window state, sky state, and attitude state recognition of the present invention. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0058] refer to Figures 1-2 A solar radiation measurement method based on window state recognition, comprising:

[0059] S1. Simultaneously acquire raw solar radiation sampling values, window images of the hemispherical transparent dome, sky images, and attitude detection data to form a synchronous sampling data set;

[0060] S2. Perform sky background suppression and dome surface partitioning mapping on the window image to generate window state information including pollution category, estimated transmittance and confidence level, and generate sky state information and attitude state information based on the sky image and the attitude detection data.

[0061] S3. Map the window partitions to the direction of the incident celestial sphere. Based on the solar radiation detection direction response, weight the transmittance loss of each window partition and the relative radiance of the corresponding sky partition to generate the window influence quantity. Generate the attitude influence quantity based on the attitude state information and the sky state information.

[0062] S4. Determine the quality of the original solar radiation sampling value based on the window influence, the attitude influence, and the confidence level, and generate a radiation sampling quality record and processing instructions.

[0063] S5. Perform the treatment according to the treatment instructions and simultaneously re-collect the window image and the original solar radiation sampling value; when the pollution category cannot be determined according to the confidence level, or when the change of the window influence amount is inconsistent with the change of the original solar radiation sampling value, use a heating pulse or airflow pulse as a diagnostic excitation, review the pollution category according to the diagnostic response information before and after the excitation, and verify the treatment effect based on the re-collection results.

[0064] In this specific embodiment, S1 includes:

[0065] In this specific embodiment, the solar radiation sensor includes a radiation detection unit, a viewing window imaging channel facing the hemispherical transparent dome, a sky imaging channel facing the entire sky, an 850nm near-infrared active illumination unit surrounding the viewing window imaging channel, a diffuse reflection transmission reference ring located on the outer edge of the dome and within the field of view of the viewing window imaging channel, an environment and attitude detection unit, a synchronous reference radiometer data interface, and a common clock. Each emission branch of the active illumination unit emits a narrow-angle beam from the inside of the dome to the corresponding zone via an extinction collimating tube. After passing through the dome, the beam reaches the diffuse reflection transmission reference ring, and the returning light passes through the same zone again before being received by the viewing window imaging channel. A light shield is placed in front of the lens, and an extinction layer is placed in the non-measurement area inside the dome to suppress direct light that does not pass through the dome and stray reflections from the inner surface. Each unit is connected to the same acquisition controller, which establishes a synchronous sampling data group at a basic sampling frequency of 10Hz and writes the sensor number, common clock sampling sequence number, Coordinated Universal Time, geographical latitude and longitude, and firmware version into the group head record.

[0066] The common clock generates a hardware trigger edge with the same sampling number in each sampling period. It first triggers the integration of the radiation detection unit and the exposure of the sky imaging channel. Then, it sends the active illumination off flag and the active illumination on flag in the two adjacent window exposure time slots respectively. The center interval of the two window exposure time slots is set to 20ms. The active illumination on frame uses the same exposure time, analog gain and lens focal length as the off frame, so that the first window frame and the second window frame have the same geometric and exposure reference.

[0067] The acquisition controller uses the common clock trigger time and the predetermined time slot offset of each channel as synchronization references, according to... Calculate the maximum time residual of the nth synchronously sampled data group. The symbol represents the maximum time residual of each channel within the nth synchronous sampling data group relative to its respective predetermined sampling time, in milliseconds. The symbol represents the hardware timestamp of the j-th acquisition channel in the n-th synchronous sampling data group, in milliseconds. The symbol indicates the corresponding common trigger time, in milliseconds (ms). The symbol represents the predetermined time slot offset of the j-th acquisition channel relative to the common trigger time, in milliseconds (ms). For the radiometric detection channel, sky imaging channel, synchronous reference radiometer channel, and active illumination off window frame, the offset is 0 ms; for the active illumination on window frame, the offset is 20 ms. The symbol represents the acquisition channel index. The symbol represents the common clock sampling sequence index; the allowed time residual is calibrated to 5ms, and the absolute error of the measured center interval between the two window frames relative to the predetermined interval of 20ms is limited to 2ms. If any threshold is exceeded or a necessary channel is missing, the group is written into the synchronization abnormal flag and is prohibited from entering the subsequent automatic correction link.

[0068] The radiation detection unit completes one integration sampling after the trigger edge, and writes the original solar radiation sample value, integration time, detector temperature, range setting, zero-point calibration version and saturation flag without window and attitude correction into the radiation record. The original solar radiation sample value is in W / m². If the analog-to-digital conversion result reaches the upper limit of the range or the zero-point calibration version is inconsistent with the current equipment configuration, the original sample value is retained and the radiation record is marked as not automatically correctable.

[0069] The synchronous reference radiometer outputs reference radiation value, hardware timestamp, calibration certificate number, traceability version, range setting, saturation flag, and device status word to the acquisition controller via an isolated data interface. The unit of reference radiation value is W / m². The acquisition controller matches the reference record with the nearest sampling sequence number according to the common clock. The synchronous reference radiometer's valid flag is set to true only when the absolute time residual between the reference record and the common trigger time is no greater than 5ms, the calibration certificate is valid, the traceability version is consistent with the current configuration, the device is not saturated, and the device status word is normal. Otherwise, the original reference reading is retained, a missing or invalid reason code is written, and clean baseline establishment, disposal verification, and baseline updates that rely on this reading are prohibited.

[0070] The window imaging channel outputs the first and second window frames in 1600×1600 pixel, 12-bit grayscale format, while the sky imaging channel outputs the sky image in 1024×1024 pixel, 12-bit grayscale format. Each image record saves the frame number, exposure start and end times, exposure time, gain, illumination status, lens temperature, bad pixel mask version, and polarization acquisition configuration flag. In this specific implementation, a normal single-channel grayscale window imaging configuration is used, the polarization acquisition configuration flag is set to unconfigured, polarization angle images are not acquired, and polarization degree is not inverted from grayscale intensity. The acquisition controller uses the factory dark field image to subtract fixed pattern noise and only performs neighborhood median replacement on the pixels marked by the bad pixel mask, without changing the original response of the other pixels.

[0071] The environment and attitude detection unit reads air temperature, relative humidity, dome surface temperature, shell temperature, and data from the three-axis accelerometer and three-axis gyroscope in the same trigger cycle. The temperature unit is ℃, the relative humidity is percentage, the acceleration unit is m / s², and the angular velocity unit is ° / s. Five consecutive inertial measurement sub-samples are written into the inertial measurement data field after median filtering, and the sensor range, temperature compensation version, and effective bitmap are saved simultaneously.

[0072] The acquisition controller calculates integrity check codes for radiation records, synchronous reference radiometer records, first window frames, second window frames, sky images, environmental records, and inertial measurement records under the same sampling sequence number, and writes them into the circular synchronization buffer using the sampling sequence number as the primary key. For channels that arrive repeatedly, the earliest record is retained according to the hardware timestamp. Data groups that fail the check or are still incomplete after more than two sampling cycles are transferred to the abnormal buffer and are not concatenated with the next sampling sequence number. When a synchronous reference radiometer record is missing, its missing reason code is allowed to be saved, but the group must not be marked as a valid reference.

[0073] The synchronous cache outputs a complete set of synchronous sampling data with valid time deviation and paired lighting status. It then transmits the data address, valid bitmap, synchronous reference radiometer record and its valid flag and synchronous anomaly flag to the status identification module of S2. At the same time, it writes the original image, original radiation value, synchronous reference radiation value and environmental attitude data into the append-only data area according to the sampling identifier, so that S4 can generate quality records and S5 can trace the same sampling chain before and after the treatment.

[0074] In this specific embodiment, S2 includes:

[0075] The state recognition module reads the synchronous sampling data group output by S1, takes the first window frame as the reference image, first calculates the homography transformation from the second window frame to the first window frame based on the fixed marker points on the edge of the dome, and then completes the registration using the sub-pixel displacement field obtained by minimizing the reprojection error of the marker points. In this specific embodiment, the root mean square error threshold of the reprojection of the marker points is calibrated to 0.6 pixels. When the threshold is exceeded, the pollution classification of the current image pair is stopped and the window state information is written as the low confidence state of registration.

[0076] Calculate the off and on pixels of the registered active illumination. To generate a background suppression map, The symbol represents the dimensionless background suppression response at pixel coordinates (x, y), with values ​​clipped to [-1, 1]. The symbol represents the registered active illumination enabled pixel response, and the unit is digital quantization value. The symbol represents the corresponding active illumination off pixel response, and the unit is digital quantization value. The symbol represents the lower limit of the response to prevent division by zero in low light conditions. In this specific embodiment, 16 digital quantization values ​​are used. and The symbols represent the column coordinate index and row coordinate index of the window image, respectively, and the two terms in the denominator have the same unit as the lower limit of the response;

[0077] The window pixel and dome surface position calibration table stores the device number, pixel coordinates, dome polar angle, dome azimuth angle, window partition number, calibration temperature and version number. This table is obtained by the turntable projecting point light sources at 2° angle intervals and fitting a third-order radial distortion model under clean dome conditions. The status recognition module uses pixel coordinates as the lookup key to map the background suppression map into 72 window partitions consisting of 12 azimuth sectors and 6 polar angle zones. Pixels that fall outside the dome boundary or have a mapping residual greater than 1.0 pixel are not included in the partition statistics.

[0078] Each window partition generates a partition feature record, recording the mean and standard deviation of the active illumination response, the local binary texture histogram, the Canny edge density, the aspect ratio of the connected components, and the structural similarity with the clean reference frame. All features are standardized using the partition mean and standard deviation of the same exposure level during the clean period and cropped to [-5,5]. Polarization is only added to the partition feature record when the polarization acquisition configuration flag indicates that the window imaging channel has 0°, 45°, 90°, and 135° micro-polarization units, the effective pixel ratio of the four polarization directions is not less than 80%, and the classification model version declares the use of polarization input. In this specific implementation, the ordinary grayscale configuration marks the polarization as not acquired and uses the corresponding model version that does not contain the polarization field in both training and inference. It is forbidden to replace the missing polarization with 0 or grayscale intensity. When the polarization configuration is inconsistent with the model version, the partition is marked as model input mismatch and output to determine the contamination category is prohibited. The clean reference frame is searched according to the equipment number, exposure level, dome temperature range, and calibration version. When no reference frame is matched, the partition is marked as reference missing and output to determine the contamination category is prohibited.

[0079] Pollution classification first uses the partition feature records as input to run a gradient boosting tree classifier, which outputs partition probability vectors for nine categories: clean, dust, fixed occlusion, condensation, frost, ice, raindrops, water film, and snow. The highest probability category is always saved as the confidence candidate category field. At the same time, the background suppression map is input into a lightweight convolution classifier in the same calibration package in 16×16 pixel blocks. The nine class probabilities of each block are output and backfilled into the effective pixels in the block to form a pixel-level category mask. Pixels with the highest probability below 0.60 are written as unknown category and are not included in any pollution category. Both classifiers were trained using samples collected from the same dome at temperatures ranging from -20°C to 50°C and relative humidity from 20% to 100%, and labeled by two people. The training set and validation set were separated by device number, and the model version was saved with the calibration package. When the difference between the probability of the highest category and the probability of the second highest category in a partition is less than 0.10, the probability vector and confidence candidate category field are retained, but the final partition pollution status label is uniquely marked as pending verification. The candidate categories are only used for distance-type candidate confidence and diagnostic records and must not be used as input for determining pollution categories or automatic correction access control.

[0080] For the i-th window partition according to Calculate the coverage of category c. The symbol represents the dimensionless coverage of category c within the i-th window partition, with values ​​in [0,1]. The symbol represents the number of valid pixels in the pixel-level category mask that are classified as category c, expressed in pixels. The symbol represents the total number of valid pixels in the partition, excluding pixels with unknown categories, pixels outside the dome boundary, and pixels with excessive mapping residuals, expressed in units of pixels. The symbol represents the window partition index. The symbol represents the pollution category index. When the effective pixel ratio is less than 80%, the coverage determination value is not output and is written to the partition occlusion anomaly flag.

[0081] The state recognition module obtains the active illumination difference map by subtracting the corresponding off pixel response from the registered active illumination on pixel response, with negative difference values ​​set to 0. The dimensionless background suppression map is only used for pollution classification and not as a digital input for estimating transmittance. In the active illumination difference map, only the reference ring imaging area that matches the two-way transmission optical path formed by the emission branch, dome partition, and diffuse reflection transmission reference ring participates in the transmission response calculation. The remaining reflection or scattering areas are excluded by the optical path mask. The clean transmission calibration table is established by applying standard samples of clean, dust accumulation, condensation, frost, ice, water film, and snow accumulation to the same dome at various incident angles, exposure levels, and dome temperature ranges. The 850nm active illumination difference response is only used as an online observation. For each standard sample to be included in the automatic correction, an independent spectrophotometer simultaneously measures the single-pass spectral transmittance before and after pollution, covering the effective spectral response range of the solar radiation detection unit. The transmittance is then weighted according to the spectral response function of the detection unit and the spectrum of the calibration light source to obtain a broadband equivalent single-pass transmittance target value. The two-pass active illumination difference response is then subjected to monotonic regression inversion of this target value by pollution category, incident angle, exposure level, and temperature range. The inversion result is then linearized into a broadband single-pass transmittance equivalent digital response. Only when the solar radiation detection unit is explicitly defined as a narrow-band object whose spectral response matches that of 850nm active illumination, the broadband equivalent target value degenerates into the 850nm single-pass transmittance measured by the independent spectrophotometer. The table stores the pollution category, zonal incident angle, exposure level, temperature range, detector unit spectral response version, calibration spectral version, monotonic effective range, 95% upper limit of wideband mapping residual, sample quantity, and version number; only the calibration range that has passed the monotonicity test, has a validation set determination coefficient of not less than 0.98, and a wideband equivalent response error of not more than 2% is retained. Only when the final zonal pollution status label is a determined category among the nine categories and the determined category confidence validity flag is true, the module looks up the table according to the determined pollution category and the aforementioned conditions, maps the current and clean two-way difference response to wideband single-way transmission equivalent digital response, and then... Determine the estimated transmittance. The symbol represents the dimensionless broadband equivalent one-way transmittance of the i-th window partition, weighted by the spectral response of the solar radiation detection unit. The symbol represents the mean of the broadband single-pass transmission equivalent difference response obtained by monotonic inversion after removing 2% of the top and bottom anomalous pixels within the optical path mask of this partition of the current active illumination difference map. The unit is digital quantization value. The symbol represents the mean of the equivalent difference response of broadband single-pass transmission under the same conditions in the clean transmission calibration table, and the unit is a digital quantized value. The symbol represents the clipping function that restricts the ratio to a given closed interval, where the index symbol i follows the complete interpretation of the i symbol in the aforementioned coverage formula; when the final partition contamination status label is pending verification, the category confidence validity flag is false, the optical path mask is mismatched, the independent calibration record is missing, the incident angle, exposure level or temperature range is mismatched, the detector unit spectral response version or calibration spectral version is mismatched, the contamination category has no wideband mapping, the current response exceeds the monotonic effective range, the wideband mapping residual exceeds the limit, the cleanliness difference response is less than 64 digital quantization values, or the current partition effective difference pixel ratio is less than 80%, the category-specific wideband mapping is not performed and the estimated transmittance is marked as unavailable, and it is transferred from S3 to unknown response weight; the saved confidence candidate categories are only used for diagnostic records, and no diagnostic estimated transmittance that can enter the S3 automatic correction link is generated, the affected direction is prohibited from participating in automatic correction, and it is not replaced by the ratio of the unmapped current response to the clean response or the 850nm single band transmittance;

[0082] Calculate based on the degree of matching between the current partition feature vector and the distribution of labeled samples for each category. , The symbol represents the dimensionless candidate confidence of the i-th window partition relative to the candidate confidence category, and takes the value (0,1]. The symbol represents an exponential function with the natural constant as its base, whose input is a dimensionless quantity and whose output is a dimensionless quantity. The symbol represents the Mahalanobis distance from the current standardized feature vector to the center of the candidate confidence class sample, with a dimension of 1. The symbol represents the Mahalanobis distance scale parameter corresponding to the candidate category of the confidence score. It has a dimension of 1 and is taken as the larger of the 90th percentile of the Mahalanobis distance of the same category in the validation set and 0.05, thus ensuring that the scale parameter is strictly greater than 0. The symbol represents the confidence candidate category index stored in the i-th window partition, where the index symbol i follows the complete interpretation of the i symbol in the aforementioned coverage formula; when the final partition contamination status label is pending review, the candidate confidence is still calculated according to the sample center and scale parameters of the stored candidate categories and written into the diagnostic record along with the probability vector, but the valid flag for determining the category confidence is set to false, S4 directly enters the pending review branch, and the candidate confidence is prohibited from participating in the dominant category gate; if the condition number of the covariance matrix exceeds the calibration upper limit, the diagonal covariance is used instead and written into the model degradation flag;

[0083] The sky state recognition branch performs dark field subtraction, fisheye distortion correction, and radiometric response linearization on sky images with the same sampling sequence. The corrected image is then fed into a lightweight U-Net segmentation model, which outputs a solar image heatmap, a cloud mask, and a clear sky mask. The model is trained using manually annotated sky images of the sun center, cloud regions, and occlusion states in all four seasons. The input size is 512×512 pixels. The actual solar image point is taken from the position of the maximum value of the heatmap. When the confidence of the actual solar image point is lower than the 0.82 threshold determined by the 95% recall rate of the validation set, the actual solar image point is marked as unusable. For the corrected pixels within the cloud mask, the brightness quantile, texture scale, edge density, and color ratio are extracted and input into the cloud classifier. The output results are thin clouds, cumulus clouds, stratiform clouds, convective clouds, and undetermined clouds. When the probability of the highest class is lower than 0.70, the undetermined cloud class is output. The class set, threshold, and validation set confusion matrix are saved with the sky model version.

[0084] The state recognition module aggregates sky pixels according to the same celestial azimuth and polar boundary as the viewport partition, and then calculates the result after conversion using camera response calibration coefficients. , The symbol represents the dimensionless composite relative radiance of the k-th sky region, with a value of [0,1]. The symbol represents the average equivalent radiance of the k-th sky zone, in units of W / (m²·sr). For zones without a solar disk, it is obtained from the average pixel value after removing saturated pixels. For zones containing a solar disk, the sky imaging channel adds short-exposure high dynamic range frames within the same common clock transaction. The module fits the point spread function with the unsaturated pixels at the outer edge of the solar disk and integrates it. Then, an independent synchronous reference radiometer with matching field of view, time, and range is used to perform absolute radiometric calibration on the integrated value to obtain the direct equivalent radiance independent of the original value of the radiometric detection to be corrected and replenish it. The original value of the radiometric detection to be corrected is only used as the correction object of S4 and the consistency comparison object of S5, and does not participate in the construction of solar directional radiance or any self-correction factor. The symbol represents the average equivalent radiance of the m-th effective sky zone, expressed in W / (m²·sr). Its calculation source is the same as that of the average equivalent radiance of the k-th sky zone mentioned earlier. The symbol represents the strict positive lower limit of the equivalent radiance of the sky, with units of W / (m²·sr), calibrated by the 95% upper limit of dark field noise, and taken as 0.1 W / (m²·sr in this specific embodiment. and The symbols all represent sky partition indices, with the latter used to calculate the maximum value. When the maximum value of the average equivalent radiance of all valid partitions is not higher than this lower limit, the relative radiance field is marked as unavailable due to low illumination. When the short exposure frame saturation or the point spread function fitting residual exceeds the calibration threshold, the independent synchronous reference radiometer is invalid, or time or field-of-view matching fails, the solar direction partition is marked as unavailable due to direct weighting, and the original values ​​of the radiometric detection to be corrected are not reused. If the number of valid sky partitions is less than 70% of the total, the relative radiance field is marked as incomplete, and automatic correction using the affected direction is prohibited in any of the aforementioned states.

[0085] The theoretical solar vector is obtained by substituting the sampled Coordinated Universal Time (UTC), sensor latitude and longitude, and altitude into the solar position algorithm. The solar position algorithm first outputs a vector in a right-handed local East-North-Sky coordinate system, where the x-axis points east, the y-axis points north, and the z-axis points to the sky. A calibration table stores a 3×3 orthogonal rotation matrix from the local East-North-Sky coordinate system to the sensor calibration coordinate system. This matrix is ​​generated based on the true north direction measured on-site and the horizontal reference given by the electronic level, and is associated with the equipment number, measurement time, true north magnetic declination correction, matrix transformation direction, and version number. The theoretical solar vector is obtained by left-multiplying the theoretical vector in the local East-North-Sky coordinate system by this rotation matrix and normalizing it to units. The observed solar vector is then calculated back to the same sensor calibration coordinate system using the sky camera's intrinsic parameters, the fisheye projection model, and the actual solar image point, and normalized to units. If the calibration table is missing, the version is mismatched, the determinant of the rotation matrix is ​​not positive, the orthogonality residual is greater than 0.01, or any unit normalized denominator is not positive, the theoretical solar vector, the solar vector consistency residual, the solar direction selection result, and the yaw correction will all be marked as unusable, and the unusable rollback path described below will be used. After completing the above common coordinate system transformation, according to Calculate the solar vector consistency residual. The symbol represents the solar vector consistency residual, in degrees. The symbol represents pi (π), with a dimension of 1. The coefficient 180 / π is used to convert the radians output by the arccosine function to degrees. The symbol represents the inverse cosine function. The input is a dimensionless vector dot product restricted to [-1, 1] by a clipping function. The output unit is rad. The symbol represents the normalized theoretical solar vector, with a dimension of 1. The symbol represents the normalized observed solar vector, with a dimension of 1. The symbol represents the vector transpose operation. The symbol indicates that the clipping function in the aforementioned estimated transmittance calculation is used and the input is restricted to the closed interval [-1,1] according to the parameter order. When the sun is covered by cloud cover or the confidence of the actual image point of the sun does not reach the threshold, the residual field is written as unavailable.

[0086] The attitude branch normalizes the acceleration vector in the inertial measurement data into the measured gravity vector after temperature compensation and a 2Hz low-pass filter, according to... The attitude deviation was obtained. The symbol represents the attitude deviation of the sensor relative to the calibration attitude, in degrees (°). The symbol represents pi (π), with a dimension of 1. The coefficient 180 / π is used to convert the radians output by the arccosine function to degrees. The symbol represents the complete interpretation of the inverse cosine function in the aforementioned solar vector consistency residual formula. Its input is also a dimensionless vector dot product restricted to [-1,1] by a pruning function, and the output unit is rad. The symbol represents the normalized measured gravity vector, with dimensions of 1. The symbol represents the normalized reference gravity vector in the sensor calibration coordinate system, with a dimension of 1. The symbol represents the vector transpose operation. The symbol indicates that the pruning function used in the aforementioned solar vector consistency residual calculation is used and the input is restricted to the closed interval [-1,1] according to the parameter order; the attitude branch uses the calibration quaternion recorded at the time of installation as the initial value, updates the quaternion using the three-axis gyro angle increment, corrects pitch and roll with gravity vector, and corrects yaw drift around the gravity axis using theoretical solar vector and observed solar vector when the solar image point is valid. Normalization is performed after each update; the absolute value of the difference between the measured acceleration modulus and the 9.80665m / s² standard gravitational acceleration is divided by the standard gravitational acceleration to obtain the dimensionless relative deviation. When the relative deviation exceeds 0.051, the gyro integral attitude is retained and the attitude confidence is reduced. When there is no solar vector correction for 60 consecutive seconds and the cumulative yaw uncertainty exceeds 2°, the azimuth attitude is marked as invalid.

[0087] The direct scattering state is determined in a mutually exclusive order by the solar shading state, the cloud cover calculated using cloud cover masking, and the relative radiance within a 15° radius around the sun: First, the case where the sun is shaded or the cloud cover is not less than 80% is identified and marked as scattering-dominant; only when the sun is not shaded and the cloud cover is less than 80% is the ratio of the peak relative radiance around the sun to the median value of the entire sky determined to be not less than the calibrated threshold of 3.0. If this condition is met, the state is marked as direct scattering-dominant, and the rest are marked as mixed states and written into the dimensionless direct scattering mixing coefficient, thus ensuring that there is only one state result in the same sampling period;

[0088] The state recognition module outputs window state information, sky state information, and attitude state information with the sampling identifier as the primary key. The window state information stores the aforementioned pollution judgment field, pixel-level category mask, coverage, transmission response field, confidence field, and valid flags in each partition. The sky state information stores the actual solar image point, the confidence level of the actual solar image point, the solar shading status, cloud area mask, cloud cover, cloud type, comprehensive relative radiance, direct sunlight weight valid flag, and direct sunlight scattering status. The attitude state information stores the theoretical solar vector, the observed solar vector, the solar vector consistency residual, the attitude deviation, the unit quaternion from the calibration coordinate system to the current coordinate system and the corresponding 3×3 rotation matrix, the yaw uncertainty, and each valid flag. The three types of state information are then transmitted to S3.

[0089] In this specific embodiment, S3 includes:

[0090] The influence quantity construction module reads the three types of status information output by S2, maps the window partition to the incident celestial sphere direction, and performs weighted processing based on the solar radiation detection direction response. The module forms a candidate record set from the calibration table of the window partition to the incident celestial sphere direction according to the device number, partition number and calibration version. The candidate records are sorted from small to large according to the pixel mapping residual. When the residuals are the same, the record with the same version number as the current device and the newer calibration time is selected first. The selected record saves the dome polar angle, azimuth angle, unit direction vector and fixed angle and forms a corresponding direction sample set. The calibration table is established by scanning each direction by the factory turntable with collimating light source. The partition that fails to find is written with the no mapping mark and does not participate in automatic correction.

[0091] For the directional samples contained in the i-th window partition, calculate the directional response function of the radiation detection unit and the fixed angle of the directional samples. , The symbol represents the dimensionless normalized response weight of the i-th window partition, and the sum of the weights of all valid partitions is 1. The symbol represents the set of incident direction samples corresponding to the i-th window partition. The symbol represents the index of the directional sample within the set. The symbol represents the dimensionless radiation detection response given by the directional response calibration table at the h-th directional sample. and The symbols represent the polar angle and azimuth angle of the h-th sample, respectively, in degrees. The symbol represents the fixed angle represented by the h-th direction sample, in sr. The symbol represents the set of all valid incident direction samples mapped by the current device. The symbol represents the index of the directional sample within the set. The symbol represents the dimensionless radiation detection response at the u-th effective direction sample, given by the response calibration table for the same direction. and The symbols represent the incident polar angle and azimuth angle of the u-th valid direction sample, respectively, in degrees. The symbol represents the fixed angle represented by the u-th valid direction sample, in units of sr; where the index symbol i follows the complete interpretation of the i symbol in the S2 coverage formula; the division is performed only when the set of all valid direction samples is not empty and the sum of the denominator response fixed angles is strictly greater than the positive lower limit determined by the repeatability of the direction response calibration and the numerical resolution, and the sum of all valid partition weights is declared to be 1. Otherwise, the division is not performed, all normalized response weights are marked as unavailable, the direction mapping error or all-zero response reason code is written, and the window influence quantity and attitude influence quantity constructed by them are prohibited from entering the automatic correction link;

[0092] The directional response calibration table uses the equipment number, temperature range, incident polar angle, and azimuth angle as search keys. It saves the normalized response, repeated measurement standard deviation, fixed angle, and version number. The table data is obtained by scanning 2° at a time under constant collimated irradiance under clean dome conditions and normalizing with zero incident angle response. When the current temperature exceeds the effective range in the table, the nearest temperature range is used but the state uncertainty flag is raised. When the directional response table is missing, the cosine response calibration table is only used if the detector has passed the cosine response acceptance.

[0093] The influence quantity construction module establishes a fixed-angle overlap index between the viewport partition and the sky partition through its direction set, and performs a fixed-angle weighted average of the comprehensive relative radiance of the overlapping sky partitions, according to... The amount of influence generated by the viewport. The symbol represents the dimensionless window effect caused by the loss of window transmittance. The value is clipped to [0,1], and the larger the value, the greater the expected radiation loss. The symbol represents the set of window partitions in this calculation that are effective in both the broadband equivalent estimated transmittance and the overall relative radiance including the independent direct input compensation. The symbol represents the window partition index within the set. The symbols follow the complete interpretation of the aforementioned normalized response weight formula and are used in the set after removing invalid partitions. Internal renormalization The symbols are retained from the complete interpretation of the broadband equivalent one-way transmittance weighted by the spectral response of the solar radiation detection unit in S2. The symbol represents the dimensionless composite relative radiance corresponding to the incident direction set of the i-th window partition, and whose available independent direct components have been replenished. This direct component does not include the original values ​​of the radiation detector to be corrected. The symbol represents the dimensionless lower limit for preventing division by zero in low radiance, which is set to 0.01 in this specific embodiment. The numerator and denominator use the same set and are both dimensionless response weighted quantities.

[0094] When the estimated transmittance or corresponding relative sky radiance of a certain window partition is unavailable, the module does not replace it with the default clean value, but instead transfers the response weight of that partition to the unknown weight field. Only when the cumulative response weight of the known partition is not less than 0.85 and the unknown partition does not cover the theoretical solar direction, will the remaining weight be normalized and the window influence amount be calculated. Otherwise, the window influence amount will be marked as unavailable for automatic correction.

[0095] The module directly reads the complete attitude quaternion and its 3×3 rotation matrix with the same sampling identifier from the S2 attitude state information, and then... Change each incident direction, The symbol represents the h-th normalized incident direction vector after the attitude change, with a dimension of 1. The symbol S2 represents a 3×3 orthogonal rotation matrix from the calibration coordinate system to the current coordinate system, output after fusing the initial attitude calibration, gyro integral, gravity correction, and solar vector yaw correction. It has a dimension of 1. The symbol represents the h-th normalized incident direction vector under the calibration attitude, with a dimension of 1. The direction sample index symbol h follows the complete interpretation of the h symbol in the aforementioned normalized response weight formula. When the orthogonal residual of the rotation matrix exceeds 0.01 or the azimuth attitude validity flag is false, the automatic attitude correction is disabled for detectors with azimuth asymmetry in the direction response. Only for detectors whose response has been proven to be rotationally symmetric about the azimuth after calibration, the system degenerates to using only the polar angle response. ;

[0096] For the direct sunlight component, the module first determines the vector selection threshold based on the 95th quantile of the solar vector consistency residual in the clean-day data of the same type of equipment. In this specific embodiment, it is limited to no more than 1.0°. When the observed solar vector is valid and the consistency residual does not exceed the threshold, the observed solar vector is used exclusively as the solar incidence direction. When the observed solar vector is invalid or the consistency residual exceeds the threshold, the theoretical solar vector is used only if the Coordinated Universal Time, geographical location, and solar position algorithm version are all valid, and a backoff reason code is written. When neither vector is available, the direct sunlight component response loss is marked as unavailable, and automatic correction relying on the direct sunlight component is prohibited. The module uses the uniquely selected solar incidence direction to search for the direction response in both the calibration attitude and the current attitude, according to... Calculate the effect of positive directional deviation as the loss. The symbol represents the dimensionless direct component response loss. The symbol represents the dimensionless radiation detection response function given in the direction response calibration table. The symbol represents the incident polar angle parameter of the response function in that direction, in degrees, and is calculated from the selected solar incidence direction in the corresponding attitude coordinate system. The symbol represents the azimuth parameter of the response function in that direction, in degrees, and is calculated from the same selected solar incidence direction in the corresponding attitude coordinate system. and The symbols represent the solar polar angle and azimuth angle at the current attitude, respectively, in degrees. and The symbols represent the solar polar angle and azimuth angle at the calibration attitude, respectively, in degrees. The symbol represents the lower limit of the directional response divided by zero, which is set to 0.01 in this specific embodiment. The symbol represents the clipping function that limits the relative loss to the closed interval [0,1] in the order of parameters; when the current attitude response is higher than the calibrated attitude response and the difference exceeds twice the standard deviation of repeated measurements of the direction response, a direct over-response flag is set and the automatic correction that only allows up-adjustment is disabled;

[0097] For the scattering component, the module substitutes the sky composite relative radiance field into the cosine response weighted integral under the calibration attitude and the current attitude, respectively, according to... Calculate the effect of the scattering component response with positive loss. The symbol represents the dimensionless scattering component response loss. The symbol represents the cumulative sum of the product of the overall relative radiance, cosine response, and fixed angle in each sky direction at the current attitude, with a dimension of 1. The symbol represents the cumulative sum of the products under the calibrated attitude, with a dimension of 1. The symbol represents the lower limit of the scattering weighting divided by zero, which is set to 0.01 in this specific embodiment. The symbol represents the clipping function that follows the aforementioned direct component formula and restricts the relative loss to the closed interval [0,1] according to the parameter order; when Higher than Furthermore, if the difference exceeds the repeatability limit of the scattering calibration, set an over-scattering response flag and disable automatic correction that only allows upward adjustment;

[0098] The direct mixing coefficient is determined based on the mutually exclusive direct scattering state written by S2, according to... Generate attitude influence quantity The symbol represents the dimensionless attitude response loss, with values ​​clipped to [0,1], and larger values ​​indicate greater response loss due to attitude. The symbol represents the dimensionless direct irradiance mixing coefficient, which is 1 when direct irradiance dominates, 0 when scattering dominates, and within (0,1) when the mixture is in a mixed state, it is determined by the ratio of the peak solar irradiance to the total sky irradiance. The notation follows the complete definition of the dimensionless direct response loss in the aforementioned direct component formula. The notation follows the complete interpretation of the dimensionless scattering response loss in the aforementioned scattering component formula; when any overresponse flag is true, the attitude influence quantity is still preserved for quality diagnosis, but it is prohibited to use it for up-adjustment correction;

[0099] When the sun is obscured or the direct scattering state output by S2 is scattering-dominant, the direct mixing coefficient is set to 0 according to the scattering-dominant rule. If only the observed solar vector is invalid or the solar vector consistency residual is unavailable and the sun is not obscured, the missing state is not equated to scattering-dominant. If the Coordinated Universal Time, geographical location, and solar position algorithm versions are all valid, the direct component response loss is calculated using the aforementioned theoretical solar vector backtracking path, and the direct mixing coefficient is still determined according to the mutually exclusive direct scattering state output by S2. If the theoretical solar vector is unavailable, or the cloud mask, cloud cover, and relative radiance data around the sun used to determine the direct scattering state and direct scattering percentage are incomplete, no definite attitude influence quantity is generated, automatic correction based on attitude influence quantity is prohibited, and the corresponding valid flag, observation missing reason code, or theoretical vector backtracking reason code, along with the unknown response weight, are written into the influence quantity record and passed to S4. When the attitude measurement is invalid, no definite attitude influence quantity is generated to avoid misinterpreting the missing attitude as zero influence.

[0100] The influence quantity construction module uses the sampling identifier as the primary key to write the window influence quantity, attitude influence quantity, direct and scattered component influence, known response weight, unknown response weight, directional response calibration version, direct over-response flag, scattered over-response flag, and each valid flag into the influence quantity record. This record, together with the original solar radiation sampling value of S1 and the partition confidence of S2, is then transferred to the S4 quality control module.

[0101] In this specific embodiment, S4 includes:

[0102] The quality control module reads the original solar radiation sample values, window influence, attitude influence, and confidence levels of all valid window partitions under the same sampling identifier. The original solar radiation sample values ​​are used only for quality assessment and constraint correction, and do not participate in the construction of window influence, attitude influence, or their upstream direct radiance. The module calculates the weighted lower quantile of the partition confidence level using response weights as the current confidence level, and this constitutes the quality state point. , The symbol represents the three-dimensional dimensionless mass state point in the nth sampling period. The symbol represents the window effect amount corresponding to the sampling period. The symbol represents the attitude influence amount for the corresponding sampling period. The symbol represents the weighted 10% positional confidence level for the corresponding sampling period. The symbol represents the vector transpose operation in the aforementioned solar vector consistency residual formula, where the sampling period index symbol n follows the complete interpretation of the n symbol in the S1 maximum time residual formula; the current confidence level is only used as the global quality state dimension across all valid partitions and does not replace the category confidence level calculated subsequently for the selected dominant pollution category.

[0103] Automatically corrected allowable zones and effective allowable zones are established based on cleanroom calibration data. A traceable cleanroom reference state is maintained, and synchronous observations are conducted using a valid independent reference radiometer. Controlled attenuators, calibrated by a spectrophotometer and with spectral weighted responses matching the radiation detector response, are sequentially installed in the window direction zones to form samples incorporating the influence of non-zero window values. Simultaneously, attitude perturbations are applied using a turntable to cover attitude influences, and confidence levels are identified by controlling changes in image signal-to-noise ratio and exposure coverage. The spectrophotometer's independent calibration residual for the attenuators does not exceed 2%, and the reference radiometer is used to provide the radiation deviation corresponding to each quality state point. Calibration data are grouped into direct-dominant, scatter-dominant, and mixed states, and a pre-defined 3D grid is constructed using window influence values ​​of 0 to 0.60 with a step size of 0.05, attitude influence values ​​of 0 to 0.30 with a step size of 0.025, and state recognition confidence values ​​of 0.60 to 1.00 with a step size of 0.05. Each supported grid cell contains at least 30 valid synchronization samples from at least 3 devices on 3 different dates, with the training and validation sets completely separated according to device number. Each record stores the quality state point, radiation deviation relative to a contemporary reference radiometer, temperature range, device number, date, and grid cell. Only in supported grid cells that meet the coverage and minimum sample size requirements are the outer envelopes of state points with absolute radiation deviation not exceeding 2% selected as the automatic correction allowable region, and the outer envelopes of state points with absolute radiation deviation not exceeding 10% selected as the effective allowable region. The automatic correction allowable region is ensured to be completely within the effective allowable region by shrinking the outer envelope.

[0104] The regional parameter table stores the direct scattering state, temperature range, allowed regional vertices for automatic correction, valid allowed regional vertices, minimum and maximum values ​​for each dimension of calibration, grid step size, index of supported grid cells, number of samples per cell, attenuator calibration version, cross-validation bias, and regional version number. It is constructed from joint observation data collected over at least 30 clean days, covering solar altitude angles from 10° to 85°. When any dimension of the mass state point exceeds the calibration range, the nearest supported grid cell does not meet the minimum sample size requirement, the interpolation simplex used for regional determination is absent, or any calibration dimension degenerates, the mass state is deemed suspicious, automatic correction is disabled, and extrapolation is prohibited. When the current mass state point falls on the boundaries of two regions, it is treated as the more conservative outer state. Automatic correction is also disabled when the regional table version does not match or the number of valid samples is insufficient.

[0105] The category determination threshold table is established based on the confidence distribution of clean status and calibration samples for each contamination category. The minimum confidence level required to ensure that the false acceptance rate of each category's validation set does not exceed 1% is used as the category determination threshold. The table records the category, direct / scattered radiation status, temperature range, threshold, minimum positive contribution lower limit, relative clean contribution determination margin, minimum effective category weight, sample quantity, and version number. The quality control module uses the same caliber to summarize the coverage, normalized response weight, and classification confidence of each clean category and each non-clean category in all effective partitions, and obtains the maximum non-clean contribution. The module first executes data sufficiency and ambiguity gates. When the proportion of unknown pixels is not less than 0.15, the effective category weight is insufficient, the valid confidence flag of any participating partition's determined category is false, or at least one non-clean contribution reaches the minimum positive contribution lower limit but the difference between the maximum and second-largest non-clean contribution does not reach the discrimination margin determined by the validation samples, a unique "pending verification" status is output, and the process transitions to diagnostic stimulus or manual branching. The pending verification gate takes precedence over clean status and dominant contamination category determination. Only when all... After passing through the pre-access control, a clean status is uniquely output and the generation of contamination disposal is prohibited only when all non-clean contributions are 0, the maximum non-clean contribution does not reach the minimum positive contribution lower limit determined by the verification sample, or does not exceed the sum of clean contribution and judgment margin. In other cases, only non-clean categories that pass through the above access control can be determined as the current dominant contamination category. The contamination category is not forcibly assigned by the tie rule. When multiple candidates pass through at the same time and have the same contribution value, they are selected in order of conservative priority: prohibited heating category, manual warning category, and safety restricted heating category. The clean contribution, the contribution of each non-clean category, the selection basis, and the conflict reason code are saved. The module only calculates the weighted 10th percentile value of the confidence of the selected dominant contamination category partition with non-zero coverage or non-zero probability contribution, using the product of the category coverage rate and the normalized response weight as the weight. When the effective category weight is lower than the lower limit in the table, the confidence is marked as unavailable. Then, the table is looked up with the selected dominant contamination category and environmental conditions. When no category record is matched, the maximum value of all contamination category thresholds is used and written into the conservative threshold flag.

[0106] For non-clean states, when the quality status point is within the automatic correction allowable area, the dominant contamination category is valid, the confidence level of the dominant category is available and reaches the category determination threshold corresponding to that dominant contamination category, and all necessary valid flags S1 to S3 are true, the module first proceeds according to... Calculate the effective response ratio. The symbol represents the dimensionless effective response ratio after the combined effect of the viewport and attitude, and its value ranges from [0,1]. The symbol represents the dimensionless window response loss. The symbol represents the dimensionless attitude response loss. Both influences are expressed in the direction of positive loss, so their complements multiplied together represent the remaining response proportion. When any influence equals 1, the effective response proportion is 0, and the automatic correction is exited directly without entering the division calculation. For clean states, the dominant contamination category or dominant category confidence level is not required. When the quality state point is within the effective allowable area and all necessary valid flags from S1 to S3 are true, the window response loss is set to 0. If the attitude response loss is valid and does not reach the attitude correction trigger threshold calibrated by the training samples within the automatic correction allowable area, the original valid output is provided without window correction. The quality status is determined by the following steps: If the attitude response loss is valid and reaches the trigger threshold, the quality status point is within the automatic correction allowable area, and all other automatic correction conditions are met, then the effective response ratio formula in this section is used to enter the limited correction branch based solely on the attitude influence amount; if the attitude response loss is unavailable, or reaches the trigger threshold but does not meet the automatic correction area or integrity conditions, the quality status is marked as suspicious and automatic correction is disabled; in the clean state, no contamination correction or contamination disposal is generated in any of the above branches; when the confidence level of the dominant category for the pending verification state or the non-clean state is unavailable or lower than the corresponding threshold, automatic correction is disabled.

[0107] The quality control module only operates when the original sampled value of solar radiation is not less than 0 W / m². Generate restricted correction values. The symbol indicates the restricted correction value, and the unit is W / m². The symbol represents the raw sampled value of solar radiation, in units of W / m². The symbols follow the complete interpretation of the aforementioned effective response ratio formula. The symbol represents the lower limit of the effective response ratio, is dimensionless, and is labeled by the minimum positive value of the training samples within the automatically corrected allowable region. The symbol represents the maximum dimensionless upward adjustment ratio allowed in a single instance, which is set to 0.15 in this specific embodiment. The symbol represents the clipping function that follows the aforementioned direct response loss formula and limits the first parameter with the lower and upper limits given by the second and third parameters. The clipping lower limit ensures that the corrected value is not lower than the original value and the upper limit limits the amplification of errors. When the original sampled value is less than 0W / m², it is marked as a zero-point anomaly, automatic correction is disabled and the original record is retained, and the clipping calculation is not performed.

[0108] If the quality status point is outside the effective allowable area, the unknown response weight is greater than 0.15, the original solar radiation sample value is saturated, the direct over-response flag or the scattered over-response flag is true, the effective response ratio is 0, or the necessary state is invalid, then the corresponding radiation sample is marked as invalid and automatic correction is disabled. In other cases, for clean states, when the quality status point is within the effective allowable area, the necessary valid flags of S1 to S3 are all true, and the attitude response loss is valid, if the attitude response loss does not reach the attitude correction trigger threshold, the quality level is marked as original valid and does not require window correction. If the attitude response loss reaches the trigger threshold and all the automatic correction conditions specific to clean states are met, the quality level is marked as attitude restricted and correctable. If the corresponding conditions are not met, it is marked as suspicious and automatic correction is disabled. For non-clean states, when the quality status point is within the effective allowable area but does not meet the automatic correction area, dominant category confidence, or integrity conditions, it is marked as suspicious and automatic correction is disabled. Only when the dominant contamination category and its category confidence access control and all other automatic correction conditions are met simultaneously is the quality level marked as restricted and correctable.

[0109] The disposal rules take pollution category, coverage, estimated transmittance, dome surface temperature, relative humidity, quality level and duration as inputs. Clean status does not generate pollution disposal and writes clean cause code. Condensation, frost, ice or snow status generate safety-restricted heating disposal candidates. Dust accumulation or fixed obstruction generates cleanliness warning. Raindrops or water film write precipitation impact cause code and automatic correction is delayed until the verification period after precipitation ends. Pending review category, dominant category confidence unavailable or window impact quantity inconsistent with radiation change status generate diagnostic excitation candidates.

[0110] Radiation sampling quality records are written according to the sampling identifier, including the original solar radiation sampling value, the restricted correction value and its valid mark, the quality level, the disposal instruction, the impact quantity record reference and the cause code. The cause code stores the dominant pollution category, the abnormal window partition, the unknown response weight, the direct scattering state, the attitude deviation level, the registration state, the low confidence state and the thermal zero bias risk state bit by bit. Any reason for disabling automatic correction takes precedence over the restricted correctable mark.

[0111] The quality control module writes radiation sampling quality records into the append-only quality database and writes disposal instructions into the closed-loop disposal queue, carrying the sampling identifier, instruction type, target window partition, safety parameter table version, and effective period, for S5 to execute disposal, diagnose stimulation, and verify disposal effects.

[0112] In this specific embodiment, S5 includes:

[0113] The closed-loop processing module reads the processing instructions output by S4 and verifies the sampling identifier, target zone, pollution category, and safety parameter table version. For condensation, frost, ice, or snow accumulation, it generates heating instructions. Instruction fields include the maximum heating power, dome allowable temperature rise, maximum duration, duty cycle, and stop temperature. For ash accumulation or fixed obstructions, it only generates a cleaning warning and keeps heating off. For raindrops or water films, it writes a precipitation impact cause code and enters a precipitation delay state. During this state, the automatic correction of the disable flag takes precedence over any restricted correction instructions. The module uses the raindrop and water film coverage and the increase in connectivity obtained from the S2 resampling as the end of precipitation observation. Precipitation is considered over only when both coverages are below 0.02 and no new raindrops or water film connectivity are added for 30 consecutive seconds. After the end, the window status, window impact, and [other parameters] are regenerated for 12 consecutive common clock sampling cycles. The quality status is deactivated only when the raindrop and water film coverage rate is consistently below 0.02, the window influence quantity is valid, the synchronous reference radiometer is valid, and the quality status meets any of the following branches in each cycle: the S4 clean status dedicated branch output is originally valid and does not require window correction; all access restrictions for the S4 clean status dedicated attitude limitation correction are met; or the S4 dominant contamination category and its category confidence access restrictions and all other automatic correction conditions are met simultaneously in the residual non-clean status. If raindrop or water film recurrence occurs in any cycle, the window influence quantity is invalid, the synchronous reference radiometer is invalid, the quality status changes to pending review, doubtful, or invalid, or any of the aforementioned branches are not met, the continuous count is immediately reset and automatic correction is disabled again. When the precipitation delay status accumulates to more than 30 minutes, it is upgraded to a manual warning and remains disabled until manual confirmation or subsequent complete verification is passed.

[0114] The heating safety parameter table is calibrated by thermal response testing of the same model dome in an environmental chamber according to temperature and pollution category. It stores the ambient temperature range, pollution category, maximum power, allowable temperature rise relative to the shell, single-cycle duration, cumulative energy limit, cooling time, heating disable flag, and version number. When the pollution category is not yet determined or multiple candidate categories exist, the module uses all non-zero candidate categories recorded in S4 to form a candidate set. For maximum power, allowable temperature rise, single-cycle duration, and cumulative energy limit, the minimum allowable value from the candidate entries is taken, and for cooling time, the maximum value is taken. The heating disable flags of any candidate are logically ORed and merged to form the most conservative safety parameter intersection. If a candidate entry is missing... Heating pulses must not be applied when there is a loss, version inconsistency, empty intersection, or any candidate is prohibited from heating. Furthermore, due to the lack of a valid thermal response vector, the process cannot directly proceed to airflow diagnosis. The module writes the heating non-execution status into the diagnostic record and uniquely generates a delayed diagnostic task. If no executable safety parameters are available for three consecutive planning cycles, the process is upgraded to a manual warning, and the candidate set, table version, and selection criteria are recorded. In this specific embodiment, even if the intersection is allowed, the heating power must not exceed 8W, the allowable temperature rise of the dome surface relative to the shell must not exceed 12°C, and the single duration must not exceed 60s. Heating must be stopped immediately and a protection reason code generated when any temperature sensor is invalid, version mismatched, or the temperature rise reaches the upper limit.

[0115] When the heating pulse is executed, the module follows Calculate the input energy. The symbol represents the input energy of the heating pulse, and the unit is J. The symbol represents the set of all control intervals in which the voltage, current, and duration of the heating pulse have all passed the validity check. The symbol represents the control interval index within this set. The symbol represents the average heating power obtained from the product of voltage and current within the k-th effective control interval, in W. The symbol represents the duration of the kth effective control interval, in seconds. Power is directly multiplied by time to obtain J. When the electrical parameters of any interval are invalid, the pulse is stopped and the energy record is marked as incomplete. Missing interval extrapolation is not used. The accumulated input energy and peak temperature rise are written into the execution record along with the disposal command.

[0116] To determine whether the change in the window's influence is consistent with the original sampled value of solar radiation, the module reads the clean baseline from the most recent clean window that is uncontaminated, has a suitable orientation, and shows stable changes in concurrent reference radiation. Calculate the radiation residual. The symbol represents the dimensionless radiation loss residual relative to the clean baseline in the nth sampling period, and a positive value indicates that the current radiation is below the baseline. The symbol represents the clean baseline obtained by interpolation based on the direct scattering state, solar altitude angle, and overall sky radiance during the same period, with units of W / m². The symbol represents the corresponding raw sample value of solar radiation, in W / m². The symbol represents the lower limit of radiation to prevent division by zero at night. In this specific embodiment, it is taken as 20W / m². The sampling period index symbol n follows the complete interpretation of the symbol n in the S1 maximum time residual formula.

[0117] The clean baseline record simultaneously saves the establishment time, effective solar altitude angle range, direct scattering status, sky model version, valid flag of the synchronous reference radiometer, and allowable usage period. In this specific implementation, if the usage period exceeds 30 days, the solar altitude angle exceeds the calibration range, the reference radiometer is missing, or the interpolation neighbor points are less than 4, the baseline is judged as unusable. When the equipment is started for the first time, when there is long-term contamination, or when the baseline is unusable, the radiation residual and consistency flag are set to unusable, automatic correction and automatic diagnostic excitation are disabled, and a manual clean initialization command is generated. Only after manual cleaning confirmation, collecting stable data for no less than 30 minutes, and the synchronous reference radiometer is valid, and the clean baseline is re-established and cross-validated, is the closed-loop link restored.

[0118] The consistency criterion is established by the joint calibration data of the clean period and the known contamination period. The joint data saves the window influence quantity, radiation residual, solar altitude angle and cloud cover according to the direct scattering state. The residual distribution of linear robust regression is used to obtain a 95% consistency tolerance band. When the radiation residual corresponding to the current window influence quantity falls outside the consistency tolerance band or the two change in opposite directions in three consecutive sampling periods, it is judged as inconsistent. The tolerance band version, deviation amount and trigger period are written into the diagnostic excitation record.

[0119] The closed-loop processing module directly reads the pending verification status, diagnostic excitation candidates, and reason codes from the S4 output. Any condition that is determined to be invalid category confidence, insufficient effective category weight, excessive unknown pixel proportion, indistinguishable contributions from multiple non-clean categories, or the dominant category confidence not reaching the category determination threshold, or triggering a consistency criterion, will be entered into the diagnostic candidate input. The module first executes the manual warning priority access control, then the maintenance freeze and safety interlock access control, and finally determines the diagnostic method: if a manual warning already exists, the maintenance strategy permanently prohibits excitation, or the safety parameter table has no executable records, automatic correction is disabled and a manual warning is directly generated; if the maintenance strategy temporarily does not allow freezing, the original solar radiation sampling value is saturated, the clean baseline is temporarily unavailable, or the safety interlock temporarily prevents excitation, the corresponding reason code is written and a delayed diagnostic task is generated at the next maintenance allowable time; in other cases, diagnostic excitation is only initiated during periods when the clean baseline is valid, the maintenance strategy allows freezing of radiation output, and the original solar radiation sampling value is not saturated, and the heating or prohibition of excitation path is selected according to the intersection of candidate category safety parameters; airflow diagnosis can only enter through the pneumatic safety access control described later after effective heating has been completed and the thermal response vector has not passed the category distinction threshold, and cannot proceed directly from the unexecuted heating state. At startup, the system latches the S4 pending verification status and all cause codes, clean baseline, pre-excitation window status information, window impact quantity, attitude impact quantity, and radiation sequence of the last 10 sampling cycles. From the start of diagnostic excitation, the quality record is set to automatic correction disabled and thermal zero bias risk cause code is written. If delayed diagnosis is still unexecutable for 3 consecutive planned cycles or the saturation state lasts for more than 60 seconds, it is upgraded to manual warning. Any manual warning takes priority over diagnosis, handling, and restricted correction instructions, so that each S4 pending verification input uniquely falls into the startup diagnosis, delayed diagnosis, or manual warning branch.

[0120] During the first 5 seconds of the heating pulse, during the pulse, and 30 seconds after it stops, the common clock synchronously collects the window image, dome surface temperature, shell temperature, voltage, current, and original solar radiation sampling values ​​at 10Hz. The status identification module regenerates the contamination mask, estimates transmittance, and confidence level according to the original window partition. During the diagnostic period, each common clock sampling group continues to use the unique sampling identifier generated by S1 as the primary key. The original sampling identifier that triggers the diagnosis is saved as a foreign key for the diagnostic transaction, and a stage field and pulse sequence number are added to ensure that multiple response records can be traced back to the same diagnostic transaction without primary key conflicts.

[0121] For the i-th window partition, the module follows Calculate the area shrinkage rate of the polluted area. The symbol represents the dimensionless area shrinkage rate of the contaminated region in the i-th partition, and a positive value indicates a decrease in the contaminated area. The symbol represents the area of ​​the contamination mask before excitation, in pixels. The symbol represents the area of ​​the contaminated mask after excitation, in pixels. The symbol represents the area divided by zero lower limit, which is taken as 1 pixel in this specific embodiment. The window partition index symbol i follows the complete interpretation of the symbol i in the S2 coverage formula.

[0122] The module calculates the edge migration velocity by optical flow tracking along the edges of a continuously contaminated mask, and sets the number of effective edge points equal to the number of elements in the set of effective edge points; it only proceeds according to the following procedure when the number of effective edge points is greater than 0 and the time interval between each adjacent frame within the set is greater than 0 seconds. Summarize the partitioning results. The symbol represents the average edge migration speed of the i-th partition, in mm / s. The symbol represents the number of valid edge points participating in the statistics, in units of points. The symbol represents the set of valid edge points. The symbol represents the edge point index. The symbol represents the two-dimensional optical flow displacement vector of the edge point between adjacent frames, in pixels. The symbol represents the two-dimensional Euclidean norm. It is obtained by taking the square root of the sum of the squares of the two pixel displacement components of the two-dimensional optical flow displacement vector, in pixels. The symbol represents the local millimeter-per-pixel conversion factor obtained from the dome mapping calibration, with the unit being mm / pixel. The symbol represents the time interval between adjacent frames, in seconds. The window partition index symbol i follows the complete interpretation of the symbol i in the S2 coverage formula. When the number of effective edge points is equal to 0 or there is a non-positive time interval, the division calculation is not performed, the edge migration speed is marked as unusable and sent to the kernel classifier through the missing value branch.

[0123] The thermal response also records the difference in partition texture entropy before and after excitation, the average optical flow of the contaminated area, the estimated transmittance recovery rate, and the optical restoration amount corresponding to a unit input energy. The estimated transmittance before excitation is the median of the effective partition transmittance within the first 5 seconds of the pulse, and the estimated transmittance after excitation is the median of the effective transmittance of the corresponding partition within the 25th to 30th seconds after stopping. The effective sample ratio in both windows must be no less than 80%. The estimated transmittance recovery rate is calculated by subtracting the median before excitation from the median after excitation as the numerator, and the larger of 1 (subtracting the median before excitation) and 0.05 as the denominator, with the result cropped to [0,1], dimensionless. If any window has insufficient effective samples or the estimated transmittance is unavailable, the recovery rate is marked as missing and the missing value branch of the verification classifier is used. Only when the input energy and estimated transmittance record belong to the same diagnostic transaction, the energy record is complete and valid, and the estimated transmittance is available before and after excitation, is the recovery rate calculated according to the specified criteria. Calculate the optical restitution. The symbol represents the estimated transmittance recovery corresponding to the unit input energy of the i-th partition, in units of 1 / J. The symbol represents the dimensionless estimated transmittance after excitation. The symbol represents the dimensionless estimated transmittance before excitation. The symbol indicates the energy input for this heating pulse, in J. The symbol represents the energy division-to-zero lower limit, which is 0.1J in this specific embodiment. The window partition index symbol i follows the complete interpretation of the symbol i in the S2 coverage formula. When the energy record is incomplete, the diagnostic transaction is mismatched, the input energy is unavailable, or any estimated transmittance is unavailable, the division calculation is not performed. The optical recovery is marked as missing and sent to the missing value branch of the verification classifier. The energy division-to-zero lower limit is only used for the numerical stabilization processing of complete and valid energy records and is not used to replace missing or incomplete input energy.

[0124] The module normalizes the area shrinkage rate, edge migration speed, texture entropy change, average optical flow, estimated transmittance recovery rate, optical recovery amount, peak temperature rise and cooling slope according to the calibrated mean and standard deviation to form a thermal response vector, and inputs it into a gradient boosting tree kernel classifier trained by known samples of dust accumulation, fixed occlusion, condensation, frost, ice, raindrops, water film and snow accumulation. The classifier outputs the class probability and class discrimination confidence. The confidence threshold of the validation set that makes the misclassification rate not exceed 2% is saved with the model version.

[0125] Heating not performed is an independent diagnostic state and is not equivalent to the thermal response vector failing the category distinction threshold. Therefore, it is not allowed to enter this airflow execution path. Only when heating has been completed, the thermal response sampling is complete, and the resulting effective thermal response vector fails the category distinction threshold, is it permitted to continue judging the pneumatic safety gate. If the thermal response vector has passed the category distinction threshold, the airflow pulse is not executed, the aerodynamic response status is marked as not executed and not included in subsequent verification, and the thermal response classification result is directly saved; if the thermal response vector has not passed the category distinction threshold and the safety parameters allow, an airflow pulse of 0.3 MPa for 0.5 s is applied to the target area. The pressure, duration and cumulative number of airflow pulses are constrained by the aerodynamic safety table. The contaminated mask re-sampled before and after the pulse is transformed to surface coordinates through dome mapping. The centroid displacement, area deformation, shedding rate and rediscovery rate within 30 s of the contaminated area are calculated and the aerodynamic response vector is formed; if the aerodynamic safety table prohibits execution, the pressure is insufficient, the target area is out of bounds, the re-sampled record is incomplete, or there is no airflow response for two consecutive times, the airflow is stopped or skipped, the aerodynamic response status is marked as unavailable and a specific reason code is written. It is prohibited to replace the unexecuted or incomplete response with an all-zero vector;

[0126] The pollution review adopts a mutually exclusive input path: when heating is not executed or the thermal response vector is missing, the pollution review classifier is not called, airflow is not executed, and the thermal response vector is not replaced with an all-zero vector. Instead, a delayed diagnosis task or manual warning is generated uniquely according to the aforementioned safety parameter rules. When the thermal response vector has passed the category distinction threshold, the review is completed based on the thermal response classification result, ambient temperature and humidity, precipitation persistence, and dome surface temperature. The aerodynamic response is not involved. When the thermal response vector has not passed the category distinction threshold and the aerodynamic response status is valid, the thermal response vector, the aerodynamic response vector, and the aforementioned environmental quantities are used for joint review. When the thermal response vector has not passed the category distinction threshold and the aerodynamic response status is unavailable, the classifier enters a dedicated branch for missing values, does not output a definite pollution category, and generates a delayed treatment or manual warning. In the joint review, when the thermal response shows a recovery in transmittance and area shrinkage, condensation, frost, ice, or snow accumulation are given priority. When the aerodynamic response shows significant displacement or detachment and the thermal response is insufficient, ash or water droplets are supported. When the review confidence is still below the threshold, a definite category is not forced, and a delayed treatment or manual warning is generated.

[0127] The closed-loop handling module generates instructions for continued heating, airflow cleaning, delayed handling, or manual warning based on the verification results. Continued heating requires re-verification of maximum power, allowable temperature rise, duration, and cumulative energy. Airflow cleaning requires verification of pressure and maximum number of times. All instructions and their execution results are written into the anomaly recovery verification record associated with the original sampling identifier.

[0128] From the start of the diagnostic excitation, the module continuously calculates the difference between the dome surface temperature and the shell temperature. Based on the detector temperature, shell temperature, integration time, and range in the S1 radiation record, it performs bilinear interpolation at the same range in the temperature-integration time-range zero-point table obtained from the factory dark field calibration. The current predicted zero point minus the predicted zero point under the same conditions before excitation is used as the online radiation zero-point drift estimate. The zero-point table covers all temperatures and ranges collected by the same type of detector under shading dark field conditions. It saves the 95% limit of the prediction residual and the version number. If the current input exceeds the calibration range, the version is mismatched, or the interpolation residual limit is greater than 5W / m², the zero-point drift is marked as unusable. Until the temperature difference returns to the ±0.5℃ allowable range corresponding to the baseline before excitation and the effective zero-point drift estimate returns to the 95% fluctuation range of the clean period in the same temperature range, the radiation sampling quality record always remains in an automatic correction disabled state and retains the thermal zero bias risk reason code. When the zero-point drift is unusable, this state must not be lifted and manual review must be initiated.

[0129] After the treatment, the system uses a common clock to regenerate the window state information, window influence, and attitude influence. It also calculates the robust average of the original solar radiation samples and the effective synchronous reference radiometer readings within each 5-second synchronous window before and after the treatment. The original solar radiation samples from both windows are only used for comparison of the treatment effect and are not included in the construction of the corresponding window influence, attitude influence, or direct sky input. A robust regression model for the cleanroom period is established using synchronous observation data from at least 30 clean days to map the reference radiometer readings to the original sampled values ​​of this sensor under cleanroom window conditions. The model is saved separately for direct sunlight-dominated, scattering-dominated, and mixed states, and uses the reference radiometer... The model takes the readings, sky zone comprehensive radiance, solar altitude angle, direct radiation mixing coefficient, temperature range, and measurement range as inputs, and the clean raw sample value of this sensor as the dependent variable. Huber loss fitting is used, and a 95% prediction interval is formed through cross-validation with one day left over. The module inputs the concurrent natural environment inputs for the pre-treatment and post-treatment windows into the model to obtain the expected clean raw sample values ​​and 95% prediction intervals for each window. The expected clean raw sample values ​​are not considered counterfactual for a contaminated state without treatment. Only after the temperature difference between the dome surface and the shell returns to the allowable range corresponding to the pre-excitation baseline, and the online radiation zero-point drift estimation is effective and within the clean period fluctuation range, will the model be considered clean. Only when the radiation change direction is consistent is the determination performed. During the determination, the actual original sampled value before treatment is based on the pre-excitation zero point, and the actual original sampled value after treatment is reduced by the effective online zero-point drift estimate for the same diagnostic transaction. The absolute residuals of the two actual values ​​relative to their respective contemporaneous clean expected original sampled values ​​are calculated separately. Radiation is considered consistent only when the reduction in the absolute residual after treatment relative to the absolute residual before treatment reaches the residual convergence threshold determined by the 95% confidence lower limit of the independent clean validation set and known contamination treatment validation samples; the actual original sampled value after treatment, compensated for zero-point caliber, falls into the corresponding 95% prediction interval; the influence of the viewing window after treatment decreases relative to before treatment; and the broadband equivalent estimated transmittance shows a recovery direction. The consistency flag is set to 1 if the direction of change is consistent, otherwise it is set to 0. If the temperature difference recovery access control is not met, online zero-point drift estimation is unavailable or exceeds the cleanroom period fluctuation range, model input is missing, the synchronous reference radiometer is invalid, the version is mismatched, or the prediction interval width exceeds 10% of the cleanroom baseline mean, the consistency flag is marked as unavailable and used as 0 in the success determination. This ensures that the positive criterion indicates that the post-treatment measurement response converges to the synchronous cleanroom expected value without recognizing thermal zero bias or abnormal over-response as recovery. Model coefficients, input standardized parameters, residual convergence threshold, residual limits, and version number are saved with the cleanroom baseline. Based on the fact that the window effect after treatment decreases relative to before treatment and exceeds the validation threshold, according to... To assess the effectiveness of the treatment, The symbol represents a Boolean value indicating whether the processing was successful. The symbol represents the dimensionless window effect quantity before treatment. The symbol represents the dimensionless window effect after treatment. The symbol represents the threshold for the decrease in the dimensionless effect quantity determined by the 95% confidence lower limit of the treatment validation sample. The symbol represents a consistent indicator of the direction of radiation change obtained according to the aforementioned calculable rules, and takes a value of 0 or 1. The symbol indicates that both conditions are true simultaneously;

[0130] After successful treatment, the one-time radiation change direction consistency flag of the treatment event is not used. Instead, a periodic direction consistency flag is regenerated cycle by cycle for each new common clock sampling flag. A 5-second synchronous rolling window up to the current time is formed for the current sampling flag, and the previous adjacent 5-second window is used as the comparison window. The robust mean of the original solar radiation sampled within the two windows, compensated by the same online zero-point aperture, the effective synchronous reference radiometer reading, the comprehensive radiance of the sky zone, the solar altitude angle, the direct mixing coefficient, the temperature range, and the range setting are used to calculate the original sampled value and absolute residual of the expected cleanliness during the same period by the aforementioned clean period robust regression model. The periodic direction consistency flag of the sampling flag is set to valid and set to 1 only when the absolute residual of the current window does not increase relative to the previous window and the decrease or stable fluctuation falls within the allowable band determined by the independent cleanliness validation set, the influence of the current window does not increase, the broadband equivalent estimated transmittance does not decrease, and the temperature difference, online zero-point drift, model input, prediction range, and synchronous reference radiometer are all valid. Otherwise, it is set to 0 when the physical direction is inconsistent, and marked as unusable when any input is invalid. The cycle flag, the sampling flag range of the two windows, the zero-point compensation value, the expected cleanliness value during the same period, and the residual are all written into the anomaly recovery verification record. The clean baseline is only updated in a robust index update manner when 12 consecutive common clock sampling flags that are different from each other after successful treatment meet the following conditions: the temperature difference is within the allowable range, the online zero-point drift estimation is valid and within the clean period fluctuation range, the regenerated cycle direction is consistent and the flag is valid and set to 1, the window influence is not higher than the verification boundary determined by the treatment verification sample, and the reference radiometer during the same period is valid. The update coefficient is selected from the candidate interval [0.01, 0.10] to minimize the mean square error of the independent clean verification set relative to the reference radiometer and is saved with the baseline version. Each time, only the baseline nodes of the current direct scattering state and the solar altitude angle neighborhood are updated. The number of consecutive cycles of 12 is obtained by the clean period radiation zero-point stability calibration. If any cycle flag is unavailable, set to 0, or other conditions are not met, the continuous count is cleared and the original clean baseline is retained.

[0131] The anomaly recovery verification record ultimately saves the window status information before and after treatment associated with the same sampling identifier, window impact amount, attitude impact amount, original solar radiation sampling value, thermal response vector, aerodynamic response vector, verification of pollution category, execution instructions, protection trigger information, treatment success judgment and clean baseline version, so that subsequent sampling cycles can read the updated clean baseline and complete the treatment closed loop.

[0132] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0133] This invention transforms local visual states into quantities that influence solar radiation measurements by using multi-source synchronization, mapping of window partitions to the celestial sphere, and weighting of detection response with sky radiance. It also incorporates window contamination, cloud conditions, solar shading, and installation attitude into the same quality judgment chain, thereby ensuring that the original values, optional correction values, quality levels, and cause codes have consistent data basis.

[0134] This invention utilizes heating and airflow devices to form a constrained diagnostic stimulus, verifies the contamination category based on thermal and aerodynamic responses, prohibits automatic correction during the recovery period of thermal zero bias risk, and verifies the treatment results and clean baseline update together with the decrease in contamination impact and radiation physics consistency. Therefore, it can improve the effective data rate while ensuring measurement traceability.

Claims

1. A method for measuring solar radiation based on window state recognition, characterized in that, include: S1. Simultaneously acquire raw solar radiation sampling values, window images of the hemispherical transparent dome, sky images, and attitude detection data to form a synchronous sampling data set; S2. Perform sky background suppression and dome surface partitioning mapping on the window image to generate window state information including pollution category, estimated transmittance and confidence level, and generate sky state information and attitude state information based on sky image and attitude detection data. S3. Map the window partitions to the direction of the incident celestial sphere. Based on the response of the solar radiation detection direction, weight the transmittance loss of each window partition and the relative radiance of the corresponding sky partition to generate the window influence quantity. Generate the attitude influence quantity based on the attitude state information and the sky state information. S4. Determine the quality of the raw solar radiation sampling values ​​based on the window influence, attitude influence, and confidence level, and generate radiation sampling quality records and processing instructions. S5. Carry out the disposal according to the disposal instructions and simultaneously re-collect the window image and the original solar radiation sampling value; when the pollution category cannot be determined based on the confidence level, or when the change in the window influence amount is inconsistent with the change in the original solar radiation sampling value, use the heating pulse or airflow pulse as the diagnostic excitation, verify the pollution category based on the diagnostic response information before and after the excitation, and verify the disposal effect based on the re-collection results.

2. The solar radiation measurement method based on window state recognition according to claim 1, characterized in that, S1 includes: A common clock sends trigger signals with the same sampling sequence number to the radiation detection unit, the window imaging channel, the sky imaging channel, the active illumination unit, and the environment and attitude detection unit; the window imaging channel acquires the first window frame and the second window frame respectively at adjacent exposure times when the active illumination unit is turned off and on; the synchronous sampling data group also records each acquisition time, exposure parameters, illumination status, ambient temperature and humidity, dome surface temperature, shell temperature, and inertial measurement data.

3. The solar radiation measurement method based on window state recognition according to claim 2, characterized in that, Step S2, generating the window state information, includes: registering the first window frame and the second window frame; generating a background suppression map based on the registered pixel difference or pixel ratio; dividing the background suppression map into multiple window partitions based on the pre-calibrated mapping relationship between image pixels and dome surface positions; extracting at least one of brightness variation, texture, edge morphology, and polarization features from each window partition; identifying pollution categories such as dust accumulation, fixed occlusion, condensation, frost, ice, raindrops, water film, and snow accumulation; determining the coverage rate based on the partition ratio of pixels of each category; determining the estimated transmittance based on the transmittance response ratio between the clean state and the current state; and determining the confidence level based on the degree of matching between the classification output and the calibrated sample distribution.

4. The solar radiation measurement method based on window state recognition according to claim 1, characterized in that, Step S2, generating the sky state information and attitude state information, includes: extracting the actual solar image point, the confidence level of the actual solar image point, the solar shading state, cloud cover, cloud type, and relative radiance of sky zones from the sky image; calculating the theoretical solar vector based on the acquisition time and the geographical location of the sensor; when the sun is not shading and the confidence level of the actual solar image point reaches the image point determination threshold determined by the calibrated sky image, comparing the theoretical solar vector with the observed solar vector calculated from the actual solar image point based on the camera calibration parameters to obtain the solar vector consistency residual; when the sun is shading or the confidence level of the actual solar image point does not reach the image point determination threshold, marking the solar vector consistency residual as unusable; comparing the gravity vector in the inertial measurement data with the reference gravity vector of the sensor calibration coordinate system to obtain the attitude deviation; and determining the direct scattering state based on the solar shading state, the cloud cover, and the relative radiance of the sky zones.

5. The solar radiation measurement method based on window state recognition according to claim 4, characterized in that, S3 includes: determining the set of incident directions corresponding to each window partition based on the calibration mapping between the window partition and the incident celestial sphere direction; determining the response weight of each incident direction set by the direction response function or cosine response calibration table of the radiation detection unit; accumulating the product of the transmittance loss obtained by subtracting the estimated transmittance of each window partition, the relative radiance of the corresponding sky partition, and the response weight to obtain the window influence amount; transforming the incident angle of the direction response function according to the attitude deviation, and calculating the direction deviation influence of the direct component and the cosine response deviation influence of the scattered component respectively in combination with the direct scattering state to obtain the attitude influence amount.

6. The solar radiation measurement method based on window state recognition according to claim 1, characterized in that, S4 includes: constructing a quality status point by the window influence, the attitude influence, and the confidence level; determining an automatic correction allowable area and an effective allowable area based on the radiation deviation distribution corresponding to different quality status points in the cleanroom calibration data, wherein the automatic correction allowable area is located within the effective allowable area; determining a category determination threshold based on the confidence distribution of cleanroom status and calibration samples of each pollution category; when the quality status point is located within the automatic correction allowable area and its confidence level reaches the category determination threshold, correcting the original solar radiation sampling value based on the window influence and the attitude influence and generating a restricted optional correction value; when the quality status point is located outside the effective allowable area, marking the corresponding radiation sample as invalid and disabling automatic correction; in other cases, marking the corresponding radiation sample as suspicious and disabling automatic correction; the radiation sampling quality record includes the original solar radiation sampling value, the restricted optional correction value, the quality level, the cause code, and the disposal instruction, and writing the pollution category, area status, and attitude deviation that triggered the quality determination into the cause code.

7. The solar radiation measurement method based on window state recognition according to claim 6, characterized in that, In step S5, heating treatment commands are generated for condensation, frost, ice, or snow accumulation, constrained by maximum power, dome allowable temperature rise, and duration; cleaning warnings are generated for dust accumulation or fixed obstructions; precipitation impact cause codes are generated for raindrops or water films and automatically corrected after a delay; a sampling identifier is generated for each of the synchronous sampling data groups by a common clock used for synchronous acquisition; after each treatment, the window status information, the window impact amount, and the attitude impact amount are regenerated, and the corresponding results before and after the treatment are written into the anomaly recovery verification record associated with the same sampling identifier.

8. The solar radiation measurement method based on window state recognition according to claim 7, characterized in that, The diagnostic excitation using heating pulses or airflow pulses includes: determining a clean baseline based on the original solar radiation sampling values ​​during the clean period, and determining a consistency criterion based on joint calibration data from the clean period and known contamination periods. The joint calibration data includes the window effect and the radiation residual of the original solar radiation sampling values ​​relative to the clean baseline. The consistency criterion is used to determine whether the changes in the window effect and the original solar radiation sampling values ​​are consistent. During the period when frozen radiation output is permitted, the clean baseline and the pre-excitation baseline are latched. A heating pulse constrained by maximum power, dome allowable temperature rise, and duration is applied. Visual data is collected before, during, and after the pulse. The window image, temperature data, and electrical parameters are used to calculate the area shrinkage rate, edge migration speed, texture entropy change, optical flow, estimated transmittance recovery rate, and optical recovery per unit input energy of the polluted area, forming a thermal response vector. Airflow pulses are applied to window partitions where the thermal response vector still fails to meet the category distinction criteria determined by the calibration samples, and the displacement, deformation, shedding rate, and re-coverage rate of the polluted area are calculated, forming an aerodynamic response vector. Combining the thermal response vector, the aerodynamic response vector, ambient temperature and humidity, precipitation persistence, and dome surface temperature, the pollution category is verified, and disposal instructions such as continued heating, airflow cleaning, delayed disposal, or artificial early warning are generated.

9. The solar radiation measurement method based on window state recognition according to claim 8, characterized in that, From the start of the diagnostic excitation until the difference between the dome surface temperature and the shell temperature returns to the allowable range corresponding to the baseline before excitation and the radiation zero-point drift returns to the clean period zero-point fluctuation range, the radiation sampling quality record remains in an automatic correction disabled state and is written with the thermal zero bias risk cause code. The treatment is considered successful when the influence of the window decreases after treatment compared to before treatment, and the original solar radiation sample value after re-collection is consistent with the direction of radiation change determined based on the direct scattering state at the same time. After confirming successful treatment, the clean baseline is updated only when the public clock sampling cycles continuously reach a preset number and all meet the allowable range, the radiation change direction is consistent, and the window influence amount remains within the verification boundary determined by the treatment verification sample. The preset number is determined by the clean period radiation zero-point stability calibration.

10. A solar radiation measurement system based on window state recognition, used to execute the solar radiation measurement method based on window state recognition as described in any one of claims 1 to 9, characterized in that, include: The synchronous acquisition module includes a radiation detection unit, a window imaging channel, a sky imaging channel, an active illumination unit, an environment and attitude detection unit, and a common clock that sends synchronous trigger signals to each unit, for forming the synchronous sampling data group; the state recognition module is used to generate the window state information, the sky state information, and the attitude state information. An influence quantity construction module is used to generate the window influence quantity and the attitude influence quantity; The quality control module is used to generate the radiation sampling quality record and generate a limited optional correction value based on the window influence amount and the attitude influence amount when the correction constraint is met; the closed-loop treatment module includes a heating device and an airflow device, used to execute the treatment command, apply the diagnostic excitation, verify the contamination category, verify the treatment effect based on the resampling results, and maintain a clean baseline for radiation change comparison.

Images