A two-point calibration intelligent anomaly detection method and system for a spaceborne microwave temperature sounder

Abstract

The application discloses a kind of two-point calibration intelligent anomaly detection method and system of spaceborne microwave temperature detector, it is related to satellite remote sensing data quality control and intelligent detection technical field, the method includes: based on the identification result carries out manual audit, and after misclassification sample is revised and labeled, it is supplemented to training set, to update the model parameter with incremental learning;Based on the identification result and O-B bias data, compare the abnormal sample and adjacent normal sample bias value, when exceeding the preset threshold range, it is marked as calibration quality anomaly;Based on the calibration quality anomaly O-B bias variation amplitude and channel spatial distribution characteristics, the severity grade of the anomaly is assessed and model iteration optimization is triggered.The application can improve the automation level of calibration anomaly detection, identification accuracy and business operation efficiency.

Description

Technical Field

[0001] This invention relates to the field of satellite remote sensing data quality control and intelligent detection technology, specifically to a two-point calibration intelligent anomaly detection method and system for a spaceborne microwave temperature detector. Background Technology

[0002] During long-term on-orbit operation of satellites, factors such as changes in the space environment, fluctuations in the thermal control system, performance drift of electronic components, abnormal calibration source status, and changes in payload operating modes can cause abnormal changes in temperature parameters, count values, and calibration coefficients during the two-point calibration process, thus affecting the calibration results and the quality of remote sensing products. Therefore, real-time monitoring and anomaly detection of calibration process parameters are crucial for ensuring the quality of remote sensing data.

[0003] Currently, the detection of calibration anomalies mainly employs methods such as manual inspection, threshold discrimination, and statistical analysis. Manual inspection relies heavily on expert experience, judging the presence of anomalies by observing changes in the time series of calibration parameters; threshold discrimination methods monitor calibration parameters by setting empirical thresholds; and statistical analysis methods identify anomalous changes using statistical indicators such as mean and standard deviation. However, these methods generally suffer from low automation, high reliance on manual intervention, low anomaly identification efficiency, and insufficient adaptability to complex anomaly patterns. As data volumes continue to grow, traditional methods struggle to meet the demands of operational processes and long-term automated quality control.

[0004] In recent years, deep learning technology has made significant progress in image recognition, time series analysis, and anomaly detection, providing new technical approaches for anomaly detection in complex remote sensing data. However, current research on anomaly detection in the two-point calibration process of spaceborne microwave temperature detectors remains relatively limited, and a comprehensive detection method that combines calibration time series characteristics, intelligent recognition capabilities, and physical consistency verification mechanisms is still lacking. Summary of the Invention

[0005] The purpose of this invention is to provide a two-point calibration intelligent anomaly detection method and system for spaceborne microwave temperature detectors, which can improve the automation level, identification accuracy and operational efficiency of calibration anomaly detection.

[0006] By adopting the above technical solutions, the timeliness and accuracy of hazard identification have been improved.

[0007] In a first aspect, a two-point calibration intelligent anomaly detection method for a spaceborne microwave temperature probe, the method comprising:

[0008] A calibration parameter database is established by extracting calibration parameters from the observation data of the spaceborne microwave temperature detector. The calibration parameters include instrument temperature data, heat source temperature data and count values ​​of each channel.

[0009] Based on the calibration parameter database, the calibrated brightness temperature is calculated, and the synchronous brightness temperature is simulated using radiative transfer mode and numerical prediction analysis field data combined with observation geometric information. The difference is then calculated to obtain OB deviation data.

[0010] Based on the calibration parameter database, each parameter is normalized according to a preset time window, and a time series image containing temperature changes, count value changes, orbital period changes and day-night changes is generated according to the time-varying law.

[0011] Based on the time-series images, normal and abnormal categories are labeled according to historical data, an abnormal sample library is established, and a training set is constructed.

[0012] A convolutional neural network is trained based on the training set, and the network extracts abnormal features from the time-series images to obtain a calibrated anomaly classification model.

[0013] Based on the classification model, the time series image corresponding to the calibration parameters to be detected is input into the model to obtain the recognition result containing normal and abnormal categories;

[0014] Based on the identification results, manual review is conducted, and misclassified samples are corrected and labeled before being added to the training set for incremental learning to update the model parameters;

[0015] Based on the identification results and OB deviation data, the deviation values ​​of the abnormal samples are compared with those of the neighboring normal samples. If the deviation exceeds a preset threshold range, the sample is marked as a calibration quality abnormality.

[0016] Based on the variation range of the calibration quality anomaly OB deviation and the spatial distribution characteristics of the channel, the severity level of the anomaly is assessed and model iterative optimization is triggered.

[0017] Secondly, a two-point calibration intelligent anomaly detection system for a spaceborne microwave temperature detector includes:

[0018] The data acquisition module is used to extract calibration parameters to build a database and to calculate OB bias data using radiative transfer model and numerical prediction analysis field data;

[0019] The feature image construction module is used to normalize the calibration parameters according to a preset time window and generate a temporal feature image.

[0020] The sample annotation and training set construction module is used to classify time-series images and construct training datasets.

[0021] The model training module is used to train a convolutional neural network with a training dataset to obtain a classification model.

[0022] The anomaly detection module is used to input the time-series image corresponding to the parameter to be detected into the classification model to obtain the recognition result;

[0023] The review and model update module is used to correct the labels of misclassified samples and update the model parameters through incremental learning.

[0024] The quality verification module is used to extract the OB deviation of abnormal samples, compare it with the deviation of normal samples, and mark the calibration quality anomalies.

[0025] The rating and iteration module is used to rate anomalies based on OB deviation amplitude and spatial distribution and trigger iterative optimization.

[0026] The above-described solution of the present invention has at least the following beneficial effects:

[0027] This invention transforms the traditional calibration quality monitoring process, which relies on manual analysis, into an automatic identification process based on deep learning. By using convolutional neural networks to automatically extract abnormal features from calibration data, it achieves automatic detection and classification of calibration anomalies, significantly reducing the workload of manual monitoring and improving operational efficiency.

[0028] Traditional thresholding and statistical analysis methods are mainly effective for large-amplitude anomalies, but their ability to identify complex anomaly patterns such as slow drift, periodic oscillations, and multi-parameter coupled changes is limited. This invention constructs a calibrated feature image that incorporates temporal variation features, orbital periodic features, and diurnal variation features. By fully utilizing the feature extraction capabilities of deep learning models, it can effectively identify various complex anomaly patterns, improving anomaly detection accuracy and recall. Attached Figure Description

[0029] Figure 1 This is a flowchart for detecting abnormal calibration parameters.

[0030] Figure 2 This is a system diagram of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0033] like Figure 1 As shown, the method in this embodiment is applicable to on-orbit spaceborne microwave temperature detectors, and specifically includes the following steps.

[0034] Step 1, calibration data acquisition and brightness temperature simulation deviation calculation, specifically includes:

[0035] Two calibration-related parameters were extracted from the observation data of the spaceborne microwave temperature probe, and a calibration parameter database was established in chronological order. The calibration parameters include instrument temperature data, heat source temperature data, and count values ​​of each channel.

[0036] The instrument temperature data specifically includes the temperature of the detector main unit, the receiver, and the calibration source. The detector main unit temperature is the real-time temperature value collected by the temperature sensor mounted on the detector main unit housing, reflecting the overall thermal state of the detector. The receiver temperature is the measured temperature value of the low-noise amplifier at the front end of each detection channel's receiver; this parameter directly affects the receiver's gain characteristics. The calibration source temperature is the arithmetic mean of the temperature values ​​measured at multiple temperature measurement points evenly distributed circumferentially on the calibration blackbody inside the detector, ensuring an accurate representation of the actual physical temperature of the blackbody's radiating surface.

[0037] The heat source temperature data includes the heat load temperature and temperature values ​​from multiple measurement points near the heat source. The heat load temperature is the temperature value collected by the temperature sensor near the heating element within the thermal calibration load mounted on the detector. The temperature values ​​from multiple measurement points near the heat source are used to monitor the temperature gradient distribution around the heat load, helping to determine the stability of the heat source's operating state.

[0038] The count values ​​for each channel are the raw digital count values ​​output by the instrument's various detection channels after analog-to-digital conversion in cold air observation mode, hot source observation mode, and Earth observation mode. Specifically, the cold air observation mode represents the observation state when the instrument's antenna is pointed towards the cold air background; the cold air count values ​​obtained in this mode are used to determine the low-end reference point of the calibration curve. The hot source observation mode represents the observation state when the instrument's antenna is pointed towards the internal thermal calibration load; the hot source count values ​​obtained in this mode are used to determine the high-end reference point of the calibration curve. The Earth observation mode represents the observation state when the instrument's antenna is pointed towards the target area on Earth; the Earth observation count values ​​obtained in this mode are the raw observation data to be calibrated.

[0039] Simultaneously, based on the aforementioned calibration parameter database, a linear conversion relationship between the count values ​​and the radiative brightness temperature is established using cold air count values, heat source count values, and their corresponding physical temperatures, and the calibrated brightness temperature is calculated. Furthermore, using radiative transfer model and numerical prediction analysis field data, combined with observational geometric information, the synchronous brightness temperature is simulated, and the difference between the calibrated brightness temperature and the simulated brightness temperature is calculated to obtain OB bias data.

[0040] The observational geometric information includes the longitude, latitude, zenith angle, and azimuth angle of the probe at the observation time. The radiative transfer model can employ either a line-by-line integral radiative transfer model or a fast radiative transfer model. The numerical prediction analysis field data includes atmospheric temperature profiles, humidity profiles, pressure profiles, and parameters such as surface temperature and surface emissivity. Specifically, the atmospheric temperature, humidity, and pressure profiles are obtained from the numerical prediction analysis field data through spatiotemporal interpolation based on the observation time and the probe's latitude and longitude. The surface temperature and surface emissivity parameters are directly extracted from the surface analysis field of the numerical prediction analysis field data. Using the radiative transfer model, with the atmospheric temperature and humidity profiles and surface parameters as input, layer-by-layer radiative transfer integration is performed along the light transmission path determined by the observation zenith angle and azimuth angle to obtain the atmospheric top exit brightness temperature of each probe channel. This atmospheric top exit brightness temperature is then used as the simulated synchrotron brightness temperature.

[0041] Step 2, construction of the calibration feature image, specifically includes:

[0042] The calibration data collected in step 1 is processed and standardized using a preset time window as a unit to generate a time-series feature image reflecting the parameter variation pattern. The duration of the preset time window must cover at least one complete orbital cycle of the probe to ensure that the constructed time-series image contains complete orbital periodic features. In this embodiment, the time window duration can be set to an integer multiple of the probe's orbital cycle, for example, covering three complete orbital cycles.

[0043] The standardization process employs a normalization approach. Specifically, the actual numerical range of each calibration parameter within the current time window is linearly mapped to a pre-defined unified numerical interval. This actual numerical range is the interval formed by the maximum and minimum values ​​of the parameter within the current time window. For example, the values ​​of each parameter can be uniformly mapped to the interval [0, 1] to eliminate interference caused by differences in units and numerical magnitudes between different parameters, allowing the temporal variation characteristics of different parameters to be compared and analyzed on a unified numerical scale.

[0044] The generated time-series images contain information on temperature changes, count changes, orbital period changes, and diurnal variations, used to comprehensively describe the operational status of the calibration process. The time-series images use time as the horizontal axis and the normalized parameter values ​​as the vertical axis, overlaying the change curves of at least two calibration parameters within the same image. Different calibration parameter change curves are distinguished by different colors or line types, and the start time of each orbital period within the time window is marked on the image to facilitate the identification of orbital periodic characteristics. For example, the change curve of the cold air count value can be drawn with a solid blue line, and the change curve of the heat source temperature can be drawn with a dashed red line. Overlaying these two types of curves in the same image not only shows their respective time-series change patterns but also visually presents the correlation or difference in change patterns between them.

[0045] In practice, when a preset time window undergoes a sliding update, there is a preset overlap ratio between adjacent time windows. For example, the overlap ratio between adjacent windows can be set to 50%, meaning that each new window covers the latter half of the data in the previous window. This overlap ratio ensures the continuity of temporal features between adjacent windows and avoids missing anomalous changes across boundaries due to window truncation. The temporal image is stored in a digital image format (e.g., PNG format) at a preset pixel resolution (e.g., 224 x 224 pixels) as input to the subsequent convolutional neural network.

[0046] Step 3, anomaly sample labeling and training set construction, specifically includes:

[0047] Based on historical operational data and expert experience, the feature images generated in step 2 are manually labeled, and the samples are divided into normal states and different types of abnormal states. A calibration abnormal sample library is established, and a training dataset for model training is constructed.

[0048] In this embodiment, the abnormal status categories include, but are not limited to, the following four types:

[0049] (1) Cold air count abnormal drift category: This refers to a situation where the cold air count value continuously shifts in a single direction over a period of time, and the shift exceeds a preset drift threshold. For example, over multiple consecutive orbital periods, the cold air count value continuously increases or decreases, and its cumulative shift exceeds several standard deviations of the historical normal fluctuation range. This type of anomaly is usually related to contamination of the cold air reflector surface or changes in the radiation environment along the cold air observation path.

[0050] (2) Heat source count value step change category: refers to the situation where the heat source count value changes abruptly between adjacent sampling points and the magnitude of the change exceeds the preset step threshold. For example, when the heating power of the heat load is adjusted or the working mode of the thermal control system is switched, the heat source count value may jump instantaneously, and the magnitude of the jump is significantly higher than the measurement noise level within the normal sampling interval.

[0051] (3) Abnormal temperature parameter fluctuation category: refers to the reciprocating oscillation changes of the instrument temperature or heat source temperature exceeding the preset fluctuation threshold within multiple consecutive sampling cycles. For example, when the feedback control parameters of the thermal control system are abnormal, the instrument temperature may exhibit continuous oscillations near the set value that exceed the normal control accuracy, and the oscillation amplitude and frequency are different from the slight temperature fluctuations caused by normal thermal control adjustment.

[0052] (4) Calibration coefficient gradual offset category: This refers to the calibration gain coefficient or calibration bias coefficient exhibiting a monotonic changing trend over multiple consecutive orbital periods, with the cumulative change exceeding the preset offset threshold. For example, due to the long-term on-orbit aging effect of electronic devices, the calibration gain coefficient may exhibit a slow monotonic decreasing trend. Although the amplitude of a single change is small, after accumulating for tens or even hundreds of days, the cumulative offset becomes non-negligible.

[0053] When labeling categories, a rectangular bounding box is used to mark the start and end time regions of anomalies on the time-series image, with the corresponding anomaly type label next to the box. The starting position of the bounding box corresponds to the first occurrence of the anomaly, and the ending position corresponds to the end of the anomaly or the end of the current time window. This labeling method not only provides anomaly type information but also clarifies the duration of the anomaly in the time dimension, helping the convolutional neural network to simultaneously learn the anomaly classification and temporal localization capabilities.

[0054] Step 4, Convolutional Neural Network Anomaly Detection, specifically includes:

[0055] The feature images in the training set constructed in step 3 are input into the convolutional neural network for training and learning, automatically extracting abnormal feature information in the images, establishing a calibration anomaly classification model, and realizing automatic identification and anomaly classification of calibration states.

[0056] The convolutional neural network used in this embodiment includes an input layer, at least three convolutional layers, at least two pooling layers, at least one fully connected layer, and an output layer connected in sequence. The following description uses a typical network structure containing three convolutional layers, two pooling layers, one fully connected layer, and one output layer as an example.

[0057] The input layer receives the temporal image generated in step 2. The width of the image corresponds to the time dimension, and the height corresponds to the normalized scaling parameter value. The first convolutional layer contains multiple convolutional kernels (e.g., 32 kernels). The width of each kernel is a preset odd number of pixels (e.g., a 3x3 two-dimensional kernel), and the convolutional stride is a preset integer value (e.g., a stride of 1). Zero-padding is applied to the boundaries of the input feature map to maintain the spatial size of the output feature map. This convolutional layer is used to extract basic edge and texture features from the image, such as local variation trends of parameter curves, abrupt changes, and periodic texture patterns.

[0058] The first pooling layer uses max pooling to downsample the feature map output from the first convolutional layer. The pooling window size is, for example, 2x2, with a stride of 2, reducing the width and height dimensions of the feature map to half their original values. After downsampling, the spatial size of the feature map is reduced, but the number of channels remains unchanged. This reduces the computational cost of subsequent layers while preserving the main feature information, enhancing the robustness of the features to small changes in the time scale.

[0059] The second and third convolutional layers have similar structures, both containing multiple convolutional kernels, but the number of kernels can be gradually increased (e.g., 64 kernels in the second layer and 128 kernels in the third layer) to extract higher-level, more abstract anomalous feature patterns. The second pooling layer also uses max pooling to further reduce the spatial dimensionality of the feature maps. Through the alternating stacking of multiple convolutional and pooling layers, the network can progressively abstract from local pixel-level features to global semantic-level features, thereby recognizing different types of complex anomalous patterns.

[0060] The fully connected layer unfolds the feature map output from the last pooling layer into a one-dimensional feature vector, and then maps this feature vector to the output nodes corresponding to each anomaly category through a fully connected operation. The number of output nodes equals the total number of anomaly categories plus the number of normal categories. The output layer uses a normalized exponential function (i.e., the Softmax function) to calculate the probability value of each output node. The sum of the probability values ​​of all categories is 1, and the category corresponding to the highest probability value is output as the final recognition result.

[0061] During the model training phase, the labeled training dataset from step 3 is input into the convolutional neural network for training. The weight parameters of each layer of the network are updated using the backpropagation algorithm, aiming to minimize the classification cross-entropy loss function, thereby gradually reducing the difference between the network's output classification results and the manually labeled class labels. Mini-batch stochastic gradient descent can be used to update parameters during training. Each time, a batch of samples is randomly selected from the training set and input into the network, the gradient of the loss function with respect to the parameters of each layer is calculated, and the parameters are updated. After several training epochs of iteration, when the classification accuracy on the validation set tends to stabilize, training is complete, and a calibrated anomaly classification model is obtained.

[0062] During the model application phase, the time-series images corresponding to the calibration parameters to be detected are input into the trained classification model. The model automatically outputs recognition results including normal and various abnormal categories. For each input image, the model also outputs the probability value of its belonging to each category, as well as the final classification judgment corresponding to the highest probability value. Operators can use this to quickly screen for calibration periods with abnormalities.

[0063] Step 5, manual review and model iterative optimization, specifically includes:

[0064] The abnormal results identified by the convolutional neural network in step 4 are manually reviewed. Professionals with experience in calibrated data analysis check and confirm the model's output results, judging whether misclassification exists for each sample. Misclassification includes, but is not limited to, the following three situations: misclassifying a normal state as an abnormal state (false positive), misclassifying an abnormal state as a normal state (missed detection), and misclassifying one type of abnormality as another (type confusion).

[0065] For samples confirmed as misclassified by manual review, they are relabeled with the correct category labels and added to the training sample library. Incremental learning is employed, where the model parameters are fine-tuned and updated using the supplemented training samples, based on the original model parameters. Specific methods of incremental learning may include: keeping the parameters of the first few layers of the network unchanged (these layers have already learned the general feature extraction capabilities of the calibrated data), and only updating the parameters of the last few layers or fully connected layers. This preserves the general feature extraction capabilities already learned by the model, while allowing the model to adapt to the correction information in the newly labeled samples, preventing a decrease in the model's ability to classify previously correctly identified samples due to the addition of a small number of samples.

[0066] This closed-loop optimization mechanism of manual review, sample correction, and model update enables the anomaly detection model to continuously accumulate experience and improve itself during long-term operation. It enhances the adaptability to changes in calibration characteristics of different detection channels, different satellite platforms, and different operational stages, gradually reducing false alarm rate and false negative rate, and improving the overall identification accuracy of anomaly detection.

[0067] Step 6, calibration brightness temperature quality assessment and verification, specifically includes:

[0068] For the abnormal samples identified in step 4, the OB deviation data calculated in step 1 is used to conduct a quality assessment and verification of the calibration brightness temperature. The core idea is: if the OB deviation value corresponding to an abnormal sample is significantly different from the OB deviation value of a nearby normal sample, it indicates that the abnormality has a real impact on the quality of the calibration brightness temperature product and should be given special attention and dealt with promptly.

[0069] Specifically, the preset threshold range is determined as follows: In the identification results, the time of occurrence of the abnormal sample is selected. Samples within a complete orbital period are extracted both forward and backward from this time of occurrence. Samples labeled as normal are selected from these extracted samples. The mean and standard deviation of the OB deviation for these normal samples are calculated: the mean is the arithmetic mean of the OB deviation values ​​of all normal samples, and the standard deviation is the sample standard deviation of the OB deviation values ​​of the normal samples. The lower threshold is set as the mean minus three times the standard deviation, and the upper threshold is set as the mean plus three times the standard deviation. The closed interval formed by these two preset thresholds is defined as the preset threshold range. Based on the statistical principle of normal distribution, for a normally operating detector, the probability that its OB deviation value falls within the range of plus or minus three times the standard deviation of the mean is approximately 99.7%. Observations exceeding this range can be considered statistically significant.

[0070] In practice, when the number of normal samples in the truncated sample is less than the preset minimum number (e.g., less than 30 samples), the truncated range is extended to two complete orbital cycles before and after the truncated sample to ensure that there are enough normal samples for reliable calculation of statistical parameters (mean and standard deviation) and to avoid statistical inference bias caused by small samples.

[0071] After determining the preset threshold range, for each identified anomalous sample, its corresponding OB deviation value is compared with the aforementioned threshold range. If the OB deviation value of the anomalous sample exceeds the threshold range (i.e., less than the lower threshold or greater than the upper threshold), the anomalous sample is marked as a calibration quality anomaly, indicating that the anomaly has a real impact on the calibration results and may cause a decline in the quality of remote sensing data products. If the OB deviation value of the anomalous sample falls within the threshold range, it indicates that although the calibration parameters have undergone abnormal changes, the anomaly has not yet had a substantial impact on the quality of the calibration brightness temperature product under the current observation conditions. Its attention priority can be appropriately reduced, but continuous monitoring is still necessary.

[0072] Step 7, Severity Level Assessment and Model Re-optimization, specifically includes:

[0073] Based on the OB deviation variation magnitude and channel spatial distribution characteristics of the samples marked as calibration quality anomalies in step 6, the severity level of the anomalies is further assessed, and iterative optimization of the model is triggered.

[0074] The severity level of the anomaly is assessed comprehensively from two dimensions: deviation magnitude and spatial distribution. In terms of deviation magnitude, the variation range of OB deviation is divided into three ranges: mild deviation, moderate deviation, and severe deviation. The mild deviation range corresponds to a variation range less than a first preset threshold (e.g., absolute deviation less than 2K), the moderate deviation range corresponds to a variation range between the first and second preset thresholds (e.g., absolute deviation between 2K and 5K), and the severe deviation range corresponds to a variation range greater than the second preset threshold (e.g., absolute deviation greater than 5K).

[0075] In terms of spatial distribution, the impact levels are divided into single-channel impact level (only a few channels show OB deviation anomalies), partial-channel impact level (several channels show OB deviation anomalies at the same time), and full-channel impact level (all or most channels show OB deviation anomalies). The impact levels are also divided into local area level (OB deviation anomalies only appear in a specific geographical area or a specific scanning angle range) and global area level (OB deviation anomalies are prevalent globally).

[0076] By considering the values ​​of amplitude range, channel impact level, and spatial range level, a preset level mapping table is used to determine the severity level of calibration quality anomalies and label the level category. For example, a situation with full-channel impact, global scope, and severe deviation should be labeled as the highest severity level (e.g., Level 1 severe anomaly), while a situation with single-channel impact, local scope, and slight deviation can be labeled as the lowest severity level (e.g., Level 3 slight anomaly). Other combinations are mapped to intermediate levels according to preset rules. Different response measures can be taken according to different severity levels: Level 1 anomalies require immediate activation of the emergency calibration plan and suspension of business distribution of affected data; Level 2 anomalies require recent updates to the calibration coefficients; and Level 3 anomalies can be included in the regular monitoring plan for continuous tracking, thereby achieving refined management of calibration anomalies.

[0077] Regarding the iterative optimization of the model, the specific process is as follows: After the network parameters of the classification model are updated, the parameters to be detected are re-input into the updated classification model for secondary recognition, and the secondary recognition results are output. The sample distribution of each category in the secondary recognition results is compared with the correctly labeled samples confirmed in the manual review, and the proportion of misclassified samples in the secondary recognition results to the total number of samples is calculated. When the proportion of misclassified samples is lower than a preset threshold (e.g., lower than 5%), the iterative optimization is confirmed to be complete and the update is stopped. When the proportion of misclassified samples is not lower than the preset threshold, the manual review and incremental learning process in step 5 continues until the condition that the proportion of misclassified samples is lower than the preset threshold is met, ensuring that the recognition performance of the model meets the requirements of business operation.

[0078] The method provided in this embodiment transforms the traditional calibration quality monitoring process, which relies on manual analysis, into an automatic identification process based on deep learning. By using a convolutional neural network to automatically extract abnormal features from the calibration data, it achieves automatic detection and classification of calibration anomalies, significantly reducing the workload of manual monitoring and improving operational efficiency.

[0079] By constructing a calibration feature image that includes temperature change features, count value change features, orbital period change features, and diurnal variation features, and by fully utilizing the multi-scale feature extraction capabilities of deep learning models, this method can effectively identify various complex anomaly patterns such as abnormal drift of cold air count values, step changes in heat source count values, abnormal fluctuations in temperature parameters, and gradual shifts in calibration coefficients. This overcomes the limitations of traditional thresholding methods and statistical analysis methods in recognizing complex anomaly patterns such as slow drift, periodic oscillations, and multi-parameter coupled changes, thereby improving the accuracy and recall of anomaly detection.

[0080] By establishing a closed-loop optimization mechanism of manual review, sample correction, and model update, new abnormal samples are continuously accumulated and misjudgment results are corrected, enabling continuous updating of model parameters. This allows the anomaly detection model to adapt to changes in calibration characteristics across different channels, satellite platforms, and operational phases, thereby improving the model's long-term stability and generalization ability.

[0081] By introducing a calibration brightness temperature quality assessment and verification mechanism based on radiative transfer simulation, and utilizing the statistical characteristics of OB deviation to determine the substantial impact of anomalies on calibration results, this mechanism can effectively distinguish between anomalies that have no significant impact on calibration results and those that have a real impact on the quality of observed products, thus improving the reliability of anomaly identification. Simultaneously, by comprehensively evaluating the severity level of anomalies from both deviation amplitude and spatial distribution dimensions, a quantitative basis is provided for graded response and handling of anomalies, transforming calibration anomaly management from passive manual investigation to proactive intelligent graded early warning.

[0082] like Figure 2 As shown, a two-point calibration intelligent anomaly detection system for a spaceborne microwave temperature probe is described. This system is used to execute the method described above, including:

[0083] The data acquisition module is used to extract calibration parameters to build a database and to calculate OB bias data using radiative transfer model and numerical prediction analysis field data;

[0084] The feature image construction module is used to normalize the calibration parameters according to a preset time window and generate a temporal feature image.

[0085] The sample annotation and training set construction module is used to classify time-series images and construct training datasets.

[0086] The model training module is used to train a convolutional neural network with a training dataset to obtain a classification model.

[0087] The anomaly detection module is used to input the time-series image corresponding to the parameter to be detected into the classification model to obtain the recognition result;

[0088] The review and model update module is used to correct the labels of misclassified samples and update the model parameters through incremental learning.

[0089] The quality verification module is used to extract the OB deviation of abnormal samples, compare it with the deviation of normal samples, and mark the calibration quality anomalies.

[0090] The rating and iteration module is used to rate anomalies based on OB deviation amplitude and spatial distribution and trigger iterative optimization.

[0091] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A two-point calibration intelligent anomaly detection method for a spaceborne microwave temperature detector, characterized in that, Includes the following steps: A calibration parameter database was established by extracting calibration parameters from the observation data of the spaceborne microwave temperature detector. The calibration parameters include instrument temperature data, heat source temperature data and count values ​​of each channel. Based on the calibration parameter database, the brightness temperature after calibration is calculated. Synchronous brightness temperature is simulated using radiative transfer model and numerical prediction analysis field data combined with observational geometric information. The difference is calculated to obtain OB bias data. Based on the calibration parameter database, each parameter is normalized according to a preset time window, and a time series image containing temperature changes, count value changes, orbital period changes and diurnal variation characteristics is generated according to the time-varying law. Based on time-series images, normal and abnormal categories are labeled according to historical data, an abnormal sample library is established, and a training set is constructed. A convolutional neural network is trained based on the training set, and the network extracts abnormal features from time-series images to obtain a calibrated anomaly classification model. Based on the classification model, the time series image corresponding to the calibration parameters to be detected is input into the model to obtain the recognition results containing normal and abnormal categories; Based on the identification results, manual review is conducted, and misclassified samples are corrected and labeled before being added to the training set for incremental learning and updating of model parameters; Based on the identification results and OB deviation data, the deviation values ​​of abnormal samples are compared with those of nearby normal samples. When the deviation exceeds the preset threshold range, it is marked as a calibration quality abnormality. Based on the variation range of calibration quality anomalies (OB deviation) and the spatial distribution characteristics of channels, the severity level of the anomalies is assessed and model iterative optimization is triggered.

2. The method according to claim 1, characterized in that, The instrument temperature data includes the temperature of the detector host, the temperature of the receiver, and the temperature of the calibration source. The temperature of the detector host is the real-time temperature value collected by the temperature sensor installed on the housing of the detector host, and the temperature of the receiver is the measured temperature value of the low-noise amplifier at the front end of each detection channel receiver. The calibration source temperature is the arithmetic mean of the temperature values ​​measured by multiple temperature measuring points evenly distributed along the circumference of the calibration blackbody inside the detector. The heat source temperature data includes the heat load temperature and the temperature values ​​of multiple temperature measurement points near the heat source. The heat load temperature is the temperature value collected by the temperature sensor near the heating element in the thermal calibration load of the detector. The count values ​​for each channel are the original digital count values ​​output by the detector channels after analog-to-digital conversion in cold air observation mode, heat source observation mode, and earth observation mode. Cold air observation mode is the observation state when the detector antenna is pointed to the cold air background, heat source observation mode is the observation state when the detector antenna is pointed to the internal thermal calibration load, and earth observation mode is the observation state when the detector antenna is pointed to the target area on earth.

3. The method according to claim 2, characterized in that, The observation geometry information includes the longitude, latitude, zenith angle, and azimuth angle of the instrument at the time of observation; the radiative transfer mode is either the line-by-line integral radiative transfer mode or the fast radiative transfer mode; the numerical prediction analysis field data includes atmospheric temperature profiles, humidity profiles, pressure profiles, and surface temperature and surface emissivity parameters. Atmospheric temperature profile, humidity profile, and pressure profile are obtained from numerical prediction analysis field data by spatiotemporal interpolation according to observation time and instrument latitude and longitude. Surface temperature and surface emissivity parameters are extracted from the surface analysis field of numerical prediction analysis field data. Using the radiative transfer model, with atmospheric temperature and humidity profiles and surface parameters as input, the radiative transfer integral is calculated layer by layer along the light transmission path determined by the observation zenith angle and observation azimuth angle to obtain the atmospheric top exit brightness temperature of each detection channel. The atmospheric top exit brightness temperature is then used as the simulated synchronous brightness temperature.

4. The method according to claim 3, characterized in that, The duration of the preset time window covers at least one complete orbital cycle of the detector; Normalization is a process that linearly maps the actual value range of each calibration parameter within the current time window to a preset uniform value range. The actual value range is the range formed by the maximum and minimum values ​​of the parameter within the current time window. The time series image uses time as the horizontal axis and the normalized parameter values ​​as the vertical axis, and overlays the variation curves of at least two calibration parameters in the same image; The curves of different calibration parameters are distinguished by different colors or line types, and the start time of each orbital period within the time window is marked in the image; When a preset time window undergoes a sliding update, there is a preset proportion of overlapping area between two adjacent time windows. The overlapping area is used to determine the continuity of temporal features between adjacent windows. The time-series images are stored as digital image formats at a preset pixel resolution.

5. The method according to claim 4, characterized in that, A convolutional neural network comprises an input layer, at least three convolutional layers, at least two pooling layers, at least one fully connected layer, and an output layer connected in sequence. Each convolutional layer contains multiple convolutional kernels, which are used to perform sliding window convolution operations on the input temporal images to extract local anomalous texture features at different scales; The width of the convolution kernel is a preset odd number of pixels, the stride is a preset integer value, and zero padding is applied to the boundaries of the input feature map to keep the spatial size of the output feature map unchanged. Each pooling layer uses max pooling to downsample the feature map output by the previous convolutional layer to reduce the width and height dimensions of the feature map. The fully connected layer unfolds the feature map output by the last pooling layer into a one-dimensional feature vector and maps the feature vector to the output node corresponding to each anomaly category; the output layer uses a normalized exponential function to calculate the probability value of each output node and takes the category corresponding to the highest probability value as the recognition result.

6. The method according to claim 5, characterized in that, The anomaly categories include cold air count value anomaly drift category, heat source count value step change category, temperature parameter fluctuation anomaly category, and calibration coefficient gradual shift category. Abnormal drift of cold air count value refers to the cold air count value continuously shifting in a single direction over a period of time and the amount of shift exceeding the preset drift threshold; A step change in the heat source count value refers to a sudden change in the heat source count value between adjacent sampling points, and the magnitude of the change exceeds a preset step threshold. Abnormal temperature parameter fluctuation refers to the reciprocating oscillation of the instrument temperature or heat source temperature exceeding the preset fluctuation threshold within multiple consecutive sampling cycles. The calibration coefficient gradual offset refers to the calibration gain coefficient or calibration bias coefficient exhibiting a monotonic changing trend over multiple consecutive orbital periods, and the cumulative change exceeding the preset offset threshold. Category labeling involves marking the start and end time ranges of an anomaly occurrence on the time series image with a rectangle, and labeling the corresponding anomaly type next to the rectangle.

7. The method according to claim 6, characterized in that, The preset threshold range is determined as follows: In the identification results, select the time of occurrence of the abnormal sample corresponding to the abnormal sample, and take the sample within a complete orbital period before and after the time of occurrence of the abnormal sample as the benchmark. Select the sample marked as normal from the sampled sample, and calculate the mean and standard deviation of the OB deviation of the normal sample. When the number of normal samples in the truncated sample is less than the preset minimum number, the truncated range is expanded to two complete orbital cycles before and after; the mean is the arithmetic mean of the OB deviation values ​​of all normal samples, and the standard deviation is the sample standard deviation of the OB deviation values ​​of normal samples. The lower threshold is set by subtracting three standard deviations from the mean, and the upper threshold is set by adding three standard deviations to the mean. The final closed interval formed by the lower and upper thresholds is used as the preset threshold range.

8. The method according to claim 7, characterized in that, The specific assessment of the severity level is as follows: The variation range of OB deviation is divided into three ranges: slight deviation, moderate deviation, and severe deviation. The variation range of the slight deviation range is less than the first preset amplitude threshold, the variation range of the moderate deviation range is between the first preset amplitude threshold and the second preset amplitude threshold, and the variation range of the severe deviation range is greater than the second preset amplitude threshold. The spatial distribution characteristics are classified into single-channel impact level, partial-channel impact level and full-channel impact level according to the number of affected detection channels, and into local area level and global area level according to the affected spatial range. By considering the values ​​of the amplitude range, channel impact level, and spatial range level, a preset level mapping table is used to determine the severity level of the calibration quality anomaly and to label the level category.

9. The method according to claim 8, characterized in that, The specific process of iterative optimization is as follows: After the network parameters of the classification model are updated, the parameters to be detected are re-input into the updated classification model for secondary recognition, and the secondary recognition results are output. The sample distribution of each category in the secondary identification results is compared with the correct labels confirmed by manual review, and the proportion of misclassified samples in the secondary identification results to the total number of samples is calculated. When the proportion of misclassified samples is lower than the preset proportion threshold, confirm that the iterative optimization is complete and stop updating; when the proportion of misclassified samples is not lower than the preset proportion threshold, continue to perform manual review and incremental learning until the conditions are met. The preset proportion threshold is the upper limit of the proportion of misclassified samples set in advance.

10. A two-point calibration intelligent anomaly detection system for a spaceborne microwave temperature probe, the system being used to execute the method as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to extract calibration parameters to build a database and to calculate OB bias data using radiative transfer model and numerical prediction analysis field data; The feature image construction module is used to normalize the calibration parameters according to a preset time window and generate a temporal feature image. The sample annotation and training set construction module is used to classify time-series images and construct training datasets. The model training module is used to train a convolutional neural network with a training dataset to obtain a classification model. The anomaly detection module is used to input the time-series image corresponding to the parameter to be detected into the classification model to obtain the recognition result; The review and model update module is used to correct the labels of misclassified samples and update the model parameters through incremental learning. The quality verification module is used to extract the OB deviation of abnormal samples, compare it with the deviation of normal samples, and mark the calibration quality anomalies. The rating and iteration module is used to rate anomalies based on OB deviation amplitude and spatial distribution and trigger iterative optimization.

Images