AI-driven low-altitude data quality intelligent verification and treatment system

Abstract

The invention relates to the technical field of data quality verification and AI intelligent treatment, in particular to an AI-driven low-altitude data quality intelligent verification treatment system, which is characterized in that a dynamic coupling coefficient accounting unit synchronizes a GPS positioning error and an image pixel offset error, and calculates a coupling coefficient of each frame through an error coupling algorithm; data continuity is guaranteed through error compensation and interpolation complementation; the ideal interval comparison unit is used for presetting a dynamic ideal interval based on the stationary flight data, calculating a deviation percentage by adopting a two-way difference value proportion method and marking positive and negative abnormal characteristics; the AI driving verification treatment unit fuses abnormal features, GPS jump and an image ambiguity threshold, constructs a multi-dimensional feature vector, identifies a recessive mode in which single-source data does not exceed a limit but coupling is abnormal through a time sequence convolutional network, outputs an alarm after triple conditions are met, drives the automatic calibration module to adjust exposure frequency, resets GPS filtering parameters, injects calibration parameters, and finally performs calibration. The data quality is managed in three stages, and the low-altitude data application accuracy is guaranteed.

Description

Technical Field

[0001] This invention relates to the field of data quality verification and AI-driven intelligent governance technology, specifically to an AI-driven intelligent verification and governance system for low-altitude data quality. Background Technology

[0002] Data quality verification and AI-powered intelligent governance is a crucial technology specifically applied to the data quality control of low-altitude equipment such as drones. Its core principle is to ensure data reliability by accurately identifying anomalies in data error coupling. This aligns with the high-quality data requirements of low-altitude data in surveying, monitoring, and other scenarios. During low-altitude data acquisition, GPS positioning errors and image pixel offset errors are dynamically correlated. These errors form a coupling relationship as flight conditions change. Since threshold verification of single-source data can only determine whether each exceeds its own limits, it cannot capture the hidden anomalies caused by the coupling of these two factors, leading to the omission of potential data quality issues and affecting the accuracy of subsequent low-altitude data applications. To address this technical problem, we provide an AI-driven intelligent verification and governance system for low-altitude data quality. Summary of the Invention

[0003] The purpose of this invention is to provide an AI-driven intelligent verification and governance system for low-altitude data quality, in order to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, an AI-driven intelligent verification and governance system for low-altitude data quality is provided, including: The dynamic coupling coefficient calculation unit synchronously acquires the GPS positioning error and image pixel offset error output by the data acquisition and processing unit frame by frame, and calculates the coupling coefficient of each frame in real time by dividing the image pixel offset error by the GPS positioning error through the error coupling algorithm. The ideal interval comparison unit presets the ideal interval of the coupling coefficient during the stable flight phase. It reads the coupling coefficient of the current frame from the coupling coefficient of each frame output by the dynamic coupling coefficient calculation unit, calculates the percentage deviation of the coupling coefficient of the current frame from the lower and upper limits of the ideal interval. If the coupling coefficient of the current frame is lower than the lower limit of the ideal interval, a negative deviation is calculated. If the coupling coefficient of the current frame is higher than the upper limit of the ideal interval, a positive deviation is calculated. The negative and positive deviation values ​​are marked as abnormal coupling coefficient features in real time. The AI-driven verification and governance unit integrates the abnormal features of the coupling coefficient as a new verification input into the AI ​​model. During the AI ​​verification feature construction stage of the AI ​​model, the deviation percentage is fused with the GPS jump threshold and the image blur threshold to form a multi-dimensional feature vector. The multi-dimensional feature vector is then input into the trained neural network model. The neural network model identifies latent anomaly patterns where the coupling coefficient deviation exceeds the preset threshold through supervised learning. When it is identified that the single source data does not exceed the threshold but the coupling coefficient deviates from the lower and upper limits of the ideal range to the preset threshold, a latent anomaly alarm signal is output, driving the automatic calibration module to adjust the data acquisition parameters to improve data quality.

[0005] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention accurately calculates the coupling coefficient for each frame through a timestamp synchronization mechanism, error compensation, and interpolation completion functions in a dynamic coupling coefficient calculation unit, ensuring the true continuity of error correlation characteristics. The ideal interval comparison unit dynamically updates the ideal interval based on stable flight data, quantifies deviation using a bidirectional difference ratio method, and clearly marks abnormal features with high and low bit alarm codes, providing accurate basis for subsequent verification. The AI-driven verification and governance unit integrates abnormal coupling coefficient features with GPS and image quality indicators to construct a multi-dimensional feature vector. A temporal convolutional network accurately identifies latent patterns in single-source data that are within limits but exhibit coupling anomalies. Triple condition verification avoids false alarms. The automatic calibration module performs three-stage governance, compensating for asynchronous time errors, resetting positioning filter parameters, and correcting historical deviations, thereby improving data quality from the root. This achieves accurate quantification of low-altitude data error coupling characteristics, intelligent identification and efficient governance of latent anomalies, filling the gap in traditional threshold verification, ensuring the accuracy of low-altitude data applications in surveying, monitoring, and other scenarios, and improving the intelligence and reliability of data quality control. Attached Figure Description

[0006] Figure 1 This is an overall block diagram of the present invention.

[0007] The meanings of the labels in the diagram are as follows: 1. Data acquisition and processing unit; 2. Dynamic coupling coefficient calculation unit; 3. Ideal interval comparison unit; 4. AI-driven verification and governance unit. Detailed Implementation

[0008] 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.

[0009] This invention provides AI-driven intelligent verification and governance of low-altitude data quality. Please refer to [link / reference]. Figure 1 As shown, it includes: The dynamic coupling coefficient calculation unit 2 synchronously acquires the GPS positioning error and image pixel offset error output by the data acquisition and processing unit 1 frame by frame, and calculates the coupling coefficient of each frame in real time by dividing the image pixel offset error by the GPS positioning error through the error coupling algorithm. The ideal interval comparison unit 3 presets the ideal interval of the coupling coefficient during the stable flight phase. It reads the coupling coefficient of the current frame from the coupling coefficient of each frame output by the dynamic coupling coefficient calculation unit 2, calculates the percentage deviation of the coupling coefficient of the current frame from the lower and upper limits of the ideal interval. If the coupling coefficient of the current frame is lower than the lower limit of the ideal interval, it calculates the negative deviation. If the coupling coefficient of the current frame is higher than the upper limit of the ideal interval, it calculates the positive deviation. The negative and positive deviation values ​​are marked as abnormal coupling coefficient features in real time. The AI-driven verification and governance unit 4 integrates the abnormal features of the coupling coefficient as a new verification input into the AI ​​model. In the AI ​​verification feature construction stage of the AI ​​model, the deviation percentage is fused with the GPS jump threshold and the image blur threshold to form a multi-dimensional feature vector. The multi-dimensional feature vector is then input into the trained neural network model. The neural network model identifies latent anomaly patterns where the coupling coefficient deviation exceeds the preset threshold through supervised learning. When it is identified that the single source data does not exceed the threshold but the coupling coefficient deviates from the lower and upper limits of the ideal range to the preset threshold, a latent anomaly alarm signal is output, driving the automatic calibration module to adjust the data acquisition parameters to improve data quality.

[0010] When the error coupling algorithm is executed in the dynamic coupling coefficient calculation unit, the time stamp synchronization mechanism ensures that the GPS positioning error and the image pixel offset error are collected at the same time. The floating-point divider is used to divide the two error values ​​to generate the coupling coefficient. When the GPS positioning error is lower than the preset minimum effective value, the error compensation module is automatically activated to inject the reference compensation amount. At the same time, a frame-level buffer queue is established to temporarily store the original error data of 5 consecutive frames. If an abnormal interruption of single frame data is detected, the interpolation algorithm of adjacent frames is called to fill in the missing values.

[0011] The ideal range of coupling coefficients preset in Ideal Range Comparison Unit 3 is obtained through training with historical stable flight phase data, specifically as follows: When the drone's attitude angle change rate is continuously lower than the set angular velocity threshold and the continuous flight time exceeds the set time window, all coupling coefficient data within the period when the continuous flight time exceeds the set time window are extracted. After removing the maximum and minimum values, the range of values ​​in the dense area of ​​the statistical distribution is taken as the ideal interval for dynamic updates. Before reading the coupling coefficient of the current frame, the data frame status flag bit output by the dynamic coupling coefficient calculation unit 2 needs to be verified. If the flag bit is abnormal, the calculation of the current frame is skipped.

[0012] The percentage deviation is calculated using the two-way difference ratio method, specifically: The difference between the current frame coupling coefficient and the lower limit of the ideal interval is divided by the lower limit of the ideal interval to obtain the negative deviation from the base value, and the difference between the current frame coupling coefficient and the upper limit of the ideal interval is divided by the upper limit of the ideal interval to obtain the positive deviation from the base value. Then, the base value is converted into a percentage form by the normalization processor. During the conversion process, the interval width weighting factor is introduced so that the deviation percentage can reflect the relative position of the current value in the ideal interval.

[0013] The positive deviation calculation is activated when the coupling coefficient of the current frame is detected to be higher than the upper limit of the ideal range. The absolute difference between the current value and the upper limit of the ideal range is calculated and divided by the upper limit of the ideal range to obtain the basic positive ratio. The basic positive ratio is input to a gain regulator that is different from the negative deviation. The gain regulator dynamically adjusts the amplification coefficient based on the UAV flight altitude parameters.

[0014] The coupling coefficient anomaly feature marker is implemented through a feature encoder, which converts the deviation percentage into a binary feature code. When the negative deviation exceeds the set warning threshold, a low-order alarm code is generated, and when the positive deviation exceeds the warning threshold, a high-order alarm code is generated. If neither deviation exceeds the limit, a zero value code is output. All feature codes are appended with timestamps and written to a circular feature buffer.

[0015] The multidimensional feature vector construction adopts the feature tensor concatenation technique. The latest coupling coefficient abnormal feature code is extracted from the circular feature buffer and combined with the real-time acquired GPS jump flag and image blur flag to form a three-dimensional feature tuple. The discrete flags are mapped into continuous vectors through the feature embedding layer, and then concatenated along the feature dimension to form a fixed-length fused feature vector.

[0016] The neural network model adopts a temporal convolutional network architecture. After receiving multi-dimensional feature vectors, the input layer sequentially extracts time-dependent features through causal convolutional layers, filters noise features through gated activation layers, focuses key abnormal segments through attention pooling layers, and connects to a sigmoid classifier through the output layer. When the output value of the neural network model exceeds the decision boundary threshold, it is determined that there is a latent abnormal pattern with an excessively high coupling coefficient deviation.

[0017] The output of the latent anomaly alarm signal must simultaneously meet three conditions: the GPS jump threshold verification flag is not exceeded, the image blur threshold verification flag is not exceeded, and the neural network model outputs an anomaly judgment signal. After the alarm signal is generated, the automatic calibration module is driven to perform three-stage governance. First, the exposure frequency of the image sensor is adjusted to compensate for the asynchronous error of the time. Second, the signal filtering parameters of the GPS receiver are reset. Finally, the calibration parameters are injected into the data fusion algorithm to compensate for historical deviations.

[0018] Further explanation is needed: after the data acquisition and processing unit outputs the GPS positioning error and image pixel offset error, the dynamic coupling coefficient calculation unit needs to use an error coupling algorithm to accurately fuse the two types of errors and calculate the coupling coefficient, ensuring that the coupling coefficient of each frame can truly reflect the error correlation characteristics of the low-altitude data. The specific implementation method is as follows: First, a timestamp synchronization mechanism is used to ensure that GPS positioning errors and image pixel offset errors are collected at the same time. This mechanism uses a unified clock source to add high-precision timestamps to the two types of error data and then matches them. Its core is to eliminate asynchronous time deviations in data acquisition. The specific process is as follows: The GPS receiver and image sensor in the data acquisition and processing unit are configured with the same GPS timing module to ensure complete clock synchronization. Each time a set of GPS positioning errors (i.e., the deviation between satellite positioning data and the actual location) and image pixel offset errors (i.e., the offset between the actual and ideal positions of image pixels) are acquired, a timestamp accurate to the millisecond level is automatically added. After receiving the data, the dynamic coupling coefficient calculation unit compares the timestamps of each set of GPS positioning errors and image pixel offset errors, calculating the time difference. If the difference is less than a preset synchronization threshold (experimentally calibrated to 1 millisecond to ensure data acquisition at the same time), synchronization is considered valid and included in subsequent calculations. If the difference exceeds the threshold, the data set is discarded, and the system waits for the next set of synchronized data, thus preventing coupling coefficient distortion caused by time asynchrony from the source. After valid synchronization, a floating-point divider is used to divide the two error values ​​to generate the coupling coefficient. The floating-point divider is a hardware module that supports high-precision decimal division, avoiding precision loss caused by integer division and adapting to the refined calculation requirements of low-altitude data errors. The specific calculation process is as follows: First, the synchronized GPS positioning error and image pixel offset error are standardized by converting them into 32-bit floating-point data to eliminate computational bias caused by inconsistent data formats. Then, the image pixel offset error is used as the dividend, and the GPS positioning error as the divisor, and a floating-point divider is used for division. During the calculation, hardware-level precision calibration is enabled to automatically correct for truncation errors in floating-point operations. The final output is the coupling coefficient for that frame, automatically retaining 4 decimal places to ensure the coefficient's accuracy meets subsequent anomaly detection requirements. For example, when the image pixel offset error is 0.5 pixels and the GPS positioning error is 2 meters, the coupling coefficient is calculated to be 0.25. When the GPS positioning error is lower than a preset minimum effective value, the error compensation module automatically injects a reference compensation amount. The preset minimum effective value is a critical value set to avoid overflow or distortion of the division result due to excessively small GPS positioning errors (close to zero). After extensive experimentation, it has been calibrated to 0.1 meters; that is, when the GPS positioning error is less than 0.1 meters, it is considered an invalid low error value. The error compensation module is a pre-installed fallback correction unit. The baseline compensation amount is a fixed value (usually set to 0.2 meters) statistically calibrated using effective error data from historical stable flight phases. This value neither excessively interferes with normal error correlations nor compromises the effectiveness of division operations. The specific injection process is as follows: The system monitors the GPS positioning error in real time. If the error is detected to be lower than the preset minimum effective value, the error compensation module is immediately triggered. The baseline compensation is added to the original GPS positioning error to obtain a corrected GPS positioning error (e.g., if the original error is 0.08 meters, the added error is 0.28 meters). Then, the corrected GPS positioning error is used as the divisor, and the image pixel offset error is used as the dividend. A floating-point divider is used to re-perform the calculation, generating an effective coupling coefficient to avoid operational anomalies caused by an excessively small divisor. Simultaneously, a frame-level buffer queue is established to temporarily store five consecutive frames of raw error data. This frame-level buffer queue is a circular data buffer using first-in-first-out logic, specifically used to store the original GPS positioning error and image pixel offset error data after synchronization. Its function is to provide a basis for data completion in case of abnormal interruption, ensuring the continuity of the operation. The specific storage process is as follows: After each set of error data is synchronized and processed, the original data (including timestamp, error value, and data status flag) is stored in a buffer queue. The queue always retains only the latest 5 frames of data. When new data is stored, the oldest frame is automatically overflowed and deleted to ensure the timeliness of the buffered data. If an abnormal interruption of a single frame of data is detected (judged by the data frame status flag; a flag of 0 indicates data acquisition interruption, transmission error, or other abnormalities, while a flag of 1 indicates normal operation), the adjacent frame interpolation algorithm is immediately invoked to fill in the missing values. The adjacent frame interpolation algorithm is based on the assumption of data continuity and uses data from normal frames before and after the missing frame to calculate the filled values. The specific filling process is as follows: The system extracts valid raw error data from the preceding and following frames of the missing frame from the frame-level buffer queue. It calculates the average GPS positioning error and the average image pixel offset error of these two frames, using these averages as the completion error value for the missing frame. A special marker (completion flag) is added to the completed data. Then, a division operation is performed based on the completed error value to generate the coupling coefficient, ensuring that even if a single frame's data is interrupted, the continuity and integrity of the overall coupling coefficient calculation are not affected. The entire process establishes a solid foundation for accurate data fusion through a timestamp synchronization mechanism, ensures computational precision through a floating-point divider, avoids computational anomalies caused by extremely low errors through an error compensation module, and solves data interruption problems using frame-level caching and interpolation algorithms. This comprehensive approach ensures the validity, accuracy, and continuity of the coupling coefficient for each frame, providing high-quality data support for subsequent ideal interval comparison units to extract the current frame's coupling coefficient and calculate the deviation percentage.

[0019] After the dynamic coupling coefficient calculation unit continuously outputs the coupling coefficient for each frame, the ideal interval comparison unit needs to first preset the ideal interval of the coupling coefficient. This interval must be generated based on real data from the stable flight phase of the UAV to ensure the objectivity and accuracy of subsequent deviation judgment. At the same time, data validity verification should be performed before reading the coupling coefficient of the current frame to avoid abnormal data interfering with the calculation. The specific implementation method is as follows: The generation and data verification of the ideal interval revolve around the logic of determining the stable phase—data extraction and filtering—interval calibration—validity verification. First, the stable flight phase is determined: the rate of change of the UAV's attitude angles refers to the real-time rate of change of the UAV's pitch, roll, and yaw angles, directly reflecting the stability of the flight state; the lower the rate of change, the more stable the flight. An angular velocity threshold is set as a critical value calibrated through numerous flight experiments (usually set to 2 degrees per second) to define whether the attitude angle changes are within a stable range. A time window is set as the minimum duration to ensure that the stable state has statistical significance (calibrated to 10 seconds), avoiding misjudging a brief period of stability as a valid phase. The specific determination process is as follows: The drone's built-in inertial measurement unit collects attitude angle change rate data in real time, recording the value every 100 milliseconds. It continuously monitors and calculates the duration for which this value remains below a set angular velocity threshold. If the duration exceeds a set time window (10 seconds), the drone is considered to have entered a stable flight phase, and a coupling coefficient data extraction command is immediately triggered. If the duration does not meet the threshold or the attitude angle change rate exceeds the threshold midway, the timing restarts, ensuring that the extracted data comes from real, stable flight scenarios. Once in the stable flight phase, all coupling coefficient data within the time period exceeding the set time window are extracted. The start and end timestamps of the stable phase are automatically recorded. Based on these timestamps, all coupling coefficient data within this time interval are filtered from the output cache of the dynamic coupling coefficient calculation unit, forming a stable phase coupling coefficient dataset. The dataset was then filtered: first, the maximum and minimum values ​​were removed using an extreme value removal algorithm. This was to eliminate potential instantaneous abnormal errors during the steady-state phase (such as sudden changes in the coupling coefficient caused by brief airflow disturbances), preventing such extreme values ​​from affecting the rationality of the ideal interval. After removing the extreme values, statistical distribution analysis was performed on the remaining data. The dense region in the statistical distribution refers to the interval range in the dataset with the highest frequency and most concentrated values. The specific determination process is as follows: The mean and standard deviation of the remaining data are calculated. Using the mean as the center, a range of 1.5 standard deviations is expanded outwards to both sides. This range contains the coupling coefficient data for the vast majority of stable phases, and this range is set as the initial ideal interval. Simultaneously, the ideal interval is dynamically updated. After each new stable flight phase data acquisition and analysis, the newly generated interval is merged with the historical ideal interval, and the intersection of the two is taken as the final dynamically updated ideal interval, ensuring that the interval can adapt to the stable state characteristics under different flight environments. Before reading the coupling coefficient of the current frame to calculate the deviation, the data frame status flag bit output by the dynamic coupling coefficient calculation unit must be verified. The data frame status flag bit is a validity indicator added by the dynamic coupling coefficient calculation unit to each frame of data (represented by 0 and 1, 0 for abnormal, 1 for normal). Its generation is based on factors such as whether the data acquisition is synchronized, whether the calculation is valid, and whether error compensation or data completion has been performed, directly reflecting the reliability of the coupling coefficient of that frame. The specific verification process is as follows: When the ideal interval comparison unit receives the coupling coefficient of the current frame, it simultaneously reads the status flag bit corresponding to the frame data. If the flag bit is 1, it indicates that the frame data acquisition is synchronized, the calculation is valid, and there are no abnormal interruptions, thus the data is deemed valid, the coupling coefficient is read normally, and the subsequent deviation calculation stage begins. If the flag bit is 0, it indicates that the frame data has problems such as time asynchrony, calculation overflow, or data completion failure, thus the data is deemed invalid, the calculation of the current frame is skipped, and it is not included in the abnormal feature marking, avoiding misjudgment of deviation caused by invalid data and ensuring the rigor of the entire verification process. The entire process ensures the reliability of the data source by accurately determining the stable flight phase, determines the rationality of the ideal interval through extreme value elimination and statistical distribution analysis, and ensures the validity of the current frame data through status flag bit verification. It not only discloses the complete technical details of the dynamic generation of the ideal interval and data verification, but also ensures that the ideal interval can truly reflect the normal error correlation level of low-altitude data through a layer-by-layer progressive logic, laying a solid foundation for subsequent deviation percentage calculation and abnormal feature marking, making the entire intelligent verification and governance process more scientific and operable.

[0020] After completing the ideal interval calibration of the coupling coefficient and the validity verification of the current frame data, the ideal interval comparison unit needs to calculate the deviation percentage using the bidirectional difference ratio method. The core is to quantify the degree of deviation between the current frame coupling coefficient and the ideal interval, providing an accurate quantitative basis for subsequent anomaly feature labeling, and ensuring that the anomaly judgment is both objective and consistent with the actual data distribution. The specific implementation method is as follows: First, it's important to clarify that the bidirectional difference ratio method addresses two scenarios: the coupling coefficient is below the lower limit of the ideal range, and the coupling coefficient is above the upper limit of the ideal range. It employs a differentiated difference ratio calculation method to accurately distinguish the degree of negative and positive deviation, avoiding misjudgments caused by a single calculation method. When the coupling coefficient of the current frame is detected to be below the lower limit of the ideal range, the calculation of the negative deviation baseline value is initiated. The lower limit of the ideal range is the minimum value of the dense coupling coefficient region calibrated using historical stable flight data, serving as the critical benchmark for determining negative deviation. The negative deviation baseline value is the original value quantifying the degree to which the coupling coefficient of the current frame is below the lower limit. Its specific calculation process is as follows: First, extract the current frame coupling coefficient, which has been verified for validity. Then, retrieve the preset lower limit value of the ideal interval. Subtract the lower limit of the ideal interval from the current frame coupling coefficient to obtain the difference (since the current value is lower than the lower limit, this difference is negative, directly reflecting the absolute magnitude of the negative deviation). Then, divide this difference by the lower limit of the ideal interval. The result is the negative deviation baseline value. For example, if the current frame coupling coefficient is 0.2 and the lower limit of the ideal interval is 0.3, the difference is -0.1. Dividing by 0.3 gives a negative deviation baseline value of approximately -0.333. The larger the absolute value of this value, the more severe the negative deviation. When the current frame coupling coefficient is detected to be higher than the upper limit of the ideal interval, the calculation of the positive deviation baseline value is initiated. The upper limit of the ideal interval is the maximum value of the dense coupling coefficient region and is the critical benchmark for determining positive deviation. The positive deviation baseline value is the original value that quantifies the degree to which the current frame coupling coefficient exceeds the upper limit. The specific calculation process is as follows: Similarly, extract the effective coupling coefficient of the current frame and the upper limit of the ideal interval. Subtract the upper limit of the ideal interval from the coupling coefficient of the current frame to get the difference (since the current value is higher than the upper limit, the difference is positive, reflecting the absolute magnitude of the positive deviation). Then divide the difference by the upper limit of the ideal interval to get the basic value of the positive deviation. For example, if the coupling coefficient of the current frame is 0.6 and the upper limit of the ideal interval is 0.5, the difference is 0.1. Dividing by 0.5 gives the basic value of the positive deviation as 0.2. The larger the value, the more significant the positive deviation. After obtaining the negative or positive deviation from the baseline value, it is converted into a percentage form by a normalization processor. The normalization processor is a functional module that realizes the standardization transformation of the original data. Its core function is to uniformly map the deviation from the baseline value in different ranges to a percentage range of 0 to 100%, making it convenient to intuitively judge the degree of deviation. The interval width weighting factor is an adjustment parameter introduced to correct the influence of the ideal interval width on the judgment of deviation. Its value is equal to the difference between the upper and lower limits of the ideal interval (i.e., the interval width) divided by the average value of the coupling coefficient in the historical stable phase. It can make the percentage deviation more accurately reflect the relative position of the current value in the ideal interval (for example, the relative deviation of the same difference in a narrow interval should be greater than that in a wide interval). The specific conversion process is as follows: First, the deviation from the base value (absolute value for negative deviations) is input into the normalization processor. The processor first multiplies the base value by the interval width weighting factor to complete the weighting correction (for example, if the negative deviation from the base value is 0.333 and the interval width weighting factor is 0.2, the weighted value is 0.0666; if the positive deviation from the base value is 0.2, the weighted value is 0.04). Then, the weighted value is multiplied by 100 to convert it into a percentage form, while keeping one decimal place. Finally, the deviation percentage is obtained. When there is a negative deviation, a negative sign is added before the percentage to indicate the direction. When there is a positive deviation, the value is presented directly. For example, the above negative case is converted to -6.7%, and the positive case is 4.0%. The entire process accurately distinguishes deviation types through direction determination, ensures the objectivity of the base value by calculating the difference ratio, and corrects the influence of distribution differences by using the interval width weighting factor. The final generated deviation percentage not only quantifies the absolute degree of deviation but also reflects the relative positional relationship. This provides a scientific quantitative standard for subsequent marking of abnormal features when the deviation exceeds the warning threshold. It perfectly connects the ideal interval calibration and coupling coefficient abnormal feature marking links, ensuring that every step of anomaly detection has traceable and verifiable technical basis.

[0021] After calculating the basic deviation value using the two-way difference ratio method, considering the special characteristics of positive deviation—that is, when the coupling coefficient is higher than the upper limit of the ideal range—the degree of impact on data quality may vary depending on the flight altitude. Therefore, a dedicated positive deviation calculation process needs to be initiated to achieve more accurate quantification through differentiated gain adjustment. The specific implementation method is as follows: The core logic of positive deviation calculation is trigger activation—basic ratio calculation—dynamic gain adjustment. First, the activation process requires precise triggering: the core condition for activation is that the current frame coupling coefficient is higher than the upper limit of the ideal range. Here, the current frame coupling coefficient is data output by the dynamic coupling coefficient calculation unit and verified as valid through the data frame status flag. The upper limit of the ideal range is the maximum value of the dense coupling coefficient region calibrated based on data from the stable flight phase, serving as the critical benchmark for determining positive deviation. After verifying the validity of the current frame coupling coefficient, it is compared with the upper limit of the ideal range in real time. If the current frame coupling coefficient is detected to be greater than the upper limit, the positive deviation calculation process is immediately activated, and the timestamp and flight status parameters of that frame are locked to ensure accurate matching of subsequent calculations and data acquisition scenarios. If the upper limit is not exceeded, the process is not initiated, and the basic positive ratio is maintained as a temporary reference to avoid unnecessary computation consuming resources. After activation, the basic positive ratio calculation begins: the current value is the verified effective coupling coefficient of the current frame. The absolute difference is the absolute value of the result after subtracting the upper limit of the ideal interval from the current value. This is done to eliminate interference from positive and negative signs, retaining only the magnitude of the deviation. For example, if the current value is 0.6 and the upper limit of the ideal interval is 0.5, the difference between the two is 0.1, and the absolute difference is still 0.1. The basic positive ratio is the proportional value that quantifies the degree of positive deviation from the original value. The specific calculation process is as follows: Dividing the obtained absolute difference by the upper limit of the ideal range yields the basic positive ratio. In the example above, 0.1 divided by 0.5 results in a basic positive ratio of 0.2. This value directly reflects the relative proportion of the current value exceeding the upper limit; a larger value indicates a more severe original degree of positive deviation. The basic positive ratio is then input to a gain regulator, which is different from the gain regulator used for negative deviation. This gain regulator is a functional module used to dynamically adjust the amplification factor of the deviation ratio based on scene parameters. It differs from the regulator used for negative deviation because positive deviation is often strongly correlated with the drone's flight altitude (at high altitudes, positive deviation may lead to larger data stitching errors), requiring targeted sensitivity adjustment. The drone's flight altitude parameters are absolute altitude data collected in real-time by the drone's built-in barometric altimeter or GPS module. The sampling frequency is consistent with the coupling coefficient calculation frequency (collected once per frame), accurate to the meter level, ensuring real-time reflection of the flight scene's altitude characteristics. The specific process of dynamically adjusting the amplification factor is as follows: Multiple preset mapping relationships between altitude ranges and magnification factors are established. These mapping relationships are calibrated based on flight experiment data at different altitudes. For example, when the flight altitude is below 100 meters, the magnification factor is set to 1.2; between 100 and 300 meters, it is 1.5; and above 300 meters, it is 1.8. The higher the altitude, the larger the magnification factor, to highlight the impact weight of positive deviation in high-altitude scenarios. The gain regulator receives the UAV's flight altitude parameters in real time, automatically matches the corresponding magnification factor, and then multiplies this factor by the basic positive ratio to obtain the final positive deviation. For example, if the basic positive ratio is 0.2 and the flight altitude is 200 meters (corresponding to a magnification factor of 1.5), the final positive deviation is 0.3 (i.e., 30%). This result retains the basic deviation ratio while incorporating the scenario characteristics of flight altitude, making the quantification of positive deviation more in line with actual application needs. The entire process ensures the targeting of the process through precise activation condition determination, guarantees the objectivity of the basic ratio through absolute difference calculation, and achieves precise quantification in a scenario-based manner through highly correlated dynamic gain adjustment. The resulting positive and negative deviations form a complete deviation quantification system, providing accurate and differentiated input data for the subsequent feature encoder to convert the deviation percentage into binary feature codes, ensuring that the abnormal feature markers can truly reflect the data quality issues under different deviation types and scenarios.

[0022] After accurately calculating the positive and negative deviations, in order for the AI-driven verification and governance unit to efficiently identify abnormal patterns, the ideal interval comparison unit needs to convert the quantified deviation percentage into a binary feature code that can be quickly parsed by the machine through a standardized feature marking process. At the same time, it generates corresponding alarm codes according to the deviation type to ensure clear transmission of abnormal information and efficient connection with subsequent processing. The specific implementation method is as follows: First, the deviation percentage is converted into binary feature code through the feature encoder. The feature encoder is a functional module that converts numerical deviation data into standardized binary code. Its core function is to unify the data format, eliminate parsing interference caused by numerical differences, and improve the feature processing efficiency of AI models. The binary feature code is an 8-bit fixed-length encoding string composed of 0 and 1. The 8-bit length ensures sufficient encoding combinations to distinguish different deviation degrees, while also controlling storage usage and adapting to real-time processing requirements. The specific conversion process is as follows: First, the deviation percentage is mapped to a range, uniformly converting it to an integer range of 0 to 255 (for example, a negative deviation percentage of -6.7% is mapped to 248, and a positive deviation percentage of 4.0% is mapped to 5). This mapping rule is calibrated through historical deviation data statistics to ensure that deviations of different magnitudes correspond to unique integers. Then, the decimal-to-binary algorithm built into the feature encoder is used to convert the mapped integers into 8-bit binary numbers. If the number is less than 8 bits, 0s are padded at the high bits to form a complete binary feature code. For example, the binary feature code corresponding to the integer 5 is 00000101, and the binary feature code corresponding to the integer 248 is 11111000, achieving a standardized conversion of numerical information to binary encoding. After the conversion is completed, a corresponding alarm code is generated based on whether the deviation exceeds the limit. First, a warning threshold is clearly set as the critical standard for determining whether an alarm needs to be triggered. Based on a comprehensive calibration of historical abnormal data and the degree of impact on data quality, the same threshold is used for both positive and negative deviations (usually set to ±5%, that is, when the absolute value of the deviation percentage exceeds 5%, it is determined to be abnormal and an alarm is required). When a negative deviation exceeds a set warning threshold, a low-order alarm code is generated. The low-order alarm code is a variant of the binary feature code specifically for negative deviation exceeding the limit. Its encoding rule is that the least significant bit (the first bit) of the basic binary feature code is fixed as 1, and the remaining bits retain the original conversion result. This is used to clearly distinguish the negative deviation exceeding the limit. For example, if the negative deviation percentage of -7% (mapped to 247, binary 11110111) exceeds the warning threshold, the generated low-order alarm code is 11110111 (the least significant bit is already 1 and remains unchanged). If the negative deviation is -4% (not exceeding the threshold), the low-order alarm code is not triggered, the basic encoding remains unchanged, and it is only stored as a normal feature code. When a positive deviation exceeds a set warning threshold, a high-order alarm code is generated. The high-order alarm code is a variant of the binary feature code corresponding to the positive deviation exceeding the limit. The encoding rule is to fix the highest bit (8th bit) of the basic binary feature code as 1, and retain the original conversion result for the remaining bits, forming a clear type distinction with the low-order alarm code. For example, if the positive deviation percentage of 6% (mapped to 6, binary 00000110) exceeds the warning threshold, the generated high-order alarm code is 10000110 (the highest bit is changed from 0 to 1). If the positive deviation does not exceed the threshold, no high-order marking is performed, and the basic encoding remains unchanged.If the detected positive and negative deviations do not exceed the set warning threshold, meaning the current frame coupling coefficient is within the ideal range or slightly deviates, no alarm is needed. Instead, a zero-value code is directly output. The zero-value code is a fixed 8-bit all-zero binary code (00000000), specifically used to mark the absence of anomalies, ensuring the integrity and consistency of the feature code and preventing missing codes when there are no anomalies. After all feature codes (including low-order alarm codes, high-order alarm codes, and zero-value codes) are generated, a timestamp needs to be appended before writing them into the circular feature buffer. The timestamp is a high-precision time stamp generated synchronously with the feature code, accurate to the millisecond level, provided by a unified GPS timing module, used to record the occurrence time of abnormal features, ensuring accurate correspondence with the drone's flight time and scenario during subsequent tracing. The circular feature buffer is a data stream buffer using a first-in-first-out circular storage mechanism, specifically used to temporarily store recently generated feature code data, preventing data overflow, and ensuring that the AI-driven verification and governance unit can quickly read the latest feature information. The specific writing process is as follows: First, the feature code is combined with the corresponding millisecond-level timestamp to form a feature code-timestamp key-value pair. Then, this key-value pair is written into a circular feature buffer in the order of generation. The buffer has a preset storage capacity of 1000 frames of data. When new data is written and the storage capacity exceeds the limit, the oldest generated key-value pair is automatically deleted, always keeping the buffer data up-to-date. This provides real-time and complete anomaly feature data support for subsequent AI model feature extraction and multi-dimensional feature vector construction. The entire process uses a feature encoder to standardize the encoding of deviation information, distinguishing anomaly types with high and low bit alarm codes and marking normal states with zero-value codes. The timestamp is then appended and stored in a circular cache to ensure data traceability and real-time performance, providing standardized and resolvable anomaly feature input for intelligent verification of low-altitude data quality.

[0023] After the timestamp is added to the abnormal feature code of the coupling coefficient and written into the circular feature buffer, the AI-driven verification and governance unit needs to fuse the single abnormal feature with other key data quality indicators to construct a multi-dimensional feature vector that adapts to the input of the neural network model. The organic integration of multi-source information is achieved through feature tensor concatenation technology to ensure that the AI ​​model can comprehensively capture potential abnormal correlations in data quality. The specific implementation method is as follows: First, the core framework is built based on feature tensor concatenation technology. Feature tensor concatenation is a technique that combines multiple feature data of different dimensions and types into a unified tensor structure according to preset rules. It can preserve the original information of each feature while eliminating format differences, providing standardized input for subsequent AI model processing. The first step is to extract core feature data: first, the latest coupling coefficient anomaly feature code is extracted from the circular feature buffer. The circular feature buffer is a circular buffer that stores feature codes with timestamps. The latest specifically refers to the feature code with the latest timestamp in the buffer (i.e., the last written anomaly feature code, which may be a low-order alarm code, a high-order alarm code, or a zero-value code). During extraction, the data of the last storage unit is directly read through the pointer positioning function of the buffer to ensure the real-time and timeliness of the features. Simultaneously, GPS jump flags and image ambiguity flags are acquired in real time: The GPS jump flag is a binary identifier generated by the GPS receiver after real-time detection of the jump amplitude of the positioning data (0 indicates that the jump has not exceeded the preset threshold, and 1 indicates that it has exceeded the limit). It is output by the GPS module at a frequency of once per frame of data acquisition, accurately reflecting the stability of the positioning data; The image ambiguity flag is a binary identifier generated by the image sensor by analyzing the image sharpness (such as edge sharpness and grayscale variance) (0 indicates that the ambiguity has not exceeded the preset threshold, and 1 indicates that it has exceeded the limit). The sampling frequency is consistent with the coupling coefficient calculation frequency to ensure time synchronization with other features. The second step is to combine them into a three-dimensional feature tuple: the extracted latest coupling coefficient anomaly feature code, the real-time acquired GPS jump flag, and the image blur flag are combined in a fixed order to form a three-dimensional feature tuple. The three-dimensional feature tuple is a basic data structure containing three types of core quality indicators, with a fixed order of (coupling coefficient anomaly feature code, GPS jump flag, image blur flag), for example (10000110, 0, 1). This structured combination ensures that the correlation of multi-source features is not lost, providing a clear data dimension division for subsequent mapping processing. The third step is to realize the continuous vector mapping of discrete flags through a feature embedding layer. The feature embedding layer is a functional module adapted to the input requirements of neural network models. Its core function is to convert discrete binary flags and short-length binary feature codes into continuously distributed vectors, because neural networks have a stronger ability to learn features from continuous data.The specific mapping process is as follows: A pre-defined mapping rule library is used to set up specific mapping schemes for different types of discrete features. GPS transition flags (0 or 1) are mapped to continuous vectors of 16 dimensions (e.g., 0 maps to a vector with values ​​in the range of 0.1-0.3, and 1 maps to a vector with values ​​in the range of 0.7-0.9). Image blur flags use the same vector length as GPS transition flags, but with different mapping value ranges (e.g., 0 maps to the range of 0.2-0.4, and 1 maps to the range of 0.6-0.8) to avoid feature confusion. The coupling coefficient abnormality feature code (8-bit binary) is mapped bit by bit according to its 0 / 1 state and then concatenated to form the same 16-dimensional continuous vector, ensuring that the vector lengths after mapping the three types of features are consistent, laying the foundation for subsequent concatenation. The fourth step is to concatenate along the feature dimensions to form a fused feature vector. Concatenation along the feature dimensions involves joining the three mapped 16-dimensional continuous vectors end-to-end according to the original three-dimensional feature tuple order to form a fixed-length fused feature vector. For example, the coupling coefficient anomaly feature code mapping vector is V1, the GPS jump flag mapping vector is V2, and the image blur flag mapping vector is V3. After concatenation, a 48-dimensional fusion feature vector is obtained, which is arranged in sequence as V1[0]-V1

[15] , V2[0]-V2

[15] , and V3[0]-V3

[15] . The fixed length means that no matter what the discrete feature state of the input is, the dimension of the final output fusion feature vector is always consistent (such as 48 dimensions), ensuring that it can adapt to the input layer structure of the neural network model. The whole process ensures real-time performance by extracting the latest features, integrates multi-dimensional quality indicators by combining three-dimensional tuples, solves the adaptation problem between discrete data and AI models by feature embedding layers, and forms a standardized input format by dimension concatenation, providing comprehensive and standardized feature support for the identification of latent anomaly patterns.

[0024] After constructing the multidimensional feature vectors, the AI-driven verification and governance unit needs to use a neural network model specifically adapted to time-series data to identify latent patterns in the fused features where the single-source data does not exceed the limits but the coupling coefficient is abnormal. These patterns cannot be captured by traditional threshold judgment and require the feature extraction and classification capabilities of time-series convolutional networks to achieve accurate identification. The specific implementation method is as follows: A temporal convolutional network architecture is adopted. This architecture is a deep learning network structure specifically designed for temporal data. Its core advantage is that it can capture long-term temporal dependencies while maintaining computational efficiency, perfectly adapting to the temporal characteristics of low-altitude data generated frame by frame, and avoiding the temporal logic confusion problem that occurs when traditional convolutional networks process temporal data. First, the input layer receives and transmits data: The input layer is the entry layer of the neural network model. The number of neurons in it perfectly matches the dimension of the multi-dimensional feature vector (e.g., for a constructed 48-dimensional fused feature vector, the input layer has 48 neurons). Its role is to completely receive the standardized multi-dimensional feature vector and convert it into a tensor format that the network can process, without any feature loss or distortion, ensuring that the input data is completely adapted to the model's requirements, laying the foundation for subsequent feature extraction. After the input data is transmitted, temporal dependent features are first extracted through a causal convolutional layer. The causal convolutional layer is the core feature extraction layer of the temporal convolutional network. Its key characteristic is causality, that is, when calculating the features at the current moment, it only relies on the feature data of the past and current moments, without utilizing data from future moments, which fully conforms to the temporal logic of low-altitude data being generated in real time frame by frame. Time-dependent features refer to the correlation between features of different frames, such as the potential correlation between anomalies in the coupling coefficient of the current frame and GPS jump flags in the previous three frames. The specific extraction process is as follows: The causal convolutional layer uses a fixed-size time window (experimentally calibrated to 5 frames) to perform convolution operations on the input temporal feature vector in a sliding manner. Each time the window slides, the features from multiple frames within the window are fused into a local feature vector. Through multiple layers of such convolution operations, the temporal dependencies from short-term to long-term are gradually captured, such as from direct correlation between two adjacent frames to indirect correlation between frames with a gap of three frames, and finally outputting a feature map containing complete temporal correlation information. Next, the feature map enters the gated activation layer to filter noisy features. The gated activation layer is a filtering module composed of reset gates and update gates. Its core function is to distinguish between valid features and noisy features. Noisy features refer to meaningless features caused by slight disturbances in the flight environment, temporary fluctuations in sensors, etc. (such as a single ambiguity flag jump caused by instantaneous airflow). The specific filtering process is as follows: The gated activation layer performs two parallel transformations on the feature map output by the causal convolutional layer. First, it generates gate weights (values ​​from 0 to 1) using the sigmoid function to determine the validity of each feature. Second, it performs a non-linear transformation on the features using the tanh function to enhance their expressive power. Then, the two transformations are multiplied. Valid features with gate weights close to 1 are retained and amplified, while noise features close to 0 are suppressed and discarded. For example, isolated features generated by sensor erroneous triggers are marked as noise and filtered to ensure the purity of features in subsequent processing. The filtered valid features are then passed to the attention pooling layer to focus on key anomalous segments. The attention pooling layer is a module that highlights important feature segments by assigning attention weights. Key anomalous segments refer to feature combinations directly related to latent anomalies (such as feature segments with positive deviations in coupling coefficients for three consecutive frames and where GPS and image markers do not exceed limits). The specific focusing process is as follows: The attention pooling layer first calculates the importance score of each feature fragment, based on the similarity between the feature and historical anomalous samples; higher similarity results in a higher score. The scores are then normalized into attention weights, and the feature fragments are weighted and summed accordingly. Key anomalous fragments with higher weights have a higher weight in the final feature set. For example, a feature fragment with 80% similarity to a historical latent anomalous sample is assigned a high weight of 0.8, while ordinary fragments are assigned a low weight of only 0.2. This method allows the model to focus on core anomalous information and avoid interference from irrelevant features. Finally, the feature vector processed by attention pooling is passed to the output layer, which is connected to a sigmoid classifier. The sigmoid classifier maps feature vectors to probability values ​​between 0 and 1. Its core function is to transform abstract features into intuitive anomaly indicators; the closer the output value is to 1, the higher the probability of a latent anomaly; the closer it is to 0, the higher the probability of normal data. The decision boundary threshold is the critical value that distinguishes between normal and abnormal data. After training with a large number of labeled samples, it was calibrated to 0.7. This threshold balances the false positive and false negative rates, ensuring accurate anomaly identification while reducing unnecessary alarms. When the probability value output by the sigmoid classifier exceeds this decision boundary threshold, the model directly determines that there is a latent anomaly pattern with an excessively high coupling coefficient deviation. This means that while the individual source data (GPS, imagery) does not exceed the threshold, the abnormal coupling coefficient constitutes a potential data quality risk, providing the core basis for subsequent latent anomaly alarm signals. The entire process adapts to the temporal characteristics of low-altitude data through a temporal convolutional network architecture. Causal convolution captures temporal correlations, gating activation filters noise, and attention pooling focuses on the core data, ultimately achieving accurate judgment through the sigmoid classifier.

[0025] After the neural network model determines that there is a latent anomaly pattern with excessive coupling coefficient deviation through the temporal convolutional network architecture, it does not directly output an alarm signal. In order to avoid misjudging the explicit anomaly caused by the single-source data exceeding the limit as a latent anomaly, it is necessary to ensure the accuracy of the alarm through triple conditional synchronous verification. Then, the automatic calibration module is driven to perform three-stage governance in sequence to compensate for data errors from the root cause. The specific implementation method is as follows: The core of the latent anomaly alarm signal output is the simultaneous fulfillment of three conditions. First, the meaning and verification logic of each condition are clarified: The GPS jump threshold verification flag is a binary identifier (0 for not exceeding the limit, 1 for exceeding the limit) generated by the GPS receiver after real-time monitoring of the jump amplitude of the positioning data. Not exceeding the limit means that the jump amplitude of the bit data does not exceed the preset threshold (0.5 meters per frame, calibrated experimentally), indicating that the GPS positioning data itself is stable and there is no obvious anomaly; The image blur threshold verification flag is a binary identifier (0 for not exceeding the limit, 1 for exceeding the limit) generated by the image sensor after analyzing the image sharpness. Not exceeding the limit means that the image blur does not exceed the preset threshold (quantified by edge sharpness, the threshold is 0.8), indicating that the image data itself is of acceptable quality; The neural network model output anomaly judgment signal is the judgment signal generated when the model output value exceeds the decision boundary threshold (0.7), indicating that there is a latent problem where the single source data is normal but the coupling coefficient is abnormal. The specific verification process is as follows: The built-in condition checker synchronously reads the current status of the GPS jump threshold check flag and the image blur threshold check flag in real time, while also receiving the output signal of the neural network model. Only when all three signals simultaneously meet the conditions that the first two are 0 (not exceeding limits) and the latter is an anomaly judgment signal, can it be determined that it meets the characteristics of a latent anomaly and immediately generate a latent anomaly alarm signal. This signal is a standardized instruction containing the timestamp of the anomaly occurrence, the deviation value of the coupling coefficient, and the current flight parameters, ensuring that the automatic calibration module can accurately identify the treatment scenario. If any condition is not met (such as the GPS jump flag being 1), it is determined to be an explicit anomaly or normal data, and no alarm signal is generated. Instead, the corresponding explicit anomaly handling process is triggered or the normal acquisition state is maintained. After the alarm signal is generated, the automatic calibration module immediately starts the three-stage treatment. The automatic calibration module is the core execution unit that integrates parameter adjustment, equipment control, and algorithm compensation functions. The three-stage treatment follows a logical progression of asynchronous time compensation, positioning noise filtering, and historical deviation correction to ensure comprehensive error coverage. The first stage involves adjusting the image sensor exposure frequency to compensate for asynchronous time errors. Asynchronous time errors are microsecond-level differences that may still exist between GPS positioning and image acquisition, even though they are synchronized via timestamps. These differences can lead to deviations in the coupling coefficient calculation. The image sensor exposure frequency is the number of times the sensor captures images per second. The adjustment process is as follows: The module calculates the required time difference to be compensated based on the coupling coefficient deviation value in the alarm signal (the larger the deviation, the larger the time difference may be). It dynamically adjusts the frequency according to the rule that the exposure frequency increases by 1 frame / second for every 0.1 millisecond increase in time difference. For example, when the deviation value is 0.3, the exposure frequency is increased from the default 30 frames / second to 33 frames / second. This increases the sampling density to offset the errors caused by asynchronous time, ensuring that the subsequent acquired images are perfectly matched with the GPS data timing. The second stage involves resetting the GPS receiver signal filtering parameters. The GPS receiver signal filtering parameters are configuration parameters used to filter satellite signal noise and stabilize positioning data (such as filter window size and smoothing coefficient). Long-term use may cause parameter drift due to environmental interference, indirectly affecting the coupling coefficient. The resetting process is as follows: The module calls a preset set of standard filtering parameters (based on optimal parameter calibration during stable flight) to cover the parameters of the current drift. Specifically, this includes resetting the filter window size to 5 frames, resetting the smoothing coefficient to 0.6, disabling temporary interference suppression, restoring basic filtering logic, ensuring the stability of GPS positioning errors, and reducing the impact of meaningless fluctuations on the coupling coefficient. The third stage injects calibration parameters into the data fusion algorithm to compensate for historical deviations: calibration parameters are error compensation values ​​obtained through statistical analysis of historical stable flight data (such as the coupling coefficient correction for different flight altitudes and speeds). The data fusion algorithm is the core algorithm for integrating GPS and image data to calculate the coupling coefficient. Historical deviations refer to errors accumulated during long-term data acquisition (such as inherent sensor offset). The specific process is as follows: Based on the current flight altitude and speed parameters, the module matches the corresponding compensation value from the preset calibration parameter library and writes it into the compensation interface of the data fusion algorithm in the form of parameter injection instructions. After receiving the compensation value, the algorithm automatically adds it to the coupling coefficient calculation process. For example, if the current flight altitude is 200 meters and the calibration parameter is 0.05, the algorithm will add 0.05 to the original calculation result to offset historical deviations and finally output a more accurate coupling coefficient.

[0026] The present invention's dynamic coupling coefficient calculation unit synchronizes GPS positioning error and image pixel offset error, calculates the coupling coefficient for each frame using an error coupling algorithm, and ensures data continuity through error compensation and interpolation. The ideal interval comparison unit presets a dynamic ideal interval based on stable flight data, calculates the deviation percentage using a two-way difference ratio method, and marks positive and negative abnormal features. The AI-driven verification and governance unit integrates abnormal features with GPS jumps and image blur thresholds to construct a multi-dimensional feature vector. A temporal convolutional network identifies latent patterns in single-source data that are not out of bounds but have abnormal coupling. After meeting three conditions, an alarm is output, driving the automatic calibration module to adjust the exposure frequency, reset GPS filter parameters, and inject calibration parameters. This three-stage governance of data quality ensures the accuracy of low-altitude data applications.

[0027] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-driven intelligent verification and governance system for low-altitude data quality, characterized in that: include: The dynamic coupling coefficient calculation unit (2) synchronously acquires the GPS positioning error and image pixel offset error output by the data acquisition and processing unit (1) frame by frame, and calculates the coupling coefficient of each frame in real time by dividing the image pixel offset error by the GPS positioning error through the error coupling algorithm. The ideal interval comparison unit (3) presets the ideal interval of the coupling coefficient during the stable flight phase. It reads the coupling coefficient of the current frame from the coupling coefficient of each frame output by the dynamic coupling coefficient calculation unit (2), calculates the percentage deviation of the coupling coefficient of the current frame from the lower and upper limits of the ideal interval. If the coupling coefficient of the current frame is lower than the lower limit of the ideal interval, it calculates the negative deviation. If the coupling coefficient of the current frame is higher than the upper limit of the ideal interval, it calculates the positive deviation. The negative and positive deviation values ​​are marked as abnormal characteristics of the coupling coefficient in real time. The AI-driven verification and governance unit (4) integrates the abnormal features of the coupling coefficient as a new verification input into the AI ​​model. In the AI ​​verification feature construction stage of the AI ​​model, the deviation percentage is fused with the GPS jump threshold and the image blur threshold to form a multi-dimensional feature vector. The multi-dimensional feature vector is then input into the trained neural network model. The neural network model identifies the latent abnormal pattern where the deviation of the coupling coefficient exceeds the preset threshold through supervised learning. When it is identified that the single source data does not exceed the threshold but the coupling coefficient deviates from the lower limit and upper limit of the ideal range to the preset threshold, a latent abnormal alarm signal is output to drive the automatic calibration module to adjust the data acquisition parameters to govern the data quality.

2. The AI-driven intelligent verification and governance system for low-altitude data quality according to claim 1, characterized in that: When the error coupling algorithm is executed in the dynamic coupling coefficient calculation unit, the time stamp synchronization mechanism ensures that the GPS positioning error and the image pixel offset error are collected at the same time. The floating-point divider is used to divide the two error values ​​to generate the coupling coefficient. When the GPS positioning error is lower than the preset minimum effective value, the error compensation module is automatically activated to inject the reference compensation amount. At the same time, a frame-level buffer queue is established to temporarily store the original error data of 5 consecutive frames. If an abnormal interruption of single frame data is detected, the adjacent frame interpolation algorithm is called to fill in the missing value.

3. The AI-driven intelligent verification and governance system for low-altitude data quality according to claim 1, characterized in that: The ideal range of coupling coefficients preset in the ideal range comparison unit (3) is obtained through training with historical stable flight phase data, specifically as follows: When the rate of change of attitude angle of the UAV is continuously lower than the set angular velocity threshold and the continuous flight time exceeds the set time window, extract all coupling coefficient data within the period when the continuous flight time exceeds the set time window, remove the maximum and minimum values, and take the numerical range of the dense area in the statistical distribution as the ideal interval for dynamic update. Before reading the coupling coefficient of the current frame, it is necessary to verify the data frame status flag bit output by the dynamic coupling coefficient calculation unit (2). If the flag bit is abnormal, skip the calculation of the current frame.

4. The AI-driven intelligent verification and governance system for low-altitude data quality according to claim 3, characterized in that: The percentage deviation is calculated using the two-way difference ratio method, specifically: The difference between the current frame coupling coefficient and the lower limit of the ideal interval is divided by the lower limit of the ideal interval to obtain the negative deviation from the base value, and the difference between the current frame coupling coefficient and the upper limit of the ideal interval is divided by the upper limit of the ideal interval to obtain the positive deviation from the base value. Then, the base value is converted into a percentage form by the normalization processor. During the conversion process, the interval width weighting factor is introduced so that the deviation percentage can reflect the relative position of the current value in the ideal interval.

5. The AI-driven intelligent verification and governance system for low-altitude data quality according to claim 3, characterized in that: The percentage deviation is calculated using the two-way difference ratio method, specifically: The difference between the current frame coupling coefficient and the lower limit of the ideal interval is divided by the lower limit of the ideal interval to obtain the negative deviation from the base value, and the difference between the current frame coupling coefficient and the upper limit of the ideal interval is divided by the upper limit of the ideal interval to obtain the positive deviation from the base value. Then, the base value is converted into a percentage form by the normalization processor. During the conversion process, the interval width weighting factor is introduced so that the deviation percentage can reflect the relative position of the current value in the ideal interval.

6. The AI-driven intelligent verification and governance system for low-altitude data quality according to claim 3, characterized in that: The positive deviation calculation is activated when the coupling coefficient of the current frame is detected to be higher than the upper limit of the ideal range. The absolute difference between the current value and the upper limit of the ideal range is calculated and divided by the upper limit of the ideal range to obtain the basic positive ratio. The basic positive ratio is input to a gain regulator that is different from the negative deviation. The gain regulator dynamically adjusts the amplification coefficient based on the UAV flight altitude parameters.

7. The AI-driven intelligent verification and governance system for low-altitude data quality according to claim 3, characterized in that: The abnormal feature marking of the coupling coefficient is implemented through a feature encoder, which converts the deviation percentage into a binary feature code. When the negative deviation exceeds the set warning threshold, a low-order alarm code is generated, and when the positive deviation exceeds the warning threshold, a high-order alarm code is generated. If neither deviation exceeds the limit, a zero value code is output. All feature codes are appended with a timestamp and written to a circular feature buffer.

8. The AI-driven intelligent verification and governance system for low-altitude data quality according to claim 3, characterized in that: The multidimensional feature vector construction adopts feature tensor splicing technology. The latest coupling coefficient abnormal feature code is extracted from the circular feature buffer and combined with the real-time collected GPS jump flag and image blur flag to form a three-dimensional feature tuple. The discrete flags are mapped into continuous vectors through the feature embedding layer, and then spliced ​​along the feature dimension to form a fixed-length fused feature vector.

9. The AI-driven intelligent verification and governance system for low-altitude data quality according to claim 1, characterized in that: The neural network model adopts a temporal convolutional network architecture. After receiving multi-dimensional feature vectors, the input layer sequentially extracts time-dependent features through a causal convolutional layer, filters noise features through a gated activation layer, and focuses key abnormal segments through an attention pooling layer. The output layer is connected to a sigmoid classifier. When the output value of the neural network model exceeds the decision boundary threshold, it is determined that there is a latent abnormal pattern with an excessively high coupling coefficient deviation.

10. The AI-driven intelligent verification and governance system for low-altitude data quality according to claim 1, characterized in that: The output of the latent anomaly alarm signal must simultaneously meet three conditions: the GPS jump threshold verification flag is not exceeded, the image blur threshold verification flag is not exceeded, and the neural network model outputs an anomaly judgment signal. After the alarm signal is generated, the automatic calibration module is driven to perform three-stage governance. First, the exposure frequency of the image sensor is adjusted to compensate for the asynchronous error of the time. Second, the signal filtering parameters of the GPS receiver are reset. Finally, the calibration parameters are injected into the data fusion algorithm to compensate for historical deviations.

Images