A method for on-site quality assessment of gravity measurement data

Abstract

The present application provides a kind of gravity measurement data field quality evaluation method, belong to gravity quality control technical field, the present application passes through establishing initial instrument parameter database and environmental parameter monitoring system, uses Kalman filter drift monitoring model real-time tracking instrument state change, uses drift correction equation set and time-frequency domain filtering algorithm to original gravity data real-time pre-processing and error compensation, constructs the quality evaluation system of quantization data quality grade based on multidimensional index, uses the abnormality detection model of the architecture of Transform to identify abnormal data and generate quality warning information, finally forms the field quality evaluation report containing data quality distribution, abnormal position and processing suggestion, realizes the real-time quality monitoring and accurate abnormality detection of gravity measurement data, solves the technical problems of gravity measurement data quality evaluation lag and insufficient accuracy.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of gravity quality control, and particularly relates to a method for evaluating the quality of gravity measurement data on site. BACKGROUND

[0002] As an important means of geophysical exploration, gravity measurement is widely used in the fields of oil and gas exploration, mineral resource investigation, engineering geological survey and geological structure research. Traditional gravity measurement data quality control mainly relies on post-processing and manual experience judgment. The reliability of data is evaluated by comparing the repeated measurement values of reference points and statistically analyzing the measurement errors. In the prior art, the gravity instrument will produce zero drift and scale factor changes in the long-term working process, and environmental factors such as temperature changes, vibration interference and air pressure fluctuations will introduce systematic errors, resulting in a large amount of noise and abnormal values in the measurement data. The traditional quality control method cannot realize real-time monitoring and accurate evaluation. That is, the prior art has the technical problems of lagging and insufficient accuracy in evaluating the quality of gravity measurement data. SUMMARY

[0003] Therefore, the application provides a method for evaluating the quality of gravity measurement data on site, which can solve the technical problems of lagging and insufficient accuracy in evaluating the quality of gravity measurement data in the prior art.

[0004] The application is implemented in the following manner. The application provides a method for evaluating the quality of gravity measurement data on site, which comprises the following steps: performing initial calibration on a gravity instrument to obtain a zero offset and a scale factor, and establishing an initial instrument parameter database; arranging environmental parameter monitoring sensors in a measurement area to collect temperature, humidity, air pressure and vibration acceleration parameter data in real time, and forming an environmental parameter time sequence; starting a drift monitoring model simultaneously when starting a gravity measurement operation, and calculating the zero drift and scale factor drift of the gravity instrument in real time by measuring the gravity value changes and environmental parameter changes of the reference points; performing real-time preprocessing on the collected original gravity measurement data, eliminating the systematic errors of the instrument by using a drift compensation function, calculating accurate drift compensation values by using a drift correction equation set, and removing high-frequency noise interference by using a time-frequency domain filtering algorithm; constructing a gravity data quality evaluation index system to calculate the data accuracy, consistency, stability and reliability indexes of each measurement point, quantifying the data quality grade by using a multi-dimensional quality comprehensive evaluation model; identifying abnormal data points by using a gravity anomaly detection model to preprocess the gravity data, marking the abnormal data points that exceed the quality threshold, and generating quality warning information; and generating a report on the quality evaluation of gravity measurement data on site.

[0005] The drift correction equation set specifically includes a zero drift correction equation and a scale factor drift correction equation; the zero drift correction equation is used to calculate the time variation characteristics and environmental response characteristics of the gravity meter zero offset, and the input includes a reference point gravity value sequence, an environmental parameter time sequence, a measurement time interval, instrument initial calibration parameters and a zero offset, and the output is a zero drift correction coefficient; the scale factor drift correction equation is used to calculate the nonlinear variation characteristics and temperature response characteristics of the gravity meter scale factor, and the input includes a scale factor drift, a temperature variation gradient, a measurement time interval, instrument initial calibration parameters and an environmental parameter time sequence, and the output is a scale factor drift correction coefficient.

[0006] The drift compensation function specifically is used to eliminate the zero offset and scale factor drift error of the gravity meter during long-term operation, and the input includes a reference point gravity value sequence, an environmental parameter time sequence, a measurement time interval and instrument initial calibration parameters, and the output is a gravity measurement value after drift compensation.

[0007] The gravity anomaly detection model specifically is a sequence analysis model based on a Transformer architecture, and contains a multi-head attention mechanism for capturing the spatio-temporal correlation of gravity data, wherein the number of attention heads is dynamically determined according to three parameters of measurement area complexity, data sampling density and environmental noise level, and the number of attention heads is increased to improve the anomaly detection precision when the measurement area has complex geological structure, high data sampling density and low environmental noise level.

[0008] The training data set establishment step of the gravity anomaly detection model specifically collects gravity measurement data under different geological environments as normal samples, artificially injects various types of instrument faults, environmental disturbances and measurement errors to generate abnormal samples, labels and classifies all samples to form a training data set containing normal data and abnormal data, and expands the number of training samples through data enhancement technology to improve the model generalization ability.

[0009] The environmental parameter monitoring sensor refers to a multi-type sensor array installed in the measurement area, which is used to monitor the changes of environmental factors affecting the gravity measurement precision in real time.

[0010] The drift monitoring model refers to a state estimation model based on a Kalman filtering algorithm, which is used to estimate and predict the drift state of the gravity meter in real time.

[0011] The multi-dimensional quality comprehensive evaluation model refers to a comprehensive evaluation algorithm that fuses multiple quality indicators, and outputs a data quality grade division result.

[0012] The attention head number adjustment function is specifically used for adjusting the multi-head attention mechanism parameters of the gravity anomaly detection model, and is calculated based on the measurement area complexity, data sampling density, and environmental noise level to obtain an attention adjustment value. When the attention adjustment value is in the range of 0 to 0.25, four attention heads are adopted, and the attention range of each attention head is 256 time steps. When the attention adjustment value is in the range of 0.25 to 0.5, six attention heads are adopted, and the attention range of each attention head is 128 time steps. When the attention adjustment value is in the range of 0.5 to 0.75, eight attention heads are adopted, and the attention range of each attention head is 64 time steps. When the attention adjustment value is in the range of 0.75 to 1, twelve attention heads are adopted, and the attention range of each attention head is 32 time steps. The multi-head attention mechanism parameters of the model are adjusted.

[0013] The attention weight threshold of each attention head is dynamically set according to the variance of the gravity data sequence. When the variance of the gravity data sequence is less than 0.01 mGal, the attention weight threshold is set to 0.8. When the variance of the gravity data sequence is between 0.01 and 0.05 mGal, the attention weight threshold is set to 0.6. When the variance of the gravity data sequence is greater than 0.05 mGal, the attention weight threshold is set to 0.4. The attention head adopts a sliding window attention method to process the gravity data time sequence and filters important features according to the attention weight threshold.

[0014] The gravity anomaly detection model training step is specifically a supervised learning method for training model parameters. A cross-entropy loss function is used to measure the difference between the model prediction results and the true labels. The model weight parameters are optimized through a back propagation algorithm. The model performance is evaluated using a validation set, and the hyperparameters are adjusted. Finally, a trained model for gravity data anomaly detection is obtained.

[0015] The time-frequency domain filtering algorithm refers to a composite filtering method combining low-pass filtering and band-stop filtering, which is used to remove high-frequency noise and frequency band interference in gravity data.

[0016] The gravity measurement data includes three main categories: normal gravity data, boundary gravity data, and abnormal gravity data. The normal gravity data is obtained by statistical analysis of the distribution of measurement values, accounting for 70% to 80% of the total data volume. The boundary gravity data is obtained by setting a quality threshold range, accounting for 15% to 20% of the total data volume. The abnormal gravity data is obtained by the gravity anomaly detection model, accounting for 5% to 10% of the total data volume.

[0017] Wherein, the proportion change of different classification data will directly affect the accuracy and reliability of the final gravity field interpretation, when the proportion of abnormal gravity data exceeds 10%, the overall quality of gravity measurement will be significantly reduced, through real-time quality monitoring and dynamic data screening mechanism to reduce the negative impact of abnormal gravity data on the final result, while adopting multiple verification and cross-validation method to improve the transition rate and delay rate control accuracy of boundary gravity data.

[0018] Wherein, the zero drift correction coefficient refers to the correction parameter for compensating the time variation of the gravity meter zero offset; the scale factor drift correction coefficient refers to the correction parameter for compensating the nonlinear variation of the gravity meter scale factor drift; the temperature variation gradient refers to the change rate parameter of the environmental temperature in the measurement time interval.

[0019] Wherein, the gravity measurement data field quality evaluation report contains data quality level distribution, abnormal data position, recommended treatment measures and measurement quality comprehensive evaluation result.

[0020] The present application realizes the real-time quality evaluation and accurate anomaly detection of gravity measurement data in the field by establishing a drift monitoring model and a gravity anomaly detection model, combining real-time monitoring of environmental parameters and a multi-dimensional quality evaluation index system. The present application uses a drift monitoring model based on Kalman filtering to track the state change of the instrument in real time, uses an anomaly detection model based on the Transformer architecture to accurately identify data anomalies, and effectively eliminates systematic errors and noise interference through a drift correction equation set and a time-frequency domain filtering algorithm, significantly improving the timeliness and accuracy of data quality evaluation. In summary, the present application solves the technical problems of lagging and insufficient accuracy of gravity measurement data quality evaluation mentioned in the background art. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The flowchart of the method of the present application. DETAILED DESCRIPTION

[0022] To make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0023] As Figure 1 shown, is a flowchart of a gravity measurement data field quality evaluation method provided by the present application, the method comprises the following steps:

[0024] S01, before the gravity measurement starts, the gravity meter is initially calibrated, the gravity meter zero offset and the scale factor are obtained by measuring the reference point with known gravity value, and an initial instrument parameter database is established;

[0025] S02, uniformly arranging environmental parameter monitoring sensors in the measurement area, collecting temperature, humidity, air pressure, vibration acceleration parameter data in real time, and forming an environmental parameter time sequence;

[0026] S03, starting the drift monitoring model when starting the gravity measurement operation, calculating the gravity meter zero drift and scale factor drift in real time by repeatedly measuring the gravity value change and environmental parameter change of the reference point;

[0027] S04, real-time preprocessing of the collected original gravity measurement data, using a drift compensation function to eliminate systematic errors of the instrument, simultaneously using a drift correction equation set to calculate accurate drift compensation values, and using time-frequency domain filtering algorithm to remove high-frequency noise interference;

[0028] S05, constructing a gravity data quality evaluation index system, calculating the data accuracy, consistency, stability and reliability index of each measurement point, and quantifying the data quality grade through a multi-dimensional quality comprehensive evaluation model;

[0029] S06, using a gravity anomaly detection model to identify abnormal values of the preprocessed gravity data, marking abnormal data points exceeding the quality threshold and generating quality warning information;

[0030] S07, generating a gravity measurement data field quality evaluation report, including data quality grade distribution, abnormal data position, recommended treatment measures and measurement quality comprehensive evaluation results.

[0031] The drift correction equation set includes a zero drift correction equation and a scale factor drift correction equation; the zero drift correction equation is used to calculate the time variation characteristics and environmental response characteristics of the gravity meter zero offset, and the input includes the reference point gravity value sequence, the environmental parameter time sequence, the measurement time interval, the instrument initial calibration parameter and the zero offset, and the output is the zero drift correction coefficient; the scale factor drift correction equation is used to calculate the nonlinear variation characteristics and temperature response characteristics of the gravity meter scale factor, and the input includes the scale factor drift, the temperature change gradient, the measurement time interval, the instrument initial calibration parameter and the environmental parameter time sequence, and the output is the scale factor drift correction coefficient.

[0032] The drift compensation function is used to eliminate the zero offset and scale factor drift error of the gravimeter during long-term operation. The input includes the reference point gravity value sequence, the environmental parameter time sequence, the measurement time interval, and the initial calibration parameters of the instrument. The output is the gravity measurement value after drift compensation. The specific structure of the gravity anomaly detection model is a sequence analysis model based on the Transformer architecture, which contains a multi-head attention mechanism to capture the spatio-temporal correlation of gravity data. The number of attention heads is dynamically determined according to three parameters: measurement area complexity, data sampling density, and environmental noise level. When the measurement area has complex geological structure, high data sampling density, and low environmental noise level, the number of attention heads is increased to improve the accuracy of anomaly detection. The training data set establishment step of the gravity anomaly detection model specifically includes collecting gravity measurement data under different geological environments as normal samples, artificially injecting various types of instrument faults, environmental disturbances, and measurement errors to generate abnormal samples, labeling and classifying all samples to form a training data set containing normal and abnormal data, and expanding the number of training samples through data augmentation techniques to improve the model's generalization ability. The gravity anomaly detection model training step specifically includes training model parameters using a supervised learning method, using a cross-entropy loss function to measure the difference between the model's predicted results and the true labels, optimizing the model's weight parameters through a backpropagation algorithm, evaluating the model's performance using a validation set and adjusting the hyperparameters, and finally obtaining a trained model for gravity data anomaly detection.

[0033] The environmental parameter monitoring sensor is a multi-type sensor array installed in the measurement area, which is used to monitor the changes of environmental factors that affect the accuracy of gravity measurement in real time. The drift monitoring model is a state estimation model based on the Kalman filter algorithm, which is used to estimate and predict the drift state of the gravimeter in real time. The time-frequency domain filtering algorithm is a composite filtering method combining low-pass filtering and band-stop filtering, which is used to remove high-frequency noise and frequency band interference in gravity data. The multi-dimensional quality comprehensive evaluation model is a comprehensive evaluation algorithm that integrates multiple quality indicators, which outputs the data quality level division result. The zero drift correction coefficient is a correction parameter used to compensate for the time variation of the zero offset of the gravimeter. The scale factor drift correction coefficient is a correction parameter used to compensate for the nonlinear variation of the scale factor drift of the gravimeter. The temperature change gradient is the rate parameter of the change of environmental temperature in the measurement time interval.

[0034] The attention head number adjustment function is used to adjust the multi-head attention mechanism parameters of the gravity anomaly detection model. The attention adjustment value is calculated based on the complexity of the measurement area, the data sampling density, and the environmental noise level. When the attention adjustment value is in the range of 0 to 0.25, 4 attention heads are used, and the attention range of each attention head is 256 time steps. When the attention adjustment value is in the range of 0.25 to 0.5, 6 attention heads are used, and the attention range of each attention head is 128 time steps. When the attention adjustment value is in the range of 0.5 to 0.75, 8 attention heads are used, and the attention range of each attention head is 64 time steps. When the attention adjustment value is in the range of 0.75 to 1, 12 attention heads are used, and the attention range of each attention head is 32 time steps. The multi-head attention mechanism parameters of the model are adjusted. The attention weight threshold of each attention head is dynamically set according to the variance of the gravity data sequence. When the variance of the gravity data sequence is less than 0.01 mGal, the attention weight threshold is set to 0.8. When the variance of the gravity data sequence is between 0.01 and 0.05 mGal, the attention weight threshold is set to 0.6. When the variance of the gravity data sequence is greater than 0.05 mGal, the attention weight threshold is set to 0.4. The attention head uses a sliding window attention method to process the gravity data time sequence and filters important features according to the attention weight threshold.

[0035] Gravity measurement data includes three main categories: normal gravity data, boundary gravity data, and abnormal gravity data. Normal gravity data is obtained by statistical analysis of measurement value distribution, accounting for 70% to 80% of the total data volume, and has a stable positive impact on measurement results. Boundary gravity data is obtained by setting a quality threshold range, accounting for 15% to 20% of the total data volume, and its contribution rate fluctuates with environmental conditions. Abnormal gravity data is obtained by a gravity anomaly detection model, accounting for 5% to 10% of the total data volume, and has a negative impact on measurement accuracy. The proportion of different classification data directly affects the accuracy and reliability of the final gravity field interpretation. When the proportion of abnormal gravity data exceeds 10%, the overall quality of gravity measurement will be significantly reduced. Real-time quality monitoring and dynamic data filtering mechanisms are used to reduce the negative impact of abnormal gravity data on the final results. Multiple verification and cross-validation methods are used to improve the transition rate and delay rate control accuracy of boundary gravity data.

[0036] The specific implementation of step S01 is to establish an instrument basic parameter database through a precisely controlled initial calibration program. First, select at least 3 national or regional reference points with known gravity values. The accuracy of the reference point gravity value should reach m / The distance between the reference points should be kept above 1000 m to ensure spatial representativeness. The gravity value is measured continuously at each reference point for not less than 10 times, the measurement interval is set to 5 min, and the least square method is used to fit the linear relationship between the gravimeter reading and the true gravity value. The zero offset is calculated by regression analysis, which reflects the systematic deviation of the instrument, and the scale factor is determined by the slope coefficient, which reflects the sensitivity characteristics of the instrument. The initial instrument parameter database established includes zero offset, scale factor, linearity error, repeatability error and other key parameters, which provides a reference for subsequent drift correction. The purpose of this step is to eliminate the inherent systematic error of the instrument, establish an accurate measurement reference, and ensure the reliability and traceability of subsequent gravity measurement data.

[0037] The specific implementation of step S02 is to construct a multi-parameter environmental monitoring network to realize real-time environmental data acquisition. The environmental parameter monitoring sensor array is uniformly deployed in the measurement area according to the grid layout principle, the sensor spacing is set to 200 m to 500 m, and the adjustment is made according to the complexity of the terrain of the measurement area. The temperature sensor uses a platinum resistance thermometer, the measurement accuracy should reach 0.01℃, and the sampling frequency is set to 1 Hz. The humidity sensor uses a capacitive humidity meter, the measurement accuracy should reach 0.5% relative humidity, and the sampling frequency is set to 0.5 Hz. The barometric pressure sensor uses a piezoresistive barometer, the measurement accuracy should reach 0.1 hPa, and the sampling frequency is set to 0.2 Hz. The vibration acceleration sensor uses a piezoelectric accelerometer, the measurement accuracy should reach m / , and the sampling frequency is set to 100 Hz to capture high-frequency vibration signals. All sensor data are collected in real time to the data acquisition center through a wireless transmission network to form a time-synchronized environmental parameter time series. The purpose of this step is to comprehensively monitor the changes of environmental factors that affect the accuracy of gravity measurement, and to provide environmental background information for drift correction and anomaly detection.

[0038] The specific implementation of step S03 is to construct a dynamic drift monitoring model based on Kalman filtering algorithm to realize real-time drift estimation. The drift monitoring model describes the drift behavior of the gravimeter using state space equations, and the state variables include the zero drift, the scale factor drift, and their change rates. The model prediction equation is established according to the physical characteristics of the instrument, and the observation equation is established by the reference point repeated measurement data. The Kalman filtering process includes a prediction step and an update step. The prediction step predicts the drift state at the current time according to the state at the last time and the system model, and the update step corrects the prediction result using the reference point measurement value. The reference point repeated measurement frequency is set to once every 30 minutes, and the average value of 5 consecutive measurements is taken each time. The process noise covariance matrix is set according to the instrument technical index, and the observation noise covariance matrix is set according to the reference point measurement accuracy. The optimal estimation value and its uncertainty of the zero drift and the scale factor drift are obtained by recursive calculation. The purpose of this step is to track the change of the instrument drift state in real time and provide accurate correction parameters for drift compensation.

[0039] The specific implementation of step S04 is to use a multi-level data processing algorithm to eliminate systematic errors and random noise. The drift compensation function is based on the drift parameters obtained in step S03, and the linear interpolation method is used to calculate the zero offset compensation value and the scale factor compensation value at each measurement point. The zero drift correction equation input includes the reference point gravity value sequence, the environmental parameter time sequence, the measurement time interval, and the instrument initial calibration parameter. The relationship model between the zero drift and the time and environmental parameters is established by multiple regression analysis, and the zero drift correction coefficient is output. The scale factor drift correction equation mainly considers the temperature response characteristics, and establishes a nonlinear relationship model between the scale factor and the temperature gradient, and outputs the scale factor drift correction coefficient. The time-frequency domain filtering algorithm first uses a Butterworth low-pass filter to remove high-frequency noise with a frequency higher than 0.1 Hz, and the cutoff frequency is set according to the spectral characteristics of the gravity signal. Then a notch filter is used to remove 50 Hz power frequency interference and its harmonic components. The filter parameters are determined by spectral analysis and signal-to-noise ratio optimization to ensure that the noise is removed while the integrity of the effective gravity signal is maintained. The purpose of this step is to obtain high-quality gravity measurement data to provide a reliable data basis for quality evaluation.

[0040] The specific implementation of step S05 is to construct a multi-dimensional quality evaluation system to quantify the data quality level. The data precision index is calculated by the standard deviation of repeated measurements, and the precision threshold is set to 0.02 mGal. Data exceeding this threshold is marked as low-precision data. The consistency index is evaluated by the gradient change of adjacent measurement points, and data with a gradient change exceeding 0.05 mGal / m is marked as consistency abnormal. The stability index is evaluated by time series variance analysis, and data with a variance exceeding 0.01 Data with a difference of more than 0.03 mGal between multiple independent measurements are marked as unreliable data. The multi-dimensional quality comprehensive evaluation model adopts a weighted comprehensive evaluation method, and the weights of each index are determined by the analytic hierarchy process. The precision index weight is 0.4, the consistency index weight is 0.3, the stability index weight is 0.2, and the reliability index weight is 0.1. After the comprehensive quality score is calculated, it is divided into four levels of excellent, good, qualified and unqualified according to the score thresholds of 90, 80 and 70. The purpose of this step is to comprehensively evaluate the quality level of gravity data and provide quantitative basis for data screening and anomaly identification.

[0041] The specific implementation of step S06 is to identify abnormal data by using a gravity anomaly detection model based on the Transformer architecture. The anomaly detection model adopts an encoder-decoder structure, the encoder part contains a multi-head attention mechanism and a feedforward neural network, and the decoder part outputs an anomaly probability. The multi-head attention mechanism is dynamically adjusted by an attention head number adjustment function, the measurement area complexity is quantified by the geological structure density and the terrain relief, the data sampling density is calculated by the number of measurement points per unit area, and the environmental noise level is evaluated by the standard deviation of environmental parameters. After normalization of the three parameters, the attention adjustment value is calculated, and the number of attention heads and the attention range are determined according to the adjustment value range. The anomaly detection threshold is determined by optimizing the receiver operating characteristic curve, the false positive rate is controlled within 5%, and the true positive rate is required to be above 95%. The detected abnormal data points are verified again, and the environmental parameter anomalies and instrument state anomalies are comprehensively judged. The quality warning information includes the location of abnormal data, the type of abnormal data, the degree of abnormal data and the processing suggestion. The purpose of this step is to accurately identify abnormal data that affect the measurement quality and provide intelligent technical support for data quality control.

[0042] The specific implementation of step S07 is to generate a comprehensive quality evaluation report to provide decision support information. The report generation module automatically summarizes the analysis results of the previous steps, and the data quality level distribution is obtained by statistical analysis to obtain the proportion of data at each level and the spatial distribution characteristics. The location of abnormal data is visualized by geographic information system, including the spatial coordinates of abnormal data, the type of abnormal data and the degree of abnormal data. The suggested processing measures are automatically generated according to the type and degree of abnormal data, including re-measurement suggestion, instrument inspection suggestion, environmental condition improvement suggestion, etc. The measurement quality comprehensive evaluation result is calculated by weighted scoring method, considering data integrity, precision level, abnormality proportion, etc. The report format adopts standardized template, including execution summary, detailed analysis, chart display, conclusion suggestion, etc. After the report is generated, it is automatically stored and can be exported in multiple formats for subsequent use. The purpose of this step is to provide comprehensive quality evaluation information to support measurement operation quality control and data processing decision.

[0043] The gravity anomaly detection model adopts a deep learning model based on the Transformer architecture, and the specific structure includes an input embedding layer, a position encoding layer, multiple encoder layers, and a classification output layer. The input embedding layer converts the gravity data sequence into a high-dimensional vector representation, with an embedding dimension of 512. The position encoding layer uses a sine-cosine position encoding method to add time position information to the input sequence. The number of encoder layers is set to 6, each containing a multi-head attention sublayer and a feedforward neural network sublayer, with residual connections and layer normalization between sublayers. In the multi-head attention mechanism, the number of attention heads is dynamically determined by an attention head number adjustment function, with each attention head having a dimension of 64. The feedforward neural network contains two linear transformation layers, with a ReLU activation function in the middle and a hidden layer dimension of 2048. The classification output layer uses a fully connected layer and a Sigmoid activation function to output anomaly probability values.

[0044] The training dataset establishment includes four main steps: normal sample collection, anomaly sample generation, sample labeling, and data augmentation. The normal samples are collected from historical gravity measurement projects in different geological environments, including sedimentary rock areas, igneous rock areas, metamorphic rock areas, and other geological backgrounds to ensure the geological representativeness of the samples. The collected normal samples require a measurement accuracy of 0.01 mGal or higher, with data integrity exceeding 95%, and no obvious anomalies are confirmed after quality inspection. The anomaly samples are generated by artificial injection, and the injected anomaly types include instrument zero mutation, scale factor anomaly, temperature shock response, vibration interference, electromagnetic interference, and other anomaly patterns. The zero mutation anomaly is simulated by superimposing a step signal on the normal data, with a mutation amplitude of 0.05 mGal to 0.5 mGal. The scale factor anomaly is simulated by changing the data scaling ratio, with an anomaly degree of 5% to 20%. The environmental disturbance anomaly is simulated by superimposing sinusoidal signals of different frequencies and amplitudes. The sample labeling adopts an expert experience combined with statistical analysis method to establish an anomaly judgment standard, and all samples are labeled with binary classification. The data augmentation techniques include time series transformation, noise injection, data translation, and other methods, which expand the number of training samples to more than 5 times the original samples, improving the model's adaptability to different measurement conditions.

[0045] The key technical ideas of the present application mainly embody in three aspects of real-time drift monitoring and compensation, abnormal detection based on Transformer, and multi-dimensional quality evaluation system. The real-time drift monitoring and compensation technology can track the time-varying characteristics of the zero offset and scale factor of the gravimeter in real time by establishing a dynamic state estimation model through Kalman filtering algorithm. Compared with the traditional linear drift hypothesis method, this technology can capture the nonlinear variation law and environmental response characteristics of the instrument drift, significantly improve the accuracy and real-time performance of the drift compensation, and effectively reduce the system error caused by the instrument drift. The abnormal detection technology based on Transformer captures the spatio-temporal correlation of gravity data by using the multi-head attention mechanism, and adjusts the number of attention heads dynamically to adapt to different complexity of the measurement environment. Compared with the traditional statistical threshold detection method, this technology has stronger feature extraction and pattern recognition capabilities, can identify complex abnormal patterns and potential data quality problems, and greatly improves the accuracy and recall rate of abnormal detection. The multi-dimensional quality evaluation system comprehensively evaluates the data quality from the aspects of accuracy, consistency, stability, and reliability. Compared with the single index evaluation method, this system can more comprehensively reflect the data quality state and provide a more scientific basis for data screening and processing decisions.

[0046] The synergistic effect of these three key technical ideas forms a complete on-site quality evaluation solution for gravity measurement data. Real-time drift monitoring provides accurate baseline data for abnormal detection, abnormal detection results provide important input for quality evaluation, and quality evaluation results guide the optimization and adjustment of drift monitoring parameters. The three form a closed-loop feedback system. Compared with the traditional offline post-processing method, this collaborative technical solution realizes real-time monitoring and dynamic optimization of gravity measurement data quality, significantly improves the efficiency and data quality of gravity measurement operation, and provides advanced technical support for high-precision gravity exploration.

[0047] It should be noted that the present application also solves the technical problem of inaccurate instrument drift compensation in the gravity measurement process. In traditional gravity measurement, instrument drift compensation mainly relies on simple linear interpolation method or empirical formula, which cannot accurately reflect the complex variation law of zero offset and scale factor. Especially in the process of long-term continuous measurement, the instrument drift presents nonlinear characteristics and is significantly affected by environmental factors. The present application can accurately calculate the drift compensation coefficient by establishing a zero drift correction equation and a scale factor drift correction equation, combining the time series of environmental parameters and the initial calibration parameters of the instrument, and accurately correcting the systematic errors of the instrument. At the same time, the drift compensation function considers the variation of the reference point gravity value sequence and the environmental response characteristics, and through real-time calculation and dynamic adjustment, the accuracy and adaptability of drift compensation are significantly improved, effectively ensuring the reliability and consistency of the gravity measurement data.

[0048] Specifically, the principle of the present application is that the root cause of solving the problems of lag and insufficient precision in gravity measurement data quality evaluation is to establish a complete real-time quality monitoring system and an accurate data processing algorithm. First, by arranging an environmental parameter monitoring sensor array in the measurement area, the key environmental factors affecting gravity measurement are collected in real time, providing accurate environmental background information for drift compensation and anomaly detection. Second, the drift monitoring model based on the Kalman filter algorithm can dynamically track the time variation characteristics of the gravity meter zero offset and scale factor, and realize real-time monitoring and compensation of instrument drift through state estimation and prediction mechanism. Third, the anomaly detection model using the Transformer architecture uses the multi-head attention mechanism to capture the spatio-temporal correlation of gravity data, and identifies complex anomaly patterns through deep learning methods, which has stronger non-linear feature extraction capability than traditional statistical methods. In addition, the drift correction equation set combines zero drift correction and scale factor drift correction, which can accurately calculate the compensation coefficients of various drifts, effectively eliminating the influence of systematic errors on measurement accuracy. Finally, the multi-dimensional quality comprehensive evaluation model integrates data accuracy, consistency, stability, reliability and other indicators, and realizes objective grading of data quality through quantitative evaluation method, ensuring the scientificity and accuracy of the quality evaluation results.

[0049] A specific embodiment 1 of the present application is provided below, and the specific implementation of each step in embodiment 1 is described in detail as follows.

[0050] The specific implementation of step S01 is to establish an instrument basic parameter database through a precisely controlled initial calibration procedure. First, select not less than 3 national or regional reference points with known gravity values, the accuracy of the reference point gravity value should reach m / order of magnitude, and the distance between the reference points should be kept above 1000m to ensure spatial representativeness. Measure the gravity value at each reference point for not less than 10 times continuously, with a measurement interval of 5min, and use the least squares method to fit the linear relationship between the gravity meter reading and the true gravity value. Calculate the zero offset through regression analysis, which reflects the systematic deviation of the instrument, with a unit of mGal, and the scale factor is determined by the slope coefficient, which reflects the sensitivity characteristics of the instrument, and is a dimensionless parameter. The established initial instrument parameter database includes key parameters such as zero offset, scale factor, linearity error , repeatability error , etc., where and have a unit of mGal, which provides a reference for subsequent drift correction. The purpose of this step is to eliminate the inherent systematic errors of the instrument, establish an accurate measurement reference, and ensure the reliability and traceability of subsequent gravity measurement data.

[0051] The specific implementation of step S02 involves constructing a multi-parameter environmental monitoring network to achieve real-time environmental data acquisition. Within the measurement area, an array of environmental parameter monitoring sensors is uniformly deployed according to a grid layout principle, with the sensor spacing... The measurement range is set from 200m to 500m, adjusted according to the complexity of the terrain in the measurement area. A platinum resistance thermometer is used as the temperature sensor, with a measurement accuracy of 0.01℃ and a sampling frequency of... Set to 1Hz. The humidity sensor uses a capacitive hygrometer, and the measurement accuracy should reach 0.5% relative humidity. Sampling frequency... The frequency is set to 0.5Hz. The barometric pressure sensor is a piezoresistive barometer, and the measurement accuracy should reach 0.1 hPa. The sampling frequency... The setting is 0.2Hz. The vibration acceleration sensor uses a piezoelectric accelerometer, and the measurement accuracy should reach [percentage missing]. m / sampling frequency The frequency was set to 100Hz to capture high-frequency vibration signals. All sensor data was collected in real time to the data acquisition center via a wireless transmission network, forming a time-synchronized environmental parameter time series. The purpose of this step was to comprehensively monitor changes in environmental factors affecting the accuracy of gravity measurements, providing environmental background information for drift correction and anomaly detection.

[0052] The specific implementation of step S03 is to construct a dynamic drift monitoring model based on the Kalman filter algorithm to achieve real-time drift estimation. The drift monitoring model uses a state-space equation to describe the drift behavior of the gravimeter, and the state variables include the zero-point drift. Scale factor drift and its rate of change and The model prediction equations are established based on the instrument's physical characteristics, while the observation equations are established using repeated measurements from benchmark points, including the zero-point drift. The unit is mGal, and the scaling factor shift is... The parameters are dimensionless. The Kalman filtering process includes a prediction step and an update step. The prediction step predicts the drift state at the current time based on the state at the previous time step and the system model. The update step corrects the prediction results using the baseline measurement values. The baseline measurement frequency is set to once every 30 minutes, with each measurement consisting of 5 consecutive measurements and the average value taken. The process noise covariance matrix... Based on the instrument's technical specifications, the observation noise covariance matrix is... Based on the accuracy setting of the benchmark point measurement, among which and All are symmetric positive definite matrices. The optimal estimates and uncertainties of the zero-point drift and scaling factor drift are obtained through recursive calculation. The purpose of this step is to track changes in the instrument's drift state in real time and provide accurate correction parameters for drift compensation.

[0053] The specific implementation of step S04 is to eliminate systematic errors and random noise by using a multi-level data processing algorithm. The drift compensation function calculates the zero offset compensation value and the scale factor compensation value of each measuring point at the moment based on the drift parameters obtained in step S03 using a linear interpolation method. The zero drift correction equation is expressed as wherein is the zero drift at time t, with the unit of mGal, is the regression coefficient of the constant term, with the unit of mGal, is the regression coefficient of the time term, with the unit of mGal / h, is the regression coefficient of the temperature term, with the unit of mGal / ℃, is the regression coefficient of the air pressure term, with the unit of mGal / hPa, is the regression coefficient of the humidity term, with the unit of mGal / %, is the measurement time, with the unit of h, is the temperature, with the unit of ℃, is the air pressure, with the unit of hPa, is the humidity, with the unit of %, a relationship model of the zero offset and the time and the environmental parameters is established by multiple regression analysis, and the zero drift correction coefficient is output, which is a dimensionless parameter. The scale factor drift correction equation is expressed as wherein is the scale factor drift at time t, which is a dimensionless parameter, is the regression coefficient of the constant term, which is dimensionless, is the regression coefficient of the temperature gradient term, with the unit of h / ℃, is the regression coefficient of the temperature square term, with the unit of 1 / ℃², is the regression coefficient of the time term, with the unit of 1 / h, is the temperature change gradient, with the unit of ℃ / h, the nonlinear relationship model of the scale factor and the temperature gradient is established by focusing on the temperature response characteristics, and the scale factor drift correction coefficient is output, which is a dimensionless parameter. The drift compensation function is expressed as wherein is the gravity measurement value after drift compensation, The values ​​are the original gravity measurements, all in mGal. This function is used to eliminate zero-point offset and scale factor drift errors that occur during long-term operation of the gravimeter. The time-frequency domain filtering algorithm first uses a Butterworth low-pass filter to remove high-frequency noise above 0.1Hz, with the cutoff frequency set based on the spectral characteristics of the gravity signal. Then, a notch filter is used to remove 50Hz power frequency interference and its harmonic components. The filter parameters are determined through spectral analysis and signal-to-noise ratio optimization to ensure that the integrity of the effective gravity signal is maintained while removing noise. The purpose of this step is to obtain high-quality gravity measurement data, providing a reliable data foundation for quality assessment.

[0054] The specific implementation of step S05 involves constructing a multi-dimensional quality assessment system to quantify data quality levels. Data accuracy is measured using the repeated measures standard deviation. Calculation, precision threshold The threshold is set to 0.02 mGal; data exceeding this threshold are marked as low-precision data. The consistency index is determined by the gradient change in gravity values ​​between adjacent measuring points. The assessment indicated that data with gradient changes exceeding 0.05 mGal / m were marked as consistency anomalies. The unit is mGal / m, and the consistency threshold is... The value is 0.05 mGal / m. The stability index is determined by the time series variance. Analysis and evaluation showed a variance exceeding 0.01. Data is marked as unstable data, stability threshold It is 0.01 Reliability metrics are evaluated by comparing multiple independent measurement results, and the differences in measurement results are considered. Data exceeding 0.03 mGal is marked as unreliable data; reliability threshold. The value is 0.03 mGal. The multi-dimensional quality comprehensive evaluation model adopts a weighted comprehensive evaluation method, and the comprehensive quality score is... Represented as ,in to Here are the weighting coefficients for each indicator, and , , , , The scores are for accuracy, consistency, stability, and reliability, each ranging from 0 to 100. The weights of each index are determined using the analytic hierarchy process (AHP). The weight of the accuracy index is as follows: Consistency index weight Stability index weight Reliability index weight The comprehensive quality score is calculated and divided into four levels: excellent, good, qualified, and unqualified, with score thresholds of 90, 80, and 70 points, respectively. The purpose of this step is to comprehensively evaluate the quality level of gravity data and provide quantitative basis for data screening and anomaly identification.

[0055] The specific implementation of step S06 is to use a gravity anomaly detection model based on the Transformer architecture to identify abnormal data. The anomaly detection model uses an encoder-decoder structure, the encoder part contains a multi-head attention mechanism and a feedforward neural network, and the decoder part outputs an anomaly probability , which ranges from 0 to 1 and represents the likelihood of data anomalies. The multi-head attention mechanism is dynamically adjusted by an attention head number adjustment function, and the attention adjustment value is represented as , where , , is a weight coefficient, and , is a measurement area complexity index, is a data sampling density index, is an environmental noise level index, and all three parameters are dimensionless normalized indices with a value range of 0 to 1. The measurement area complexity is quantified by the geological structure density and the terrain relief, the data sampling density is calculated by the number of measurement points per unit area, and the environmental noise level is evaluated by the standard deviation of environmental parameters. After normalizing the three parameters, the attention adjustment value is calculated, and the number of attention heads and the attention range are determined according to the adjustment value range, where is a positive integer, is the number of time steps. The anomaly detection threshold is determined by the receiver operating characteristic curve optimization, with a false positive rate controlled within 5% and a true positive rate required to be above 95%. The detected abnormal data points are verified again, combined with environmental parameter anomalies and instrument state anomalies for comprehensive judgment. The quality warning information includes abnormal data location, abnormal type, abnormal degree, and processing suggestions. The purpose of this step is to accurately identify abnormal data that affect measurement quality and provide intelligent technical support for data quality control.

[0056] The specific implementation of step S07 is to generate a comprehensive quality evaluation report to provide decision support information. The report generation module automatically summarizes the analysis results of the previous steps, and the data quality level distribution is obtained by statistical analysis to obtain the proportion and spatial distribution characteristics of each level data. The abnormal data location is visualized by geographic information system, including the spatial coordinates of abnormal data, abnormal type identification and abnormal degree classification , where For the coordinates of the measurement points, For the type of anomaly, For the degree of anomaly, taking values from 1 to 5. The recommended treatment measures are automatically generated according to the type and degree of anomaly, including re-measurement recommendations, instrument inspection recommendations, environmental condition improvement recommendations, etc. The comprehensive evaluation result of measurement quality is calculated by a weighted scoring method, considering factors such as data integrity , precision level , anomaly proportion , etc., where and are percentage values, is the precision level. The report format uses a standardized template, including an executive summary, detailed analysis, chart display, conclusion and recommendations, etc. After the report is generated, it is automatically stored and can be exported in multiple formats for subsequent use. The purpose of this step is to provide comprehensive quality evaluation information to support measurement operation quality control and data processing decisions.

[0057] The gravity anomaly detection model uses a deep learning model based on the Transformer architecture, with a specific structure including input embedding layer, position encoding layer, multi-layer encoder, and classification output layer. The input embedding layer converts the gravity data sequence into a high-dimensional vector representation, with an embedding dimension of 512. The position encoding uses a sine-cosine function form, and the position encoding function is represented as and , where is the position index, taking values from 0 to the sequence length minus 1, is the dimension index, taking values from 0 to minus 1, is the embedding dimension, here , adding time position information to the input sequence. The number of encoder layers is set to 6, each containing multi-head attention sublayers and feedforward neural network sublayers, with residual connections and layer normalization between sublayers. In the multi-head attention mechanism, the attention weight calculation function is represented as , where , , are the query matrix, key matrix, and value matrix, respectively, with dimensions of sequence length multiplied by , is the key vector dimension, here , the number of attention heads is dynamically determined by the number of attention head adjustment function, and the dimension of each attention head is 64 dimensions. The feedforward neural network includes two linear transformation layers, the intermediate activation function adopts the ReLU function, and the hidden layer dimension is 2048 dimensions. The classification output layer adopts a fully connected layer and a Sigmoid activation function, and outputs an abnormal probability value. The feedforward neural network includes two linear transformation layers, the intermediate activation function adopts the ReLU function, and the hidden layer dimension is 2048 dimensions. The classification output layer adopts a fully connected layer and a Sigmoid activation function, and outputs an abnormal probability value.

[0058] The training data set establishment includes four main steps of normal sample collection, abnormal sample generation, sample labeling and data enhancement. The normal sample collection is obtained from historical gravity measurement projects in different geological environments, including gravity data of sedimentary rock area, igneous rock area, metamorphic rock area and various geological backgrounds, to ensure the geological representativeness of the samples. The collected normal samples require a measurement accuracy of 0.01 mGal or more, a data integrity of more than 95%, and no obvious abnormality after quality inspection. The abnormal samples are generated by artificial injection, and the injected abnormal types include instrument zero mutation, scale factor abnormality, temperature impact response, vibration interference, electromagnetic interference and other abnormal modes. The zero mutation abnormality is simulated by superimposing a step signal on the normal data, and the mutation amplitude is set to 0.05 mGal to 0.5 mGal. The scale factor abnormality is simulated by changing the data scaling ratio, and the abnormal degree is set to 5% to 20%. The environmental interference abnormality is simulated by superimposing sinusoidal signals of different frequencies and amplitudes , wherein ranges from 0.01 Hz to 10 Hz, ranges from 0.01 mGal to 0.1 mGal. The sample labeling adopts the method of expert experience combined with statistical analysis to establish the abnormal judgment standard, and all samples are labeled by binary classification. The data enhancement technology includes time series transformation, noise injection, data translation and other methods, which expands the number of training samples to more than 5 times of the original samples, and improves the adaptability of the model to different measurement conditions.

[0059] In order to better understand and implement the present application, the following provides an embodiment 2 of a specific application scenario of the present application: a technical team performs a seabed gravity measurement task in the northern continental slope area of the South China Sea, and the measurement area is about 500 , the water depth ranges from 200m to 1500m, the seafloor topography is complex and varied, including submarine canyons, seamounts and sedimentary plains and other geomorphic units. The geological structure of the region is active, with multiple fault zones and magmatic intrusions, which have a significant impact on the gravity field. The technical team used the gravity measurement data field quality evaluation method of the invention to lay out 120 gravity measurement points on the seafloor, with a point spacing of 2000m, and planned to complete a 30-day continuous gravity observation operation.

[0060] Before the start of gravity measurement, the technical team selected three seafloor reference points with known gravity values for initial calibration. The reference point gravity value accuracy reached m / , the distance between reference points was kept above 1200m. At each reference point, the gravity value was measured continuously for 15 times, with a measurement interval of 5min, and the linear relationship between the gravimeter reading and the true gravity value was obtained by least squares fitting. The initial zero offset was calculated to be 0.025mGal, the scale factor was 0.9987, the linearity error was 0.008mGal, and the repeatability error was 0.012mGal. The established initial instrument parameter database provided an accurate reference for subsequent drift correction.

[0061] The technical team uniformly deployed 75 environmental parameter monitoring sensors in the measurement area at an interval of 400m, constructing a multi-parameter environmental monitoring network. The temperature sensor measurement accuracy reached 0.008℃, and the sampling frequency was set to 1.2Hz. The humidity sensor measurement accuracy reached 0.3% relative humidity, and the sampling frequency was set to 0.6Hz. The air pressure sensor measurement accuracy reached 0.08hPa, and the sampling frequency was set to 0.25Hz. The vibration acceleration sensor measurement accuracy reached m / , and the sampling frequency was set to 120Hz. All sensor data was transmitted in real time to the sea surface data acquisition center through the seafloor optical cable, forming a time-synchronized environmental parameter time series.

[0062] During the formal measurement process, the technical team started the drift monitoring model based on the Kalman filter algorithm. The model state variables include the zero drift , the scale factor drift , and their change rates and . The reference point repeated measurement frequency was set to every 25min, and each time 6 continuous measurements were taken to obtain the average value. The process noise covariance matrix and observation noise covariance matrix The measurement accuracy of the instrument and the fiducial point was determined. After 48 hours of continuous monitoring, real-time estimates of the zero drift and scale factor drift were obtained, as shown in Table 1:

[0063] Table 1. Gravity meter drift parameter monitoring results

[0064]

[0065] The technical team used multi-level data processing algorithms to real-time preprocess the original gravity measurement data. The regression coefficients of the zero drift correction equation were determined by multiple regression analysis, in which is 0.018 mGal, is mGal / h, is mGal / ℃, is mGal / hPa, is mGal / %. In the regression coefficients of the scale factor drift correction equation, is , is h / ℃, is / ℃², is / h. The systematic errors of the instrument were eliminated by the drift compensation function, while the high-frequency noise with a frequency higher than 0.12 Hz was removed by the Butterworth low-pass filter, and the 50 Hz power frequency interference was removed by the notch filter.

[0066] The technical team constructed a multi-dimensional quality evaluation system to calculate the data quality indicators of each measurement point. The data precision indicator was calculated by the standard deviation of repeated measurements , and the precision threshold was set to 0.018 mGal. The consistency indicator was evaluated by the gradient change of the gravity values of adjacent measurement points , and the consistency threshold was set to 0.055 mGal / m. The stability indicator was analyzed by the time series variance , and the stability threshold was set to 0.012 . The reliability indicator was compared by multiple independent measurement results, and the reliability threshold was set to 0.025 mGal. The quality evaluation results are shown in Table 2:

[0067] Table 2. Gravity data quality evaluation statistical results

[0068]

[0069] The technical team used a gravity anomaly detection model based on the Transformer architecture to identify anomalous data. Based on three parameters—the geological complexity of the measurement area, data sampling density, and environmental noise level—the attention adjustment value was calculated. The value is 0.68, which determines the number of attention heads. There are 8 attention points, and the attention range of each attention point is... The time step is 64. The model takes a 512-dimensional gravity data sequence as input, processes it through a 6-layer encoder, and outputs the anomaly probability. Anomaly detection threshold The receiver operating characteristic curve was optimized to a value of 0.75, resulting in a false positive rate of 4.2% and a true positive rate of 96.8%. Anomaly detection results are shown in Table 3.

[0070] Table 3. Gravity Anomaly Detection Model Identification Results

[0071]

[0072] The technical team generated a comprehensive quality assessment report, which included the distribution of data quality levels, the location of outliers, recommended remedial measures, and the overall evaluation results of measurement quality. The report also included the spatial coordinates of the outliers. Exception type identifier and abnormality level classification Visualized through a geographic information system. The comprehensive evaluation of measurement quality takes into account data integrity. The accuracy rate is 97.5%. Excellent rating, abnormal proportion The overall quality score is 2.5%. The score reached 92.3, and the overall measurement quality was rated as excellent.

[0073] Throughout the 30-day measurement operation, the statistical results of gravity measurement data classification showed that normal gravity data accounted for 74.2% of the total data volume, boundary gravity data accounted for 23.3%, and abnormal gravity data accounted for 2.5%. The proportion of abnormal gravity data was far below the 10% warning line, indicating that the measurement quality control was effective. Through real-time quality monitoring and dynamic data filtering mechanisms, the negative impact of abnormal gravity data on the final results was effectively reduced. The use of multiple validation and cross-validation methods improved the conversion rate and latency control accuracy of boundary gravity data.

[0074] It should be noted that the variables involved in this invention are explained in detail in Tables 4 and 5.

[0075] Table 4. Variable Explanation Table (Part 1)

[0076]

[0077] Table 5. Variable explanation table (second part)

[0078]

[0079] The above description is merely that of a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, and all such changes or replacements should be encompassed within the protection scope of the present application.

Claims

1. A method for on-site quality assessment of gravity measurement data, characterized in that, include: The gravimeter is initially calibrated to obtain the zero offset and scaling factor to establish an initial instrument parameter database; environmental parameter monitoring sensors are deployed in the measurement area to collect temperature, humidity, air pressure, and vibration acceleration parameter data in real time to form an environmental parameter time series. When starting a gravity measurement operation, a drift monitoring model is simultaneously activated. By repeatedly measuring changes in gravity values ​​at reference points and changes in environmental parameters, the zero-point drift and scaling factor drift of the gravimeter are calculated in real time. The raw gravity measurement data is preprocessed in real time, and a drift compensation function is used to eliminate systematic errors in the instrument. At the same time, a drift correction equation system is used to calculate accurate drift compensation values, and a time-frequency domain filtering algorithm is used to remove high-frequency noise interference. A gravity data quality evaluation index system is constructed to calculate the data accuracy, consistency, stability, and reliability indicators for each measurement point. The data quality level is quantified through a multi-dimensional comprehensive quality evaluation model. An outlier detection model is used to identify outliers in the preprocessed gravity data, mark outlier data points that exceed the mass threshold, and generate mass warning information. Generate an on-site quality assessment report for gravity measurement data.

2. The method for on-site quality assessment of gravity measurement data according to claim 1, characterized in that, The drift correction equation set specifically includes the zero-point drift correction equation and the scaling factor drift correction equation. The zero-point drift correction equation is used to calculate the time variation characteristics and environmental response characteristics of the gravimeter's zero-point offset. The inputs include the reference point gravity value sequence, the environmental parameter time sequence, the measurement time interval, the instrument's initial calibration parameters, and the zero-point offset. The output is the zero-point drift correction coefficient. The scaling factor drift correction equation is used to calculate the nonlinear variation characteristics and temperature response characteristics of the gravimeter's scaling factor. The inputs include the scaling factor drift, temperature change gradient, measurement time interval, instrument initial calibration parameters, and environmental parameter time series. The output is the scaling factor drift correction coefficient.

3. The method for on-site quality assessment of gravity measurement data according to claim 2, characterized in that, The drift compensation function is specifically used to eliminate the zero offset and scale factor drift errors generated by the gravimeter during long-term operation. The inputs include the reference point gravity value sequence, environmental parameter time sequence, measurement time interval and instrument initial calibration parameters, and the output is the drift-compensated gravity measurement value.

4. The method for on-site quality assessment of gravity measurement data according to claim 3, characterized in that, The gravity anomaly detection model is specifically a sequence analysis model based on the Transformer architecture, which includes a multi-head attention mechanism to capture the spatiotemporal correlation of gravity data. The number of attention heads is dynamically determined based on three parameters: the complexity of the measurement area, the data sampling density, and the environmental noise level. When the geological structure of the measurement area is complex, the data sampling density is high, and the environmental noise level is low, the number of attention heads is increased to improve the anomaly detection accuracy.

5. The method for on-site quality assessment of gravity measurement data according to claim 4, characterized in that, The steps for establishing the training dataset for the gravity anomaly detection model are as follows: gravity measurement data under different geological environments are collected as normal samples; various types of instrument malfunctions, environmental interference, and measurement errors are artificially injected to generate abnormal samples; all samples are labeled and classified to form a training dataset containing both normal and abnormal data; and data augmentation techniques are used to expand the number of training samples to improve the model's generalization ability.

6. The method for on-site quality assessment of gravity measurement data according to claim 5, characterized in that, The environmental parameter monitoring sensor refers to a multi-type sensor array installed within the measurement area, used to monitor changes in environmental factors that affect the accuracy of gravity measurement in real time.

7. The method for on-site quality assessment of gravity measurement data according to claim 6, characterized in that, The drift monitoring model refers to a state estimation model based on the Kalman filtering algorithm, which is used to estimate and predict the drift state of the gravimeter in real time.

8. The method for on-site quality assessment of gravity measurement data according to claim 7, characterized in that, The multi-dimensional quality comprehensive evaluation model refers to a comprehensive evaluation algorithm that integrates multiple quality indicators and outputs data quality level classification results.

9. The method for on-site quality assessment of gravity measurement data according to claim 8, characterized in that, The attention head number adjustment function is specifically used to adjust the multi-head attention mechanism parameters of the gravity anomaly detection model. The attention adjustment value is calculated based on three data points: measurement area complexity, data sampling density, and environmental noise level. When the attention adjustment value is in the range of 0 to 0.25, four attention heads are used, each with an attention range of 256 time steps. When the attention adjustment value is in the range of 0.25 to 0.5, six attention heads are used, each with an attention range of 128 time steps. When the attention adjustment value is in the range of 0.5 to 0.75, eight attention heads are used, each with an attention range of 64 time steps. When the attention adjustment value is in the range of 0.75 to 1, twelve attention heads are used, each with an attention range of 32 time steps, adjusting the multi-head attention mechanism parameters of the model.

10. The method for on-site quality assessment of gravity measurement data according to claim 9, characterized in that, The attention weight threshold for each attention head is dynamically set based on the variance of the gravity data sequence. When the variance of the gravity data sequence is less than 0.01 mGal, the attention weight threshold is set to 0.8; when the variance of the gravity data sequence is between 0.01 and 0.05 mGal, the attention weight threshold is set to 0.6; and when the variance of the gravity data sequence is greater than 0.05 mGal, the attention weight threshold is set to 0.

4. The attention head uses a sliding window attention method to process the gravity data time series and filters important features based on the attention weight threshold.

Images