Multi-source data collaborative deformation prediction method and system based on deep reinforcement learning

Abstract

This invention discloses a multi-source data collaborative deformation prediction method and system based on deep reinforcement learning, belonging to the field of geological disaster monitoring technology. The method includes: acquiring multi-source data of a target area and preprocessing it to obtain preprocessed multi-source data; obtaining the corresponding signal-to-noise ratio (SNR) based on the preprocessed multi-source data; inputting the standardized preprocessed multi-source data into a prediction model to obtain prediction results corresponding to each data source; performing deep reinforcement learning based on the historical prediction error and SNR of the prediction model as input states to output the current optimal weights; weighted fusion of the current optimal weights and the corresponding prediction results to obtain the final deformation prediction result; and obtaining a risk assessment index based on the final deformation prediction result and performing graded early warning. This method overcomes the limitations of existing methods in terms of accuracy, timeliness, and reliability, thereby improving the accuracy and timeliness of geological disaster early warning.

Description

Technical Field

[0001] This invention relates to the field of geological disaster monitoring technology, and more specifically to a method and system for multi-source data collaborative deformation prediction based on deep reinforcement learning. Background Technology

[0002] Currently, crustal deformation monitoring is a key technological support for the early identification and warning of geological disasters (such as landslides, earthquakes, and ground subsidence). Current mainstream monitoring methods include Global Navigation Satellite Systems (GNSS), Synthetic Aperture Radar Interferometry (InSAR), and gravity observations. While these methods have achieved some success in large-scale crustal movement monitoring, they still face multiple technical bottlenecks in practical applications requiring high precision and timely geological disaster warnings. These bottlenecks are specifically reflected in the following four aspects:

[0003] (1) Insufficient spatial resolution: Traditional GNSS monitoring relies on sparsely deployed ground stations, with a station spacing of 5–10 km, resulting in an overall spatial resolution of no more than 5 km. This resolution makes it difficult to effectively capture local deformation gradient features within a 1–3 km scale, easily creating monitoring blind spots and thus missing potential disaster sources, weakening the coverage and sensitivity of the early warning system. (2) Inconsistent spatiotemporal references for multi-source data: Different monitoring methods exhibit significant heterogeneity in spatiotemporal dimensions. For example, InSAR can provide a spatial resolution of approximately 0.5–1 km, but its revisit period is typically 12–35 days; while GNSS sampling frequencies can reach 1–10 Hz, providing high temporal resolution. Furthermore, the processing model for gravity data is independent of InSAR / GNSS, and different data sources often use different coordinate reference systems and time bases. This inconsistency in spatiotemporal references introduces systematic biases during multi-source data fusion, significantly increasing deformation prediction errors and making it difficult to meet the needs of high-precision early warning for geological disasters. (3) Dynamic response lag: Most existing fusion algorithms are based on offline batch processing mode. The overall delay from data acquisition, transmission, processing to the final deformation assessment result output is usually no less than 6 hours. In the case of sudden geological disasters, such delay seriously restricts the real-time performance and effectiveness of early warning, and misses the key emergency response window. (4) High false alarm rate under noise interference: Monitoring data is susceptible to various noise interferences, including atmospheric delay, changes in surface vegetation, instrument drift and fluctuations in environmental temperature and humidity. Existing fusion and prediction algorithms generally lack the ability to adaptively suppress complex noise environments, resulting in insufficient robustness in abnormal signal identification, high false alarm rate, and easy waste of emergency resources or public panic.

[0004] Therefore, how to overcome the limitations of existing methods in terms of accuracy, timeliness, and reliability, and thus improve the accuracy and timeliness of geological disaster early warning, is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, this invention is proposed to provide a multi-source data collaborative deformation prediction method and system based on deep reinforcement learning to overcome or at least partially solve the above problems. It breaks through the limitations of existing methods in terms of accuracy, timeliness and reliability, thereby improving the accuracy and timeliness of geological disaster early warning.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, embodiments of the present invention provide a multi-source data collaborative deformation prediction method based on deep reinforcement learning, comprising: Acquire multi-source data of the target area and preprocess it to obtain preprocessed multi-source data; The corresponding signal-to-noise ratio is obtained based on the preprocessed multi-source data; Based on the preprocessed multi-source data, after standardization, it is input into the prediction model to obtain the prediction results corresponding to each data source; Deep reinforcement learning is performed based on the historical prediction error of the prediction model and the signal-to-noise ratio as input states, and the current optimal weights are output. The final deformation prediction result is obtained by weighted fusion of the current optimal weight and the corresponding prediction result; Based on the final deformation prediction results, a risk assessment index is obtained and a graded early warning is issued.

[0007] In another embodiment, the method for preprocessing multi-source data acquisition is as follows: The multi-source data includes: raw GNSS observation data, InSAR data, and gravimeter observation data; After time and spatial alignment of the raw GNSS observation data, the InSAR data, and the gravimeter observation data, preprocessed GNSS data, preprocessed InSAR data, and preprocessed gravity data with synchronized timestamps and unified spatial coordinates are obtained, which together constitute the preprocessed multi-source data.

[0008] In another embodiment, obtaining the corresponding signal-to-noise ratio specifically includes: The useful signal variance and the comprehensive noise variance are obtained based on the preprocessed GNSS data. The GNSS signal-to-noise ratio is based on the ratio of the useful signal variance to the comprehensive noise variance. The deformation signal amplitude and spatial correlation noise amplitude are obtained based on the preprocessed InSAR data. The InSAR signal-to-noise ratio is based on the ratio of the deformation signal amplitude to the spatial correlation noise amplitude. Based on the preprocessed gravity data, the gravity anomaly amplitude and the total gravity noise amplitude are obtained; The gravity signal-to-noise ratio is defined as the ratio of the gravity anomaly amplitude to the total gravity noise amplitude.

[0009] In another embodiment, the method for obtaining the useful signal variance and the combined noise variance is as follows: Based on the preprocessed GNSS data, gross error removal, ionospheric and tropospheric delay correction, Kalman filtering or moving average filtering are performed to extract long-term trend terms and periodic terms reflecting crustal deformation as useful signals. The remaining random residual terms are the combined noise of receiver noise floor and multipath effect. Based on the sequence corresponding to the useful signal, the variance of the useful signal is obtained; Based on the sequence corresponding to the comprehensive noise, the variance of the comprehensive noise is obtained; The method for obtaining the amplitude of the deformation signal and the amplitude of the spatial correlation noise is as follows: The deformation signal amplitude is extracted based on the preprocessed InSAR data; The DEM error noise amplitude and atmospheric delay noise amplitude are obtained based on the preprocessed InSAR data; The spatial correlation noise amplitude is obtained based on the DEM error noise amplitude and the atmospheric delay noise amplitude.

[0010] In another embodiment, the method for obtaining the gravity anomaly amplitude and the total gravity noise amplitude is as follows: The gravity anomaly amplitude is obtained based on the preprocessed gravity data, which is used to reflect the intensity of the gravity anomaly signal associated with crustal deformation. The instrument noise amplitude is obtained based on the accuracy parameters of the gravimeter itself; The amplitude of environmental impact noise is obtained based on the residual of the preprocessed gravity data. The total gravity noise amplitude is obtained based on the instrument noise amplitude and the environmental noise amplitude.

[0011] In another embodiment, the method for obtaining the prediction results corresponding to each data source is as follows: The preprocessed GNSS data, preprocessed InSAR data, and preprocessed gravity data are respectively standardized to obtain standardized GNSS data, standardized InSAR data, and standardized gravity data. Based on the standardized GNSS data, the standardized InSAR data, and the standardized gravity data, respectively, the prediction model is input to obtain the first prediction result, the second prediction result, and the third prediction result. The first prediction result includes a first deformation rate prediction result and a first cumulative deformation prediction result; The second prediction result includes the second deformation rate prediction result and the second cumulative deformation prediction result; The third prediction result includes the third deformation rate prediction result and the third cumulative deformation prediction result.

[0012] In another embodiment, the deep reinforcement learning specifically includes: Based on the input state, the data is fed into the deep reinforcement learning controller, and the output action is a set of weight adjustment instructions. The deep reinforcement learning controller learns a set of weight adjustment strategies that maximize long-term cumulative rewards through continuous interaction with the prediction model and the environment, and outputs the current optimal weights. The reward function R of the deep reinforcement learning controller is: R=α×(1-w)+β×(1-c); Where α represents the accuracy weight, w represents the current prediction error, β represents the timeliness weight, and c represents the response delay; Heuristic constraints are introduced in the deep reinforcement learning controller: When the corresponding signal-to-noise ratio is greater than or equal to the first threshold, the corresponding weight is increased; When the corresponding signal-to-noise ratio is less than the second threshold, the corresponding weight is reduced.

[0013] In another embodiment, the method for obtaining the final deformation prediction result is as follows: The current optimal weights include: GNSS optimal weights, InSAR optimal weights, and gravity optimal weights; The crustal deformation rate at the current moment is obtained by multiplying the GNSS optimal weight, the InSAR optimal weight, and the gravity optimal weight with the first deformation rate prediction result, the second deformation rate prediction result, and the third deformation rate prediction result, respectively. The cumulative deformation at the current time is obtained by multiplying the GNSS optimal weight, the InSAR optimal weight, and the gravity optimal weight by the first cumulative deformation prediction result, the second cumulative deformation prediction result, and the third cumulative deformation prediction result, respectively. The current crustal deformation rate and the current cumulative deformation are used together as the final deformation prediction result.

[0014] In another embodiment, obtaining a risk assessment index and performing tiered early warning specifically includes: Based on the current crustal deformation rate and the current cumulative deformation, data normalization is performed to obtain the normalized crustal deformation rate and the normalized cumulative deformation; The Risk Assessment Index (CSI) is derived based on the normalized crustal deformation rate and the normalized cumulative deformation. Set the first preset value Q1 and the second preset value Q2; When CSI < Q1, it is considered a normal state and no warning is issued; When Q1≤CSI<Q2, a yellow alert is issued and monitoring is automatically intensified; When CSI≥Q2, a red alert is issued, an emergency response is initiated, and corresponding emergency response recommendations are output.

[0015] Secondly, embodiments of the present invention provide a multi-source data collaborative deformation prediction system based on deep reinforcement learning, comprising: a multi-source data processing module, a signal-to-noise ratio acquisition module, an initial result output module, an optimal weight acquisition module, a final result output module, and a hierarchical early warning module; The multi-source data processing module is used to acquire multi-source data of the target area and perform preprocessing to obtain preprocessed multi-source data; The signal-to-noise ratio (SNR) acquisition module is used to acquire the corresponding SNR based on the preprocessed multi-source data; The initial result output module is used to input the standardized preprocessed multi-source data into the prediction model to obtain the prediction results corresponding to each data source. The optimal weight acquisition module is used to perform deep reinforcement learning based on the historical prediction error of the prediction model and the signal-to-noise ratio as input states, and output the current optimal weights. The final result output module is used to obtain the final deformation prediction result by weighted fusion of the current optimal weight and the corresponding prediction result; The graded early warning module is used to obtain a risk assessment index based on the final deformation prediction result and to conduct graded early warning.

[0016] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a multi-source data collaborative deformation prediction method and system based on deep reinforcement learning, which has the following beneficial effects: 1. By fusing GNSS, InSAR and gravity sensor multi-source geodetic data, and using standardized multi-source data after spatiotemporal alignment as the core input, a deep reinforcement learning controller dynamically optimizes the contribution of each data source in real time, driving the deformation predictor to generate high-precision estimates of crustal deformation rate and accumulation. The weights of each source data are dynamically adjusted according to the data signal-to-noise ratio (SNR) to improve the accuracy of deformation prediction. Then, a comprehensive deformation risk assessment index is constructed to output graded early warning signals.

[0017] 2. This invention introduces an error feedback mechanism, which feeds back historical prediction errors to the deep reinforcement learning controller, forming a closed loop of continuous self-optimization. While ensuring prediction accuracy, it achieves low data processing latency and low false alarm rate, and can be widely used in the field of early monitoring and emergency response to geological disasters such as earthquakes and landslides.

[0018] 3. After fusing InSAR and GNSS data, the spatial resolution of the monitoring is greatly improved, and small-scale deformation gradients can be captured, thus improving the spatial resolution.

[0019] 4. The deep reinforcement learning controller of the present invention not only considers the real-time quality of the data, but also takes the historical prediction error as a key input. This composite state space design of "data quality + model performance" makes the weight adjustment decision more forward-looking and robust, which is far superior to static weights or simple attention mechanisms based on a single data feature.

[0020] 5. In the learning process of the deep reinforcement learning controller, this invention introduces heuristic constraints based on physics and engineering experience. This design accelerates the convergence of the model and ensures its physical rationality in extreme cases. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0022] Figure 1 This is a flowchart of a multi-source data collaborative deformation prediction method based on deep reinforcement learning provided in an embodiment of the present invention.

[0023] Figure 2 This is a flowchart of the signal-to-noise ratio acquisition method provided in an embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of the structure of a multi-source data collaborative deformation prediction system based on deep reinforcement learning provided in an embodiment of the present invention. Detailed Implementation

[0025] 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.

[0026] Example 1 like Figure 1 As shown, this invention discloses a multi-source data collaborative deformation prediction method based on deep reinforcement learning, including the following steps. For ease of description, these steps are numbered S1 to S6, and these numbers are not used to limit the sequential relationship between the various steps of this invention: S1 acquires multi-source data of the target area and performs preprocessing to obtain preprocessed multi-source data.

[0027] Furthermore, the method for preprocessing multi-source data acquisition is as follows: Multi-source data includes: raw GNSS observation data, InSAR data, and gravimeter observation data; After performing time and spatial alignment on the raw GNSS observation data, InSAR data, and gravimeter observation data respectively, preprocessed GNSS data, preprocessed InSAR data, and preprocessed gravity data with synchronized timestamps and unified spatial coordinates are obtained, which together constitute preprocessed multi-source data.

[0028] Furthermore, GNSS (Global Navigation Satellite System) raw observation data refers to the low-level measurement information that the receiver directly extracts from satellite signals, without processing or only after preliminary decoding; Raw GNSS observation data includes at least one or more of the following: pseudorange observations, carrier phase observations, Doppler shift observations, signal-to-noise ratio, satellite ephemeris and clock bias information; In this embodiment, five GNSS receivers supporting the output of raw observation data are used to collect raw GNSS observation data. The receivers are Trimble R10 models with a sampling rate of 1Hz and positioning accuracy of ±2.5mm+0.5ppm horizontally and ±5mm+0.5ppm vertically.

[0029] Furthermore, InSAR data includes: InSAR interferograms or InSAR deformation maps; InSAR interferograms are complex images generated by conjugate multiplication of two SAR (Synthetic Aperture Radar) images acquired at different times. Their phase information includes: true surface deformation, phase caused by topographic relief, atmospheric delay, orbital error, and temporal / spatial decorrelation noise. InSAR deformation maps are spatial distribution maps of the true surface deformations that are inverted from interferograms after removing non-deformation components such as topography, atmosphere, and orbit. In this embodiment, InSAR data is acquired by Sentinel-1 SAR, an Earth observation satellite equipped with a SAR (Synthetic Aperture Radar) sensor, with a resolution of 0.5km × 0.5km; band: C-band; revisit period of 12 days.

[0030] Furthermore, gravimeter observation data refers to physical quantity data obtained by measuring minute changes in the Earth's gravitational field in space and time using a gravimeter; In this embodiment, two MEMS gravimeters are used to collect gravimeter observation data.

[0031] Furthermore, the InSAR data employs piecewise linear interpolation to interpolate discrete deformation results onto GNSS timestamps, generating pseudo-continuous time series and achieving time alignment. This method, while ensuring computational efficiency, can better preserve the macroscopic deformation trend of InSAR. The gravimeter observation data was resampled from 0.1 Hz to 1 Hz using cubic spline interpolation. This method better preserves high-frequency details in the gravity data, avoids signal distortion, and achieves time alignment.

[0032] Furthermore, spatial alignment and coordinate unification were achieved: using the International Earth Reference Frame (ITRF2014) as a unified benchmark, the raw GNSS observation data (WGS84), InSAR data (local coordinate system), and gravimeter observation data (CGCS2000) were all converted to the ITRF2014 coordinate system through a seven-parameter Helmert transformation.

[0033] Resolution unification: Using the 0.5km×0.5km grid of InSAR data as a benchmark, the discrete GNSS point data and gravity point data are interpolated to generate corresponding 0.5km×0.5km grid data through Kriging interpolation. Kriging interpolation can optimize estimation by utilizing spatial correlation and is a classic method in the field of geosciences.

[0034] Furthermore, raw data from different sensors with different spatiotemporal characteristics are transformed into a standardized dataset with unified spatiotemporal reference, standardized format, and consistent physical units, providing high-quality input for subsequent fusion prediction.

[0035] S2 obtains the corresponding signal-to-noise ratio based on preprocessed multi-source data.

[0036] Furthermore, such as Figure 2 As shown, obtaining the corresponding signal-to-noise ratio specifically includes: Useful signal variance and comprehensive noise variance are obtained from preprocessed GNSS data; The ratio of useful signal variance to comprehensive noise variance is used as the GNSS signal-to-noise ratio. SNR GNSS ; Deformation signal amplitude and spatial correlation noise amplitude are obtained from preprocessed InSAR data; The InSAR signal-to-noise ratio is based on the ratio of deformation signal amplitude to spatial correlation noise amplitude. SNR InSAR ; Gravity anomaly amplitude and total gravity noise amplitude are obtained based on preprocessed gravity data; The gravity signal-to-noise ratio is calculated as the ratio of the gravity anomaly amplitude to the total gravity noise amplitude. SNR Gravity .

[0037] Furthermore, the methods for obtaining the useful signal variance and the overall noise variance are as follows: Based on preprocessed GNSS data, gross error removal, ionospheric and tropospheric delay correction, Kalman filtering or moving average filtering are used to extract long-term trend terms and periodic terms reflecting crustal deformation as useful signals. The remaining random residual terms are the combined noise of receiver noise floor and multipath effect. Based on the sequence corresponding to the useful signal, the variance of the useful signal is obtained. : ; Where n represents the number of consecutive observation samples; The i-th sample of the signal sequence refers to the signal value corresponding to the "long-term trend term + periodic term" obtained after separation at the i-th observation time. It represents the mean of the useful signal sequence; Based on the sequence corresponding to the comprehensive noise, the variance of the comprehensive noise is obtained. : ; in, The i-th sample in the integrated noise sequence refers to the noise value corresponding to the random residual remaining after separating the useful signal at the i-th observation time. This represents the mean of the combined noise sequence.

[0038] Furthermore, the variance of the useful signal Used to quantify the wave characteristics of "crustal deformation-related useful signals" in raw GNSS observation data, it is essential for calculating the GNSS signal-to-noise ratio. SNR GNSS The numerator; the combined noise variance Used to quantify the fluctuation characteristics of "receiver noise floor + multipath effect" in raw GNSS observation data, it is used to calculate the GNSS signal-to-noise ratio. SNR GNSS The denominator term.

[0039] Furthermore, raw GNSS observation data is the sole core data source for calculating these two variances. The raw data contains useful signals related to crustal deformation, target noise such as receiver noise floor, and systematic errors that need to be corrected in advance; it forms the basis for separating signal and noise. The completeness and accuracy of the raw data directly determine the purity of signal-noise separation, thus affecting the accuracy of the two variance calculations, and ultimately determining... SNR GNSS Whether it can truly reflect the real-time quality of the data. To verify the reliability of the total noise variance calculation and to clarify the sources of noise, based on the "additivity of independent noise variances," the total noise is decomposed into two categories: receiver noise floor and multipath effects. ; in, Indicates the overall noise variance. This represents the receiver's noise floor variance, estimated using the GNSS receiver's technical parameters. The variance of the multipath effect is expressed by the following formula extracted from the residuals after preprocessing the raw GNSS observation data: ; in, This represents the separated single-path effect noise value. This represents the mean of multipath noise. The splitting result and the direct calculation The data are cross-referenced to ensure accurate noise quantification.

[0040] Furthermore, the methods for obtaining the amplitude of the deformation signal and the amplitude of spatial correlation noise are as follows: The amplitude of the deformation signal was extracted from preprocessed InSAR data; DEM error noise amplitude and atmospheric delay noise amplitude were obtained based on preprocessed InSAR data; Spatial correlation noise amplitude is obtained based on DEM error noise amplitude and atmospheric delay noise amplitude.

[0041] Furthermore, the amplitude of the deformation signal and the amplitude of the spatial correlation noise are directly extracted from the InSAR interferogram / deformation map, which is the only core data source for calculating the InSAR SNR: the InSAR interferogram records the phase difference information corresponding to the surface deformation, and the deformation map is a visual expression of the deformation after the interferogram is unwrapped and geocoded.

[0042] Furthermore, spatial correlation noise A Noise : ; in, A DEM Indicates the amplitude of DEM error noise. AAtm This indicates the amplitude of atmospheric delay noise.

[0043] Furthermore, the amplitude of DEM error noise A DEM Extract from the preprocessed InSAR interferogram: By performing a difference operation between the interferogram and the high-precision reference DEM, the DEM error distribution map is obtained, and the statistical mean of its pixel-level amplitude is used as the DEM error noise amplitude; Atmospheric delay noise amplitude A Atm The atmospheric correction model based on InSAR interferograms is inverted to obtain the following: the atmospheric delay component is extracted from the phase of the interferogram using the phase separation method, and after converting the component into the deformation equivalent amplitude, the statistical mean is taken as the atmospheric delay noise amplitude.

[0044] Furthermore, the methods for obtaining the amplitude of gravity anomalies and the amplitude of total gravity noise are as follows: Gravity anomaly amplitudes are obtained from preprocessed gravity data to reflect the intensity of gravity anomaly signals associated with crustal deformation. The instrument noise amplitude is obtained based on the accuracy parameters of the gravimeter itself; The amplitude of environmental impact noise is obtained from the residual of preprocessed gravity data; The total gravity noise amplitude is obtained based on the instrument noise amplitude and the environmental noise amplitude.

[0045] Furthermore, gravity signal-to-noise ratio SNR Gravity Specifically: ; in, This indicates the magnitude of gravity anomalies, reflecting the intensity of gravity anomaly signals associated with crustal deformation. This represents the total noise amplitude due to gravity.

[0046] Furthermore, the amplitude of gravity anomalies for: ; in, This represents the instrument noise amplitude, estimated from the gravimeter's own accuracy parameters. The magnitude of environmental noise is extracted from the residuals after preprocessing the original gravimeter observation data.

[0047] S3 is based on the preprocessed multi-source data that has been standardized and then input into the prediction model to obtain the prediction results corresponding to each data source.

[0048] Furthermore, the methods for obtaining the prediction results corresponding to each data source are as follows: Standardization processes are performed on preprocessed GNSS data, preprocessed InSAR data, and preprocessed gravity data respectively to obtain standardized GNSS data, standardized InSAR data, and standardized gravity data. Based on standardized GNSS data, standardized InSAR data, and standardized gravity data, respectively, the prediction model is input to obtain the first prediction result, the second prediction result, and the third prediction result. The first prediction result includes the first deformation rate prediction result. v GNSS and the first cumulative deformation prediction results S GNSS ; The second prediction result includes the second deformation rate prediction result. v InSAR Second cumulative deformation prediction results S InSAR ; The third prediction result includes the third deformation rate prediction result. v Gravity and the third cumulative deformation prediction results S Gravity .

[0049] Furthermore, standardization: Format standardization: unify preprocessed GNSS data, preprocessed InSAR data, and preprocessed gravity data into scientific data formats such as NetCDF or HDF5, including necessary metadata (such as time, coordinates, and units). Unit standardization: Ensure that the unit of all deformation-related data is consistent in millimeters (mm), and the unit of rate is consistent in millimeters per day (mm / day).

[0050] Furthermore, in this embodiment, the prediction model (Predictor) adopts a Long Short-Term Memory (LSTM) network as the core prediction model. LSTM is good at processing time series data and can capture long-range dependencies of deformation.

[0051] S4 uses the historical prediction error and signal-to-noise ratio of the prediction model as input states for deep reinforcement learning, and outputs the current optimal weights.

[0052] Furthermore, deep reinforcement learning specifically includes: The input state is fed into the deep reinforcement learning controller, and the output action is a set of weight adjustment instructions. The deep reinforcement learning controller learns a set of weight adjustment strategies that maximize long-term cumulative rewards through continuous interaction with the prediction model and the environment, and outputs the current optimal weights. The reward function R of the deep reinforcement learning controller is: R=α×(1-w)+β×(1-c); Where α represents the accuracy weight, w represents the current prediction error, β represents the timeliness weight, and c represents the response delay; Heuristic constraints are introduced in deep reinforcement learning controllers: When the corresponding signal-to-noise ratio is greater than or equal to the first threshold, the corresponding weight is increased; When the corresponding signal-to-noise ratio is less than the second threshold, the corresponding weight is reduced.

[0053] Furthermore, in this embodiment, the deep reinforcement learning controller (DRL controller) employs a deep Q-network; State (S): The input state of a deep reinforcement learning controller (DRL) is composed of two parts: the signal-to-noise ratio and the historical prediction error of the prediction model. (1) Current data quality vector composed of the signal-to-noise ratio of each data source: [SNR] GNSS SNR InSAR SNR Gravity ]; (2) The historical performance vector composed of the historical prediction errors of the prediction model: [Error] t-1 Error t-2 Error t-3 ]; Action (A): The output action of a deep reinforcement learning controller (DRL) is a set of weight adjustment instructions, such as adjusting the current weight [w]. GNSS , w InSAR , w Gravity They were to be increased slightly, kept unchanged, or decreased slightly, respectively. Reward function (R): The reward function is the "command stick" for the learning of the deep reinforcement learning controller (DRL). It is designed as: R = α × (1 - w) + β × (1 - c). In this embodiment, α is set to 0.7 and β is set to 0.3. High prediction accuracy and low response latency will be rewarded, while large prediction errors or long processing delays will be penalized. Learning and Constraints: The Deep Reinforcement Learning Controller (DRL) learns a set of weight adjustment strategies that maximize long-term cumulative rewards through continuous interaction with the prediction model and the environment. To accelerate convergence and ensure physical plausibility, heuristic constraints are introduced: when the corresponding signal-to-noise ratio (SNR) is greater than or equal to a first threshold, the source weight is increased; when the corresponding SNR is less than a second threshold, the source weight is significantly decreased. In this embodiment, the first threshold is set to 10 dB, and the second threshold is set to 30 dB. These constraints serve as guidance in the DRL exploration process, rather than replacing its learning.

[0054] Furthermore, the single-step prediction error of the prediction model in the current processing cycle is Error. t Specifically: ; Among them, Predict t This represents the predicted deformation value at the current moment: that is, the result of a certain deformation index output after dynamic weight fusion; Truth t This represents the high-precision true value at the current moment; Error t-1 Error t-2..n It is a sequence of single-step errors from the previous n consecutive processing cycles, arranged in reverse chronological order.

[0055] S5 calculates the final deformation prediction result by weighting and fusing the current optimal weight with the corresponding prediction result.

[0056] Furthermore, the final deformation prediction result is obtained as follows: The current optimal weights include: GNSS optimal weights InSAR optimal weights And optimal weight of gravity ; Based on GNSS optimal weights InSAR optimal weights And optimal weight of gravity Comparing with the first deformation rate prediction results v GNSS Second deformation rate prediction results v InSAR and the third deformation rate prediction results v Gravity Multiply and sum to obtain the crustal deformation rate at the current moment. v : ; Based on GNSS optimal weights InSAR optimal weights And optimal weight of gravity Comparing with the first cumulative deformation prediction results respectively S GNSS Second cumulative deformation prediction results S InSAR and the third cumulative deformation prediction results S Gravity Multiply and sum to obtain the cumulative deformation at the current moment. S : ; Current crustal deformation rate v and cumulative deformation at the current momentS Together, they form the final deformation prediction result.

[0057] S6 derives a risk assessment index based on the final deformation prediction results and conducts graded early warnings.

[0058] Furthermore, a risk assessment index is obtained and a tiered early warning system is implemented, specifically including: Data normalization is performed based on the current crustal deformation rate and the current cumulative deformation to obtain the normalized crustal deformation rate and the normalized cumulative deformation; The Risk Assessment Index (CSI) is derived based on normalized crustal deformation rate and normalized cumulative deformation. Set the first preset value Q1 and the second preset value Q2; When CSI < Q1, it is considered a normal state and no warning is issued; When Q1≤CSI<Q2, a yellow alert is issued and monitoring is automatically intensified; When CSI≥Q2, a red alert is issued, an emergency response is initiated, and corresponding emergency response recommendations are output.

[0059] Furthermore, this embodiment employs the Min-Max Scaling method to normalize the crustal deformation rate at the current moment. v and cumulative deformation at the current moment S Mapping to the [0,1] interval yields the normalized crustal deformation rate v. norm and normalized cumulative deformation S norm : v norm =(v-v min ) / (v max -v min ), Where v min =0mm / day, v max =20mm / day; S norm =(S-S) min ) / (S max -S min ); Where S min =0mm, S max =100mm; When v - v min =0 means "the difference between the current deformation rate and the indeformed baseline value". This setting is in line with the industry practice of geological disaster deformation monitoring: taking "indeformed" as the baseline, the deviation of the current deformation from the baseline is quantified by molecular weight. The denominators are the deformation rates v. max v min=20mm / day and cumulative deformation S max S min =100mm critical range width, the combination of the two can unify the scaling of deformation data of different magnitudes and dimensions (rate of mm / day, cumulative deformation of mm) to the [0,1] interval.

[0060] Furthermore, the Risk Assessment Index (CSI) is specifically defined as follows: CSI = 0.4 × v norm +0.6×S norm ; The weights of 0.4 and 0.6 are empirical values ​​derived from a large amount of historical disaster data, reflecting the higher contribution of deformation rate to short-term early warning.

[0061] Furthermore, in this embodiment, the first preset value Q1 and the second preset value Q2 are set to 0.6 and 0.8 respectively; When CSI < 0.6, it is considered a normal state and no warning is issued; When CSI is between 0.6 and 0.8, a yellow alert is issued, monitoring is automatically intensified, and the GNSS sampling rate is automatically increased to 2 Hz. When CSI ≥ 0.8, a red alert is issued, an emergency response is initiated and corresponding emergency response recommendations are output, and an emergency imaging request for InSAR can be triggered.

[0062] Furthermore, emergency response recommendations include: monitoring key areas and evacuation zones, etc.

[0063] Furthermore, this invention solves the problems of low spatial resolution, large error, delayed response, and high false alarm rate in traditional crustal deformation monitoring by using multi-source data spatiotemporal alignment, DRL dynamic weight allocation, and hierarchical early warning mechanism. It significantly improves the accuracy and timeliness of geological disaster early warning and can be extended to geological disaster monitoring scenarios such as major active fault zones and landslide-prone areas around the world, and has important engineering application value.

[0064] Example 2 like Figure 3 As shown, based on the same inventive concept, this embodiment of the invention also provides a multi-source data collaborative deformation prediction system based on deep reinforcement learning, including: a multi-source data processing module, a signal-to-noise ratio acquisition module, an initial result output module, an optimal weight acquisition module, a final result output module, and a hierarchical early warning module; The multi-source data processing module is used to acquire multi-source data from the target area and preprocess it to obtain preprocessed multi-source data. The signal-to-noise ratio (SNR) acquisition module is used to obtain the corresponding SNR based on preprocessed multi-source data. The initial result output module is used to input the preprocessed multi-source data after standardization into the prediction model to obtain the prediction results corresponding to each data source. The optimal weight acquisition module is used to perform deep reinforcement learning based on the historical prediction error and signal-to-noise ratio of the prediction model as input states, and output the current optimal weights. The final result output module is used to perform weighted fusion based on the current optimal weight and the corresponding prediction result to obtain the final deformation prediction result; The graded early warning module is used to obtain a risk assessment index based on the final deformation prediction results and to provide graded early warnings.

[0065] Furthermore, in this embodiment, the functional implementation methods of each functional module correspond one-to-one with the methods described above, and will not be repeated here.

[0066] Example 3 Based on the same inventive concept, the present invention also provides an electronic device, which includes a processor and a memory, wherein the memory stores instructions, characterized in that the instructions are loaded and executed by the processor to implement the multi-source data collaborative deformation prediction method based on deep reinforcement learning as in Embodiment 1.

[0067] Based on the same inventive concept, the present invention also provides a computer device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When the processor executes a program stored in memory, it is able to implement a multi-source data collaborative deformation prediction method based on deep reinforcement learning, as described in Example 1.

[0068] The electronic device may include a processor, a communications interface, a memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions in the memory to execute the multi-source data collaborative deformation prediction method based on deep reinforcement learning in Embodiment 1.

[0069] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0070] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0071] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-source data collaborative deformation prediction method based on deep reinforcement learning, characterized in that, include: Acquire multi-source data of the target area and preprocess it to obtain preprocessed multi-source data; The corresponding signal-to-noise ratio is obtained based on the preprocessed multi-source data; Based on the preprocessed multi-source data, after standardization, it is input into the prediction model to obtain the prediction results corresponding to each data source; Deep reinforcement learning is performed based on the historical prediction error of the prediction model and the signal-to-noise ratio as input states, and the current optimal weights are output. The final deformation prediction result is obtained by weighted fusion of the current optimal weight and the corresponding prediction result; Based on the final deformation prediction results, a risk assessment index is obtained and a graded early warning is issued.

2. The multi-source data collaborative deformation prediction method based on deep reinforcement learning according to claim 1, characterized in that, The method for preprocessing multi-source data acquisition is as follows: The multi-source data includes: raw GNSS observation data, InSAR data, and gravimeter observation data; After time and spatial alignment of the raw GNSS observation data, the InSAR data, and the gravimeter observation data, preprocessed GNSS data, preprocessed InSAR data, and preprocessed gravity data with synchronized timestamps and unified spatial coordinates are obtained, which together constitute the preprocessed multi-source data.

3. The multi-source data collaborative deformation prediction method based on deep reinforcement learning according to claim 2, characterized in that, To obtain the corresponding signal-to-noise ratio, the specific steps include: The useful signal variance and the comprehensive noise variance are obtained based on the preprocessed GNSS data. The GNSS signal-to-noise ratio is based on the ratio of the useful signal variance to the comprehensive noise variance. The deformation signal amplitude and spatial correlation noise amplitude are obtained based on the preprocessed InSAR data. The InSAR signal-to-noise ratio is based on the ratio of the deformation signal amplitude to the spatial correlation noise amplitude. Based on the preprocessed gravity data, the gravity anomaly amplitude and the total gravity noise amplitude are obtained; The gravity signal-to-noise ratio is defined as the ratio of the gravity anomaly amplitude to the total gravity noise amplitude.

4. The multi-source data collaborative deformation prediction method based on deep reinforcement learning according to claim 3, characterized in that, The methods for obtaining the useful signal variance and the comprehensive noise variance are as follows: Based on the preprocessed GNSS data, gross error removal, ionospheric and tropospheric delay correction, Kalman filtering or moving average filtering are performed to extract long-term trend terms and periodic terms reflecting crustal deformation as useful signals. The remaining random residual terms are the combined noise of receiver noise floor and multipath effect. Based on the sequence corresponding to the useful signal, the variance of the useful signal is obtained; Based on the sequence corresponding to the comprehensive noise, the variance of the comprehensive noise is obtained; The method for obtaining the amplitude of the deformation signal and the amplitude of the spatial correlation noise is as follows: The deformation signal amplitude is extracted based on the preprocessed InSAR data; The DEM error noise amplitude and atmospheric delay noise amplitude are obtained based on the preprocessed InSAR data; The spatial correlation noise amplitude is obtained based on the DEM error noise amplitude and the atmospheric delay noise amplitude.

5. The multi-source data collaborative deformation prediction method based on deep reinforcement learning according to claim 3, characterized in that, The method for obtaining the gravity anomaly amplitude and the total gravity noise amplitude is as follows: The gravity anomaly amplitude is obtained based on the preprocessed gravity data, which is used to reflect the intensity of the gravity anomaly signal associated with crustal deformation. The instrument noise amplitude is obtained based on the accuracy parameters of the gravimeter itself; The amplitude of environmental impact noise is obtained based on the residual of the preprocessed gravity data. The total gravity noise amplitude is obtained based on the instrument noise amplitude and the environmental noise amplitude.

6. The multi-source data collaborative deformation prediction method based on deep reinforcement learning according to claim 1, characterized in that, The methods for obtaining prediction results for each data source are as follows: The preprocessed GNSS data, preprocessed InSAR data, and preprocessed gravity data are respectively standardized to obtain standardized GNSS data, standardized InSAR data, and standardized gravity data. Based on the standardized GNSS data, the standardized InSAR data, and the standardized gravity data, respectively, the prediction model is input to obtain the first prediction result, the second prediction result, and the third prediction result. The first prediction result includes a first deformation rate prediction result and a first cumulative deformation prediction result; The second prediction result includes the second deformation rate prediction result and the second cumulative deformation prediction result; The third prediction result includes the third deformation rate prediction result and the third cumulative deformation prediction result.

7. The multi-source data collaborative deformation prediction method based on deep reinforcement learning according to claim 1, characterized in that, The deep reinforcement learning specifically includes: Based on the input state, the data is fed into the deep reinforcement learning controller, and the output action is a set of weight adjustment instructions. The deep reinforcement learning controller learns a set of weight adjustment strategies that maximize long-term cumulative rewards through continuous interaction with the prediction model and the environment, and outputs the current optimal weights. The reward function R of the deep reinforcement learning controller is: R=α×(1-w)+β×(1-c); Where α represents the accuracy weight, w represents the current prediction error, β represents the timeliness weight, and c represents the response delay; Heuristic constraints are introduced in the deep reinforcement learning controller: When the corresponding signal-to-noise ratio is greater than or equal to the first threshold, the corresponding weight is increased; When the corresponding signal-to-noise ratio is less than the second threshold, the corresponding weight is reduced.

8. The multi-source data collaborative deformation prediction method based on deep reinforcement learning according to claim 6, characterized in that, The method for obtaining the final deformation prediction result is as follows: The current optimal weights include: GNSS optimal weights, InSAR optimal weights, and gravity optimal weights; The crustal deformation rate at the current moment is obtained by multiplying the GNSS optimal weight, the InSAR optimal weight, and the gravity optimal weight with the first deformation rate prediction result, the second deformation rate prediction result, and the third deformation rate prediction result, respectively. The cumulative deformation at the current time is obtained by multiplying the GNSS optimal weight, the InSAR optimal weight, and the gravity optimal weight by the first cumulative deformation prediction result, the second cumulative deformation prediction result, and the third cumulative deformation prediction result, respectively. The current crustal deformation rate and the current cumulative deformation are used together as the final deformation prediction result.

9. The multi-source data collaborative deformation prediction method based on deep reinforcement learning according to claim 8, characterized in that, Obtain a risk assessment index and implement tiered early warning systems, specifically including: Based on the current crustal deformation rate and the current cumulative deformation, data normalization is performed to obtain the normalized crustal deformation rate and the normalized cumulative deformation; The Risk Assessment Index (CSI) is derived based on the normalized crustal deformation rate and the normalized cumulative deformation. Set the first preset value Q1 and the second preset value Q2; When CSI < Q1, it is considered a normal state and no warning is issued; When Q1≤CSI<Q2, a yellow alert is issued and monitoring is automatically intensified; When CSI≥Q2, a red alert is issued, an emergency response is initiated, and corresponding emergency response recommendations are output.

10. A multi-source data collaborative deformation prediction system based on deep reinforcement learning, used to execute the multi-source data collaborative deformation prediction method based on deep reinforcement learning as described in any one of claims 1-9, characterized in that, include: The system includes a multi-source data processing module, a signal-to-noise ratio acquisition module, an initial result output module, an optimal weight acquisition module, a final result output module, and a graded early warning module. The multi-source data processing module is used to acquire multi-source data of the target area and perform preprocessing to obtain preprocessed multi-source data; The signal-to-noise ratio (SNR) acquisition module is used to acquire the corresponding SNR based on the preprocessed multi-source data; The initial result output module is used to input the standardized preprocessed multi-source data into the prediction model to obtain the prediction results corresponding to each data source. The optimal weight acquisition module is used to perform deep reinforcement learning based on the historical prediction error of the prediction model and the signal-to-noise ratio as input states, and output the current optimal weights. The final result output module is used to obtain the final deformation prediction result by weighted fusion of the current optimal weight and the corresponding prediction result; The graded early warning module is used to obtain a risk assessment index based on the final deformation prediction result and to conduct graded early warning.

Images