Power system geological disaster identification monitoring and early warning method based on multi-modal data fusion

Abstract

This invention provides a multimodal data fusion method for identifying, monitoring, and warning of geological disasters in power systems, belonging to the field of power system safety monitoring technology. By collecting multimodal observation data from the land surface, remote sensing, microseismic data, meteorological precipitation, and historical disaster data, a geological disaster perception space is constructed. A unified spatiotemporal framework is built to complete the spatiotemporal alignment and structured representation of the data, generating a standard dataset. Features from each modality are extracted and adaptively weighted and fused, relying on the fused features to achieve geological disaster identification and dynamic early warning. This method solves the problems of existing data having single dimensions, insufficient spatiotemporal matching, and static fusion mechanisms. It can integrate multi-source heterogeneous data, perform cross-modal adaptive fusion, and dynamically output risks, significantly improving the accuracy of disaster identification and the timeliness of early warning.

Description

Technical Field

[0001] This application relates to the field of power system safety monitoring technology, and in particular to a method for identifying, monitoring and early warning of geological disasters in power systems through multimodal data fusion. Background Technology

[0002] As a core infrastructure supporting socio-economic operations, the safe and stable operation of the power system directly affects the reliability of regional power supply and the resilience of the power grid. Under the combined influence of complex terrain, geological tectonic activity, meteorological and hydrological changes, and human engineering activities, areas surrounding key power system equipment and transmission lines are susceptible to geological hazards such as landslides, collapses, debris flows, ground subsidence, and ground fissures. These hazards are characterized by their highly concealed gestation process, highly complex triggering mechanisms, unpredictable evolution speed, and the potential for cascading effects. They often cause severe consequences in a short period, including structural instability, deformation of structures, tilting of towers or frames, damage to roads and transmission lines, and loss of electrical safety distances. This can lead to line faults, unplanned equipment outages, and even large-scale power outages, and may even trigger secondary disasters and safety accidents. Therefore, constructing a geological hazard monitoring and early warning system with high-precision identification capabilities, strong interpretability, and continuous adaptability has become a key technical requirement for ensuring the safe operation of the power system.

[0003] Current practices in power system geological hazard identification and early warning mainly rely on three technical approaches: engineering experience-based discrimination methods based on single-point sensors, statistical learning models based on limited sensor data, and surface change detection technologies based on remote sensing imagery. These methods generally suffer from limitations such as limited data dimensions and incomplete information coverage. Furthermore, they often employ offline analysis and static threshold alarm mechanisms, making it difficult to continuously and dynamically depict the gestation and evolution of hazard events. Specifically, some solutions utilize only field sensors such as displacement gauges, inclinometers, crack gauges, and rain gauges for localized monitoring, triggering alarms through preset fixed thresholds. While these methods can play a limited role in areas with dense sensor deployment and stable geological conditions, the significant spatial heterogeneity and cross-scale impact of geological hazards mean that point sensors can only capture local response information, failing to effectively cover the spatial evolution of upstream catchment areas, overall slope stability changes, and surface deformation at the transmission line level. When sensors drift, malfunction, experience data loss, or are deployed off-center from critical deformation zones, false alarms or missed alarms are highly likely, making it difficult to meet the stringent requirements of the power grid for high-reliability identification and continuous monitoring.

[0004] Another type of technology focuses on using optical remote sensing, synthetic aperture radar, and interferometry to acquire information on surface deformation and cover changes for identifying potential disaster areas. Although remote sensing technology has advantages such as wide coverage and recurring observation periods, it still has significant shortcomings in the refined monitoring of power system equipment and transmission channels. Optical remote sensing data is easily affected by cloud and rain weather, topographic shadows, and seasonal vegetation cover changes, severely compromising the continuity and consistency of time-series observation data. While synthetic aperture radar and interferometry have some cloud and fog penetration capabilities, they are often limited by factors such as coherence attenuation, geometric distortion, phase unwrapping errors, and changes in ground object scattering characteristics, often introducing noise interference and spurious deformation signals. Furthermore, remote sensing observations are constrained by satellite orbits and revisit periods, making it difficult to match the temporal resolution to the real-time operation and maintenance needs of power systems. Under conditions of extreme rainfall or sudden geological disturbances, a disaster can rapidly develop from initial incubation to structural instability within hours. Relying solely on periodic remote sensing data cannot provide continuous and stable criteria for judgment, let alone support high-frequency monitoring and immediate early warning response for critical transmission channels.

[0005] Power system operational status data includes crucial information such as load fluctuations, power flow changes, protection device action signals, equipment vibration spectra, and environmental monitoring parameters. This data implicitly reflects abnormal response patterns of equipment and lines under the influence of geological disasters. For example, foundation deformation may lead to increased vibration of electrical equipment, structural displacement may cause changes in insulation distance, and abnormal environmental parameters often foreshadow accumulated disaster risks. However, in the existing technological system, operational data, remote sensing imagery, and field monitoring data have long been isolated, lacking a unified spatiotemporal benchmark framework and cross-modal correlation modeling mechanism. This makes it impossible to effectively uncover the intrinsic coupling relationship between operational anomalies and the evolution of geological disasters. This data silo phenomenon severely restricts the construction of comprehensive judgment capabilities for power system safety risks, making it difficult to effectively connect early warning results with operation and maintenance response strategies.

[0006] From an algorithmic architecture perspective, the integration of existing technologies with deep learning is clearly insufficient. Most methods still rely on traditional machine learning models or empirical threshold rules, with feature extraction heavily dependent on manual design, making it difficult to adapt to the nonlinear, multi-source coupled feature representation requirements of complex geological environments. Even with the introduction of deep learning frameworks, existing implementations are often limited to shallow feature processing of single-modal data, such as performing convolution operations or simple classification and segmentation on remote sensing images, lacking a systematic fusion architecture design for multimodal data. In power system scenarios, multi-source data exhibit significant heterogeneity: remote sensing images are primarily spatial raster structures, field monitoring data exhibits time-series characteristics, operational data often shows event-triggered or periodic sampling patterns, and geological and topographical data possess static spatial attributes. Without strict spatiotemporal alignment mechanisms and consistency processing strategies, simple feature stitching or coarse fusion will inevitably introduce erroneous associations, not only weakening model learning performance but also reducing the credibility and interpretability of engineering applications.

[0007] Existing fusion methods generally employ static feature stitching or fixed weighting strategies, failing to dynamically adjust the contribution of each mode based on the disaster's development stage, changes in the observation environment, and fluctuations in data quality. For example, under conditions of heavy rainfall and cloud cover, the effective information of optical remote sensing modes decays sharply, while ground-based sensing and synthetic aperture radar deformation modes should bear higher weights. When on-site monitoring equipment fails or data is abnormal, remote sensing imagery and geological background modes need to play a dominant role. If the fusion mechanism lacks adaptive response capabilities, the model will be unable to dynamically allocate the reliability of multi-source observations, leading to unstable judgment results or even systematic misjudgments, severely impacting the reliable execution of early warning strategies. Furthermore, geological disaster risk identification requires not only single-time judgment capabilities but also attention to the continuous evolution trend of risks and the stability of early warnings. Existing technologies mostly output discrete-time classification results or isolated alarm signals, lacking a mechanism design to integrate continuous observation data into dynamic risk accumulation, making it difficult to guarantee the temporal continuity of early warning outputs and operational operability. In engineering practice, operation and maintenance departments urgently need to understand whether the risk level is continuously escalating, whether there are phased acceleration characteristics, and whether the risk level has reached the critical point for triggering response strategies. If the warning results fluctuate drastically in adjacent moments, it will make it difficult to implement the response strategy and may even lead to alarm fatigue.

[0008] From a closed-loop system perspective, existing technologies lack a unified optimization objective function to achieve end-to-end overall optimization from data processing, feature extraction, cross-modal coupling to risk output. Without clear optimization constraints, multimodal feature extraction and fusion modules are prone to learning local correlations or noisy patterns, making it difficult to guarantee overall recognition performance and long-term generalization ability. Furthermore, during long-term system operation, data distribution may shift due to seasonal changes, land cover variations, sensor performance drift, and adjustments to operation and maintenance strategies. The lack of a sustainable optimization mechanism will lead to significant performance degradation over time, failing to meet the application requirements of power systems for long-term online monitoring and continuous early warning.

[0009] In summary, existing technologies for identifying and monitoring geological hazards in power systems face several core challenges, including insufficient multimodal data coverage, fragmented use of operational data, remote sensing data, and field monitoring data, weak spatiotemporal consistency processing capabilities, lack of cross-modal coupling mechanisms, deep learning fusion limited to shallow single-modal processing, insufficient dynamic risk evolution modeling capabilities, and an imperfect system-level closed-loop optimization mechanism. These deficiencies make it difficult for existing solutions to achieve stable, high-precision, and sustainable geological hazard identification and risk assessment under complex geological environments and variable weather conditions. Therefore, there is an urgent need to develop a novel monitoring and early warning method that can integrate multi-source heterogeneous observation data, construct a unified spatiotemporal reference framework, achieve cross-modal adaptive fusion, and possess dynamic risk output and overall optimization capabilities.

[0010] To address the aforementioned issues, existing technologies urgently need improvement. Summary of the Invention

[0011] The purpose of this application is to provide a method for identifying, monitoring and early warning of geological disasters in power systems based on multimodal data fusion. This method has the advantages of effectively integrating multi-source heterogeneous observation data, constructing a unified spatiotemporal reference framework, realizing cross-modal adaptive fusion and having dynamic risk output capability, thereby significantly improving the accuracy of geological disaster identification and the timeliness of early warning.

[0012] The multimodal data fusion method for identifying, monitoring, and providing early warning of geological disasters in power systems, as provided in this application, adopts the following technical solution: A method for identifying, monitoring, and providing early warning of geological hazards in power systems through multimodal data fusion, comprising: To acquire multimodal raw observation data to construct a multimodal geological disaster perception space, the multimodal raw observation data includes at least surface condition data, remote sensing observation data, microseismic monitoring structure data, meteorological and precipitation data, and historical disaster data; A unified spatiotemporal reference framework is established to perform spatiotemporal alignment and structural characterization on the original multimodal observation data, generating a standardized multimodal dataset; Feature extraction and adaptive weighted fusion are performed on the modal data in the standardized multimodal dataset to obtain fused features; Geological disaster identification and dynamic early warning are based on the fused features.

[0013] Optionally, the surface condition data includes surface deformation data, surface displacement data, and surface strain data; The remote sensing observation data includes optical remote sensing image data and synthetic aperture radar (SAR) time series data. The microseismic monitoring structural data includes microseismic event location data, source mechanism parameters, and structural response monitoring data. The meteorological rainfall data includes rainfall amount, rainfall intensity, and cumulative rainfall data; The historical disaster data includes records of historical disaster events, disaster trigger threshold parameters, and disaster evolution patterns.

[0014] Optionally, the spatiotemporal alignment of the original multimodal observation data includes: Using the location coordinates of power system facilities as a spatial reference and combining the timestamp information of multimodal data acquisition, a unified spatiotemporal reference system is constructed. To address the spatial resolution differences among different modal data, spatial resampling and georegistration techniques are employed to map each modal data onto a unified spatial grid. To address the temporal frequency differences among different modal data, a time interpolation and alignment algorithm is employed to align the time series of each modal data to a unified time node. And the spatiotemporally aligned modal data are organized according to spatial location and time dimension to form the standardized multimodal dataset.

[0015] Optionally, the structural characterization includes: For surface state data, surface deformation field interpolation and mechanical parameter transformation are used to generate a surface state characterization vector with a unified structure. For remote sensing observation data, standardized remote sensing image representation tensors are generated through radiometric correction, geometric correction, and time phase normalization. For microseismic monitoring structural data, a structural state characterization vector is generated through source parameter inversion and structural response feature extraction. A meteorological environment characterization matrix is ​​generated by using time-series accumulation and spatial distribution interpolation of meteorological precipitation data; For historical disaster data, a knowledge representation vector of historical disasters is generated by encoding disaster events and parameterizing triggering conditions.

[0016] Optionally, the feature extraction of each modality data in the standardized multimodal dataset includes: An independent feature extraction channel is constructed for each modality of data. Based on the channel attention mechanism, the importance weights of the features of each modality channel are weighted and learned to obtain the high-level semantic feature representation of each modality.

[0017] Optionally, the adaptive weighted fusion includes: Based on the acquisition quality, data integrity, and historical reliability of each modality data at the current moment, calculate the initial confidence score for each modality; Based on the initial confidence score, adaptive fusion weights for each modality feature are dynamically generated through a learnable gating network. During the fusion process, when a certain modality's data is missing or its confidence level is lower than a preset threshold, the fusion weight of that modality is automatically reduced, and the weight contribution of other reliable modalities is correspondingly increased. The weighted modal features are then deeply fused to obtain the fused features.

[0018] Optionally, the geological hazard identification based on the fused features includes: The fused features are input into the geological disaster identification model to identify the types, levels, and impact range of geological disasters around power system facilities.

[0019] Optionally, the dynamic early warning includes: The comprehensive disaster risk index is calculated based on the geological disaster identification results, and the current warning level is determined by combining the preset multi-level warning threshold range; The early warning level is dynamically adjusted based on the disaster evolution trend prediction results; Generate early warning information that includes disaster risk level, affected area, and evolution trend.

[0020] Optionally, it also includes: based on the feedback results of the early warning information and the prediction deviation of the identification model, constructing a system-level optimization objective function that includes a geological disaster identification loss term, an early warning consistency loss term, and a spatiotemporal smoothing loss term, and performing end-to-end collaborative optimization on the network parameters of the geological disaster identification model and the weight parameters of the adaptive weighted fusion, so as to realize the adaptive update of the model to the dynamic changes in the reliability of multi-source observation data.

[0021] Optionally, the method is applied to the monitoring of the surrounding area of ​​power system facilities, including transmission lines, substations or power utility tunnels. The multimodal raw observation data is obtained by multi-source collaborative collection of surface conditions, remote sensing images, microseismic activity, meteorological environment and historical disaster information within a preset monitoring range around the power system facilities.

[0022] In summary, this application constructs a multimodal geological hazard perception space by acquiring multimodal raw observation data. The multimodal raw observation data includes at least surface condition data, remote sensing observation data, microseismic monitoring structural data, meteorological and precipitation data, and historical hazard data. A unified spatiotemporal reference framework is established to perform spatiotemporal alignment and structural characterization on the multimodal raw observation data, generating a standardized multimodal dataset. Feature extraction and adaptive weighted fusion are performed on the modal data within the standardized multimodal dataset to obtain fused features. Geological hazard identification and dynamic early warning are then performed based on these fused features. This approach achieves unified spatiotemporal benchmark processing and dynamic weight allocation for multi-source heterogeneous data, effectively overcoming the problems of single data dimension, insufficient spatiotemporal alignment, and static fusion mechanisms in existing technologies. It possesses the advantages of effectively integrating multi-source heterogeneous observation data, constructing a unified spatiotemporal reference framework, achieving cross-modal adaptive fusion, and having dynamic risk output capabilities, thereby significantly improving the accuracy of geological hazard identification and the timeliness of early warning. Attached Figure Description

[0023] Figure 1 This is a flowchart illustrating the first embodiment of the power system geological disaster identification, monitoring, and early warning method based on multimodal data fusion of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this application.

[0025] Traditional power system geological disaster identification and early warning technologies generally suffer from limitations such as single data sources, insufficient information dimension coverage, reliance on offline analysis and threshold-based alarms, and difficulty in continuously depicting the disaster's gestation and evolution process. Point-based sensors cannot provide global coverage, and remote sensing data is affected by the environment and lacks sufficient temporal resolution. Operational data and monitoring data are fragmented, lacking a unified spatiotemporal reference and cross-modal correlation. Existing algorithms suffer from insufficient deep fusion, poor handling of multimodal data heterogeneity, and fixed fusion weights that are difficult to adaptively adjust. Early warning results lack continuity and stability, and the absence of a system-level closed-loop optimization mechanism leads to low identification accuracy, unreliable early warnings, and difficulty in continuously updating the model.

[0026] To address this, this application provides a method for identifying, monitoring, and issuing early warning of geological disasters in power systems based on multimodal data fusion, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the power system geological disaster identification, monitoring, and early warning method based on multimodal data fusion of this application.

[0027] In this embodiment, the power system geological disaster identification, monitoring, and early warning method based on multimodal data fusion includes the following steps: Step S10: Acquire multimodal raw observation data to construct a multimodal geological disaster perception space. The multimodal raw observation data includes at least surface condition data, remote sensing observation data, microseismic monitoring structure data, meteorological and precipitation data, and historical disaster data.

[0028] For ease of understanding, the following explains some key terms in this embodiment: Multimodal raw observation data refers to a collection of raw data reflecting the geological environment surrounding power system facilities, acquired from different sources and through different sensors or observation methods. It encompasses multiple dimensions, including surface conditions, remote sensing observations, microseismic activity, meteorological precipitation, and historical disasters, aiming to provide comprehensive geological hazard information input.

[0029] Multimodal geological hazard perception space refers to a virtual information space constructed by integrating multimodal raw observation data, which can comprehensively reflect the formation, occurrence, and evolution of geological hazards. This space provides a unified background and context for subsequent data processing and analysis.

[0030] A unified spatiotemporal reference frame refers to providing a common temporal and spatial coordinate system for multimodal data from different sources and in different formats. By establishing this framework, it is possible to ensure that all data are comparable and consistent in time and space, laying the foundation for subsequent data fusion.

[0031] Spatiotemporal alignment refers to the process of calibrating and matching data from different modalities in both time and spatial dimensions. Its purpose is to eliminate data inconsistencies caused by differences in acquisition time, sampling frequency, spatial resolution, etc., enabling the effective integration of heterogeneous data within a unified framework.

[0032] Structural representation refers to the process of converting raw, heterogeneous, multimodal data into a unified, structured data format or feature vector. Through structural representation, key information related to geological hazards can be extracted from each modality of data, making it computable and fusionable.

[0033] A standardized multimodal dataset refers to a collection of data where all modal data meet consistency requirements in format, time, and space after a unified spatiotemporal reference frame has been established, spatiotemporally aligned, and structurally represented. This dataset serves as the direct input for feature extraction and fusion.

[0034] Feature extraction refers to the process of identifying and quantifying representative and discriminative information from raw or structured data. In geological hazard identification, feature extraction aims to extract deep patterns closely related to hazard type, severity, and evolution trend from various modal data.

[0035] Adaptive weighted fusion refers to the process of dynamically adjusting the contribution weight of each modality of data in the fusion process based on its real-time quality, reliability, or importance, and then integrating the weighted features. This mechanism aims to improve the robustness and accuracy of the fusion results, especially when some data is of poor quality or missing.

[0036] Fusion features refer to comprehensive feature representations that integrate the advantages of multimodal data after adaptive weighted fusion processing. These features contain richer and more comprehensive geological hazard discrimination information than single-modal data.

[0037] Geological hazard identification refers to the process of determining whether there are geological hazards around power system facilities based on fusion characteristics and model analysis, and further determining the type, level and scope of impact of the hazard.

[0038] Dynamic early warning refers to the process of adjusting early warning levels and generating early warning information in real time based on geological hazard identification results and disaster evolution trend predictions. Its aim is to provide continuous, real-time risk assessment to support power system operation and maintenance departments in taking timely countermeasures.

[0039] It should be noted that the surface condition data includes surface deformation data, surface displacement data, and surface strain data; The remote sensing observation data includes optical remote sensing image data and synthetic aperture radar (SAR) time series data. The microseismic monitoring structural data includes microseismic event location data, source mechanism parameters, and structural response monitoring data. The meteorological rainfall data includes rainfall amount, rainfall intensity, and cumulative rainfall data; The historical disaster data includes records of historical disaster events, disaster trigger threshold parameters, and disaster evolution patterns.

[0040] In practical implementation, a geological hazard perception space is constructed with the monitored object s as the core, and its impact range is defined as the spatial domain. ,in, This represents the spatial extent relevant to the safe operation of the power system and the effects of geological hazards. Simultaneously, a unified analysis time domain is defined to describe the gestation, development, and evolution of geological hazards, expressed as: In the spatial domain With the time domain Within the system, multi-source heterogeneous observation data are collected to form a multimodal dataset for monitoring geological disasters in the power system, which is represented as follows: Among them, superscript These represent operational mode, optical remote sensing, SAR deformation, environmental and structural sensing, topographic and geological, meteorological, and historical disaster prior modes, respectively.

[0041] Through the above technical solution, this embodiment clarifies the specific composition of multimodal raw observation data, ensuring the relevance and effectiveness of the collected data. This detailed data classification and definition enables subsequent spatiotemporal alignment and structural characterization processes to more accurately process different types of data and extract key features closely related to geological hazards. For example, surface deformation, displacement, and strain data can comprehensively characterize the deformation state of geological bodies; optical remote sensing and SAR time-series data monitor surface changes from different dimensions; microseismic data reveals internal activities of geological bodies; meteorological and precipitation data provides external inducing factors; and historical disaster data provides empirical knowledge and triggering conditions. This multi-dimensional and multi-level refined data input significantly improves the quality and information content of fused features, enabling geological hazard identification models to more accurately determine the type, level, and impact range of hazards, and achieve more timely and reliable dynamic early warnings, providing a solid data foundation and decision support for the safe operation of power system facilities.

[0042] Step S20: Establish a unified spatiotemporal reference framework, perform spatiotemporal alignment and structural characterization on the original multimodal observation data, and generate a standardized multimodal dataset.

[0043] In the process of spatiotemporally aligning raw multimodal observation data to generate standardized multimodal datasets, significant differences exist between different modal data in terms of acquisition sources, sensor types, sampling frequencies, and spatial coverage, resulting in inconsistent spatial resolution and temporal frequencies. Furthermore, the lack of a unified spatiotemporal benchmark makes direct data fusion and subsequent analysis difficult, affecting the accuracy and real-time performance of geological disaster identification and early warning.

[0044] To address this, this embodiment further proposes a specific method for spatiotemporal alignment of the original multimodal observation data, including: constructing a unified spatiotemporal reference system using the location coordinates of power system facilities as a spatial reference and combining the timestamp information of multimodal data acquisition; mapping each modal data to a unified spatial grid by employing spatial resampling and georegistration techniques to address the spatial resolution differences of different modal data; aligning the time series of each modal data to a unified time node by employing temporal interpolation and alignment algorithms to address the temporal frequency differences of different modal data; and organizing the spatiotemporally aligned modal data according to spatial location and time dimension to form the standardized multimodal dataset.

[0045] Through the above technical solutions, this embodiment effectively solves the heterogeneity problem of multimodal raw observation data in terms of spatial resolution and temporal frequency by establishing a unified spatiotemporal reference system. Specifically, using power system facilities as a spatial benchmark and combining timestamp information, a common reference system is provided for all data, ensuring data consistency. Spatial resampling and georegistration techniques enable observation data at different spatial scales to be accurately mapped to a unified grid, eliminating spatial misalignment. Temporal interpolation and alignment algorithms compensate for the differences in acquisition frequency between different modal data, achieving synchronization of time series. Finally, the spatiotemporally aligned data is structured and organized to form a standardized multimodal dataset, greatly simplifying the complexity of subsequent feature extraction and fusion, and significantly improving the accuracy, reliability, and real-time response capability of geological disaster identification and dynamic early warning.

[0046] It should be noted that the structural representation includes: for surface state data, surface deformation field interpolation and mechanical parameter transformation are used to generate a surface state representation vector with a unified structure; for remote sensing observation data, radiometric correction, geometric correction and time phase normalization are used to generate a standardized remote sensing image representation tensor; for microseismic monitoring structural data, source parameter inversion and structural response feature extraction are used to generate a structural state representation vector; for meteorological and precipitation data, time series accumulation and spatial distribution interpolation are used to generate a meteorological environment representation matrix; for historical disaster data, disaster event coding and trigger condition parameterization are used to generate a historical disaster knowledge representation vector.

[0047] Through the above technical solutions, this embodiment employs customized structural characterization methods for different modalities of raw observation data, including surface state data, remote sensing observation data, microseismic monitoring structural data, meteorological and precipitation data, and historical disaster data. Specifically, surface deformation field interpolation and mechanical parameter transformation can convert discrete surface state information into a unified mechanical characterization vector, revealing the internal stress and deformation state of geological bodies; radiometric correction, geometric correction, and time-phase normalization ensure the comparability and standardization of remote sensing images across different time and space, facilitating the extraction of macroscopic surface change characteristics; source parameter inversion and structural response feature extraction can extract key information on internal geological activity and structural damage from microseismic data; temporal accumulation and spatial distribution interpolation enable meteorological and precipitation data to comprehensively reflect their triggering impact on geological disasters; and disaster event coding and trigger condition parameterization transform historical experience into quantifiable knowledge. These targeted structural characterization methods effectively overcome the challenges posed by the heterogeneity of multimodal data, transforming raw and complex observation data into a standardized characterization form that is structurally unified, semantically clear, and easily computationally processed. This not only lays a solid foundation for subsequent feature extraction and adaptive weighted fusion, significantly improving the effectiveness and fusionability of deep features of various modal data, but also avoids the decline in recognition and early warning accuracy caused by inconsistent data formats or information loss, thereby ensuring the accuracy and reliability of geological disaster recognition and dynamic early warning.

[0048] In practice, after completing multimodal data acquisition, data from different sources, scales, and sampling frequencies are uniformly organized and reconstructed. By establishing a unified spatiotemporal reference framework, remote sensing image data, sensor time-series data, and power operation status data are aligned in time series and registered spatially, ensuring a one-to-one correspondence between various data types within the same time node and spatial unit. Simultaneously, multimodal data undergoes standardization and scale-consistent mapping to reduce the uncertainties introduced by spatial resolution differences, inconsistent physical dimensions, and observation noise, thereby constructing a multimodal feature representation foundation with good spatiotemporal consistency.

[0049] Specifically, due to differences in sampling frequency, time reference, and spatial reference frame among different modal data, it is necessary to uniformly organize and spatiotemporally consistentize multimodal data. Regarding the first... Modal data First, time alignment is achieved using a time consistency operator, which is represented as: in, Indicates the first The temporal alignment operator corresponding to the modality. Subsequently, the temporally aligned data is mapped to a unified spatial grid using a spatial registration operator, expressed as: in, Indicates the first The time alignment operator corresponding to the modality.

[0050] Through the above time and space consistency processing, all modal data are made to be at the same time node. and spatial location A one-to-one correspondence is formed.

[0051] Step S30: Perform feature extraction and adaptive weighted fusion on each modality data in the standardized multimodal dataset to obtain fused features.

[0052] Because different modal data are heterogeneous, diverse, and may have redundant or missing information, if a unified or simple feature extraction method is used, it may not be able to fully capture the unique information and deep correlations of each modal data, resulting in insufficient feature expression ability. This will affect the effectiveness of subsequent multimodal data fusion and reduce the accuracy and robustness of geological disaster identification and early warning.

[0053] In this regard, this embodiment further proposes that the feature extraction of each modality data in the standardized multimodal dataset includes: constructing an independent feature extraction channel for each modality data, and performing weighted learning of the importance weights of the features of each modality channel based on the channel attention mechanism to obtain the high-level semantic feature representation of each modality.

[0054] By constructing independent feature extraction channels for each modality of data using the above technical solution, the heterogeneity of different modalities can be fully considered, avoiding the limitations of a single general-purpose model in processing all modalities and enabling the effective capture of unique information from each modality. Based on this, a channel attention mechanism is used to weight the importance weights of features in each modality channel, allowing the model to dynamically adjust the level of attention given to different modalities according to the characteristics of the data and the current task requirements. This effectively suppresses interference from redundant or low-quality modal information and enhances the expression of key modal information. Ultimately, the resulting high-level semantic feature representation not only possesses stronger abstraction and discriminative capabilities but also better reflects the deep mechanisms and evolutionary patterns of geological hazards. This significantly improves the quality and efficiency of multimodal data fusion, thereby enhancing the accuracy of geological hazard identification and the reliability of dynamic early warning, enabling the system to more accurately assess hazard risks and issue timely warnings.

[0055] It should be noted that the adaptive weighted fusion includes: calculating the initial confidence score of each modality based on the acquisition quality, data completeness, and historical reliability of the data at the current moment; dynamically generating adaptive fusion weights for each modality feature based on the initial confidence score through a learnable gating network; during the fusion process, when a modality's data is missing or its confidence score is lower than a preset threshold, automatically reducing the fusion weight of that modality and correspondingly increasing the weight contribution of other reliable modalities; and performing deep fusion on the weighted modality features to obtain the fused feature.

[0056] Through the above technical solution, this embodiment can dynamically assess the reliability of each modality's data based on its actual acquisition quality, data completeness, and historical reliability at the current moment, and adaptively adjust the weights of each modality's features in the fusion process accordingly. This mechanism effectively solves the challenges brought about by the dynamic changes in the quality, completeness, and reliability of different modal data, and avoids the adverse effects on the overall identification and early warning performance caused by the decline or loss of quality of a single modality's data. Specifically, when a modality's data is abnormal or its confidence level is lower than a preset threshold, the system can intelligently reduce its weight contribution and enhance the role of other reliable modalities, thereby ensuring the robustness and accuracy of the fused features. This significantly improves the reliability and adaptability of geological disaster identification and dynamic early warning, enabling the early warning system to maintain efficient and stable operation in complex and ever-changing environments, and providing a more solid technical guarantee for the safe operation of power system facilities.

[0057] In practical implementation, after the multimodal data undergoes spatiotemporal consistency processing, structured feature representations are reconstructed based on the physical attributes and information characteristics inherent in different modal data. For remote sensing optical data, the focus is on characterizing land cover changes, landform evolution, and surface anomalies; for synthetic aperture radar data, the focus is on extracting surface deformation information and the temporal evolution characteristics of minute displacements; for environmental and structural sensor data, the focus is on describing the local structural response of the power system, changes in foundation stability, and environmental disturbance characteristics; and for power operation data, the focus is on reflecting the intrinsic response relationship between equipment operating status and abnormal operating conditions. Through these differentiated feature representation methods, each modal data retains its own physical meaning while forming a feature representation with a unified expression form.

[0058] Specifically, after completing the spatiotemporal consistency processing, a mode-specific feature mapping function is introduced to reconstruct the structured representation of the original data, based on the physical attributes and information characteristics contained in different modal data. The feature representation of modal data is defined as follows: in, Indicates that for the first A nonlinear feature mapping function for modal design. This step transforms the raw observation data into feature representations with a uniform expression form while retaining physical meaning.

[0059] Building upon the reconstruction of multimodal feature representations, a multi-channel convolutional neural network is introduced to process various modal features in parallel. Independent feature extraction channels are set up for each type of modality data, and spatial and temporal evolution features related to geological hazards are extracted step-by-step through multi-layer convolutional operations, thereby forming a high-level semantic feature representation with high discriminative power. In this process, each channel, while maintaining sufficient modeling of its own modality information, provides a structurally consistent and scale-alignable feature foundation for subsequent cross-modal feature fusion.

[0060] Specifically, after completing the reconstruction of multimodal feature representations, a multi-channel convolutional neural network is constructed to process different modal features in parallel. For the first... Modal features, with independent feature extraction channels set up, in the first... The feature extraction process of a layer is represented as follows: in, , and They represent the first Layer convolution kernel parameters and bias terms, This represents a non-linear activation function.

[0061] After obtaining the high-level features output from each modal channel, a cross-modal feature coupling mechanism is constructed to jointly model the features from different modalities. This process does not employ simple feature concatenation; instead, it uses an adaptive weight allocation and feature response mapping mechanism to characterize the relative contributions of different modal features in the geological hazard identification task. This enables the system to automatically learn the intrinsic correlation structure between multimodal data based on changes in the hazard development stage and observation conditions. Through cross-modal feature fusion, a holistic understanding of multi-source geological hazard observation information is achieved.

[0062] Specifically, after obtaining the high-level feature representations of each modality, cross-modal coupling modeling is performed on the features of different modalities. First, the first... Modal in spatiotemporal location The response value at that location is represented as: in, This represents the learnable modal response parameter vector. Subsequently, adaptive fusion weights for each modality are obtained through normalized mapping, expressed as: Based on the above weights, the multimodal features are weighted and fused to obtain a joint feature representation: .

[0063] Step S40: Geological disaster identification and dynamic early warning based on the fused features.

[0064] It should be noted that the geological hazard identification based on the fused features includes: inputting the fused features into the geological hazard identification model to identify the type, level and scope of geological hazards around power system facilities.

[0065] It should be noted that the dynamic early warning includes: calculating a comprehensive disaster risk index based on the geological disaster identification results, determining the current early warning level by combining it with a preset multi-level early warning threshold range; dynamically adjusting the early warning level based on the disaster evolution trend prediction results; and generating early warning information that includes the disaster risk level, affected area, and evolution trend.

[0066] It should be noted that this embodiment also includes: based on the feedback results of the early warning information and the prediction deviation of the identification model, constructing a system-level optimization objective function that includes a geological disaster identification loss term, an early warning consistency loss term, and a spatiotemporal smoothing loss term, and performing end-to-end collaborative optimization on the network parameters of the geological disaster identification model and the weight parameters of the adaptive weighted fusion, so as to realize the adaptive update of the model to the dynamic changes in the reliability of multi-source observation data.

[0067] In practical implementation, after completing cross-modal feature fusion, the resulting joint features are input into the geological hazard identification model for discriminant analysis. This model infers the hazard status of the area surrounding the power system based on a deep feature space, outputting the probability distribution of hazard occurrence or the hazard type identification result. Through continuous updates to the identification results, dynamic identification and evolutionary perception of geological hazard risks are achieved, providing reliable data support for subsequent early warning issuance, operational adjustments, and prevention and control decisions.

[0068] The combined features after fusion The input is fed into the geological hazard identification model, and the output is the probability of geological hazard occurrence, which is expressed as follows: in, Indicates a location in time and space. The probability of a geological disaster occurring at a certain location. The sigmoid activation function maps any real number to the range 0-1. Let be the weight vector, and b be the bias term, which adjusts the overall baseline level of the model output. To characterize the continuous evolution of geological hazard risk, a dynamic risk accumulation is introduced, defined as: in, This represents the cumulative dynamic risk at that moment. The risk memory factor is used to characterize the duration of a disaster's impact over a given time scale. Representing the time domain The moment before going up, The total amount represents the cumulative dynamic risk at the previous moment.

[0069] Based on the dynamic risk accumulation results, the geological hazard status of the power system and its surrounding areas is identified, and corresponding hazard risk levels or early warning information are output. During the model training phase, the model parameters are adaptively optimized by minimizing the difference between the predicted results and the actual hazard status, thereby continuously improving the accuracy of geological hazard identification and the robustness of the system.

[0070] Specifically, based on the results of dynamic risk accumulation This function identifies the geological hazard status of the power system and its surrounding areas, and outputs the corresponding hazard risk level or early warning result. The hazard risk level determination function is defined as follows: in, This represents a risk mapping function used to map continuous risk accumulation values ​​to discrete risk levels or warning states.

[0071] The risk mapping function can be expressed in a threshold determination form: in, and For risk assessment thresholds that are pre-set or adaptively updated, different risk levels correspond to different operational handling or early warning strategies.

[0072] During model training and system operation, to continuously improve the accuracy of geological hazard identification and system stability, the model parameters are adaptively optimized by minimizing the difference between the predicted results and the actual hazard state. The overall optimization objective function is defined as follows: in, The classification loss term, representing the disaster identification result, is used to constrain the predicted probability. Consistency with actual disaster labels; This represents a time smoothing constraint term, used to suppress non-physical fluctuations in risk identification results over time. This represents the regularization term for the model parameters; and Here, represents the weighting coefficients. In one implementation, the classification loss term can be expressed as: in, Labels indicating the actual disaster status This represents the disaster state label predicted by the model. Through the above-mentioned risk level determination and objective function optimization process, a closed-loop linkage between the identification result output and the model parameter update is achieved, thereby enabling the power system geological disaster identification system to have continuous adaptive optimization capabilities during long-term operation.

[0073] It is understood that the method is applied to the monitoring of the surrounding area of ​​power system facilities, including transmission lines, substations or power pipelines. The multimodal raw observation data is obtained by multi-source collaborative collection of surface conditions, remote sensing images, microseismic activity, meteorological environment and historical disaster information within a preset monitoring range around the power system facilities.

[0074] This embodiment constructs a multimodal geological hazard perception space by acquiring multimodal raw observation data. The multimodal raw observation data includes at least surface condition data, remote sensing observation data, microseismic monitoring structure data, meteorological and precipitation data, and historical disaster data. A unified spatiotemporal reference framework is established to perform spatiotemporal alignment and structural characterization on the multimodal raw observation data, generating a standardized multimodal dataset. Feature extraction and adaptive weighted fusion are performed on the modal data in the standardized multimodal dataset to obtain fused features. Geological hazard identification and dynamic early warning are performed based on the fused features. This realizes unified spatiotemporal benchmark processing and dynamic weight allocation of multi-source heterogeneous data, effectively overcoming the problems of single data dimension, insufficient spatiotemporal alignment, and static fusion mechanism in existing technologies. It has the advantages of effectively integrating multi-source heterogeneous observation data, constructing a unified spatiotemporal reference framework, realizing cross-modal adaptive fusion, and having dynamic risk output capabilities, thereby significantly improving the accuracy of geological hazard identification and the timeliness of early warning.

[0075] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this application. In practical applications, those skilled in the art can select some or all of it to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0076] In addition, for technical details not described in detail in this embodiment, please refer to the method for identification, monitoring and early warning of geological disasters in power systems based on multimodal data fusion provided in any embodiment of this application, which will not be repeated here.

[0077] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0078] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0079] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application. The above are only preferred embodiments of this application and do not limit the patent scope of this application. All equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for identifying, monitoring, and providing early warning of geological disasters in power systems through multimodal data fusion, characterized in that, include: To acquire multimodal raw observation data to construct a multimodal geological disaster perception space, the multimodal raw observation data includes at least surface condition data, remote sensing observation data, microseismic monitoring structure data, meteorological and precipitation data, and historical disaster data; A unified spatiotemporal reference framework is established to perform spatiotemporal alignment and structural characterization on the original multimodal observation data, generating a standardized multimodal dataset; Feature extraction and adaptive weighted fusion are performed on the modal data in the standardized multimodal dataset to obtain fused features; Geological disaster identification and dynamic early warning are based on the fused features.

2. The method according to claim 1, characterized in that, The surface condition data includes surface deformation data, surface displacement data, and surface strain data; The remote sensing observation data includes optical remote sensing image data and synthetic aperture radar (SAR) time series data. The microseismic monitoring structural data includes microseismic event location data, source mechanism parameters, and structural response monitoring data. The meteorological rainfall data includes rainfall amount, rainfall intensity, and cumulative rainfall data; The historical disaster data includes records of historical disaster events, disaster trigger threshold parameters, and disaster evolution patterns.

3. The method according to claim 1, characterized in that, The spatiotemporal alignment of the original multimodal observation data includes: Using the location coordinates of power system facilities as a spatial reference and combining the timestamp information of multimodal data acquisition, a unified spatiotemporal reference system is constructed. To address the spatial resolution differences among different modal data, spatial resampling and georegistration techniques are employed to map each modal data onto a unified spatial grid. To address the temporal frequency differences among different modal data, a time interpolation and alignment algorithm is employed to align the time series of each modal data to a unified time node. And the spatiotemporally aligned modal data are organized according to spatial location and time dimension to form the standardized multimodal dataset.

4. The method according to claim 1, characterized in that, The structural characterization includes: For surface state data, surface deformation field interpolation and mechanical parameter transformation are used to generate a surface state characterization vector with a unified structure. For remote sensing observation data, standardized remote sensing image representation tensors are generated through radiometric correction, geometric correction, and time phase normalization. For microseismic monitoring structural data, a structural state characterization vector is generated through source parameter inversion and structural response feature extraction. A meteorological environment characterization matrix is ​​generated by using time-series accumulation and spatial distribution interpolation of meteorological precipitation data; For historical disaster data, a knowledge representation vector of historical disasters is generated by encoding disaster events and parameterizing triggering conditions.

5. The method according to claim 1, characterized in that, The feature extraction of each modality data in the standardized multimodal dataset includes: An independent feature extraction channel is constructed for each modality of data. Based on the channel attention mechanism, the importance weights of the features of each modality channel are weighted and learned to obtain the high-level semantic feature representation of each modality.

6. The method according to claim 1, characterized in that, The adaptive weighted fusion includes: Based on the acquisition quality, data integrity, and historical reliability of each modality data at the current moment, calculate the initial confidence score for each modality; Based on the initial confidence score, adaptive fusion weights for each modality feature are dynamically generated through a learnable gating network. During the fusion process, when a certain modality's data is missing or its confidence level is lower than a preset threshold, the fusion weight of that modality is automatically reduced, and the weight contribution of other reliable modalities is correspondingly increased. The weighted modal features are then deeply fused to obtain the fused features.

7. The method according to claim 1, characterized in that, The geological hazard identification based on the fused features includes: The fused features are input into the geological disaster identification model to identify the types, levels, and impact range of geological disasters around power system facilities.

8. The method according to claim 1, characterized in that, The dynamic early warning includes: The comprehensive disaster risk index is calculated based on the geological disaster identification results, and the current warning level is determined by combining the preset multi-level warning threshold range; The early warning level is dynamically adjusted based on the disaster evolution trend prediction results; Generate early warning information that includes disaster risk level, affected area, and evolution trend.

9. The method according to claim 1, characterized in that, Also includes: Based on the feedback results of the early warning information and the prediction deviation of the identification model, a system-level optimization objective function is constructed, which includes a geological disaster identification loss term, an early warning consistency loss term, and a spatiotemporal smoothing loss term. The network parameters of the geological disaster identification model and the weight parameters of the adaptive weighted fusion are optimized end-to-end to achieve adaptive updating of the model in response to the dynamic changes in the reliability of multi-source observation data.

10. The method according to claim 1, characterized in that, The method is applied to the monitoring of the surrounding area of ​​power system facilities, including transmission lines, substations or power utility tunnels. The multimodal raw observation data is obtained by multi-source collaborative acquisition of surface conditions, remote sensing images, microseismic activity, meteorological environment and historical disaster information within a preset monitoring range around the power system facilities.

Images