Coal mine operation scene anomaly detection method and system based on causal wavelet resonance network

Abstract

This invention relates to the field of industrial safety technology, and in particular to an anomaly detection method and system for coal mine operation scenarios based on a causal wavelet resonance network. The method includes: acquiring multimodal sensing data and constructing a high-dimensional spatiotemporal feature field; transforming the high-dimensional spatiotemporal feature field using a causal wavelet resonance network to extract a real-time causal resonance spectrum; retrieving a pre-learned baseline harmonic spectrum and its covariance matrix, calculating the Mahalanobis distance between the baseline harmonic spectrum and the real-time causal resonance spectrum, and generating a spatiotemporal detuning map; determining whether the detuning value at each spatiotemporal location in the spatiotemporal detuning map exceeds a preset threshold; if it does, it is considered an anomaly; analyzing the causal wavelet type and parameters that contribute most to the detuning, and generating interpretable anomaly alarm information, thereby improving the anomaly detection accuracy and enabling highly sensitive perception and interpretable judgment of weak, unknown, and causal-logically implied anomalies.

Description

Technical Field

[0001] This invention relates to the field of industrial safety technology, and in particular to an anomaly detection method and system for coal mine operation scenarios based on causal wavelet resonance networks. Background Technology

[0002] As a typical process industry, coal mines present unique challenges in their underground operations, characterized by confined spaces, variable lighting, pervasive dust, a high degree of mixing of personnel, machinery, and materials, and numerous potential risks. Achieving comprehensive, accurate, and real-time safety monitoring of underground work environments is crucial for preventing violations of regulations (illegal command, illegal operation, and violation of labor discipline), mitigating accidents, and promoting intelligent construction in the coal mining industry. Currently, mainstream intelligent safety monitoring technologies for coal mines primarily rely on deep learning-based computer vision algorithms. While these solutions, including object detection and segmentation methods based on convolutional neural networks (CNNs) and behavior recognition technologies based on recurrent neural networks (RNNs) or Transformers, have achieved some commercial success in identifying abnormal events in specific types and known scenarios, their inherent technological limitations increasingly highlight their limitations in the high-risk and dynamic underground working environments of coal mines.

[0003] In other words, current mainstream intelligent safety monitoring technologies for coal mines mainly suffer from several problems, including the fundamental bottleneck of "pattern matching," a lack of causal reasoning ability, a trust crisis caused by the "black box" nature of the technology, and insufficient sensitivity to anomalies in their early stages. Among these, the core essence of existing deep learning models is a powerful function fitter and pattern matcher. These models learn from massive amounts of labeled sample data, essentially memorizing the visual appearance patterns of "normal states" and "abnormal states." This often results in their generalization ability being severely limited by the coverage of the training dataset, making it difficult to cope with the complex and ever-changing risk patterns in coal mine scenarios. Furthermore, coal mine safety accidents are not instantaneous events, but rather a causal progression from the emergence of risk to its gradual evolution and eventual manifestation. Existing deep learning models primarily rely on statistical correlations between data for judgment, lacking the ability to understand and capture such underlying causal relationships, thus failing to achieve early prediction of risks in their nascent stages. The decision-making process of deep neural networks is highly nonlinear and opaque, and their abnormal alarm outputs lack logical explanations that can be understood and trusted by human operators, making it difficult to gain the genuine trust of on-site managers and workers. Many early signs of major coal mine safety accidents often manifest as extremely weak changes at the signal level, and existing traditional monitoring models based on macroscopic visual features lack the ability to effectively capture such weak, early abnormal signals. They often only detect these signals when they have accumulated to a certain extent and are close to an accident, missing the best opportunity for risk intervention.

[0004] Therefore, traditional deep learning-based coal mine safety monitoring technology has significant shortcomings in its ability to detect unknown anomalies, early risks, and complex causal anomalies, and it is difficult to provide interpretable decision-making basis, failing to meet the high requirements for safety monitoring in high-risk underground coal mine scenarios; it often suffers from low detection accuracy and difficulty in interpreting detection results. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, a method and system for anomaly detection in coal mine operation scenarios based on causal wavelet resonance networks is provided. This method can improve the accuracy of anomaly detection and can perform highly sensitive perception and interpretable judgment of weak, unknown and causal anomaly buds.

[0006] A method for anomaly detection in coal mine operation scenarios based on causal wavelet resonance networks, the method comprising:

[0007] Video data, positioning data, and sensor data from coal mine operation scenarios are acquired as multimodal sensing data, and the multimodal sensing data is uniformly mapped onto a discrete four-dimensional spatiotemporal grid to construct a high-dimensional spatiotemporal feature field.

[0008] A causal wavelet resonance network is used to transform the high-dimensional spatiotemporal feature field to extract the real-time causal resonance spectrum characterizing the current coal mine operation scenario. The causal wavelet resonance network includes at least one causal wavelet transform layer, which uses a set of predefined causal wavelet basis functions that are sensitive to specific causal patterns.

[0009] Retrieve the baseline harmonic spectrum and its covariance matrix obtained through learning from standard operation scenario data without anomalies, calculate the Mahalanobis distance between the baseline harmonic spectrum and the real-time causal resonance spectrum, and generate a spatiotemporal detuning map.

[0010] Determine whether the detuning value at each spatiotemporal location in the spatiotemporal detuning diagram exceeds a preset threshold. If it does, it is determined to be an abnormal state. Analyze the causal wavelet type and parameters that contribute the most to the detuning and generate interpretable abnormal alarm information that includes the abnormal location, degree, and causality.

[0011] In one embodiment, the high-dimensional spatiotemporal feature field is a tensor defined on a discrete four-dimensional spatiotemporal grid;

[0012] The multimodal sensing data is uniformly mapped onto a discrete four-dimensional spatiotemporal grid to construct a high-dimensional spatiotemporal feature field, including:

[0013] The video data is calibrated by camera and reconstructed in 3D. The 2D pixel coordinates are mapped to 3D physical space and discretized and synchronized in the time dimension. Pixel intensity, optical flow vector and semantic segmentation label are extracted as feature channels.

[0014] A feature field channel is set for the positioning data, and velocity and acceleration are set as supplementary feature channels;

[0015] Based on the installation location and sampling time of the sensor data, the sensor data is filled into the corresponding feature channels, and the spatially sparse data in the sensor data is smoothed by Kriging interpolation to form a continuous feature field.

[0016] In one embodiment, the causal wavelet basis function The formula is expressed as:

[0017] ;

[0018] in, It is a standard four-dimensional mother wavelet function, selected from the four-dimensional Mexican hat wavelet or the four-dimensional Morlet wavelet; This is a spatiotemporal scale parameter used to control the size of the analytical field of view; As a spatiotemporal translation vector, it enables the sliding analysis of wavelets in a four-dimensional spatiotemporal field; For the direction parameter The determined four-dimensional rotation matrix is ​​used to impart directional selectivity to the wavelet; For specific causal concepts The associated causal kernel function, wherein the causal kernel function is a learnable or predefined mathematical model; This represents a voxel in a discrete four-dimensional spatiotemporal grid.

[0019] In one embodiment, the causal concept is a risk event in the safety rules of a coal mine operation scenario; the causal kernel function is set as a mathematical model corresponding to the risk event.

[0020] In one embodiment, a causal wavelet resonance network is used to transform the high-dimensional spatiotemporal feature field to extract a real-time causal resonance spectrum characterizing the current state of the coal mine operation scenario, including:

[0021] By performing a convolution integral operation on the high-dimensional spatiotemporal feature field and the causal wavelet basis functions in the causal wavelet resonance network, a real-time causal resonance spectrum is obtained.

[0022] The formula for the real-time causal resonance spectrum is expressed as follows:

[0023] ;

[0024] in, Representation and the concept of causation The most relevant feature channels; It is a high-dimensional tensor, each element The value represents the scene at a point in time and space. Nearby, on scale ,direction Above, and the concept of cause and effect The degree of conformity; Represents the complex conjugate operation; It represents the complete spacetime consisting of three spatial dimensions (x, y, z) and one time dimension t; It is a voxel in a discrete four-dimensional spacetime grid. ; yes The abbreviated form, used as a measure of integration in the formula for the real-time causal resonance spectrum, indicates that the integration operation is performed on the integrand over a continuous four-dimensional manifold consisting of three spatial dimensions and one time dimension. The values ​​at each infinitesimal spacetime translation vector are summed.

[0025] In one embodiment, the method further includes:

[0026] Obtain labeled normal and abnormal samples, and use a contrastive learning loss function to maximize the distance between the normal and abnormal samples in the causal resonance spectrum space;

[0027] The learnable parameters in the causal kernel function are optimized based on the maximized distance.

[0028] In one embodiment, the method further includes:

[0029] Collect various operation data in a standard operation scenario without abnormalities, and calculate the causal resonance spectrum of each operation data.

[0030] The baseline harmonic spectrum and its covariance matrix are obtained by statistical averaging and covariance calculation of the causal resonance spectrum of each operation.

[0031] In one embodiment, the formula for the spatiotemporal detuning diagram is expressed as:

[0032] ;

[0033] in, Indicates the baseline harmonic spectrum; Represents the real-time causal resonance spectrum; Represents the covariance matrix; each spatiotemporal location , take all other parameters The corresponding resonance spectral coefficients are organized into a long vector. Spatiotemporal Disharmony Diagram It is a scalar field, at a point The larger the value of , the further the state deviates; T represents the transpose matrix.

[0034] An anomaly detection system for coal mine operation scenarios based on causal wavelet resonance networks, the system comprising:

[0035] The high-dimensional spatiotemporal feature field construction module is used to acquire video data, positioning data, and sensor data of coal mine operation scenarios as multimodal perception data, and to uniformly map the multimodal perception data to a discrete four-dimensional spatiotemporal grid to construct a high-dimensional spatiotemporal feature field.

[0036] The causal resonance spectrum extraction module is used to transform the high-dimensional spatiotemporal feature field using a causal wavelet resonance network to extract the real-time causal resonance spectrum characterizing the current coal mine operation scenario. The causal wavelet resonance network includes at least one causal wavelet transform layer, which uses a set of predefined causal wavelet basis functions that are sensitive to specific causal patterns.

[0037] The spatiotemporal detuning graph generation module is used to retrieve the baseline harmonic spectrum and its covariance matrix obtained by learning through data from a standard operation scenario without anomalies, calculate the Mahalanobis distance between the baseline harmonic spectrum and the real-time causal resonance spectrum, and generate a spatiotemporal detuning graph.

[0038] An anomaly detection module is used to determine whether the detuning value at each spatiotemporal location in the spatiotemporal detuning diagram exceeds a preset threshold. If it does, it is determined to be an abnormal state. The module also analyzes the causal wavelet type and parameters that contribute the most to the detuning and generates interpretable anomaly alarm information that includes the location, degree and causality of the anomaly.

[0039] In one embodiment, the high-dimensional spatiotemporal feature field is a tensor defined on a discrete four-dimensional spatiotemporal grid. The high-dimensional spatiotemporal feature field construction module is further used to map the two-dimensional pixel coordinates to three-dimensional physical space through camera calibration and three-dimensional reconstruction of the video data, and to discretize and synchronize it according to the time dimension, extracting pixel intensity, optical flow vector, and semantic segmentation labels as feature channels; setting feature field channels for the positioning data, and setting velocity and acceleration as supplementary feature channels; based on the installation location and sampling time of the sensor data, filling the sensor data into the corresponding feature channels, and smoothing the spatially sparse data in the sensor data through Kriging interpolation to form a continuous feature field.

[0040] The aforementioned anomaly detection method and system for coal mine operation scenarios based on causal wavelet resonance networks constructs a high-dimensional spatiotemporal feature field, mapping multimodal sensing data onto a unified four-dimensional grid, thus avoiding the problems of fragmented and difficult-to-coordinate multi-source data. By extracting causal resonance spectra through causal wavelet resonance networks and calculating detuning values ​​by comparing with baseline harmonic spectra, it can comprehensively capture anomalies that cause abnormal fluctuations in the spatiotemporal feature field without relying on anomaly sample training. Furthermore, by using causal wavelet basis functions, it can perceive underlying causal relationships and achieve prevention, thereby improving the accuracy of anomaly detection. It can also perform highly sensitive perception and interpretable judgment of weak, unknown, and causal-logically implied anomalies. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the application environment and hardware deployment architecture of a coal mine operation scene anomaly detection method based on causal wavelet resonance network in one embodiment.

[0042] Figure 2 This is a flowchart illustrating an anomaly detection method for coal mine operation scenarios based on causal wavelet resonance networks in one embodiment.

[0043] Figure 3 This is a schematic diagram of the high-dimensional spatiotemporal feature field (SFF) construction process in one embodiment;

[0044] Figure 4 This is a schematic diagram of the multi-level structure of a causal wavelet resonance network (CWRN) in one embodiment;

[0045] Figure 5 This is a schematic diagram illustrating the working principle of generating a spatiotemporal detuning map based on resonance spectrum comparison in one embodiment;

[0046] Figure 6 This is a schematic diagram of the overall architecture and process of an anomaly detection method for coal mine operation scenarios based on causal wavelet resonance networks in one embodiment;

[0047] Figure 7 This is a block diagram of a coal mine operation scene anomaly detection system based on a causal wavelet resonance network in one embodiment.

[0048] Figure 8 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0050] The anomaly detection method for coal mine operation scenes based on causal wavelet resonance networks provided in this application can be applied to, for example... Figure 1 The application environment shown. For example... Figure 1 As shown, the application environment includes an underground operation site - edge computing device 110 and a ground dispatch center - central server device 120, with data transmission between the edge computing device 110 and the central server device 120 via industrial Ethernet. The edge computing device 110 can collect video data, positioning data, and sensor data from the coal mine operation scene as multimodal sensing data and send them to the central server device 120. The central server device 120 can acquire the video data, positioning data, and sensor data from the coal mine operation scene as multimodal sensing data and uniformly map the multimodal sensing data onto a discrete four-dimensional spatiotemporal grid to construct a high-dimensional spatiotemporal feature field. The central server device 120 can use a causal wavelet resonance network to transform the high-dimensional spatiotemporal feature field and extract the real-time causal resonance spectrum characterizing the current state of the coal mine operation scene. The causal wavelet resonance network includes... At least one causal wavelet transform layer is included, employing a set of predefined causal wavelet basis functions sensitive to specific causal patterns. The central server device 120 can retrieve the baseline harmonic spectrum and its covariance matrix learned from data of a standard operating scenario without anomalies, calculate the Mahalanobis distance between the baseline harmonic spectrum and the real-time causal resonance spectrum, and generate a spatiotemporal detuning map. The central server device 120 can determine whether the detuning value at each spatiotemporal location in the spatiotemporal detuning map exceeds a preset threshold; if it does, it is considered an abnormal state. It also analyzes the causal wavelet type and parameters that contribute most to the detuning, generating interpretable anomaly alarm information including the location, degree, and causality of the anomaly. The edge computing device 110 can be, but is not limited to, various explosion-proof high-definition cameras, positioning base stations, and various sensors; the central server device 120 can be, but is not limited to, various personal computers, laptops, smartphones, robots, unmanned aerial vehicles, tablets, and other devices.

[0051] In one embodiment, such as Figure 2 As shown, a method for anomaly detection in coal mine operation scenarios based on causal wavelet resonance networks is provided, including the following steps:

[0052] Step 202: Acquire video data, positioning data, and sensor data of the coal mine operation scene as multimodal perception data, and map the multimodal perception data uniformly into a discrete four-dimensional spatiotemporal grid to construct a high-dimensional spatiotemporal feature field.

[0053] Edge computing devices 110 can be evenly distributed in a coal mine operation scenario to collect multimodal sensing data. Specifically, the edge computing devices 110 may include explosion-proof high-definition cameras, positioning base stations, and various sensors. The explosion-proof high-definition cameras can be used to collect video data of the coal mine operation scenario, the positioning base stations can be used to collect positioning data of the coal mine operation scenario, and the various sensors can be used to collect sensor data of the coal mine operation scenario. The edge computing devices 110 can transmit the collected multimodal sensing data to the central server device 120. After obtaining the multimodal sensing data of the coal mine operation scenario, the central server device 120 performs further data processing.

[0054] In one embodiment, the provided method for anomaly detection in coal mine operation scenarios based on causal wavelet resonance networks may further include the process of constructing a high-dimensional spatiotemporal feature field. The specific process includes: calibrating and reconstructing video data using a camera, mapping two-dimensional pixel coordinates to three-dimensional physical space, and discretizing and synchronizing them according to the time dimension; extracting pixel intensity, optical flow vector, and semantic segmentation labels as feature channels; setting feature field channels for positioning data, and setting velocity and acceleration as supplementary feature channels; filling sensor data into the corresponding feature channels based on the installation location and sampling time of sensor data, and smoothing the spatially sparse data in the sensor data through Kriging interpolation to form a continuous feature field.

[0055] The central server equipment needs to treat the coal mine operation scenario as a measurable physical field. To achieve this, it is first necessary to unify multimodal sensing data of varying sources, formats, and dimensions under a common mathematical framework. In this embodiment, a high-dimensional spatiotemporal feature field... It is a tensor defined on a discrete four-dimensional spacetime grid, where For voxels in a discrete four-dimensional spatiotemporal grid, These are the feature channels, which include pixel intensity and optical flow information extracted from video data, as well as physical quantity readings obtained from various sensors. This is achieved by constructing a discretized four-dimensional spatiotemporal grid. The three spatial dimensions of a grid Corresponding to the physical space of the monitored area, the time dimension Then it is discretized at a fixed frame rate or sampling rate. Each voxel on the grid... Each feature vector is associated with a single feature vector, and all feature vectors together constitute a high-dimensional spatiotemporal feature field. .

[0056] Specifically, after constructing the four-dimensional spatiotemporal grid, the multimodal sensing data needs to be mapped to form a high-dimensional spatiotemporal feature field. For example... Figure 3As shown, multimodal raw data sources can include camera data, location tag data, and environmental and equipment sensor data. Through mapping and unified processing, a high-dimensional spatiotemporal feature field is ultimately formed in a unified expression form. Specifically, for multiple cameras deployed in a coal mine operation scenario, their two-dimensional pixel coordinates are first mapped to three-dimensional physical space coordinates through camera calibration and 3D reconstruction techniques. Regarding time... In each frame, the RGB value, optical flow vector, or more advanced semantic segmentation label of each pixel is assigned to its corresponding... Voxels form different characteristic channels The optical flow vector can be calculated using optical flow algorithms such as Farnback; semantic segmentation labels, such as people, devices, and ceilings, can be obtained through pre-trained segmentation models. For individuals or mobile devices wearing UWB positioning tags, their location at time... precise three-dimensional coordinates Once acquired, a dedicated "existence" channel can be established in the high-dimensional spacetime feature field S to store voxels. The values ​​of the sensor and its neighbors are set to 1, while other locations are set to 0. Simultaneously, information such as velocity and acceleration can constitute additional feature channels. Sensor data can include gas sensors, temperature and humidity sensors, equipment vibration sensors, roof delamination meters, etc.; the readings of each sensor are determined based on its fixed physical installation location. and sampling time It is filled into the corresponding voxel. In the corresponding characteristic channels; for spatially sparse sensor data, methods such as Kriging interpolation can be used to perform smooth interpolation in space to form a continuous field.

[0057] By mapping multimodal sensing data onto a four-dimensional spatiotemporal grid, a unified, dense, high-dimensional spatiotemporal feature field that can comprehensively describe the scene state can be obtained. .

[0058] Step 204: A causal wavelet resonance network is used to transform the high-dimensional spatiotemporal feature field to extract the real-time causal resonance spectrum that characterizes the current state of the coal mine operation scenario. The causal wavelet resonance network contains at least one causal wavelet transform layer, which uses a set of predefined causal wavelet basis functions that are sensitive to specific causal modes.

[0059] The Causal Wavelet Resonance Network (CWRN) is the core computing engine. It is not a traditional weighted neural network, but a large-scale, parallel signal processing framework whose core function is to perform causal wavelet transforms. The central server can employ a causal wavelet resonance network to transform high-dimensional spatiotemporal feature fields, extracting a real-time causal resonance spectrum characterizing the current state of the coal mine operation.

[0060] The multi-level structure of the causal wavelet resonance network (CWRN) is as follows: Figure 4 As shown, the causal wavelet resonance network consists of at least one causal wavelet transform layer. The causal wavelet transform layer is equipped with a convolution kernel to perform convolution integration operations, thereby outputting a real-time causal resonance spectrum multidimensional tensor.

[0061] In one embodiment, a causal wavelet basis function consists of multiple parts. The formula is expressed as: ;in, It is a standard four-dimensional mother wavelet function, selected from the four-dimensional Mexican hat wavelet or the four-dimensional Morlet wavelet. The Mexican hat wavelet is sensitive to point singularities, while the Morlet wavelet is sensitive to directional textures. The spatiotemporal scale parameter is used to control the size of the analytical field of view. Small scales focus on high-frequency, transient details, such as rapid sparks, while large scales focus on low-frequency, slowly changing trends, such as the slow sinking of the roof. Using the spatiotemporal translation vector, we can perform wavelet sliding analysis in a four-dimensional spatiotemporal field to analyze each local region. For the direction parameter The determined four-dimensional rotation matrix is ​​used to give the wavelet direction selectivity. In four-dimensional spacetime, rotation is more complex than in three-dimensional spacetime. The role of this four-dimensional rotation matrix is ​​to orient the mother wavelet toward a specific spacetime direction, thereby enabling the detection of events in a specific direction, such as whether an object is moving upward or downward. For specific causal concepts The associated causal kernel function, which is a learnable or predefined mathematical model, is key to injecting causal concepts into wavelet analysis. The causal kernel function is a modulation function that reshapes the form of the mother wavelet, making it more responsive to the implied causal concepts. The spatiotemporal pattern produces the strongest response. This represents a voxel in a discrete four-dimensional spatiotemporal grid.

[0062] In one embodiment, the concept of causality is a risk event in the safety rules of a coal mine operation scenario; the causal kernel function is set as a mathematical model corresponding to the risk event.

[0063] Specifically, the design of causal kernel functions is knowledge-driven. The concept of causality... The basic events selected from the coal mine safety rules include, but are not limited to: rapid approach of personnel and equipment, abnormal high-frequency vibration of equipment, abnormal changes in roof stress, and personnel leaving the safe area.

[0064] The central server equipment can set the causal kernel function as a mathematical model corresponding to a risk event, that is, to transform the key risk points in coal mine safety regulations into mathematical form. The risk event is the rapid approach of personnel and equipment; the corresponding causal kernel function is... It can be designed as the product of two spatiotemporally separate Gaussian functions whose trajectories predict collision, to enhance the response to patterns where the trajectories of two targets intersect at some future scale. When a wavelet is modulated by this kernel, it will only produce a strong response when analyzing a region and simultaneously there are human and equipment signals conforming to this collision trajectory within that region. The corresponding causal kernel function is used when the risk event is abnormal high-frequency vibration of equipment. It can be designed as a function that is a high-frequency sine wave in the time dimension and concentrated on the equipment location in the spatial dimension, making the wavelet a bandpass filter of a specific frequency, specifically used to detect abnormal equipment resonance. When the risk event is an abnormal change in roof stress, the corresponding causal kernel function... Based on an elasticity model, it can be designed to be sensitive to shear stress or tensile stress modes. When applied to a characteristic channel constructed from data from a top plate delamination instrument and microseismic sensors, it can effectively detect signs of impending collapse.

[0065] In one embodiment, a method for detecting anomalies in coal mine operation scenarios based on causal wavelet resonance networks may further include the process of calculating a real-time causal resonance spectrum. Specifically, the process includes performing a convolution integral operation on the high-dimensional spatiotemporal feature field and the causal wavelet basis functions in the causal wavelet resonance network to obtain the real-time causal resonance spectrum.

[0066] The central server equipment can process high-dimensional spatiotemporal feature fields. To perform a transformation is to calculate The inner product (convolution) with the defined, large library of causal wavelet basis functions produces the real-time causal resonance spectrum (CRS), denoted as […]. .

[0067] Real-time causal resonance spectrum The formula is expressed as: ;in, Representation and the concept of causation The most relevant feature channels; It is a high-dimensional tensor, each element The value represents the scene at a point in time and space. Nearby, on scale ,direction Above, and the concept of cause and effect The degree of conformity; Represents the complex conjugate operation; It represents the complete spacetime consisting of three spatial dimensions (x, y, z) and one time dimension t; It is a voxel in a discrete four-dimensional spacetime grid. ; yes The abbreviated form, used as a measure of integration in the formula for the real-time causal resonance spectrum, indicates that the integration operation is performed on the integrand over a continuous four-dimensional manifold consisting of three spatial dimensions and one time dimension. The values ​​at each infinitesimal spacetime translation vector are summed.

[0068] Step 206: Retrieve the baseline harmonic spectrum and its covariance matrix obtained through learning from the standard operation scenario data without anomalies, calculate the Mahalanobis distance between the baseline harmonic spectrum and the real-time causal resonance spectrum, and generate a spatiotemporal detuning map.

[0069] like Figure 5 As shown, the central server device can compare the real-time causal resonance spectrum with a pre-learned baseline harmonic spectrum representing a normal operating scenario to calculate a spatiotemporal detuning map. Specifically, the central server device can obtain the resonance vector at the current moment based on the real-time causal resonance spectrum, as a real-time data stream; simultaneously, it can calculate the mean vector and covariance matrix based on the pre-learned baseline harmonic spectrum, as a historical baseline stream; then, it can perform Mahalanobis distance calculation to calculate the deviation between the real-time vector and the mean, and normalize it using the covariance matrix, thereby generating a spatiotemporal detuning map quantitative field, and then quantify the degree of detuning, with larger energy values ​​indicating greater anomalies. The core idea of ​​the central server device in detecting anomalies is to compare the difference between the real-time causal resonance spectrum and the normal state; therefore, a learning phase is required initially.

[0070] In one embodiment, the provided method for anomaly detection in coal mine operation scenarios based on causal wavelet resonance networks may further include a baseline harmonic spectrum learning process. The specific process includes: collecting various operation data in a standard operation scenario without anomalies, and calculating the operation causal resonance spectrum for each operation data; performing statistical averaging and covariance calculation on each operation causal resonance spectrum to obtain the baseline harmonic spectrum and its covariance matrix.

[0071] Among them, the baseline harmonic spectrum and its covariance matrix It is obtained by calculating the real-time causal resonance spectrum and determining its statistical average and covariance on a large amount of anomaly-free, standard operating scenario data. Specifically, by collecting hundreds or even thousands of hours of operational data, manually confirmed to be free of any anomalies, covering various normal operating conditions (such as normal tunneling, support, ventilation, etc.), the causal resonance spectrum of this data is calculated; then, the baseline harmonic spectrum (HRS) is obtained through statistical averaging. and its resonance covariance matrix: . This represents the perfectly normal average resonance mode, while This represents the reasonable fluctuation range of resonance under normal conditions.

[0072] During real-time monitoring, the real-time causal resonance spectrum calculated at each moment... The central server equipment can calculate its harmonic spectrum with the baseline. The degree of deviation. Simple Euclidean distance is insufficient to describe statistical bias; therefore, in this embodiment, Mahalanobis distance is preferred because it considers the correlation between features and performs scale normalization. In one embodiment, a spatiotemporal dissonance map (SDM) is used. The formula is expressed as: ;in, Indicates the baseline harmonic spectrum; Represents the real-time causal resonance spectrum; Represents the covariance matrix; each spatiotemporal location , take all other parameters The corresponding resonance spectral coefficients are organized into a long vector. Spatiotemporal Disharmony Diagram It is a scalar field, at a point The larger the value, the further the scene deviates from its state at that point in time and space, and the more discordant it is; T represents the transpose matrix.

[0073] Step 208: Determine whether the detuning value at each spatiotemporal location in the spatiotemporal detuning diagram exceeds a preset threshold. If it does, it is determined to be an abnormal state. Analyze the causal wavelet type and parameters that contribute the most to the detuning and generate interpretable abnormal alarm information that includes the abnormal location, degree, and causality.

[0074] When the detuning diagram If the value at any point exceeds a preset dynamic threshold, it can be determined that... An anomaly occurred. The preset threshold can be determined based on FROC curve analysis. In this embodiment, interpretability can be achieved through contribution analysis: for an anomaly... Going back to the calculation of the Mahalanobis distance, we need to identify which component(s) are involved. The one that contributes the most to the final distance value; this directly corresponds to which causal concept it is. At what scale Which direction The most serious detuning occurred.

[0075] The central server equipment can generate highly interpretable alarms in the following format: [Level 3 Alarm: Potential Risk] At coordinates (10.2, 5.5, 3.1), time (14:32:05.2), a spatiotemporal misalignment of 7.8 was detected (threshold 5.0); Main causes: abnormal energy in the causal pattern of rapid approach of personnel and equipment, contributing 65%; abnormal energy in the causal pattern of high-speed movement of personnel, contributing 20%; Recommendation: Please immediately verify the relative position and speed of personnel and tunneling machines in this area.

[0076] In one embodiment, the provided anomaly detection method for coal mine operation scenarios based on causal wavelet resonance networks may further include a process of optimizing the learnable parameters in the causal kernel function. The specific process includes: acquiring labeled normal samples and abnormal samples; using a contrastive learning loss function to maximize the distance between normal samples and abnormal samples in the causal resonance spectrum space; and optimizing the learnable parameters in the causal kernel function based on the maximized distance.

[0077] The performance of causal kernel functions can be further improved by introducing a small amount of labeled data to perform supervised fine-tuning of the learnable parameters in the causal kernel function.

[0078] In this embodiment, a contrastive learning framework can be used. Assume there is some labeled normal scene data. and abnormal scenario data That is, normal samples and abnormal samples. For a given anchor point sample, its real-time causal resonance spectrum is denoted as... ,hope The resonance spectrum with all normal samples in the spectral space. The (positive sample) is closer, while the resonance spectrum is closer to all the anomalous samples. The (negative samples) are farther away. Therefore, the contrastive loss function can be defined as: ;in, For real-time causal resonance spectrum, The resonance spectrum of a normal sample. The resonance spectrum of the anomalous sample. It is cosine similarity. It is a temperature hyperparameter. While most of the causal wavelet basis functions are predefined, the causal kernel function among them... It can be designed to be learnable; for example, it can be parameterized using a small neural network. By minimizing the aforementioned contrast loss, the optimal causal kernel morphology can be learned end-to-end, making it more discriminative in distinguishing between normal and abnormal.

[0079] In one embodiment, the overall architecture and process structure of a coal mine operation scene anomaly detection method based on causal wavelet resonance networks are as follows: Figure 6 As shown, the process mainly includes the data perception and field construction stage, the causal resonance analysis stage, the detuning diagnosis stage, and the decision-making and alarm stage. The data perception and field construction stage mainly includes: acquiring multimodal data, primarily video / sensor / location data; data mapping and synchronization; and constructing a high-dimensional spatiotemporal feature field. The causal resonance analysis stage mainly includes: inputting the high-dimensional spatiotemporal feature field into a causal wavelet resonance network; performing transformation processing using causal wavelet basis functions; and extracting the real-time causal resonance spectrum. The detuning diagnosis stage mainly includes: calling the baseline and harmonic spectrum statistical values ​​for normal operating scenarios; calculating Mahalanobis distance for comparison; and generating a spatiotemporal detuning map. The decision-making and alarm stage mainly includes: determining whether the detuning value exceeds a threshold; if not, it is judged as a normal state; if so, it is judged as an abnormal state, and the contribution of the causal wavelet type is analyzed to calculate the contribution, outputting interpretable alarm information including location / degree / cause.

[0080] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0081] In one embodiment, such as Figure 7 As shown, a coal mine operation scene anomaly detection system based on causal wavelet resonance network is provided, including: a high-dimensional spatiotemporal feature field construction module 710, a causal resonance spectrum extraction module 720, a spatiotemporal detuning map generation module 730, and an anomaly detection module 740, wherein:

[0082] The high-dimensional spatiotemporal feature field construction module 710 is used to acquire video data, positioning data, and sensor data of coal mine operation scenarios as multimodal perception data, and to uniformly map the multimodal perception data to a discrete four-dimensional spatiotemporal grid to construct a high-dimensional spatiotemporal feature field.

[0083] The causal resonance spectrum extraction module 720 is used to transform the high-dimensional spatiotemporal feature field using a causal wavelet resonance network to extract the real-time causal resonance spectrum characterizing the current coal mine operation scenario. The causal wavelet resonance network contains at least one causal wavelet transform layer, which uses a set of predefined causal wavelet basis functions that are sensitive to specific causal modes.

[0084] The spatiotemporal detuning graph generation module 730 is used to retrieve the baseline harmonic spectrum and its covariance matrix learned from the data of the standard operation scenario without anomalies, calculate the Mahalanobis distance between the baseline harmonic spectrum and the real-time causal resonance spectrum, and generate the spatiotemporal detuning graph.

[0085] The anomaly detection module 740 is used to determine whether the detuning value at each spatiotemporal location in the spatiotemporal detuning diagram exceeds a preset threshold. If it does, it is determined to be an abnormal state. The module also analyzes the causal wavelet type and parameters that contribute the most to the detuning and generates interpretable anomaly alarm information that includes the location, degree and causality of the anomaly.

[0086] In one embodiment, the high-dimensional spatiotemporal feature field is a tensor defined on a discrete four-dimensional spatiotemporal grid. The high-dimensional spatiotemporal feature field construction module 710 is also used to map the two-dimensional pixel coordinates to three-dimensional physical space through camera calibration and three-dimensional reconstruction of video data, and to discretize and synchronize them according to the time dimension, extracting pixel intensity, optical flow vector, and semantic segmentation label as feature channels; setting feature field channels for positioning data, and setting velocity and acceleration as supplementary feature channels; filling the corresponding feature channels with sensor data based on the installation location and sampling time of sensor data, and smoothing the spatially sparse data in the sensor data through Kriging interpolation to form a continuous feature field.

[0087] In one embodiment, the causal resonance spectrum extraction module 720 is further used to perform convolution integration operations on the high-dimensional spatiotemporal feature field and the causal wavelet basis functions in the causal wavelet resonance network to obtain the real-time causal resonance spectrum.

[0088] In one embodiment, the spatiotemporal detuning diagram generation module 730 is also used to collect various operation data in a standard operation scenario without anomalies, and calculate the causal resonance spectrum of each operation data; perform statistical averaging and covariance calculation on each operation causal resonance spectrum to obtain the baseline harmonic spectrum and its covariance matrix.

[0089] In one embodiment, a coal mine operation scene anomaly detection system based on a causal wavelet resonance network may further include a parameter optimization module, which is used to acquire labeled normal samples and abnormal samples, maximize the distance between normal samples and abnormal samples in the causal resonance spectrum space using a contrastive learning loss function, and optimize the learnable parameters in the causal kernel function based on the maximized distance.

[0090] In one embodiment, such as Figure 1 As shown, the deployment architecture in the actual scenario is divided into the downhole operation site - edge computing device 110 and the ground dispatch center - central server device 120.

[0091] Edge computing device 110 is deployed underground on an explosion-proof server or a powerful PLC, directly connecting to various devices in the monitoring area, including high-definition cameras, UWB positioning base stations, dust / gas sensors, and equipment PLC interfaces. Its main tasks are: real-time acquisition and synchronization of all multimodal data; construction of a spatiotemporal feature field with relatively low computational cost; and execution of a small-scale, low-computational-cost causal wavelet transform for preliminary and rapid risk screening. The processed spatiotemporal feature field S is then transmitted to the surface via an industrial Ethernet ring network.

[0092] Central Server Equipment 120 is a high-performance GPU server cluster deployed in the ground dispatch and command center, responsible for core computing tasks: receiving and caching spatiotemporal feature field data from underground; running a complete causal wavelet resonance network to calculate comprehensive, multi-scale, multi-directional, and multi-causal real-time causal resonance spectra. ; Execution and baseline harmonic spectrum The comparison generates a spatiotemporal misharmonicity map. Perform threshold judgment, contribution analysis, and generate the final interpretable alarm.

[0093] In this embodiment, a monitoring terminal can also be provided, with a dispatcher's operating interface on the monitoring terminal. Through a 3D visualization engine, the real-time underground scene, spatiotemporal detuning diagram (which can be overlaid on the 3D model in the form of a heat map), and detailed alarm text information are displayed on the same screen to the dispatcher, achieving a global safety situation awareness in a single graphical format.

[0094] In one embodiment, an application scenario simulation is performed on the coal mine operation scene anomaly detection method and system based on causal wavelet resonance network provided in this application, mainly simulating the detection of precursors to minor roof collapses. The specific simulation process is as follows:

[0095] Normal state: The roof of the tunneling face is stable. The spatiotemporal characteristic field is composed of stress, vibration, and texture features extracted from the roof delamination instrument, microseismic sensors, and high-definition camera. It is stable. In fact, it is a real-time causal resonance spectrum. Harmony spectrum with baseline Highly matched, mismatched graph A patch of "blue" (low detuning value).

[0096] Anomaly initiation: Due to changes in geological stress, microscopic cracks, invisible to the naked eye, begin to appear in a certain part of the roof, accompanied by weak acoustic emission signals. This change generates a weak, transient disturbance in specific characteristic channels of S (such as stress gradient, high-frequency acoustics).

[0097] CWRN Response: Although this weak perturbation energy is small, its spatiotemporal pattern happens to coincide with one or more causal wavelet bases specifically designed for "rock mass fracturing" or "shear stress". Highly matched. Therefore, after convolution calculation, the real-time causal resonance spectrum... In the corresponding In terms of location, a spike was generated that far exceeded the normal fluctuation range.

[0098] Alarm generation: This spike caused a spatiotemporal detuning in the calculation of Mahalanobis distance. A distinct "red dot" (high detuning value) appeared at the corresponding location. The system immediately triggered an alarm and reported to the dispatcher: "[Level 1 Alarm: Major Risk] Severe detuning was detected in the roof area at [coordinates]. Main cause: The energy of the rock mass fracture causal mode exceeded the standard by 3.5 standard deviations. Strong recommendation: Immediately evacuate personnel from this area and conduct a special support inspection!"

[0099] In this way, this application can capture extremely faint precursory information during the golden window before a catastrophic accident occurs, enabling true predictive maintenance and proactive safety management.

[0100] In one embodiment, a central server device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 8 As shown, the central server device includes a processor, memory, network interface, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for detecting anomalies in coal mine operation scenarios based on causal wavelet resonance networks. The input devices of the central server device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0101] Those skilled in the art will understand that Figure 8The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the central server equipment to which the present application is applied. The specific central server equipment may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0102] In one embodiment, a central server device is provided, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of a coal mine operation scene anomaly detection method based on a causal wavelet resonance network.

[0103] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, the steps of a method for detecting anomalies in coal mine operation scenarios based on causal wavelet resonance networks are implemented.

[0104] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0105] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0106] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for anomaly detection in coal mine operation scenarios based on causal wavelet resonance networks, characterized in that, The method includes: Video data, positioning data, and sensor data from coal mine operation scenarios are acquired as multimodal sensing data, and the multimodal sensing data is uniformly mapped onto a discrete four-dimensional spatiotemporal grid to construct a high-dimensional spatiotemporal feature field. A causal wavelet resonance network is used to transform the high-dimensional spatiotemporal feature field, extracting a real-time causal resonance spectrum characterizing the current coal mine operation scenario. The causal wavelet resonance network includes at least one causal wavelet transform layer, which employs a set of predefined causal wavelet basis functions sensitive to specific causal patterns. The formula is expressed as: ;in, It is a standard four-dimensional mother wavelet function, selected from the four-dimensional Mexican hat wavelet or the four-dimensional Morlet wavelet; This is a spatiotemporal scale parameter used to control the size of the analytical field of view; As a spatiotemporal translation vector, it enables the sliding analysis of wavelets in a four-dimensional spatiotemporal field; For the direction parameter The determined four-dimensional rotation matrix is ​​used to impart directional selectivity to the wavelet; For the concept of causality The associated causal kernel function, wherein the causal kernel function is a learnable or predefined mathematical model; The step of transforming the high-dimensional spatiotemporal feature field using a causal wavelet resonance network to extract the real-time causal resonance spectrum characterizing the current coal mine operation scenario includes: performing a convolution integral operation on the high-dimensional spatiotemporal feature field and the causal wavelet basis functions in the causal wavelet resonance network to obtain the real-time causal resonance spectrum; the formula for the real-time causal resonance spectrum is expressed as: ;in, Representation and the concept of causation The most relevant feature channels; It is a high-dimensional tensor, each element The value represents the scene's translation vector in a specific time and space. Nearby, at a specific spatiotemporal scale parameter Specific directional parameters Above, with specific causal concepts The degree of conformity; For spatiotemporal scale parameters; Represents the complex conjugate operation; It represents the complete spacetime consisting of three spatial dimensions (x, y, z) and one time dimension t; It is a voxel in a discrete four-dimensional spacetime grid. ; yes The abbreviated form, used as a measure of integration in the formula for the real-time causal resonance spectrum, indicates that the integration operation is performed on a continuous four-dimensional manifold consisting of three spatial dimensions and one time dimension, with respect to the integrand. Summ the values ​​at each infinitesimal spacetime translation vector; Retrieve the baseline harmonic spectrum and its covariance matrix obtained through learning from standard operating procedure data without anomalies, calculate the Mahalanobis distance between the baseline harmonic spectrum and the real-time causal resonance spectrum, and generate a spatiotemporal detuning map; the formula for the spatiotemporal detuning map is expressed as: ;in, Indicates the baseline harmonic spectrum; Represents the covariance matrix; each spatiotemporal translation vector , take all other parameters The corresponding resonance spectral coefficients are organized into a long vector. Spatiotemporal Disharmony Diagram It is a scalar field, a vector translated in a specific spacetime. The larger the value of , the further the state deviates; T represents the transpose matrix; Determine whether the detuning value at each spatiotemporal location in the spatiotemporal detuning diagram exceeds a preset threshold. If it does, it is determined to be an abnormal state. Analyze the causal wavelet type and parameters that contribute the most to the detuning and generate interpretable abnormal alarm information that includes the abnormal location, degree, and causality.

2. The method for anomaly detection in coal mine operation scenarios based on causal wavelet resonance networks according to claim 1, characterized in that, The high-dimensional spatiotemporal feature field is a tensor defined on a discrete four-dimensional spatiotemporal grid; The multimodal sensing data is uniformly mapped onto a discrete four-dimensional spatiotemporal grid to construct a high-dimensional spatiotemporal feature field, including: The video data is calibrated by camera and reconstructed in 3D. The 2D pixel coordinates are mapped to 3D physical space and discretized and synchronized in the time dimension. Pixel intensity, optical flow vector and semantic segmentation label are extracted as feature channels. A feature field channel is set for the positioning data, and velocity and acceleration are set as supplementary feature channels; Based on the installation location and sampling time of the sensor data, the sensor data is filled into the corresponding feature channels, and the spatially sparse data in the sensor data is smoothed by Kriging interpolation to form a continuous feature field.

3. The method for anomaly detection in coal mine operation scenarios based on causal wavelet resonance networks according to claim 1, characterized in that, The causal concept refers to the risk event in the safety rules of coal mine operation scenarios; the causal kernel function is set as a mathematical model corresponding to the risk event.

4. The method for anomaly detection in coal mine operation scenarios based on causal wavelet resonance networks according to claim 1, characterized in that, The method further includes: Obtain labeled normal and abnormal samples, and use a contrastive learning loss function to maximize the distance between the normal and abnormal samples in the causal resonance spectrum space; The learnable parameters in the causal kernel function are optimized based on the maximized distance.

5. The method for anomaly detection in coal mine operation scenarios based on causal wavelet resonance networks according to claim 1, characterized in that, The method further includes: Collect various operation data in a standard operation scenario without abnormalities, and calculate the causal resonance spectrum of each operation data. The baseline harmonic spectrum and its covariance matrix are obtained by statistical averaging and covariance calculation of the causal resonance spectrum of each operation.

6. A coal mine operation scene anomaly detection system based on causal wavelet resonance network, applied to the coal mine operation scene anomaly detection method based on causal wavelet resonance network as described in any one of claims 1-5, characterized in that, The system includes: The high-dimensional spatiotemporal feature field construction module is used to acquire video data, positioning data, and sensor data of coal mine operation scenarios as multimodal perception data, and to uniformly map the multimodal perception data to a discrete four-dimensional spatiotemporal grid to construct a high-dimensional spatiotemporal feature field. The causal resonance spectrum extraction module is used to transform the high-dimensional spatiotemporal feature field using a causal wavelet resonance network to extract the real-time causal resonance spectrum characterizing the current coal mine operation scenario. The causal wavelet resonance network includes at least one causal wavelet transform layer, which uses a set of predefined causal wavelet basis functions that are sensitive to specific causal modes. The spatiotemporal detuning graph generation module is used to retrieve the baseline harmonic spectrum and its covariance matrix learned from data of a standard operation scenario without anomalies, calculate the Mahalanobis distance between the baseline harmonic spectrum and the real-time causal resonance spectrum, and generate a spatiotemporal detuning graph. An anomaly detection module is used to determine whether the detuning value at each spatiotemporal location in the spatiotemporal detuning diagram exceeds a preset threshold. If it does, it is determined to be an abnormal state. The module also analyzes the causal wavelet type and parameters that contribute the most to the detuning and generates interpretable anomaly alarm information that includes the location, degree, and causality of the anomaly.

7. The coal mine operation scene anomaly detection system based on causal wavelet resonance network according to claim 6, characterized in that, The high-dimensional spatiotemporal feature field is a tensor defined on a discrete four-dimensional spatiotemporal grid; the high-dimensional spatiotemporal feature field construction module is also used to map the two-dimensional pixel coordinates to three-dimensional physical space through camera calibration and three-dimensional reconstruction of the video data, and to discretize and synchronize them according to the time dimension, and extract pixel intensity, optical flow vector, and semantic segmentation label as feature channels. A feature field channel is set for the positioning data, and velocity and acceleration are set as supplementary feature channels. Based on the installation location and sampling time of the sensor data, the sensor data is filled into the corresponding feature channels, and the spatially sparse data in the sensor data is smoothed by Kriging interpolation to form a continuous feature field.

Images