A coal mine underground abnormal event rapid identification and hierarchical alarm method

Abstract

This invention provides a method for rapid identification and hierarchical alarm of abnormal events in coal mines, including the acquisition of mine operation images and environmental data and the construction of an anomaly perception dataset, the design of a mine anomaly perception generation model based on anomaly decoupling, the design of a mine anomaly event hierarchical alarm system based on a hierarchical risk decision tree, and the deployment of the mine anomaly perception and hierarchical alarm system. This application constructs a mine anomaly perception dataset for anomaly perception and risk reasoning by collecting operation images and environmental data from key mine operation areas in a time-series manner and annotating the start and end times of abnormal events. By integrating multi-level feature encoding, anomaly-driven feature decoupling, and time-series evolution modeling, it achieves joint perception of environmental anomalies, equipment anomalies, and personnel anomalies, and generates anomaly event hypotheses. Furthermore, by utilizing the combination relationships of abnormal events in terms of temporal continuity, consistency of type evolution, and progressive time span, it realizes hierarchical risk determination and alarm triggering of mine anomaly events.

Description

Technical Field

[0001] This invention relates to the field of mine abnormal event detection technology, specifically to a method for rapid identification and hierarchical alarm of abnormal events in coal mines. Background Technology

[0002] The underground working environment in mines is enclosed and complex, with equipment operation, personnel work, and environmental conditions interacting with each other, resulting in multi-source and dynamically changing safety risks. During mining, transportation, and equipment operation, some anomalies often do not initially manifest themselves, exhibiting a degree of concealment and phased development. If not detected and continuously tracked in a timely manner, they can accumulate over time, eventually leading to significant safety hazards. Furthermore, due to insufficient underground lighting, dust and smoke interference, and the complex structure of the working area, anomalies often appear in images with diverse shapes and unclear boundaries. Relying solely on manual inspections or fixed threshold rules makes it difficult to accurately judge the anomaly state and its development trend. Simultaneously, different anomalies are often temporally related and semantically progressive; detecting only a single anomaly cannot comprehensively reflect the overall risk level of the mine, easily leading to false alarms, missed alarms, or unreasonable alarm levels. Currently, methods for detecting anomalies in mines include:

[0003] (1) Deep learning-based method for detecting anomalies in mines

[0004] This method typically employs convolutional neural networks or their improved models to detect anomalies in mine operation images. By learning the spatial features of the images, it identifies equipment malfunctions, abnormal personnel behavior, or environmental anomalies. While this type of method possesses certain detection capabilities in specific scenarios, it often focuses on a single anomaly category or a short time window, lacking systematic modeling of the persistence and evolution of anomalous events. Consequently, it struggles to accurately distinguish between sporadic anomalies and persistent anomalies with potential risks.

[0005] (2) Rule-based or threshold-based mine alarm methods

[0006] Some mine safety systems employ rule-based or experience-based threshold-based alarm strategies, triggering alarms based on the frequency, duration, or simple statistical indicators of anomalies. While these methods are simple to implement and computationally inexpensive, they lack characterization of the temporal correlations and type evolution between anomalies. Alarm results rely on manual experience settings, making them difficult to adapt to complex and ever-changing mine operation scenarios and prone to problems such as alarm lag or unreasonable alarm levels.

[0007] (3) Mine risk assessment method based on multi-source information fusion

[0008] These methods attempt to improve risk identification capabilities by integrating video, sensor, and environmental monitoring data for comprehensive risk assessment, taking a multi-dimensional information perspective. However, these methods have complex system structures, rely heavily on data synchronization and model training, and have high deployment and maintenance costs, making them difficult to promote on a large scale in actual mine sites. Furthermore, they lack a clear, structured expression of the evolutionary logic of abnormal events themselves.

[0009] Therefore, existing technologies still have shortcomings in the continuous modeling, evolutionary relationship characterization, and hierarchical alarm mechanisms for abnormal events in mines, making it difficult to achieve highly reliable anomaly detection and risk decision-making in complex mine environments. Summary of the Invention

[0010] To address the shortcomings of existing mine anomaly detection technologies in terms of continuous modeling, evolutionary relationship characterization, and hierarchical alarm mechanisms for mine anomalies, which make it difficult to achieve highly reliable anomaly perception and risk decision-making in complex mine environments, this invention provides a method for rapid identification and hierarchical alarm of underground coal mine anomalies, thereby improving the intelligence and practicality of mine safety monitoring systems.

[0011] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0012] A method for rapid identification and hierarchical alarm of abnormal events in coal mines includes the following steps:

[0013] S1. Mine Operation Image and Environmental Data Acquisition and Anomaly Perception Dataset Construction: Through sensors fixedly deployed in key underground mine operation areas, operation image data and environmental data of mining faces, transport roadways, equipment operating areas, and personnel passage areas are continuously collected; the collected data are organized in time windows, and an anomaly labeling strategy is adopted to label anomalies that reflect deviations from the normal operating mode. The anomalies include environmental anomalies, equipment status anomalies, and personnel behavior anomalies. Based on this, an anomaly perception dataset for complex mine operation scenarios is constructed.

[0014] S2. Design of Mine Anomaly Perception Generation Model Based on Anomaly Decoupling: Using the anomaly perception dataset sample input constructed in step S1 as the input of the mine anomaly perception generation model, multi-scale feature representations of scene structure, motion state, and environmental changes are extracted through multi-level feature encoding; based on the anomaly driving mechanism, the features are decoupled to generate environmental anomaly factors, equipment anomaly factors, and personnel anomaly factors, and the change characteristics of each anomaly factor within a continuous time window are characterized by time series modeling; further, based on the combination relationship of anomaly factors in the dimensions of time, space, and evolution trend, multiple parallel anomaly event units are constructed to form a set of mine anomaly events;

[0015] S3. Design of a hierarchical alarm system for mine abnormal events based on a hierarchical risk decision tree: For the set of mine abnormal events, a training-free hierarchical risk decision tree algorithm is constructed to screen and determine the risk of abnormal events layer by layer; in different decision layers, the abnormal events are assessed from coarse to fine risk by combining the persistence characteristics, abnormal factor combination relationship and evolution trend, and general abnormal, risk abnormal and high-risk abnormal events are identified step by step, and alarm signals of the corresponding level are triggered according to the judgment results, so as to realize the progressive multi-level alarm from abnormal perception to risk intervention;

[0016] S4. Deployment of Mine Anomaly Detection and Hierarchical Alarm System: The mine anomaly detection generation model designed in step S2 and the mine anomaly event hierarchical alarm system based on hierarchical risk decision tree designed in step S3 are deployed to edge computing nodes and local servers at the mine site. Real-time image processing and anomaly event generation are completed at the edge, and anomaly aggregation analysis and multi-level alarm decision are completed at the central end. The system continuously outputs anomaly events and their corresponding risk level information, pushes the alarm results to the mine safety management platform and links the audible and visual alarm devices to realize real-time detection, hierarchical early warning and safety response of mine anomaly events.

[0017] Furthermore, step S1, the acquisition of mine operation images and environmental data, and the construction of an anomaly perception dataset, includes:

[0018] S11. Sensor Deployment: Visual sensors and environmental sensors are fixedly deployed within the mine operation area according to the operation process and risk distribution characteristics. The visual sensors and environmental sensors are synchronized using a unified time reference. The visual sensors are used to collect operation image data. The environmental sensor is used to collect operational environment data. ;

[0019] S12. Input Time Series Data Acquisition: Based on the sensors deployed in step S11, data is collected during the time period to be detected. A sequence of images of mine operations at each time step. and environmental sequence ,Right now , ;

[0020] S13. Anomaly Labeling: Anomaly labels are applied to the time-series data collected in step S12. The anomaly labels include the type of anomaly event. and the start time of the abnormal event and termination time A complete label for an exception event is: Since multiple abnormal events may exist within the detection period, the complete abnormal label is: ,in This represents the total number of abnormal events within that time period.

[0021] S14. Anomaly detection dataset creation: using the image sequences acquired in step S12. and environmental sequence As input data, the anomaly labels constructed in step S13 As output data, a dataset is obtained; repeat the data collection and labeling process to finally obtain the anomaly perception dataset used for model training in step S2.

[0022] Further, the design of the mine anomaly perception generation model based on anomaly decoupling in step S2 includes: a mine anomaly perception generation model and model training, consisting of a multi-level anomaly perception feature encoding module, an anomaly-driven anomaly factor decoupling module, an anomaly factor temporal evolution modeling module, and an anomaly recognition and generation module connected in sequence. The multi-level anomaly perception feature encoding module is used to perform multi-level feature encoding on continuously acquired mine operation image sequences and environmental sequences, extracting multi-scale feature representations that reflect changes in the mine scene structure, target state, and environment. The anomaly-driven anomaly factor decoupling module is used to eliminate semantic overlap between scene, target, and environmental features. The anomaly factor temporal evolution modeling module is used to capture the nonlinear trend of various factors evolving from normal working conditions to abnormal states. The anomaly recognition and generation module is used to transform the evolutionary features into an interpretable set of mine anomaly events. The model training is based on the anomaly perception dataset constructed in step S1, performing end-to-end training on the mine anomaly perception generation model based on anomaly decoupling. The optimizer, combined with a cosine annealing strategy, ultimately yields the trained mine anomaly perception generation model.

[0023] Furthermore, the mine multi-level anomaly perception feature encoding module includes:

[0024] With image sequences As input, the local transmittance is estimated using the dark channel prior enhancement algorithm and atmospheric veil compensation is applied. After mapping with the Sigmoid activation function, the visual transparency enhancement feature is output. Subsequently, through parallel processing of the scene geometry channel and the target dynamic semantic channel, the scene geometry channel extracts the edge frequency response of the roadway support and track through Gabor convolutional layers. Combined with residual connection layers and Group Normalization, 12 sets of this structure are stacked in succession to extract the roadway skeleton structure features that reflect the fixed benchmark of the mine. The target dynamic semantic channel is mapped to a patch sequence through a segmented embedding layer, utilizing... The encoding layer captures the long-distance dependence between the human and the machine, and combines this with a temporal translation layer to exchange features between adjacent frames. The temporal translation layer achieves inter-frame interaction with zero parameter overhead by shifting 1 / 4 of the channel dimension forward and backward. Similarly, 12 sets of this structure are stacked consecutively to output key target state features reflecting the dynamic poses of the human and the machine. ;

[0025] With environmental sequence For input, via The activation function performs linear rectification; then, extreme numerical fluctuations are captured through a global max pooling layer, and the logical connections between sensors are modeled by combining 1D convolutional layers and fully connected layers. By stacking three sets of this structure in succession, dynamic environmental characterization features reflecting the risk background of the working environment are obtained. ;

[0026] Finally, the multi-level anomaly detection feature encoding module in the mine outputs features. , , .

[0027] Furthermore, the anomaly-driven anomaly factor decoupling module includes:

[0028] For features respectively , , Perform global average pooling and fully connected mapping to obtain vectorized features. , , Subsequently, in response to and By performing a Schmidt orthogonal projection to remove the structural reference from the target vector, ensuring that the two are orthogonal in geometric space, the decoupling structural factor is obtained. and After that, regarding , , Perform feature alignment based on cosine similarity to obtain aligned high-purity anomaly factors. , , ;

[0029] Finally, the anomaly-driven anomaly factor decoupling module outputs high-purity anomaly factors. , , .

[0030] Furthermore, the anomalous factor temporal evolution modeling module includes:

[0031] First, regarding high-purity anomalous factors , , Channel splicing is performed to generate a comprehensive mine characterization sequence. Subsequently, the multi-expert temporal modeling unit designed in this invention performs temporal feature purification to obtain spatiotemporal enhancement features. ;

[0032] The multi-expert temporal modeling unit comprises N parallel expert heads. Each expert head utilizes a 1D dilated convolutional layer to capture local trends with different receptive fields. Subsequently, long-range temporal dependencies are modeled through a bidirectional recurrent gating unit, and then... The activation function performs nonlinear compression; finally, the outputs of N expert heads are dynamically weighted and fused through an attention gating mechanism to output spatiotemporal enhancement features. .

[0033] Furthermore, the anomaly identification and generation module includes:

[0034] Enhanced features through spatiotemporal deepening The input is processed through a fully connected layer and the LeakyReLU activation function to reduce the dimensionality of the features, resulting in an anomaly event vector. Then, the following three detection heads are used to detect abnormal events:

[0035] Abnormal event classification branches: via fully connected layer and The activation function outputs the specific exception event type. ;

[0036] Evolutionary origin prediction branch: via fully connected layers and Activation function, predicting the start timestamp of the initial appearance of abnormal signs. ;

[0037] Evolutionary endpoint prediction branch: via fully connected layers and Activation function, predicting the termination timestamp of an event's outbreak or the reaching of a critical state. ;

[0038] Finally, abnormal event types Start timestamp End timestamp Together they constitute a risk anomaly event The model detected a total of An abnormal risk event ,Will As a collection of abnormal events .

[0039] Furthermore, step S3, the design of a hierarchical alarm system for mine abnormal events based on a hierarchical risk decision tree, includes:

[0040] S31. Normalized Time Coding: This involves encoding a set of abnormal events... Mapped to a unified time axis representation space, using time interval encoding operators Each anomalous event is mapped to an occupancy vector on a unified time axis. ,in For indicator functions, Indicates the first The occupancy vector of each risky and abnormal event, and , For event stamps, and , For time-step logic and operations, The starting timestamp, For the end timestamp; Used to describe the duration of an abnormal event over a time dimension;

[0041] Furthermore, based on the occupied vector corresponding to each abnormal event... The initial activation vector will be the set of anomalous events. The events are sequentially arranged to form a set of abnormal event sequences that satisfy temporal consistency. ;

[0042] S32. Determining the Temporal Continuity of Abnormal Events: As the first decision-making layer, this step is used to determine whether abnormal events exhibit persistent or clustered characteristics from a temporal structure perspective, in order to initially distinguish between scattered anomalies and potential risk anomalies; specifically,

[0043] For the set of abnormal event sequences obtained in step S31 Define the operator for the temporal continuity relationship between adjacent abnormal events. Calculate the degree of overlap between two anomalous events on a unified timeline. ,in Indicates the relationship with the first The occupancy vectors of neighboring anomalies of a risky anomaly; based on Abnormal events are classified into the following two types:

[0044] (1) If If the value is 0, it means that the event only occurs in a single time period and has no time overlap with adjacent events, which is defined as a time-isolated anomalous event;

[0045] (2) If If the value is not 0, it means that the event consists of two or more consecutive or overlapping abnormal events on the timeline, which is defined as a time-continuous abnormal event;

[0046] All events defined as time-continuous anomalies constitute the key anomaly subset of the first decision layer's output. This allows us to obtain the set of occupied vectors corresponding to the key anomaly subsets. ;

[0047] S33. Consistency Determination of Abnormal Event Type Evolution: As the second decision-making layer, it is used to determine, from the semantic level of abnormalities, whether time-continuous abnormalities constitute a risk process with a clear evolutionary direction, focusing on a subset of abnormalities. Extract the predicted risk category for each event. ;

[0048] Then, based on the predefined prior evolutionary reachability relationships between anomaly types, the first... The event is directed to the first Heterogeneity determination factor of event evolution ; where a subsequent event can be derived from the preceding event type through temporal evolution, The value is 1 if it is true and 0 otherwise; specifically, the evolutionary reachability relationship is determined by the causes of anomalies, physical constraints, and visual degradation mechanisms.

[0049] Therefore, a subset of anomalies can be identified for focused attention. Heterogeneity criterion factors among all events are used to ultimately obtain a set of heterogeneity criterion factors. ;

[0050] S34. Progressive Judgment of Abnormal Event Time Span: As the third decision-making layer, it is used to make a final judgment on whether an abnormal process constitutes a risky abnormal event requiring an alarm, based on the determination of the temporal continuity and consistency of the type evolution of the abnormal event, and from the perspective of the completeness and directional stability of the abnormal event's evolution along a unified time axis; specifically,

[0051] Heterogeneity determination factor set for the output of the second decision layer On a unified timeline, the following progressive determination process is performed on the sequence of anomalous events: For each pair of adjacent anomalous events, the corresponding anomalous type is read from the set of heterogeneity determination factors. In the context of positive and negative reachability relationships, if there exists only a one-way reachability relationship from the first occurrence of an anomaly to the second occurrence of an anomaly, and no negative reachability relationship exists, then the anomaly event pair is determined to satisfy the consistency of unidirectional type evolution. If both positive and negative reachability relationships exist simultaneously, then the anomaly event pair is determined to have undergone semantic backtracking, thus obtaining the set of heterogeneity determination factors. The set of evolutionary relationships of all anomalous event pairs in the middle ;

[0052] Based on the set of evolutionary relationships When all adjacent abnormal events corresponding to any abnormal event satisfy the consistency of unidirectional type evolution, the abnormal process is determined to have the characteristic of time span progression, and such abnormal events are marked as significant risk abnormal events with continuous evolution risk characteristics.

[0053] Finally, an alarm signal is triggered for the significant risk anomaly, and the corresponding risk anomaly type, as well as the start and end times of the anomaly, are returned.

[0054] Furthermore, step S4, the deployment of the mine anomaly detection and hierarchical alarm system, includes:

[0055] S41. Deployment of Model and Decision Module: The mine anomaly perception generation model trained in step S2, as well as the hierarchical risk decision tree algorithm and multi-level alarm system built based on S3, are uniformly deployed to the edge computing nodes and local servers at the mine site.

[0056] S42. Real-time access to multi-source data on site: During the mining operation, visual sensors and environmental sensors fixedly deployed in key operation areas are used to continuously collect mining operation image sequences and environmental data sequences that are time-aligned with them, and transmit them in real time to the locally deployed anomaly perception and risk decision-making system according to a unified time reference.

[0057] S43. Abnormal Event Perception and Hierarchical Risk Adjudication: The real-time data obtained in step S42 is input into the mine abnormal perception generation model deployed on the local server to complete the automatic identification and time location of abnormal events; then, based on the hierarchical risk decision tree algorithm, joint reasoning is performed on the detected abnormal events in terms of time continuity, consistency of type evolution and time span progression, and the corresponding risk abnormal events are output.

[0058] S44. Alarm Triggering and Risk Information Feedback: When the system determines a significant risk anomaly, it automatically triggers the corresponding audible and visual alarm device and feeds back the anomaly type, start time, and end time of the anomaly to the mine site safety management platform in real time to support timely intervention and safety decision-making for mine operation risks.

[0059] Compared with existing technologies, the rapid identification and hierarchical alarm method for abnormal events in coal mines provided by this invention has the following advantages:

[0060] 1. Achieve stable perception and accurate modeling of mine anomalies: By collecting time-series images of key mine operation areas and annotating the start and end times of anomalies, a mine anomaly perception dataset for anomaly perception and risk reasoning is constructed. By building a mine anomaly perception generation model based on anomaly decoupling, the model integrates multi-level perception feature encoding, anomaly-driven feature decoupling, and time-series evolution modeling to achieve decoupled modeling of environmental anomalies, equipment anomalies, and personnel anomalies in mine operation images, thereby improving the stability and accuracy of anomaly event identification in complex underground environments.

[0061] 2. Enhance the temporal consistency and evolution discrimination capability of abnormal events: Perform time-series modeling of abnormal factors and introduce a judgment mechanism for temporal continuity, type evolution consistency and time span progression at the level of abnormal events to avoid misjudgment and omission caused by relying only on single frame or short-term information.

[0062] 3. Hierarchical risk assessment and multi-level alarm without additional training: A hierarchical risk decision tree algorithm is constructed that does not require the introduction of additional model training. It utilizes the combination relationship between abnormal events in time and semantic space to complete the step-by-step screening and adjudication of risks, reducing the complexity of system deployment and improving the interpretability and reliability of risk assessment at the mine site.

[0063] 4. Improve the rationality and engineering usability of alarm results: By classifying abnormal events into hierarchical categories, the problem of a single abnormality directly triggering a high-level alarm is avoided, and the interference of invalid alarms on work organization and scheduling systems is reduced. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of the method for rapid identification and hierarchical alarm of abnormal events in coal mines provided by the present invention.

[0065] Figure 2 This is a network structure diagram of the mine anomaly perception generation model based on anomaly decoupling provided by the present invention.

[0066] Figure 3 This is a schematic diagram of the hierarchical alarm process for abnormal events in mines based on a hierarchical risk decision tree, provided by the present invention. Detailed Implementation

[0067] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below with reference to specific illustrations.

[0068] Please refer to Figure 1 As shown, this invention provides a method for rapid identification and hierarchical alarm of abnormal events in coal mines, comprising the following steps:

[0069] S1. Mine Operation Image and Environmental Data Acquisition and Anomaly Perception Dataset Construction: Through sensors fixedly deployed in key underground mine operation areas, operation image data and environmental data of mining faces, transport roadways, equipment operating areas, and personnel passage areas are continuously collected; the collected data are organized in time windows, and an anomaly labeling strategy is adopted to label anomalies that reflect deviations from the normal operating mode. The anomalies include environmental anomalies, equipment status anomalies, and personnel behavior anomalies. Based on this, an anomaly perception dataset for complex mine operation scenarios is constructed.

[0070] S2. Design of Mine Anomaly Perception Generation Model Based on Anomaly Decoupling: Using the anomaly perception dataset sample input constructed in step S1 as the input of the mine anomaly perception generation model, multi-scale feature representations of scene structure, motion state, and environmental changes are extracted through multi-level feature encoding; based on the anomaly driving mechanism, the features are decoupled to generate environmental anomaly factors, equipment anomaly factors, and personnel anomaly factors, and the change characteristics of each anomaly factor within a continuous time window are characterized by time series modeling; further, based on the combination relationship of anomaly factors in the dimensions of time, space, and evolution trend, multiple parallel anomaly event units are constructed to form a set of mine anomaly events;

[0071] S3. Design of a hierarchical alarm system for mine abnormal events based on a hierarchical risk decision tree: For the set of mine abnormal events, a training-free hierarchical risk decision tree algorithm is constructed to screen and determine the risk of abnormal events layer by layer; in different decision layers, the abnormal events are assessed from coarse to fine risk by combining the persistence characteristics, abnormal factor combination relationship and evolution trend, and general abnormal, risk abnormal and high-risk abnormal events are identified step by step, and alarm signals of the corresponding level are triggered according to the judgment results, so as to realize the progressive multi-level alarm from abnormal perception to risk intervention;

[0072] S4. Deployment of Mine Anomaly Detection and Hierarchical Alarm System: The mine anomaly detection generation model designed in step S2 and the mine anomaly event hierarchical alarm system based on hierarchical risk decision tree designed in step S3 are deployed to edge computing nodes and local servers at the mine site. Real-time image processing and anomaly event generation are completed at the edge, and anomaly aggregation analysis and multi-level alarm decision are completed at the central end. The system continuously outputs anomaly events and their corresponding risk level information, pushes the alarm results to the mine safety management platform and links the audible and visual alarm devices to realize real-time detection, hierarchical early warning and safety response of mine anomaly events.

[0073] As a specific embodiment, step S1, mine operation image and environmental data acquisition and anomaly perception dataset construction, includes:

[0074] S11. Sensor Deployment: Visual sensors and environmental sensors are fixedly deployed within the mine operation area according to the operation process and risk distribution characteristics. The visual sensors and environmental sensors are synchronized using a unified time reference. The visual sensors are used to collect operation image data. The operational image data includes information such as equipment operating status and personnel operational behavior; the environmental sensor is used to collect operational environment data. Work environment data includes dust concentration ,temperature ,temperature Equipment vibration frequency Equipment operating speed ,Right now ;

[0075] S12. Input Time Series Data Acquisition: Based on the sensors deployed in step S11, data is collected during the time period to be detected. A sequence of images of mine operations at each time step. and environmental sequence ,Right now , ;

[0076] S13. Anomaly Labeling: Anomaly labels are applied to the time-series data collected in step S12. The anomaly labels include the type of anomaly event. and the start time of the abnormal event and termination time The types of abnormal events include abnormal brightness, abnormal lighting fluctuations, abnormal partial occlusion, abnormal target loss (including personnel or critical mechanical parts), and abnormal target inconsistency. Therefore, a complete label for an abnormal event is... Since multiple abnormal events may exist within the detection period, the complete abnormal label is: ,in This represents the total number of abnormal events within that time period.

[0077] S14. Anomaly detection dataset creation: using the image sequences acquired in step S12. and environmental sequence As input data, the anomaly labels constructed in step S13 As output data, a dataset is obtained; repeat the data collection and labeling process to finally obtain the anomaly perception dataset used for model training in step S2.

[0078] As a specific embodiment, this invention addresses the characteristics of complex lighting conditions, strong dust interference, diverse operational targets, and temporal evolution of anomalies in mining scenarios. It designs a mine anomaly perception and generation model based on anomaly decoupling. The model uses a sequence of mine operation images... and the environmental sequence that is time-aligned with it As input, output risk and abnormal events. This is used to support subsequent hierarchical risk decision-making and multi-level alarms. Accordingly, the design of the mine anomaly perception generation model based on anomaly decoupling in step S2 includes: a mine anomaly perception generation model composed of a multi-level anomaly perception feature encoding module, an anomaly-driven anomaly factor decoupling module, an anomaly factor temporal evolution modeling module, and an anomaly identification and generation module connected in sequence, as well as model training.

[0079] Please refer to Figure 2 As shown, the multi-level anomaly perception feature encoding module for mines is used to perform multi-level feature encoding on continuously acquired mine operation image sequences and environmental sequences, extracting multi-scale feature representations that can reflect the mine scene structure, target state, and environmental changes. Specifically, it includes:

[0080] With image sequences As input, the local transmittance is estimated using the dark channel prior enhancement algorithm and atmospheric veil compensation is applied. After mapping with the Sigmoid activation function, the visual transparency enhancement feature is output. Subsequently, the process involves parallel processing through two heterogeneous dual channels: a scene geometry channel and a target dynamic semantic channel. The scene geometry channel extracts the edge frequency response of the roadway support and track using Gabor convolutional layers. Combined with residual connection layers and Group Normalization, 12 sets of this structure are stacked consecutively to extract the roadway skeleton structure features that reflect the fixed benchmark of the mine. The target dynamic semantic channel is mapped to a patch sequence through a segmented embedding layer, utilizing... The encoding layer captures the long-distance dependence between the human and the machine, and combines this with a temporal translation layer to exchange features between adjacent frames. The temporal translation layer achieves inter-frame interaction with zero parameter overhead by shifting 1 / 4 of the channel dimension forward and backward. Similarly, 12 sets of this structure are stacked consecutively to output key target state features reflecting the dynamic poses of the human and the machine. ;

[0081] With environmental sequence For input, via The activation function performs linear rectification; then, extreme numerical fluctuations are captured through a global max pooling layer, and the logical connections between sensors are modeled by combining 1D convolutional layers and fully connected layers. By stacking three sets of this structure in succession, dynamic environmental characterization features reflecting the risk background of the working environment are obtained. ;

[0082] Finally, the multi-level anomaly detection feature encoding module in the mine outputs features. , , .

[0083] Please refer to Figure 2 As shown, the anomaly-driven anomaly factor decoupling module is used to eliminate semantic overlap between scene, target, and environmental features, specifically including:

[0084] For features respectively , , Perform global average pooling and fully connected mapping to obtain vectorized features. , , Subsequently, in response to and By performing a Schmidt orthogonal projection to remove the structural reference from the target vector, ensuring that the two are orthogonal in geometric space, the decoupling structural factor is obtained. and After that, regarding , , Perform feature alignment based on cosine similarity to obtain aligned high-purity anomaly factors. , , ;

[0085] Specifically, the feature alignment operation constructs an independence constraint loss function by calculating the sum of the cosine similarities between each pair of the three elements. The goal is to minimize this loss and ensure that environmental noise, structural changes and target behavior are completely decoupled in terms of physical meaning in the latent space.

[0086] Finally, the anomaly-driven anomaly factor decoupling module outputs high-purity anomaly factors. , , .

[0087] Please refer to Figure 2 As shown, the anomalous factor time-series evolution modeling module is used to capture the nonlinear trend of various factors evolving from normal operating conditions to anomalous states, specifically including:

[0088] First, regarding high-purity anomalous factors , , Channel splicing is performed to generate a comprehensive mine characterization sequence. Subsequently, the multi-expert temporal modeling unit designed in this invention performs temporal feature purification to obtain spatiotemporal enhancement features. ;

[0089] The multi-expert temporal modeling unit comprises N parallel expert heads. Each expert head utilizes a 1D dilated convolutional layer to capture local trends with different receptive fields. Long-range temporal dependencies are then modeled using a bidirectional recurrent gated unit (GRU). The activation function performs nonlinear compression; finally, the outputs of N expert heads are dynamically weighted and fused through an attention gating mechanism to output spatiotemporal enhancement features. .

[0090] Please refer to Figure 2 As shown, the anomaly identification and generation module is used to transform evolutionary features into an interpretable set of mine anomaly events, specifically including:

[0091] Enhanced features through spatiotemporal deepening The input is processed through a fully connected layer and the LeakyReLU activation function to reduce the dimensionality of the features, resulting in an anomaly event vector. Then, the following three detection heads are used to detect abnormal events:

[0092] (1) Abnormal event classification branch: via fully connected layer and The activation function outputs the specific exception event type. ;

[0093] (2) Evolutionary starting point prediction branch: via fully connected layers and Activation function, predicting the start timestamp of the initial appearance of abnormal signs. ;

[0094] (3) Evolutionary endpoint prediction branch: via fully connected layers and Activation function, predicting the termination timestamp of an event's outbreak or the reaching of a critical state. ;

[0095] Finally, abnormal event types Start timestamp End timestamp Together they constitute a risk anomaly event The model detected a total of An abnormal risk event ,Will As a collection of abnormal events .

[0096] Model Training: Based on the anomaly perception dataset constructed in step S1, the mine anomaly perception generation model based on anomaly decoupling is trained end-to-end. The optimizer, combined with a cosine annealing strategy, ultimately yields the trained mine anomaly perception generation model.

[0097] As a specific embodiment, the detection results based on step S2 Treat it as a set of abnormal events This leads to the development of a hierarchical risk decision-making algorithm based on the internal relationships within a set of abnormal events. Without introducing any additional observation data or parameter training, it achieves multi-level risk assessment and alarm triggering through the combination relationships of abnormal events in time and semantic space. Please refer to [link / reference needed] for details. Figure 3 As shown, step S3, the design of the mine abnormal event hierarchical alarm system based on a hierarchical risk decision tree, includes:

[0098] S31. Normalized Time Coding: This involves encoding a set of abnormal events... Mapped to a unified time axis representation space, using time interval encoding operators Each anomalous event is mapped to an occupancy vector on a unified time axis. ,in For indicator functions, Indicates the first The occupancy vector of each risky and abnormal event, and , For event stamps, and , For time-step logic and operations, The starting timestamp, For the end timestamp; Used to describe the duration of an abnormal event over a time dimension;

[0099] Furthermore, based on the occupied vector corresponding to each abnormal event... The initial activation vector will be the set of anomalous events. The events are sequentially arranged to form a set of abnormal event sequences that satisfy temporal consistency. .

[0100] S32. Determining the Temporal Continuity of Abnormal Events: As the first decision-making layer, this step is used to determine whether abnormal events exhibit persistent or clustered characteristics from a temporal structure perspective, in order to initially distinguish between scattered anomalies and potential risk anomalies; specifically,

[0101] For the set of abnormal event sequences obtained in step S31 Define the operator for the temporal continuity relationship between adjacent abnormal events. Calculate the degree of overlap between two anomalous events on a unified timeline. ,in Indicates the relationship with the first The occupancy vectors of neighboring anomalies of a risky anomaly; based on Abnormal events are classified into the following two types:

[0102] (1) If If the value is 0, it means that the event only occurs in a single time period and has no time overlap with adjacent events, which is defined as a time-isolated anomalous event;

[0103] (2) If If the value is not 0, it means that the event consists of two or more consecutive or overlapping abnormal events on the timeline, which is defined as a time-continuous abnormal event;

[0104] All events defined as time-continuous anomalies constitute the key anomaly subset of the first decision layer's output. This allows us to obtain the set of occupied vectors corresponding to the key anomaly subsets. .

[0105] S33. Consistency Determination of Abnormal Event Type Evolution: As the second decision-making layer, it is used to determine, from the semantic level of abnormalities, whether time-continuous abnormalities constitute a risk process with a clear evolutionary direction, focusing on a subset of abnormalities. Extract the predicted risk category for each event. ;

[0106] Then, based on the predefined prior evolutionary reachability relationships between anomaly types, the first... The event is directed to the first Heterogeneity determination factor of event evolution ; where a subsequent event can be derived from the preceding event type through temporal evolution, The value is 1 if the condition is met, and 0 otherwise. Specifically, the evolutionary reachability relationship is determined by the causes of anomalies, physical constraints, and visual degradation mechanisms, for example:

[0107] (1) When the mine lighting is insufficient or the light changes frequently, the contrast of local areas in the image is easily reduced or the area is covered by dark areas, which in turn manifests as abnormal local occlusion.

[0108] (2) When the key parts of the target are continuously obscured, the target’s appearance or structural information is incomplete, which can easily lead to unstable target state estimation, manifested as inconsistent and abnormal target state.

[0109] (3) When the degree of occlusion is high or the duration is long, the anomaly can skip the intermediate state and directly evolve into the target missing anomaly.

[0110] Therefore, a subset of anomalies can be identified for focused attention. Heterogeneity criterion factors among all events are used to ultimately obtain a set of heterogeneity criterion factors. .

[0111] S34. Progressive Judgment of Abnormal Event Time Span: As the third decision-making layer, it is used to make a final judgment on whether an abnormal process constitutes a risky abnormal event requiring an alarm, based on the determination of the temporal continuity and consistency of the type evolution of the abnormal event, and from the perspective of the completeness and directional stability of the abnormal event's evolution along a unified time axis; specifically,

[0112] Heterogeneity determination factor set for the output of the second decision layer On a unified timeline, the following progressive determination process is performed on the sequence of anomalous events: For each pair of adjacent anomalous events, the corresponding anomalous type is read from the set of heterogeneity determination factors. In the context of positive and negative reachability relationships, if there exists only a one-way reachability relationship from the first occurrence of an anomaly to the second occurrence of an anomaly, and no negative reachability relationship exists, then the anomaly event pair is determined to satisfy the consistency of unidirectional type evolution. If both positive and negative reachability relationships exist simultaneously, then the anomaly event pair is determined to have undergone semantic backtracking, thus obtaining the set of heterogeneity determination factors. The set of evolutionary relationships of all anomalous event pairs in the middle ;

[0113] Based on the set of evolutionary relationships When all adjacent abnormal events corresponding to any abnormal event satisfy the consistency of unidirectional type evolution, the abnormal process is determined to have the characteristic of time span progression, and such abnormal events are marked as significant risk abnormal events with continuous evolution risk characteristics.

[0114] Finally, an alarm signal is triggered for the significant risk anomaly, and the corresponding risk anomaly type, as well as the start and end times of the anomaly, are returned.

[0115] As a specific embodiment, to achieve real-time perception, hierarchical risk assessment, and multi-level alarm triggering of abnormal events at the mine operation site, this invention, based on the aforementioned mine anomaly perception generation model and hierarchical risk decision tree algorithm, deploys and operates the system on-site to achieve continuous monitoring and risk response of the mine operation process. Specifically, step S4, the deployment of the mine anomaly perception and hierarchical alarm system, includes:

[0116] S41. Deployment of Model and Decision Module: The mine anomaly perception generation model trained in step S2, as well as the hierarchical risk decision tree algorithm and multi-level alarm system built based on S3, are uniformly deployed to the edge computing nodes and local servers at the mine site.

[0117] S42. Real-time access to multi-source data on site: During the mining operation, visual sensors and environmental sensors fixedly deployed in key operation areas are used to continuously collect mining operation image sequences and environmental data sequences that are time-aligned with them, and transmit them in real time to the locally deployed anomaly perception and risk decision-making system according to a unified time reference.

[0118] S43. Abnormal Event Perception and Hierarchical Risk Adjudication: The real-time data obtained in step S42 is input into the mine abnormal perception generation model deployed on the local server to complete the automatic identification and time location of abnormal events; then, based on the hierarchical risk decision tree algorithm, joint reasoning is performed on the detected abnormal events in terms of time continuity, consistency of type evolution and time span progression, and the corresponding risk abnormal events are output.

[0119] S44. Alarm Triggering and Risk Information Feedback: When the system determines a significant risk anomaly, it automatically triggers the corresponding audible and visual alarm device and feeds back key information such as the anomaly type, start time, and end time of the anomaly to the mine site's safety management platform or dispatch system in real time to support timely intervention and safety decision-making for mine operation risks.

[0120] Compared with existing technologies, the rapid identification and hierarchical alarm method for abnormal events in coal mines provided by this invention has the following advantages:

[0121] 1. Achieve stable perception and accurate modeling of mine anomalies: By collecting time-series images of key mine operation areas and annotating the start and end times of anomalies, a mine anomaly perception dataset for anomaly perception and risk reasoning is constructed. By building a mine anomaly perception generation model based on anomaly decoupling, the model integrates multi-level perception feature encoding, anomaly-driven feature decoupling, and time-series evolution modeling to achieve decoupled modeling of environmental anomalies, equipment anomalies, and personnel anomalies in mine operation images, thereby improving the stability and accuracy of anomaly event identification in complex underground environments.

[0122] 2. Enhance the temporal consistency and evolution discrimination capability of abnormal events: Perform time-series modeling of abnormal factors and introduce a judgment mechanism for temporal continuity, type evolution consistency and time span progression at the level of abnormal events to avoid misjudgment and omission caused by relying only on single frame or short-term information.

[0123] 3. Hierarchical risk assessment and multi-level alarm without additional training: A hierarchical risk decision tree algorithm is constructed that does not require the introduction of additional model training. It utilizes the combination relationship between abnormal events in time and semantic space to complete the step-by-step screening and adjudication of risks, reducing the complexity of system deployment and improving the interpretability and reliability of risk assessment at the mine site.

[0124] 4. Improve the rationality and engineering usability of alarm results: By classifying abnormal events into hierarchical categories, the problem of a single abnormality directly triggering a high-level alarm is avoided, and the interference of invalid alarms on work organization and scheduling systems is reduced.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for rapid identification and hierarchical alarm of abnormal events in coal mines, characterized in that, Includes the following steps: S1. Mine Operation Image and Environmental Data Acquisition and Anomaly Perception Dataset Construction: Through sensors fixedly deployed in key underground mine operation areas, operation image data and environmental data of mining faces, transport roadways, equipment operating areas, and personnel passage areas are continuously collected; the collected data are organized in time windows, and an anomaly labeling strategy is adopted to label anomalies that reflect deviations from the normal operating mode. The anomalies include environmental anomalies, equipment status anomalies, and personnel behavior anomalies. Based on this, an anomaly perception dataset for complex mine operation scenarios is constructed. S2. Design of Mine Anomaly Perception Generation Model Based on Anomaly Decoupling: Using the anomaly perception dataset sample input constructed in step S1 as the input of the mine anomaly perception generation model, multi-scale feature representations of scene structure, motion state, and environmental changes are extracted through multi-level feature encoding; based on the anomaly driving mechanism, the features are decoupled to generate environmental anomaly factors, equipment anomaly factors, and personnel anomaly factors, and the change characteristics of each anomaly factor within a continuous time window are characterized by time series modeling; further, based on the combination relationship of anomaly factors in the dimensions of time, space, and evolution trend, multiple parallel anomaly event units are constructed to form a set of mine anomaly events; S3. Design of a hierarchical alarm system for mine abnormal events based on a hierarchical risk decision tree: For the set of mine abnormal events, a hierarchical risk decision tree algorithm that does not require training is constructed to screen and determine the risk of abnormal events layer by layer. At different decision-making levels, the risk assessment of anomalies is carried out from coarse to fine by combining the continuous characteristics of anomalies, the combination relationship of anomaly factors and the evolution trend. General anomalies, risk anomalies and high-risk anomalies are identified step by step, and alarm signals of the corresponding level are triggered according to the judgment results, so as to realize the progressive multi-level alarm from anomaly perception to risk intervention. S4. Deployment of Mine Anomaly Detection and Hierarchical Alarm System: The mine anomaly detection generation model designed in step S2 and the mine anomaly event hierarchical alarm system based on hierarchical risk decision tree designed in step S3 are deployed to edge computing nodes and local servers at the mine site. Real-time image processing and anomaly event generation are completed at the edge, and anomaly aggregation analysis and multi-level alarm decision are completed at the central end. The system continuously outputs anomaly events and their corresponding risk level information, pushes the alarm results to the mine safety management platform and links the audible and visual alarm devices to realize real-time detection, hierarchical early warning and safety response of mine anomaly events.

2. The method for rapid identification and hierarchical alarm of abnormal events in coal mines according to claim 1, characterized in that, Step S1, which involves the acquisition of mine operation images and environmental data and the construction of an anomaly perception dataset, includes: S11. Sensor Deployment: Visual sensors and environmental sensors are fixedly deployed within the mine operation area according to the operation process and risk distribution characteristics. The visual sensors and environmental sensors are synchronized using a unified time reference. The visual sensors are used to collect operation image data. The environmental sensor is used to collect operational environment data. ; S12. Input Time Series Data Acquisition: Based on the sensors deployed in step S11, data is collected during the time period to be detected. A sequence of images of mine operations at each time step. and environmental sequence ,Right now , ; S13. Anomaly Labeling: Anomaly labels are applied to the time-series data collected in step S12. The anomaly labels include the type of anomaly event. and the start time of the abnormal event and termination time A complete label for an exception event is: Since multiple abnormal events may exist within the detection period, the complete abnormal label is: ,in This represents the total number of abnormal events within that time period. S14. Anomaly detection dataset creation: using the image sequences acquired in step S12. and environmental sequence As input data, the anomaly labels constructed in step S13 As output data, a dataset is obtained; repeat the data collection and labeling process to finally obtain the anomaly perception dataset used for model training in step S2.

3. The method for rapid identification and hierarchical alarm of abnormal events in coal mines according to claim 1, characterized in that, The design of the mine anomaly perception generation model based on anomaly decoupling in step S2 includes: a mine anomaly perception generation model and model training, consisting of a multi-level anomaly perception feature encoding module, an anomaly-driven anomaly factor decoupling module, an anomaly factor temporal evolution modeling module, and an anomaly recognition and generation module connected in sequence. The multi-level anomaly perception feature encoding module performs multi-level feature encoding on continuously acquired mine operation image sequences and environmental sequences, extracting multi-scale feature representations that reflect changes in the mine scene structure, target state, and environment. The anomaly-driven anomaly factor decoupling module eliminates semantic overlap between scene, target, and environmental features. The anomaly factor temporal evolution modeling module captures the nonlinear trend of various factors evolving from normal working conditions to abnormal states. The anomaly recognition and generation module transforms the evolutionary features into an interpretable set of mine anomaly events. The model training is based on the anomaly perception dataset constructed in step S1, performing end-to-end training on the mine anomaly perception generation model based on anomaly decoupling. The optimizer, combined with a cosine annealing strategy, ultimately yields the trained mine anomaly perception generation model.

4. The method for rapid identification and hierarchical alarm of abnormal events in coal mines according to claim 3, characterized in that, The mine multi-level anomaly detection feature encoding module includes: With image sequences As input, the local transmittance is estimated using the dark channel prior enhancement algorithm and atmospheric veil compensation is applied. After mapping with the Sigmoid activation function, the visual transparency enhancement feature is output. Subsequently, through parallel processing of the scene geometry channel and the target dynamic semantic channel, the scene geometry channel extracts the edge frequency response of the roadway support and track through Gabor convolutional layers. Combined with residual connection layers and Group Normalization, 12 sets of this structure are stacked in succession to extract the roadway skeleton structure features that reflect the fixed benchmark of the mine. The target dynamic semantic channel is mapped to a patch sequence through a segmented embedding layer, utilizing... The encoding layer captures the long-distance dependence between the human and the machine, and combines this with a temporal translation layer to exchange features between adjacent frames. The temporal translation layer achieves inter-frame interaction with zero parameter overhead by shifting 1 / 4 of the channel dimension forward and backward. Similarly, 12 sets of this structure are stacked consecutively to output key target state features reflecting the dynamic poses of the human and the machine. ; With environmental sequence For input, via The activation function performs linear rectification; then, extreme numerical fluctuations are captured through a global max pooling layer, and the logical connections between sensors are modeled by combining 1D convolutional layers and fully connected layers. By stacking three sets of this structure in succession, dynamic environmental characterization features reflecting the risk background of the working environment are obtained. ; Finally, the multi-level anomaly detection feature encoding module in the mine outputs features. , , .

5. The method for rapid identification and hierarchical alarm of abnormal events in coal mines according to claim 4, characterized in that, The anomaly-driven anomaly factor decoupling module includes: For features respectively , , Perform global average pooling and fully connected mapping to obtain vectorized features. , , Subsequently, in response to and By performing a Schmidt orthogonal projection to remove the structural reference from the target vector, ensuring that the two are orthogonal in geometric space, the decoupling structural factor is obtained. and After that, regarding , , Perform feature alignment based on cosine similarity to obtain aligned high-purity anomaly factors. , , ; Finally, the anomaly-driven anomaly factor decoupling module outputs high-purity anomaly factors. , , .

6. The method for rapid identification and hierarchical alarm of abnormal events in coal mines according to claim 5, characterized in that, The anomalous factor temporal evolution modeling module includes: First, regarding high-purity anomalous factors , , Channel splicing is performed to generate a comprehensive mine characterization sequence. Subsequently, the multi-expert temporal modeling unit designed in this invention performs temporal feature purification to obtain spatiotemporal enhancement features. ; The multi-expert temporal modeling unit comprises N parallel expert heads. Each expert head utilizes a 1D dilated convolutional layer to capture local trends with different receptive fields. Subsequently, long-range temporal dependencies are modeled through a bidirectional recurrent gating unit, and then... The activation function performs nonlinear compression; finally, the outputs of N expert heads are dynamically weighted and fused through an attention gating mechanism to output spatiotemporal enhancement features. .

7. The method for rapid identification and hierarchical alarm of abnormal events in coal mines according to claim 6, characterized in that, The anomaly identification and generation module includes: Enhanced features through spatiotemporal deepening The input is processed through a fully connected layer and the LeakyReLU activation function to reduce the dimensionality of the features, resulting in an anomaly event vector. Then, the following three detection heads are used to detect abnormal events: Abnormal event classification branches: via fully connected layer and The activation function outputs the specific exception event type. ; Evolutionary origin prediction branch: via fully connected layers and Activation function, predicting the start timestamp of the initial appearance of abnormal signs. ; Evolutionary endpoint prediction branch: via fully connected layers and Activation function, predicting the termination timestamp of an event's outbreak or the reaching of a critical state. ; Finally, abnormal event types Start timestamp End timestamp Together they constitute a risk anomaly event The model detected a total of An abnormal risk event ,Will As a collection of abnormal events .

8. The method for rapid identification and hierarchical alarm of abnormal events in coal mines according to claim 1, characterized in that, Step S3, the design of a hierarchical alarm system for mine abnormal events based on a hierarchical risk decision tree, includes: S31. Normalized Time Coding: This involves encoding a set of abnormal events... Mapped to a unified time axis representation space, using time interval encoding operators Each anomalous event is mapped to an occupancy vector on a unified time axis. ,in For indicator functions, Indicates the first The occupancy vector of each risky and abnormal event, and , For event stamps, and , For time-step logic and operations, The starting timestamp, For the end timestamp; Used to describe the duration of an abnormal event over a time dimension; Furthermore, based on the occupied vector corresponding to each abnormal event... The initial activation vector will be the set of anomalous events. The events are sequentially arranged to form a set of abnormal event sequences that satisfy temporal consistency. ; S32. Determining the Temporal Continuity of Abnormal Events: As the first decision-making layer, this step is used to determine whether abnormal events exhibit persistent or clustered characteristics from a temporal structure perspective, in order to initially distinguish between scattered anomalies and potential risk anomalies; specifically, For the set of abnormal event sequences obtained in step S31 Define the operator for the temporal continuity relationship between adjacent abnormal events. Calculate the degree of overlap between two anomalous events on a unified timeline. ,in Indicates the relationship with the first The occupancy vectors of neighboring anomalies of a risky anomaly; based on Abnormal events are classified into the following two types: (1) If If the value is 0, it means that the event only occurs in a single time period and has no time overlap with adjacent events, which is defined as a time-isolated anomalous event; (2) If If the value is not 0, it means that the event consists of two or more consecutive or overlapping abnormal events on the timeline, which is defined as a time-continuous abnormal event; All events defined as time-continuous anomalies constitute the key anomaly subset of the first decision layer's output. This allows us to obtain the set of occupied vectors corresponding to the key anomaly subsets. ; S33. Consistency Determination of Abnormal Event Type Evolution: As the second decision-making layer, it is used to determine, from the semantic level of abnormalities, whether time-continuous abnormalities constitute a risk process with a clear evolutionary direction, focusing on a subset of abnormalities. Extract the predicted risk category for each event. ; Then, based on the predefined prior evolutionary reachability relationships between anomaly types, the first... The event is directed to the first Heterogeneity determination factor of event evolution ; where a subsequent event can be derived from the preceding event type through temporal evolution, The value is 1 if it is true and 0 otherwise; specifically, the evolutionary reachability relationship is determined by the causes of anomalies, physical constraints, and visual degradation mechanisms. Therefore, a subset of anomalies can be identified for focused attention. Heterogeneity criterion factors among all events are used to ultimately obtain a set of heterogeneity criterion factors. ; S34. Progressive Judgment of Abnormal Event Time Span: As the third decision-making layer, it is used to make a final judgment on whether an abnormal process constitutes a risky abnormal event requiring an alarm, based on the determination of the temporal continuity and consistency of the type evolution of the abnormal event, and from the perspective of the completeness and directional stability of the abnormal event's evolution along a unified time axis; specifically, Heterogeneity determination factor set for the output of the second decision layer On a unified timeline, the following progressive determination process is performed on the sequence of anomalous events: For each pair of adjacent anomalous events, the corresponding anomalous type is read from the set of heterogeneity determination factors. In the context of positive and negative reachability relationships, if there exists only a one-way reachability relationship from the first occurrence of an anomaly to the second occurrence of an anomaly, and no negative reachability relationship exists, then the anomaly event pair is determined to satisfy the consistency of unidirectional type evolution. If both positive and negative reachability relationships exist simultaneously, then the anomaly event pair is determined to have undergone semantic backtracking, thus obtaining the set of heterogeneity determination factors. The set of evolutionary relationships of all anomalous event pairs in the middle ; Based on the set of evolutionary relationships When all adjacent abnormal events corresponding to any abnormal event satisfy the consistency of unidirectional type evolution, the abnormal process is determined to have the characteristic of time span progression, and such abnormal events are marked as significant risk abnormal events with continuous evolution risk characteristics. Finally, an alarm signal is triggered for the significant risk anomaly, and the corresponding risk anomaly type, as well as the start and end times of the anomaly, are returned.

9. The method for rapid identification and hierarchical alarm of abnormal events in coal mines according to claim 1, characterized in that, The deployment of the mine anomaly detection and hierarchical alarm system in step S4 includes: S41. Deployment of Model and Decision Module: The mine anomaly perception generation model trained in step S2, as well as the hierarchical risk decision tree algorithm and multi-level alarm system built based on S3, are uniformly deployed to the edge computing nodes and local servers at the mine site. S42. Real-time access to multi-source data on site: During the mining operation, visual sensors and environmental sensors fixedly deployed in key operation areas are used to continuously collect mining operation image sequences and environmental data sequences that are time-aligned with them, and transmit them in real time to the locally deployed anomaly perception and risk decision-making system according to a unified time reference. S43. Abnormal Event Perception and Hierarchical Risk Adjudication: The real-time data obtained in step S42 is input into the mine abnormal perception generation model deployed on the local server to complete the automatic identification and time location of abnormal events; then, based on the hierarchical risk decision tree algorithm, joint reasoning is performed on the detected abnormal events in terms of time continuity, consistency of type evolution and time span progression, and the corresponding risk abnormal events are output. S44. Alarm Triggering and Risk Information Feedback: When the system determines a significant risk anomaly, it automatically triggers the corresponding audible and visual alarm device and feeds back the anomaly type, start time, and end time of the anomaly to the mine site safety management platform in real time to support timely intervention and safety decision-making for mine operation risks.

Images