Artificial intelligence large model driven spatio-temporal information identification method and system

Abstract

The application discloses an artificial intelligence large model driven spatio-temporal information pair identification method and system, relates to the technical field of multi-source information identification, and comprises the following steps: acquiring mine multi-source spatio-temporal data, performing cleaning and denoising and spatio-temporal alignment processing, and generating a standard spatio-temporal data sequence; inputting the standard spatio-temporal data sequence into a pre-trained spatio-temporal correlation large model, performing multi-modal spatio-temporal feature extraction, and obtaining comprehensive spatio-temporal feature representation; based on the comprehensive spatio-temporal feature representation, constructing a dynamic anomaly identification network, respectively performing equipment abnormal displacement identification, personnel track deviation from a safety area identification and geographical area and safety production data anomaly correlation identification, and generating an initial anomaly information pair set; performing spatio-temporal confidence compensation and risk level fusion on the initial anomaly information pair set, and outputting an identification result containing an abnormal position coordinate, a time node and a risk level. The application solves the technical problem of insufficient mine multi-source spatio-temporal data correlation analysis in the prior art.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of multi-source information recognition, and in particular to a spatio-temporal information pair recognition method and system driven by an artificial intelligence large model. BACKGROUND

[0002] Mine safety production is the core concern in the mining field. With the popularization of the Internet of Things and sensing technology, various monitoring devices such as environmental monitoring sensors and personnel positioning systems are deployed in mine environments, generating massive multi-source spatio-temporal data. In the prior art, mine safety monitoring systems usually analyze these data independently, using simple threshold alarms or rule engines for anomaly detection, such as triggering device displacement alarms or personnel boundary warnings by preset fixed thresholds. However, these methods have significant defects: first, multi-source data is difficult to form a comprehensive safety situation awareness, and second, due to the complexity and dynamic changes of mine environments, simple threshold-based rules are difficult to adapt to changing conditions, leading to false positives and false negatives. SUMMARY

[0003] The present application provides a spatio-temporal information pair recognition method and system driven by an artificial intelligence large model, which solves the technical problem of insufficient correlation analysis of mine multi-source spatio-temporal data in the prior art.

[0004] In view of the above problems, the present application provides a spatio-temporal information pair recognition method and system driven by an artificial intelligence large model.

[0005] In a first aspect, the present application provides a spatio-temporal information pair recognition method driven by an artificial intelligence large model, the method comprising:

[0006] obtaining mine multi-source spatio-temporal data, cleaning and denoising the multi-source spatio-temporal data, and performing spatio-temporal alignment processing to generate a standard spatio-temporal data sequence;

[0007] inputting the standard spatio-temporal data sequence into a pre-trained spatio-temporal correlation large model to extract multi-modal spatio-temporal features and obtain a comprehensive spatio-temporal feature representation, wherein the comprehensive spatio-temporal feature representation includes a trend feature sequence of device position changes over time, a spatial matching feature sequence of personnel trajectories and roadway environments, and a coupling feature sequence of safety production data and geographic regions;

[0008] based on the comprehensive spatio-temporal feature representation, constructing a dynamic anomaly recognition network to respectively identify device abnormal displacement, personnel trajectory deviation from safety regions, and abnormal correlation between geographic regions and safety production data, and generating an initial abnormal information pair set;

[0009] performing spatio-temporal confidence compensation and risk level fusion on the initial abnormal information pair set, and outputting an identification result containing abnormal position coordinates, time nodes, and risk levels.

[0010] In a second aspect, the present application provides an artificial intelligence large model driven spatio-temporal information pair identification system, comprising:

[0011] a data acquisition module, configured to acquire mine multi-source spatio-temporal data, clean and denoise the multi-source spatio-temporal data, and perform spatio-temporal alignment processing to generate a standard spatio-temporal data sequence;

[0012] a feature extraction module, configured to input the standard spatio-temporal data sequence into a pre-trained spatio-temporal correlation large model, perform multi-modal spatio-temporal feature extraction, and obtain a comprehensive spatio-temporal feature representation, wherein the comprehensive spatio-temporal feature representation comprises a trend feature sequence of device position change over time, a spatial matching feature sequence of personnel trajectory and roadway environment, and a coupling feature sequence of safety production data and geographic region;

[0013] an anomaly identification module, configured to construct a dynamic anomaly identification network based on the comprehensive spatio-temporal feature representation, respectively perform device abnormal displacement identification, personnel trajectory deviation from safety region identification, and geographic region and safety production data abnormal association identification, and generate an initial anomaly information pair set;

[0014] an identification result output module, configured to perform spatio-temporal confidence compensation and risk level fusion on the initial anomaly information pair set, and output an identification result containing abnormal position coordinates, time nodes, and risk levels.

[0015] One or more technical solutions provided in the present application have at least the following technical effects or advantages:

[0016] The present application provides an artificial intelligence large model driven spatio-temporal information pair identification method and system, which acquires mine multi-source spatio-temporal data, cleans and denoises the data, and performs spatio-temporal alignment to generate a standard sequence, then uses a pre-trained spatio-temporal correlation large model to extract a comprehensive spatio-temporal feature representation containing device position trend, personnel trajectory matching, and safety production data coupling relationship, constructs a dynamic anomaly identification network based on this to perform parallel identification of device abnormal displacement, personnel trajectory deviation from safety region, and geographic region and safety production data abnormal association, and generates an initial anomaly information pair set, finally performs spatio-temporal confidence compensation and risk level fusion on the set, significantly improving the accuracy and comprehensiveness of multi-source anomaly situation awareness in mine safety production monitoring. Compared with traditional methods, the technical solution provided in the present application significantly overcomes the false alarm and missed alarm problems caused by isolated data source analysis and static rules, and achieves the technical effect of improving the real-time monitoring accuracy of multi-dimensional, cross-scale abnormal events in a complex dynamic environment. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0018] Figure 1 A flowchart of the artificial intelligence large model driven spatio-temporal information pair identification method provided by the embodiments of the present application is shown.

[0019] Figure 2 A structural diagram of the artificial intelligence large model driven spatio-temporal information pair identification system provided by the embodiments of the present application is shown.

[0020] In the drawings, the components represented by the respective reference numerals are described as follows.

[0021] The data acquisition module 100, the feature extraction module 200, the abnormality identification module 300, and the identification result output module 400. DETAILED DESCRIPTION

[0022] The present application provides an artificial intelligence large model driven spatio-temporal information pair identification method and system, which is used to solve the technical problem of insufficient correlation analysis of mine multi-source spatio-temporal data in the prior art.

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings of the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without any creative effort belong to the scope of protection of the present application.

[0024] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server comprising a series of steps or units need not be limited to only those steps or units clearly listed, but can include other steps or modules not clearly listed or inherent to the process, method, product or device.

[0025] Embodiment one, as shown in the present application provides an artificial intelligence large model driven spatio-temporal information pair identification method, wherein the method comprises: Figure 1

[0026] S10: Obtain mine multi-source spatio-temporal data, clean and denoise the multi-source spatio-temporal data, and perform spatio-temporal alignment processing to generate a standard spatio-temporal data sequence.

[0027] ​In the mine safety production monitoring, the effective use of multi-source spatio-temporal data is the basis for accurate identification of abnormalities. However, in practical applications, the raw data directly collected from sensors, positioning devices and monitoring systems have significant quality problems, which restricts the reliability of subsequent analysis. Multi-source data comes from heterogeneous systems, and their collection timestamps and spatial coordinate systems are not unified. For example, device location data, personnel trajectory and environmental parameters each use independent spatio-temporal reference bases, making it difficult to directly correlate and compare between data. This inconsistency at the data level can cause inherent bias in the input of subsequent feature extraction and anomaly identification models, making it difficult to capture real dynamic change patterns and even magnifying errors, resulting in false warning signals.

[0028] The step S10 in the method provided by the embodiment of the present application comprises:

[0029] Obtaining mine multi-source spatio-temporal data, wherein the multi-source spatio-temporal data comprises a geographic coordinate sequence, environmental monitoring data, device operating state data and personnel positioning trajectory data;

[0030] Respectively performing noise filtering and outlier rejection on each type of data, using a spatio-temporal interpolation method to complete missing data, and obtaining a cleaned data sequence;

[0031] Based on a unified spatio-temporal reference system, performing coordinate alignment and timestamp synchronization on the cleaned data sequence to generate a standard spatio-temporal data sequence.

[0032] In the embodiment of the present application, mine multi-source spatio-temporal data is obtained, and the multi-source spatio-temporal data is cleaned and denoised and spatio-temporally aligned to generate a standard spatio-temporal data sequence.

[0033] Specifically, first, mine multi-source spatio-temporal data is collected and obtained. The multi-source spatio-temporal data comprises a geographic coordinate sequence, environmental monitoring data, device operating state data and personnel positioning trajectory data. For example, a plurality of latitude and longitude coordinates of a mine roadway and mine equipment are collected to obtain a geographic coordinate sequence; temperature, humidity and gas concentration of the mine roadway are collected using temperature and humidity sensors and gas sensors to integrate and obtain environmental monitoring data; the on-off state and operating power of mine equipment such as a coal mining machine are recorded to integrate and obtain device operating state data; and personnel positioning trajectory data is collected using a GPS positioning system.

[0034] The noise filtering and outlier removal are performed on each type of data respectively, and the missing data is completed by using the space-time interpolation method to obtain the cleaned data sequence. Specifically, the sliding average filtering method is used for noise filtering on data with fluctuations, such as environmental monitoring data sequence. Further, the box plot method is used to identify and remove outliers that deviate significantly from the normal range, for example, records in the device running state data that are much higher than the rated speed are judged as abnormal and removed. Further, for missing values in the data sequence caused by transmission failure, the space-time Kriging interpolation method is used for completion. Finally, the data sequence that is free of noise interference, has no obvious abnormal points, and is continuous and complete is obtained.

[0035] Based on the unified space-time reference system, the coordinate alignment and timestamp synchronization of the cleaned data sequence are performed to generate the standard space-time data sequence. Specifically, in space, all collected geographic coordinates, device position coordinates and personnel trajectory coordinates are converted by using the GDAL library to unify all geographic coordinates to the WGS84 coordinate system. In time, the timestamps of all data records are converted to UTC time, and the data from different devices but representing the same time are time-matched to ensure that all data are synchronized on the timeline, and finally the standard space-time data sequence that is completely consistent in space and time reference and regular is obtained.

[0036] By performing targeted cleaning, denoising and outlier removal operations on multi-source space-time data, the noise introduced by environmental interference and device failure is effectively filtered out, the data missing is repaired, and the integrity of the original data is significantly improved. Further, by using the space-time interpolation method and the unified space-time reference system for coordinate alignment and timestamp synchronization, the originally dispersed and heterogeneous data streams are integrated into a standard sequence with consistent space-time dimensions. The data of different modalities such as device coordinates, personnel trajectories and environmental parameters can be accurately associated and compared in a unified space-time framework, greatly enhancing the internal consistency and comparability of the data.

[0037] S20: inputting the standard space-time data sequence into the pre-trained space-time correlation large model to perform multi-modal space-time feature extraction and obtaining comprehensive space-time feature representation, wherein the comprehensive space-time feature representation includes trend feature sequence of device position changing with time, spatial matching feature sequence of personnel trajectory and roadway environment, and coupling feature sequence of safety production data and geographic area.

[0038] The existing analysis method is usually limited to independent investigation of a single data source or shallow features, such as analyzing only the equipment displacement trajectory or monitoring the environmental parameters alone, and lacks a mechanism for coordinating and fusing multi-modal spatio-temporal information. This processing method cannot capture the interaction of equipment operation trends and roadway geographical constraints, the matching degree of personnel movement paths and environmental safety states, and the coupling characteristics of specific regional safety production data and their spatial positions, and it is difficult to identify potential risks caused by the synergistic action of multiple factors, such as composite failures of abnormal displacement of equipment accompanied by changes in environmental parameters in a specific region, or abnormal aggregation behavior of personnel trajectories in key geographical areas.

[0039] The step S20 in the method provided by the embodiment of the present application comprises:

[0040] The construction step of the spatio-temporal correlation large model comprises:

[0041] Collecting historical mine multi-source spatio-temporal data samples to construct a sample spatio-temporal data sequence set;

[0042] Labeling the label data of the equipment position trend, personnel trajectory matching degree, and safety production data and geographical area correlation in the sample to obtain a sample feature sequence set;

[0043] Based on machine learning, constructing a network architecture of the spatio-temporal correlation large model, and using the sample spatio-temporal data sequence set and the sample feature sequence set for supervised training, and obtaining the spatio-temporal correlation large model after training convergence.

[0044] In the embodiment of the present application, the standard spatio-temporal data sequence is input into the pre-trained spatio-temporal correlation large model for multi-modal spatio-temporal feature extraction to obtain a comprehensive spatio-temporal feature representation, wherein the comprehensive spatio-temporal feature representation includes a trend feature sequence of the equipment position changing over time, a spatial matching feature sequence of the personnel trajectory and the roadway environment, and a coupling feature sequence of the safety production data and the geographical area.

[0045] Specifically, first, the spatio-temporal correlation large model is constructed.

[0046] Collecting historical mine multi-source spatio-temporal data samples to construct a sample spatio-temporal data sequence set. The historical mine multi-source spatio-temporal data samples are pre-processed historical mine multi-source spatio-temporal data samples, for example, based on equipment operation logs and personnel positioning records, obtaining historical geographical coordinate sequences, historical environmental monitoring data, historical equipment operation state data, and historical personnel positioning trajectory data, and pre-processing to obtain historical mine multi-source spatio-temporal data samples, and constructing a sample spatio-temporal data sequence set.

[0047] Further, the label data of the device position trend, the personnel trajectory matching degree, the safety production data and the geographical area correlation in the labeled sample are obtained to obtain a sample feature sequence set. Specifically, the device position trend can be labeled by using "stable" and "offset" labels, the personnel trajectory matching degree can be labeled by using "match" and "not match" labels, and the safety production data and the geographical area correlation can be labeled by using "safe" and "risk" labels to obtain the sample feature sequence set.

[0048] Based on machine learning, a network architecture of the spatiotemporal correlation large model is constructed. Illustratively, a spatiotemporal correlation model including a three-layer network structure is constructed. The first layer is a time feature extraction layer, which is constructed by using a long short-term memory network, and 128 neurons are set to capture the time dependence in the data; the second layer is a spatial feature extraction layer, which is constructed by using a two-dimensional convolutional neural network, and a 3x3 convolution kernel and 64 filters are used to extract spatial features; and the third layer is a multi-modal fusion layer, which is constructed by using a full connection layer and includes 256 neurons to fuse the spatiotemporal features. The input dimension of the model is set to (sequence length, feature dimension), wherein the sequence length can be set to 24 according to the actual data, representing 24 time points, and the feature dimension is set according to the coordinates, environmental parameters, etc. Further, the sample spatiotemporal data sequence set and the sample feature sequence set are used for supervised training, the sample spatiotemporal data sequence set is used as the input, the sample feature sequence set is used as the target output, the Adam optimizer is used, the learning rate is set to 0.001, the batch size is set to 32, and the mean square error loss is used as the loss function, and the loss value = average value x (model predicted value-target value) 2 . The training process continues until the loss function converges, such as after 100 rounds of training, the loss function does not decrease significantly, and the trained spatiotemporal correlation large model is obtained.

[0049] Further, the preprocessed standard spatiotemporal data sequence is input into the trained spatiotemporal correlation large model. The model first extracts the time sequence features through the LSTM layer to obtain the trend feature sequence of the device position change over time; then processes the spatial data through the CNN layer to obtain the spatial matching feature sequence of the personnel trajectory and the roadway environment; and finally fuses the multi-modal information through the full connection layer to obtain the coupling feature sequence of the safety production data and the geographical area. These feature sequences together constitute a comprehensive spatiotemporal feature representation, wherein the trend feature sequence reflects the motion law of the device, the spatial matching feature sequence reveals the adaptation degree of personnel activity and the environment, and the coupling feature sequence shows the correlation strength of the safety parameters and the regional features.

[0050] The standard spatiotemporal data sequence is input into a pre-trained large model to extract comprehensive feature representations covering the dynamic trend of equipment position, the matching relationship between personnel trajectory and roadway space structure, and the coupling characteristics of safety production indicators and geographic regions from complex data. The comprehensive spatiotemporal feature representation breaks through the limitations of traditional single feature description, and integrates originally isolated data dimensions into an information carrier that can fully reflect the running state of the mine system. It not only depicts the dynamic changes of individual elements, but more importantly, reveals the internal relationship and interaction rules of equipment, personnel, environment and other elements in the spatiotemporal dimension. This high-level and integrated feature representation provides more abundant and discriminative information input for the subsequent anomaly recognition network, enabling the recognition model to perceive and judge anomalies from the overall perspective of the system rather than the conventional local perspective, thereby laying a key feature foundation for significantly improving the detection capability of complex anomaly patterns, especially multi-factor coupled risks.

[0051] S30: Based on the comprehensive spatiotemporal feature representation, a dynamic anomaly recognition network is constructed to respectively perform equipment abnormal displacement recognition, personnel trajectory deviation from a safety region recognition, and geographic region and safety production data anomaly association recognition, and an initial anomaly information pair set is generated.

[0052] The existing technology often uses serial or isolated anomaly detection modules, such as first detecting equipment anomalies and then analyzing personnel trajectories. This processing method is difficult to capture the potential association between different anomaly types and is low in efficiency. More importantly, the definition of anomalies in conventional methods often relies on fixed thresholds or simple rules, which cannot adapt to the complex and nonlinear interactions between equipment displacement, personnel behavior and regional safety state in the mine dynamic environment. For example, a small change in equipment position may be normal in a particular environment, but when coupled with a particular personnel trajectory or regional environmental data, it may constitute a high risk.

[0053] The method provided in the embodiments of the present application includes step S30:

[0054] Based on the comprehensive spatiotemporal feature representation, a dynamic anomaly recognition network is constructed, wherein the dynamic anomaly recognition network includes an equipment abnormal displacement recognition subnetwork, a personnel trajectory deviation recognition subnetwork, and an anomaly association region recognition subnetwork;

[0055] The equipment abnormal displacement recognition subnetwork is configured to fit an equipment motion trajectory based on the trend feature sequence, calculate a deviation amplitude of an actual position from the fitted equipment motion trajectory, and recognize equipment abnormal displacement.

[0056] The personnel trajectory deviation recognition subnetwork is configured to extract a personnel real-time trajectory point sequence based on the spatial matching feature sequence, and recognize trajectory deviation from a safety region based on pre-defined electronic fence data.

[0057] The abnormal association area identification sub-network is configured to calculate the association strength of the geographic area and the safety production data based on the coupling feature sequence, and identify an abnormal association area.

[0058] The method comprises the following steps:

[0059] The real-time association strength indicator of the geographic area and the safety production data in the coupling feature sequence is extracted, and compared with the association strength benchmark under the historical normal working condition to calculate the association deviation degree.

[0060] The method comprises the following steps:

[0061] The real-time association strength indicator of the target geographic area and the corresponding safety production data in the current period is extracted based on the coupling feature sequence.

[0062] The association strength benchmark sequence corresponding to the same geographic area under the historical normal working condition is obtained, wherein the association strength benchmark sequence comprises association strength data of a plurality of historical normal periods.

[0063] The statistical distribution difference between the real-time association strength indicator and the plurality of association strength benchmark sequences is calculated.

[0064] The overall deviation degree of the real-time association strength indicator of the current period and the adjacent period relative to the association strength benchmark sequence is calculated based on a sliding time window mechanism.

[0065] The statistical distribution difference and the overall deviation degree are weighted and fused to generate a comprehensive association deviation degree.

[0066] When the association deviation degree continuously exceeds a preset abnormal threshold, it is determined that the current geographic area has an abnormal association, and is identified as an abnormal association area.

[0067] The identification results of the equipment abnormal displacement identification sub-network, the personnel trajectory deviation identification sub-network, and the abnormal association area identification sub-network are fused to generate an initial abnormal information pair set.

[0068] In the embodiments of the present application, a dynamic abnormality identification network is constructed based on a comprehensive spatio-temporal feature representation, and the dynamic abnormality identification network is used to identify equipment abnormal displacement, personnel trajectory deviation from a safety area, and abnormal association between a geographic area and safety production data, and generate an initial abnormal information pair set.

[0069] Specifically, first, based on the integrated spatio-temporal feature representation, a dynamic anomaly recognition network is constructed, wherein the dynamic anomaly recognition network includes a device abnormal displacement recognition sub-network, a personnel trajectory deviation recognition sub-network, and an abnormal association area recognition sub-network.

[0070] The device abnormal displacement recognition sub-network is configured to fit a device motion trajectory based on a trend feature sequence, calculate a deviation amplitude of an actual position from the fitted device motion trajectory, and recognize device abnormal displacement. Specifically, a linear regression method is used to fit the device motion trajectory to obtain a fitted position sequence, and a deviation amplitude of the actual position from the fitted position at each time point is calculated, such as calculating the absolute distance between the actual position and the fitted position as the deviation amplitude. At the same time, the abnormal threshold is dynamically adjusted based on the historical displacement data, for example, the historical average deviation amplitude multiplied by a safety factor is used as the dynamic threshold. For example, the historical average deviation amplitude is 0.2 meters, and the safety factor is set to 1.1, then the dynamic threshold = 0.2 x 1.1 = 0.22. When the deviation amplitude exceeds the dynamic abnormal threshold, the device position at this time point is marked as abnormal displacement. For example, the deviation amplitude is 0.4, which is greater than the dynamic threshold, and it is determined to be abnormal. Finally, the recognition result of the device abnormal displacement is output, wherein the device abnormal displacement recognition result includes the abnormal position and the time point.

[0071] The personnel trajectory deviation recognition sub-network is configured to extract a personnel real-time trajectory point sequence based on a spatial matching feature sequence, and recognize trajectory deviation from a safety area based on predefined electronic fence data. Specifically, first, the predefined electronic fence data is obtained. The electronic fence data is boundary data used to define a safety area that is set in advance to prevent personnel from entering a dangerous area, and is exemplarily defined in the form of a polygon boundary point set. From the spatial matching feature sequence, the personnel real-time trajectory point sequence is extracted, and the trajectory points not in the electronic fence are filtered out, and the shortest Euclidean distance from the trajectory points to the boundary of the electronic fence is calculated. The distance is compared with a preset safety distance threshold, and if the Euclidean distance of the trajectory point continuously or repeatedly exceeds the safety distance threshold, it is determined that the trajectory deviates from the safety area. For example, the preset safety distance threshold is 1 meter, and the distance from a trajectory point not in the electronic fence to the boundary is 1.2 meters, and the Euclidean distance of the trajectory point to the boundary at three consecutive time points exceeds the safety distance threshold, and it is determined to deviate. Finally, the recognition result of the personnel trajectory deviation is output, including the deviation position and the time sequence. This filtering can exclude incidental deviations and obtain the personnel trajectory deviation recognition result that actually exists safety hazards.

[0072] The abnormal association region identification sub-network is configured to calculate the association strength of the geographic region and the safety production data based on the coupling feature sequence, and identify an abnormal association region. Specifically, the real-time association strength index of the geographic region and the safety production data in the coupling feature sequence is extracted, and compared with the association strength benchmark in the historical normal working condition to calculate the association deviation degree.

[0073] Based on the coupling feature sequence, the real-time association strength index of the target geographic region and the corresponding safety production data in the current period is extracted. Illustratively, a pre-acquired feature decoder can be used for mapping to obtain the real-time association strength index of the target geographic region and the corresponding safety production data. The feature decoder can be constructed based on a linear decoder, and the feature decoding is performed by a linear regression method. Finally, the coupling feature sequence is input into the decoder, and the real-time association strength index can be accurately output.

[0074] Further, the association strength benchmark sequence corresponding to the same geographic region in the historical normal working condition is obtained, wherein the association strength benchmark sequence comprises the association strength data of a plurality of historical normal periods.

[0075] The statistical distribution difference between the real-time association strength index and the statistical distribution of the plurality of association strength benchmark sequences is calculated. Illustratively, the statistical distribution difference = | (current real-time association strength index - mean of association strength benchmark sequence) ÷ standard deviation of association strength benchmark sequence |.

[0076] Based on the sliding time window mechanism, the overall deviation degree of the real-time association strength index of the current period and the adjacent period relative to the association strength benchmark sequence is calculated. Illustratively, a sliding time window with a size of 5 is used to calculate the absolute value of the difference between the real-time association strength index of the current time point and the mean of the association strength benchmark sequence, to obtain the absolute deviation value, and the absolute deviation value = | current real-time association strength index - mean of association strength benchmark sequence |. The absolute deviation values of all time points in the time window are summed to obtain the total absolute deviation value of the time window. The average absolute deviation is obtained by dividing the total absolute deviation value of the time window by the size of the sliding time window. Based on the average absolute deviation, the overall deviation degree is obtained, and the overall deviation degree = average absolute deviation ÷ standard deviation of association strength benchmark sequence.

[0077] The statistical distribution difference and the overall deviation degree are weighted and fused to generate a comprehensive association deviation degree. The association deviation degree = W1 × statistical distribution difference + W2 × overall deviation degree, wherein the weights W1 + W2 = 1, and W1 and W2 can be obtained based on the importance of the instantaneous deviation and the recent trend in the analysis of the association deviation degree. For example, more emphasis is placed on analyzing the instantaneous deviation, and W1 can be taken as 0.6, and W2 = 1 - 0.6 = 0.4.

[0078] When the correlation deviation continuously exceeds the preset abnormal threshold, such as continuously exceeding the preset abnormal threshold at three consecutive time points, it is determined that the current geographic area is abnormally correlated, and the current geographic area is identified as an abnormally correlated area.

[0079] Further, the identification results of the device abnormal displacement identification sub-network, the personnel trajectory deviation identification sub-network and the abnormal correlation area identification sub-network are fused to generate an initial abnormal information pair set. Each abnormal information pair includes a device abnormal displacement identification result, a personnel trajectory deviation identification result set and an abnormal correlation area identification result.

[0080] Based on the comprehensive spatio-temporal feature representation, the present application synchronously analyzes the device abnormal displacement, the personnel trajectory deviation from the safety area and the abnormal correlation between the geographic area and the safety production data through special sub-networks. This parallel processing architecture not only improves the identification efficiency, but also guarantees the specificity and depth of detection of different types of abnormalities, providing structured and multi-source abnormal event input for subsequent risk comprehensive assessment, and significantly enhancing the ability to capture complex and cross-dimensional risks.

[0081] S40: Spatio-temporal confidence compensation and risk level fusion are performed on the initial abnormal information pair set to output an identification result containing abnormal position coordinates, time nodes and risk levels.

[0082] Existing methods usually directly output initial identification results or perform simple weighted averaging, lacking consideration of the spatio-temporal context of abnormal events such as historical contemporaneous patterns and surrounding area states, leading to risk assessment results that may be overly sensitive to short-term fluctuations or not alert enough to gradual risks. At the same time, the potential risks of different abnormal information pairs are not uniformly quantified and graded, making it difficult for management personnel to quickly judge the severity and priority of events, thereby affecting emergency response efficiency.

[0083] The method provided in the embodiments of the present application includes step S40:

[0084] Extract the spatio-temporal context features of each abnormal information pair in the initial abnormal information pair set, and calculate the matching degree with the historical abnormal pattern to generate a spatio-temporal confidence weight;

[0085] Wherein, extracting the spatio-temporal context features of each abnormal information pair in the initial abnormal information pair set and calculating the matching degree with the historical abnormal pattern to generate a spatio-temporal confidence weight includes:

[0086] Obtain the spatio-temporal context features corresponding to the abnormal information pair, wherein the spatio-temporal context features include time period features and spatial distribution features;

[0087] Calculate the time matching degree of the time period features with the historical abnormal pattern to obtain a time period matching degree;

[0088] calculate a spatial matching degree of the spatial distribution feature and the historical abnormal pattern, to obtain a spatial distribution matching degree;

[0089] weight and fuse the time period matching degree and the spatial distribution matching degree, to generate the spatio-temporal confidence weight;

[0090] compensate and correct an abnormal degree of a corresponding abnormal information pair based on the spatio-temporal confidence weight, to obtain a weighted abnormal score, wherein the abnormal degree is a weighted fusion value of a device abnormal displacement deviation amplitude, a track deviation safety region distance and an associated deviation degree;

[0091] match the weighted abnormal score with a preset risk level threshold interval, to determine a final risk level;

[0092] combine the position coordinates and the time nodes in the abnormal information pair, to output an identification result containing abnormal position coordinates, time nodes and the final risk level.

[0093] In the embodiments of the present application, the initial abnormal information pair set is subjected to spatio-temporal confidence compensation and risk level fusion, to output an identification result containing abnormal position coordinates, time nodes and a risk level.

[0094] Specifically, the spatio-temporal context features of each abnormal information pair in the initial abnormal information pair set are extracted, and the matching degrees with the historical abnormal patterns are calculated, to generate the spatio-temporal confidence weight.

[0095] First, the abnormal information pair is acquired to obtain the corresponding spatio-temporal context features, wherein the spatio-temporal context features include time period features and spatial distribution features. Exemplarily, the time period features are used to represent whether the abnormality occurs in a peak period or a maintenance period, and the spatial distribution features are used to represent whether the abnormal position is located in a main mine roadway trunk area or an edge area. The time period matching degree is determined by calculating the similarity between the current abnormal occurrence time period and the historical abnormal high-occurrence time period. For example, if the current time is 9:00 am and the historical data shows that the abnormality occurs frequently at this time period, the time period matching degree is high, and the time period matching degree = 1- | (current time point - historical high-occurrence time point) ÷ 24 |. The spatial distribution matching degree is determined by calculating the Euclidean distance between the current abnormal position and the historical abnormal high-occurrence area. For example, if the current abnormal position is close to the center point of the historical abnormal high-occurrence area, the spatial distribution matching degree is high, and the spatial distribution matching degree = 1- (distance between the current position and the center of the historical high-occurrence area ÷ maximum effective distance). The time period matching degree and the spatial distribution matching degree are weighted and fused to generate a spatio-temporal confidence weight, and the spatio-temporal confidence weight = A1 x time period matching degree + A2 x spatial distribution matching degree. Wherein, A1 + A2 = 1. By default, A1 = A2 = 0.5, and A1 and A2 are dynamically based on the importance of the time period matching degree and the spatial distribution matching degree in the spatio-temporal confidence weight. For example, the spatial distribution matching degree is more important in the analysis of the spatio-temporal confidence weight, so A2 = 0.6, and A1 = 1-0.6 = 0.4.

[0096] Based on the spatio-temporal confidence weight, the abnormality degree of the corresponding abnormal information pair is compensated and corrected to obtain a weighted abnormality score, wherein the abnormality degree is a weighted fusion value of the device abnormal displacement deviation amplitude, the trajectory deviation safety region distance and the associated deviation degree. Specifically, for the device abnormal displacement, the abnormality degree is the Euclidean distance between the actual position and the expected position of the device, i.e. the device abnormal displacement deviation amplitude; for the personnel trajectory deviation, the abnormality degree is the shortest Euclidean distance from the personnel position to the safety region boundary, i.e. the trajectory deviation safety region distance; for the regional abnormality association, the abnormality degree is the associated deviation degree. The abnormality degree is obtained by weighted fusion. Abnormality degree = B1 x device abnormal displacement deviation amplitude + B2 x trajectory deviation safety region distance + B3 x associated deviation degree. Wherein, B1 + B2 + B3 = 1. The weights B1, B2 and B3 can be obtained based on the abnormality importance, for example, the personnel trajectory deviation has a high risk degree and a high abnormality importance, then B2 can be taken as 0.4, B1 and B3 can be taken as 0.3, and the original abnormality degree is obtained by weighted calculation.

[0097] The original abnormality degree is compensated using the spatio-temporal confidence weight to obtain a weighted abnormality score. Weighted abnormality score = original abnormality degree x spatio-temporal confidence weight. For example, the original abnormality degree is 0.8, and the spatio-temporal confidence weight is 0.9, then the weighted abnormality score = 0.8 x 0.9 = 0.72.

[0098] The weighted abnormality score is matched with a preset risk level threshold interval to determine the final risk level. Illustratively, four risk level threshold intervals are set, wherein the low risk is 0-0.2, the medium risk is 0.2-0.6, the high risk is 0.6-0.8, and the super high risk is 0.8-1. The weighted abnormality score of each abnormal information pair is compared with these threshold intervals to determine the corresponding risk level. For example, the abnormal information pair with a weighted abnormality score of 0.72 is determined as a high risk level.

[0099] In combination with the position coordinates and time nodes in the abnormal information pair, an identification result containing the abnormal position coordinates, time nodes and final risk level is output. For example, the identification result can be [device abnormal displacement, (112.5, 35.8), 2024-05-01 14:30:00, medium risk].

[0100] By extracting the spatio-temporal context features of each abnormal information pair and matching them with historical abnormal patterns, a spatio-temporal confidence weight representing the credibility of each abnormality is generated, the initial abnormality degree is compensated and corrected, false alarms caused by accidental interference are effectively suppressed, the significance of real risks consistent with historical abnormal patterns is strengthened, and the accuracy of the results in the spatio-temporal dimension is enhanced. Further, by mapping the weighted abnormality score to the preset risk level threshold, the unified quantification and grading of abnormal risks of different sources and different types are realized, and a clear final risk level is generated. The final output identification result not only contains the position coordinates and time nodes of the abnormality, but also adds the risk level information, so that the output result is no longer a simple list of original signals, but an intelligent refined decision support information, greatly improving the early warning accuracy and decision support capability of the mine safety monitoring system.

[0101] Embodiment two, as shown in Figure 2 based on the same inventive concept of the artificial intelligence large model driven spatio-temporal information pair identification method provided in embodiment one, the present embodiment also provides an artificial intelligence large model driven spatio-temporal information pair identification system, comprising:

[0102] A data acquisition module 100 is configured to acquire mine multi-source spatio-temporal data, clean and denoise the multi-source spatio-temporal data, and perform spatio-temporal alignment processing to generate a standard spatio-temporal data sequence;

[0103] The feature extraction module 200 is configured to input the standard spatio-temporal data sequence into a pre-trained spatio-temporal correlation large model, perform multi-modal spatio-temporal feature extraction, and obtain a comprehensive spatio-temporal feature representation, wherein the comprehensive spatio-temporal feature representation includes a trend feature sequence of device position change over time, a spatial matching feature sequence of personnel trajectory and tunnel environment, and a coupling feature sequence of safety production data and geographic area.

[0104] The anomaly identification module 300 is configured to construct a dynamic anomaly identification network based on the comprehensive spatio-temporal feature representation, respectively perform device abnormal displacement identification, personnel trajectory deviation from safety area identification, and geographic area and safety production data anomaly association identification, and generate an initial anomaly information pair set.

[0105] The identification result output module 400 is configured to perform spatio-temporal confidence compensation and risk level fusion on the initial anomaly information pair set, and output an identification result including abnormal position coordinates, time nodes, and risk levels.

[0106] In one embodiment, the data acquisition module 100 is further configured to:

[0107] acquire mine multi-source spatio-temporal data, wherein the multi-source spatio-temporal data includes geographic coordinate sequences, environment monitoring data, device operation state data, and personnel positioning trajectory data;

[0108] perform noise filtering and outlier rejection on each type of data respectively, and use a spatio-temporal interpolation method to complete missing data to obtain cleaned data sequences;

[0109] based on a unified spatio-temporal reference system, perform coordinate alignment and timestamp synchronization on the cleaned data sequences to generate standard spatio-temporal data sequences.

[0110] In one embodiment, the feature extraction module 200 is further configured to:

[0111] The construction steps of the spatio-temporal correlation large model include:

[0112] collect historical mine multi-source spatio-temporal data samples, and construct a sample spatio-temporal data sequence set;

[0113] label device position trend, personnel trajectory matching degree, and safety production data and geographic area association in the samples to obtain a sample feature sequence set;

[0114] based on machine learning, construct a network architecture of the spatio-temporal correlation large model, use the sample spatio-temporal data sequence set and the sample feature sequence set for supervised training, and obtain the spatio-temporal correlation large model after training convergence.

[0115] In one embodiment, the anomaly identification module 300 is further configured to:

[0116] construct a dynamic anomaly recognition network based on the comprehensive spatio-temporal feature representation, wherein the dynamic anomaly recognition network comprises a device abnormal displacement recognition sub-network, a personnel trajectory deviation recognition sub-network, and an abnormal association area recognition sub-network;

[0117] The device abnormal displacement recognition sub-network is configured to fit a device motion trajectory based on the trend feature sequence, calculate a deviation amplitude of an actual position from the fitted device motion trajectory, and recognize device abnormal displacement.

[0118] The personnel trajectory deviation recognition sub-network is configured to extract a personnel real-time trajectory point sequence based on the spatial matching feature sequence, and recognize trajectory deviation from a safety area based on predefined electronic fence data.

[0119] The abnormal association area recognition sub-network is configured to calculate an association strength of a geographic area and safety production data based on the coupling feature sequence, and recognize an abnormal association area.

[0120] The abnormal association area recognition sub-network is configured to calculate an association strength of a geographic area and safety production data based on the coupling feature sequence, and recognize an abnormal association area.

[0121] The coupling feature sequence is extracted, and the real-time association strength indicator of the geographic area and the safety production data is compared with an association strength benchmark under a historical normal working condition to calculate an association deviation degree.

[0122] The coupling feature sequence is extracted, and the real-time association strength indicator of the geographic area and the safety production data is compared with an association strength benchmark under a historical normal working condition to calculate an association deviation degree.

[0123] The coupling feature sequence is extracted, and the real-time association strength indicator of the geographic area and the safety production data is compared with an association strength benchmark under a historical normal working condition to calculate an association deviation degree.

[0124] The coupling feature sequence is extracted, and the real-time association strength indicator of the geographic area and the safety production data is compared with an association strength benchmark under a historical normal working condition to calculate an association deviation degree.

[0125] The coupling feature sequence is extracted, and the real-time association strength indicator of the geographic area and the safety production data is compared with an association strength benchmark under a historical normal working condition to calculate an association deviation degree.

[0126] The coupling feature sequence is extracted, and the real-time association strength indicator of the geographic area and the safety production data is compared with an association strength benchmark under a historical normal working condition to calculate an association deviation degree.

[0127] The coupling feature sequence is extracted, and the real-time association strength indicator of the geographic area and the safety production data is compared with an association strength benchmark under a historical normal working condition to calculate an association deviation degree.

[0128] When the correlation deviation continues to exceed the preset abnormal threshold, it is determined that the current geographic area has an abnormal correlation, and the abnormal correlation area is identified.

[0129] Fusing the recognition results of the device abnormal displacement recognition sub-network, the personnel trajectory deviation recognition sub-network and the abnormal correlation area recognition sub-network, an initial abnormal information pair set is generated.

[0130] In one embodiment, the recognition result output module 400 is further configured to:

[0131] Extracting the spatio-temporal context features of each abnormal information pair in the initial abnormal information pair set, and calculating the matching degree with the historical abnormal pattern to generate a spatio-temporal confidence weight;

[0132] The extraction of the spatio-temporal context features of each abnormal information pair in the initial abnormal information pair set, and the calculation of the matching degree with the historical abnormal pattern to generate a spatio-temporal confidence weight, includes:

[0133] Obtaining the spatio-temporal context features corresponding to the abnormal information pair, wherein the spatio-temporal context features include time period features and spatial distribution features;

[0134] Calculating the time matching degree of the time period features with the historical abnormal pattern to obtain a time period matching degree;

[0135] Calculating the spatial matching degree of the spatial distribution features with the historical abnormal pattern to obtain a spatial distribution matching degree;

[0136] Weightedly fusing the time period matching degree and the spatial distribution matching degree to generate the spatio-temporal confidence weight;

[0137] Compensating and correcting the abnormal degree of the corresponding abnormal information pair based on the spatio-temporal confidence weight to obtain a weighted abnormal score, wherein the abnormal degree is a weighted fusion value of device abnormal displacement deviation amplitude, trajectory deviation safety area distance and correlation deviation;

[0138] Matching the weighted abnormal score with a preset risk level threshold interval to determine a final risk level;

[0139] Outputting the recognition result containing the abnormal position coordinates, the time nodes and the final risk level in combination with the position coordinates and the time nodes in the abnormal information pair.

[0140] In summary, the embodiments of the present application have at least the following technical effects:

[0141] The application proposes an artificial intelligence large model driven spatio-temporal information pair identification method and system. By acquiring mine multi-source spatio-temporal data and performing cleaning, denoising and spatio-temporal alignment to generate a standard sequence, a pre-trained spatio-temporal correlation large model is used to extract comprehensive spatio-temporal feature representation containing device location trend, personnel trajectory matching and safety production data coupling relationship. On this basis, a dynamic anomaly identification network is constructed to perform parallel identification of device abnormal displacement, personnel trajectory deviation from the safety area and abnormal association of the geographic area and safety production data, and generate an initial anomaly information pair set. Finally, the set is subjected to spatio-temporal confidence compensation and risk level fusion, significantly improving the accuracy and comprehensiveness of multi-source anomaly situation awareness in mine safety production monitoring. Specifically, through deep feature extraction and dynamic network identification driven by a large model, the adaptive learning and accurate description of the complex spatio-temporal correlation between devices, personnel and the environment are realized, so that potential trend deviations and collaborative risks can be captured from massive heterogeneous data. At the same time, by introducing a spatio-temporal confidence compensation mechanism, the historical anomaly patterns and real-time context features are effectively combined to enhance the discrimination ability of instantaneous anomalies and progressive risks, avoiding judgment distortion caused by environmental fluctuations or data noise, so that the risk assessment results are more in line with the actual working conditions. Compared with traditional methods, the technical scheme provided by the application significantly overcomes the false alarm and missed alarm problems caused by isolated data source analysis and static rules, achieving the technical effect of improving the real-time monitoring accuracy of multi-dimensional and cross-scale abnormal events in the complex dynamic environment of the mine safety monitoring system.

[0142] It should be noted that the above sequence of the embodiments of the application is only for description, and does not represent the advantages and disadvantages of the embodiments. Moreover, the above describes specific embodiments of the present application. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or can be advantageous.

[0143] The above only describes the preferred embodiments of the application and does not limit the application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the application shall be included in the protection scope of the application.

[0144] The present application and the drawings are only exemplary descriptions of the application, and any and all modifications, changes, combinations or equivalents within the scope of the application are considered to be covered by the present application. Obviously, those skilled in the art can make various modifications and changes to the application without departing from the scope of the application. Thus, if these modifications and changes of the application belong to the scope of the application and its equivalents, the application intends to include these modifications and changes.

Claims

1. A spatiotemporal information pair recognition method driven by a large artificial intelligence model, characterized in that, The method includes: Acquire multi-source spatiotemporal data from the mine, and perform cleaning, noise reduction, and spatiotemporal alignment processing on the multi-source spatiotemporal data to generate a standard spatiotemporal data sequence; The standard spatiotemporal data sequence is input into a pre-trained spatiotemporal correlation model to perform multimodal spatiotemporal feature extraction and obtain a comprehensive spatiotemporal feature representation. The comprehensive spatiotemporal feature representation includes a trend feature sequence of equipment location changes over time, a spatial matching feature sequence of personnel trajectory and roadway environment, and a coupling feature sequence of safety production data and geographical area. Based on the comprehensive spatiotemporal feature representation, a dynamic anomaly identification network is constructed to identify abnormal equipment displacement, personnel trajectory deviation from safe areas, and abnormal correlation between geographical areas and safety production data, generating an initial set of anomaly information pairs, including: Based on the comprehensive spatiotemporal feature representation, a dynamic anomaly identification network is constructed, wherein the dynamic anomaly identification network includes an equipment abnormal displacement identification subnetwork, a personnel trajectory deviation identification subnetwork, and an anomaly associated region identification subnetwork; The device abnormal displacement identification subnetwork is used to fit the device motion trajectory based on the trend feature sequence, calculate the deviation between the actual position and the fitted device motion trajectory, and identify the device abnormal displacement. The personnel trajectory deviation identification subnetwork is used to extract the real-time trajectory point sequence of personnel based on the spatial matching feature sequence, and to identify trajectory deviations from the safe area based on predefined electronic fence data. The abnormal association region identification sub-network is used to calculate the association strength between geographical regions and safety production data based on the coupled feature sequence, and to identify abnormal association regions. By integrating the recognition results of the equipment abnormal displacement recognition subnetwork, the personnel trajectory deviation recognition subnetwork, and the abnormal associated region recognition subnetwork, an initial abnormal information pair set is generated; The initial anomaly information set is subjected to spatiotemporal confidence compensation and risk level fusion to output an identification result containing anomaly location coordinates, time nodes, and risk levels, including: Extract the spatiotemporal context features of each anomaly information pair in the initial anomaly information pair set, calculate the matching degree with historical anomaly patterns, and generate spatiotemporal confidence weights. The degree of anomaly of the corresponding abnormal information pairs is compensated and corrected based on the spatiotemporal confidence weight to obtain a weighted anomaly score, wherein the degree of anomaly is the weighted fusion value of the abnormal displacement deviation of the equipment, the distance of the trajectory from the safe area, and the correlation deviation. The weighted anomaly score is matched with a preset risk level threshold range to determine the final risk level; Combining the location coordinates and time nodes of the anomaly information pair, the output includes the anomaly location coordinates, time nodes, and the final risk level.

2. The method for identifying spatiotemporal information pairs driven by a large artificial intelligence model according to claim 1, characterized in that, Acquire multi-source spatiotemporal data from the mine, perform cleaning, denoising, and spatiotemporal alignment processing on the multi-source spatiotemporal data, and generate a standard spatiotemporal data sequence, including: Acquire multi-source spatiotemporal data of the mine, wherein the multi-source spatiotemporal data includes geographic coordinate sequences, environmental monitoring data, equipment operating status data, and personnel positioning trajectory data; Noise filtering and outlier removal are performed on each type of data, and spatiotemporal interpolation is used to fill in missing data to obtain a cleaned data sequence. Based on a unified spatiotemporal reference system, coordinate alignment and timestamp synchronization are performed on the cleaned data sequence to generate a standard spatiotemporal data sequence.

3. The method for identifying spatiotemporal information pairs driven by a large artificial intelligence model according to claim 1, characterized in that, The construction steps of a large-scale spatiotemporal correlation model include: Collect multi-source spatiotemporal data samples from historical mines and construct a set of sample spatiotemporal data sequences; Label data on equipment location trends, personnel trajectory matching degree, and correlation between safety production data and geographical areas in the sample are used to obtain a set of sample feature sequences; Based on machine learning, a network architecture for a spatiotemporal correlation model is constructed. The model is trained under supervision using the set of sample spatiotemporal data sequences and the set of sample feature sequences. After training convergence, the spatiotemporal correlation model is obtained.

4. The method for identifying spatiotemporal information pairs driven by a large artificial intelligence model according to claim 1, characterized in that, Based on the aforementioned coupling feature sequence, the correlation strength between geographical regions and safety production data is calculated, and abnormal correlation regions are identified, including: Extract the real-time correlation strength index between geographical regions and safety production data from the coupled feature sequence, and compare it with the correlation strength benchmark under historical normal working conditions to calculate the correlation deviation. When the correlation deviation continuously exceeds the preset abnormal threshold, it is determined that there is an abnormal correlation in the current geographical area and it is marked as an abnormal correlation area.

5. The method for identifying spatiotemporal information pairs driven by a large artificial intelligence model according to claim 4, characterized in that, Extract the real-time correlation strength index between geographical regions and safety production data from the coupled feature sequence, and compare it with the correlation strength benchmark under historical normal working conditions to calculate the correlation deviation, including: Based on the coupling feature sequence, extract the real-time correlation strength index between the target geographical area and the corresponding safety production data in the current time period; Obtain the correlation strength benchmark sequence corresponding to the same geographical area under historical normal operating conditions, wherein the correlation strength benchmark sequence contains correlation strength data for multiple historical normal periods; Calculate the statistical distribution difference between the real-time correlation strength index and multiple correlation strength benchmark sequences; Based on the sliding time window mechanism, the overall deviation of the real-time correlation strength index of the current time period and adjacent time periods from the correlation strength benchmark sequence is calculated. The statistical distribution differences are weighted and fused with the overall deviation to generate a comprehensive correlation deviation.

6. The method for identifying spatiotemporal information pairs driven by a large artificial intelligence model according to claim 1, characterized in that, Extract the spatiotemporal context features of each anomaly pair in the initial anomaly information pair set, calculate the matching degree with historical anomaly patterns, and generate spatiotemporal confidence weights, including: Obtain the spatiotemporal context features corresponding to the anomaly information, wherein the spatiotemporal context features include time periodic features and spatial distribution features; Calculate the time period matching degree between the time period features and historical anomaly patterns to obtain the time period matching degree; Calculate the spatial matching degree between the spatial distribution features and historical anomaly patterns to obtain the spatial distribution matching degree; The time period matching degree and the spatial distribution matching degree are weighted and fused to generate the spatiotemporal confidence weight.

7. A spatiotemporal information pair recognition system driven by a large artificial intelligence model, characterized in that, The system is used to implement the artificial intelligence large model-driven spatiotemporal information pair identification method according to any one of claims 1-6, the system comprising: The data acquisition module is used to acquire multi-source spatiotemporal data from the mine, clean and denoise the multi-source spatiotemporal data and perform spatiotemporal alignment processing to generate a standard spatiotemporal data sequence. The feature extraction module is used to input the standard spatiotemporal data sequence into a pre-trained spatiotemporal correlation model to perform multimodal spatiotemporal feature extraction and obtain a comprehensive spatiotemporal feature representation. The comprehensive spatiotemporal feature representation includes a trend feature sequence of equipment location changes over time, a spatial matching feature sequence of personnel trajectory and roadway environment, and a coupling feature sequence of safety production data and geographical area. Anomaly identification module is used to construct a dynamic anomaly identification network based on the comprehensive spatiotemporal feature representation, and to perform anomaly identification of equipment displacement, personnel trajectory deviation from safe areas, and abnormal correlation between geographical areas and safety production data, generating an initial set of anomaly information pairs, including: Based on the comprehensive spatiotemporal feature representation, a dynamic anomaly identification network is constructed, wherein the dynamic anomaly identification network includes an equipment abnormal displacement identification subnetwork, a personnel trajectory deviation identification subnetwork, and an anomaly associated region identification subnetwork; The device abnormal displacement identification subnetwork is used to fit the device motion trajectory based on the trend feature sequence, calculate the deviation between the actual position and the fitted device motion trajectory, and identify the device abnormal displacement. The personnel trajectory deviation identification subnetwork is used to extract the real-time trajectory point sequence of personnel based on the spatial matching feature sequence, and to identify trajectory deviations from the safe area based on predefined electronic fence data. The abnormal association region identification sub-network is used to calculate the association strength between geographical regions and safety production data based on the coupled feature sequence, and to identify abnormal association regions. By integrating the recognition results of the equipment abnormal displacement recognition subnetwork, the personnel trajectory deviation recognition subnetwork, and the abnormal associated region recognition subnetwork, an initial abnormal information pair set is generated; The identification result output module is used to perform spatiotemporal confidence compensation and risk level fusion on the initial anomaly information set, and output the identification result including anomaly location coordinates, time nodes, and risk levels, including: Extract the spatiotemporal context features of each anomaly information pair in the initial anomaly information pair set, calculate the matching degree with historical anomaly patterns, and generate spatiotemporal confidence weights. The degree of anomaly of the corresponding abnormal information pairs is compensated and corrected based on the spatiotemporal confidence weight to obtain a weighted anomaly score, wherein the degree of anomaly is the weighted fusion value of the abnormal displacement deviation of the equipment, the distance of the trajectory from the safe area, and the correlation deviation. The weighted anomaly score is matched with a preset risk level threshold range to determine the final risk level; Combining the location coordinates and time nodes of the anomaly information pair, the output includes the anomaly location coordinates, time nodes, and the final risk level.

Images