An open-pit mine intelligent mine safety monitoring and early warning system based on AI identification

Abstract

The application relates to the technical field of mine safety early warning, in particular to an open-pit mine intelligent mine safety monitoring and early warning system based on AI recognition, which comprises a hierarchical identification unit used for constructing a multi-task hierarchical AI recognition neural network, the multi-task hierarchical AI recognition neural network comprising a type identification subnetwork and a risk factor identification subnetwork; the corrected risk weight is embedded into a cross attention layer of the operation type identification subnetwork and the risk factor identification subnetwork as an input feature, and operation type classification results and risk factor detection results are output; by introducing the multi-task hierarchical AI recognition neural network and a risk score change rate mechanism into the hierarchical identification unit and the self-adaptive switching unit, the AI recognition submodels corresponding to two operation states can be operated in parallel when the operation state mutates, and an effective AI recognition submodel can be selected according to the real-time risk score change rate, so that the delay problem caused by periodic processing is reduced.

Description

Technical Field

[0001] This application relates to the field of mine safety early warning technology, specifically to an AI-based intelligent mine safety monitoring and early warning system for open-pit mines. Background Technology

[0002] Open-pit mines operate in complex environments, influenced by factors such as weather changes, equipment operating status, and personnel distribution, posing certain safety risks. In recent years, artificial intelligence-based safety monitoring methods have been increasingly applied to mine safety management. These technologies typically use video surveillance footage and sensor data to identify operational status and then match the identification results with appropriate risk assessment models to provide early warnings of potential risks during mine operations.

[0003] However, existing related technologies still have certain limitations in practical applications, specifically: First, the risk modeling process often relies on preset mapping functions or historical data statistics to allocate risk weights, which are static empirical models. When encountering new risk events that have never occurred before, such as new equipment failures or sudden extreme weather conditions, the risk weights often cannot be adjusted in time, resulting in biases in risk estimation and making it difficult to effectively identify and respond to new risks.

[0004] Secondly, existing model switching mechanisms are typically based on periodic batch processing, relying on the average recognition results within a preset time period to trigger model switching, which lacks real-time capability. In scenarios where job types change rapidly or job states change abruptly, the system may struggle to update the recognition model in a timely manner, potentially leading to delays in risk monitoring response.

[0005] Therefore, an AI-based intelligent mine safety monitoring and early warning system for open-pit mines is proposed to address the above issues. Summary of the Invention

[0006] To achieve the above objectives, this application provides an AI-based intelligent open-pit mine safety monitoring and early warning system, comprising the following: The data acquisition unit is used to acquire multi-source data from the open-pit mine operation area and perform alignment processing on the multi-source data.

[0007] The environment construction unit is used to construct the dynamic environment tensor field of open-pit mining operations based on the aligned multi-source data through a data fusion algorithm.

[0008] The risk estimation unit is used to extract risk factors and estimate risk weights based on the dynamic environmental tensor field of open-pit mining operations. When a new risk event occurs, the initial risk weights are generated through the meta-learning unit, and the initial risk weights are dynamically corrected by combining the online gradient update mechanism to obtain the corrected risk weights.

[0009] The hierarchical recognition unit is used to construct a multi-task hierarchical AI recognition neural network, which includes a task type recognition sub-network and a risk factor recognition sub-network. The modified risk weights are embedded as input features into the cross-attention layer of the task type recognition sub-network and the risk factor recognition sub-network, and the task type classification results and risk factor detection results are output.

[0010] The adaptive switching unit is used to calculate the risk score change rate based on the job type classification results and risk factor detection results. When the job status changes abruptly, it runs the AI ​​recognition sub-models corresponding to the two job statuses in parallel, selects the effective AI recognition sub-model based on the risk score change rate, and outputs the risk recognition result after adaptive switching.

[0011] The instruction output unit is used to classify the open-pit mine operation area according to the risk identification results after adaptive switching, and push the corresponding preset instruction after the rating reaches the preset level.

[0012] Preferably, based on the aligned multi-source data, a dynamic environmental tensor field for open-pit mining operations is constructed using a data fusion algorithm, specifically including: Deploy fixed video surveillance cameras, drone inspection image acquisition terminals, meteorological sensors, mine equipment operation status sensors, and personnel positioning and trajectory acquisition devices to acquire multi-source data within the open-pit mine operation area.

[0013] The acquired multi-source data includes video surveillance data, meteorological environmental data, mining equipment operation data, and personnel positioning data. The video surveillance data includes image sequences of open-pit mine operation sites. The acquired meteorological environmental data includes parameters such as temperature, humidity, wind speed, wind direction, and precipitation. The acquired mining equipment operation data includes mining equipment temperature data, mining equipment vibration data, and mining equipment current data. The acquired personnel positioning data includes personnel location information and personnel trajectory information.

[0014] The acquired video surveillance data, meteorological environmental data, mining equipment operation data, and personnel positioning data are processed with unified timestamp synchronization and spatial coordinate mapping. Alignment is completed under a unified time reference and a unified coordinate reference, and the aligned multi-source data is output.

[0015] Preferably, based on the aligned multi-source data, a dynamic environmental tensor field for open-pit mining operations is constructed using a data fusion algorithm, specifically including: After alignment, the multi-source data undergoes quality assessment and noise suppression; video surveillance data undergoes frame deblurring and illumination normalization; meteorological and environmental data undergoes outlier removal; mining equipment operation data undergoes sensor drift correction; and personnel positioning data undergoes trajectory continuity verification.

[0016] Based on a unified time reference, time resampling is performed on video surveillance frame sequence data, meteorological environment time series data, mining equipment operation time series data, and personnel positioning time series data, and a unified sampling period is configured to generate a time index sequence.

[0017] A three-dimensional grid coordinate system for the open-pit mine operation area is established based on a unified coordinate benchmark. The grid resolution is configured, and the field of view of video surveillance data, meteorological and environmental observation locations, mine equipment installation locations, and personnel positioning locations are mapped to the three-dimensional grid coordinate system to generate a spatial index sequence.

[0018] Video feature vectors are formed by extracting the detection results, motion vectors, and regional occlusion probabilities of the work objects from video surveillance frame sequence data; meteorological feature vectors are formed by extracting temperature, humidity, wind speed, wind direction, and precipitation parameters from meteorological environmental time series data; equipment feature vectors are formed by extracting mine equipment temperature statistics, mine equipment vibration frequency band energy, and mine equipment current fluctuation amplitude from mine equipment operation time series data; and personnel feature vectors are formed by extracting personnel density, personnel speed, and personnel dwell time from personnel positioning time series data.

[0019] Based on the time index sequence and spatial index sequence, the video feature vector, meteorological feature vector, equipment feature vector and personnel feature vector are subjected to spatiotemporal weighted fusion processing. The weighting factor is calculated based on the data alignment integrity score, sensor self-inspection status score and historical stability score to generate a dynamic environment tensor unit containing a multi-source feature channel set and a quality mask. The dynamic environment tensor units are stacked according to the time index sequence and spatial index sequence to form the dynamic environment tensor field of open-pit mining operations.

[0020] Preferably, risk factor extraction and risk weight estimation are performed based on the dynamic environmental tensor field of open-pit mining operations, specifically including: Meteorological environmental feature factors, mining equipment operation feature factors, personnel distribution feature factors, and work scene image feature factors are extracted from the dynamic environment tensor field. The meteorological environmental feature factors, including temperature, humidity, wind speed, wind direction, and precipitation parameters, are normalized. The mining equipment operation feature factors, including temperature statistics, vibration frequency band energy, and current fluctuation amplitude, are standardized. The personnel distribution feature factors, including personnel density, personnel speed, and personnel dwell time, are statistically modeled. The work scene image feature factors, including the detection results of work objects, motion vectors, and regional occlusion probabilities, are convolutionally encoded. The processed data are then summarized to form a risk factor set.

[0021] Risk weights are estimated using a set of risk factors. An initial risk weight is assigned to each risk factor set using a probability calculation method based on Bayesian inference, and the risk factor weight distribution is output.

[0022] Preferably, when a new type of risk event occurs, initial risk weights are generated through a meta-learning unit, and the initial risk weights are dynamically corrected using an online gradient update mechanism to obtain corrected risk weights, specifically including: The meta-learning unit is used to adaptively model the input distribution of the risk factor set under conditions of distribution shift or small sample size, through task-level parameter extraction, fast parameter update and cross-task transfer learning mechanism, to generate initial risk weights that can reflect the characteristics of new risk events.

[0023] When a new type of risk event appears in the dynamic environment tensor field, the meta-learning unit is invoked to generate the initial risk weight of the new risk event. The meta-learning unit establishes the matching relationship between risk factors and risk levels under small sample conditions through a fast parameter adaptive algorithm, and outputs the initial risk weight corresponding to the new risk event.

[0024] The initial risk weights of new risk events are fused with the historical risk factor weight distributions. An online gradient update mechanism is used to iteratively correct the fusion result. The online gradient update mechanism calculates and corrects the gradient based on the risk event feedback error and gradually updates the risk weights under the control of the learning rate parameter, outputting the corrected risk weight vector.

[0025] Preferably, a multi-task hierarchical AI recognition neural network is constructed, which includes a task type recognition sub-network and a risk factor recognition sub-network, specifically including: A multi-task hierarchical AI recognition neural network is constructed, consisting of an input layer and a feature extraction layer. The input layer receives the corrected risk weight vector and the feature vector of the dynamic environment tensor field, while the feature extraction layer performs convolutional encoding and sequence modeling on the feature vector of the dynamic environment tensor field.

[0026] In the task type identification subnetwork, a feature extraction structure composed of convolutional neural units and recurrent neural units is set up to output the feature vector of candidate task types; in the risk factor identification subnetwork, a feature extraction structure composed of convolutional neural units and graph convolutional units is set up to output the feature vector of candidate risk factors.

[0027] A cross-attention layer is introduced between the job type identification subnetwork and the risk factor identification subnetwork. The cross-attention layer is used to establish the correlation mapping between the candidate feature vectors of job types and the candidate feature vectors of risk factors, and to realize the feature cross constraint through the attention weight matrix.

[0028] Preferably, the modified risk weights are embedded as input features into the cross-attention layer of the job type identification subnetwork and the risk factor identification subnetwork, outputting job type classification results and risk factor detection results, specifically including: The corrected risk weights are embedded as input features into the attention weight matrix of the cross-attention layer. The cross-attention layer adjusts the weight distribution of the candidate feature vectors for job types and candidate feature vectors for risk factors according to the corrected risk weights, and generates a fused feature tensor.

[0029] The fused feature tensor is fed into the classification output layer of the job type identification subnetwork and the detection output layer of the risk factor identification subnetwork, respectively. The job type identification subnetwork outputs the job type classification result, and the risk factor identification subnetwork outputs the risk factor detection result.

[0030] Preferably, the risk score change rate is calculated based on the job type classification results and risk factor detection results, specifically including: A job status score is generated based on the job type classification results. The job status score is obtained by calculating the classification probability of the dynamic environment tensor field feature vector through the job type recognition subnetwork.

[0031] Risk factor scores are generated based on the risk factor detection results. These scores are obtained by calculating the confidence level of the risk factor set through a risk factor identification sub-network.

[0032] The work status score and risk factor score are weighted and fused to configure a risk score sequence, and the rate of change of the risk score sequence is calculated based on continuous time segments.

[0033] The risk score change rate is used to determine the stability of the operation status. When the risk score change rate exceeds the preset change rate threshold, it is determined that the open-pit mine operation status has changed abruptly, and the parallel operation mechanism of the AI ​​recognition sub-model is triggered.

[0034] Preferably, when the work status changes abruptly, the AI ​​recognition sub-models corresponding to the two work states are run in parallel, and the effective AI recognition sub-model is selected based on the risk score change rate, outputting the risk recognition result after adaptive switching, specifically including: The AI ​​recognition sub-models corresponding to the two types of operation states are the normal operation state AI recognition sub-model and the emergency risk operation state AI recognition sub-model.

[0035] The AI ​​recognition sub-model for normal operating conditions adopts a hybrid architecture combining convolutional neural networks and recurrent neural networks, which is used for classifying operating types and detecting risk factors in stable environments.

[0036] The AI ​​identification sub-model for sudden risk operation status adopts a hybrid architecture combining convolutional neural networks and graph convolutional networks, which is used for the detection of highly sensitive risk factors and the determination of risk levels in extreme environments.

[0037] After a sudden change in the work status, the AI ​​identification sub-model for normal work status and the AI ​​identification sub-model for sudden risk work status run in parallel, respectively outputting work type classification results and risk factor detection results.

[0038] By comparing the risk score change rate and feedback error distribution of the two AI recognition sub-models, the one with a better risk score change rate and feedback error distribution than the other sub-model is selected as the effective AI recognition sub-model, and the risk recognition result after adaptive switching is output.

[0039] Preferably, the open-pit mine operation area is rated and divided according to the risk identification results after adaptive switching. After the rating reaches a preset level, the corresponding preset instruction is pushed, specifically including: A comprehensive risk score is calculated based on the risk identification results after adaptive switching. The comprehensive risk score is obtained by fusing the job status score of the job type classification result and the risk factor score of the risk factor detection result according to a weighted coefficient.

[0040] Five risk levels are configured within the range of 0 to 100 based on the comprehensive risk score: the first risk level corresponds to 0–20, the second risk level corresponds to 21–40, the third risk level corresponds to 41–60, the fourth risk level corresponds to 61–80, and the fifth risk level corresponds to 81–100.

[0041] The open-pit mine operation area is rated and divided according to the risk level corresponding to the comprehensive risk score, and the corresponding risk level label is output.

[0042] Preset instructions for each risk level are set: the first risk level pushes the operation instruction to maintain normal operation status; the second risk level pushes the local operation area inspection instruction; the third risk level pushes the operation instruction to reduce the intensity of operation and dispatch backup equipment; the fourth risk level pushes the control instruction to divert personnel, reduce equipment load and implement local control; and the fifth risk level pushes the linkage instruction to shut down the entire area, carry out emergency evacuation and emergency rescue.

[0043] When the risk level of an open-pit mine operation area reaches level three or above, a preset instruction push mechanism is automatically triggered, and corresponding preset instructions are issued through the mine dispatch center control terminal, personnel positioning terminal and equipment control terminal.

[0044] The beneficial effects of this application are as follows: 1. With the support of the data acquisition unit and the environment construction unit, a dynamic environmental tensor field for open-pit mining operations is constructed based on multi-source data after alignment processing. A meta-learning unit and an online gradient update mechanism are introduced into the risk estimation unit, enabling risk weights to be dynamically adjusted according to new risk events, avoiding excessive reliance on historical experience data for risk estimation. This allows for timely identification and response to risks such as new equipment failures and extreme weather conditions, improving the system's risk adaptability in dynamic environments.

[0045] 2. By introducing a multi-task hierarchical AI recognition neural network and a risk score change rate calculation mechanism into the hierarchical recognition unit and the adaptive switching unit, the AI ​​recognition sub-models corresponding to the two job states can run in parallel when the job state changes abruptly. The effective AI recognition sub-model is selected based on the real-time risk score change rate, thereby reducing the delay problem caused by periodic processing, realizing real-time response to the scenario of rapid job type switching, and improving the timeliness and accuracy of risk recognition results.

[0046] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is an overall framework diagram of an AI-based intelligent mine safety monitoring and early warning system for open-pit mines, provided as an embodiment of this application. Detailed Implementation

[0049] The technical solutions in the embodiments of this application will now be described with reference to the accompanying drawings.

[0050] Please see Figure 1 , Figure 1 This is an overall framework diagram of an AI-based intelligent mine safety monitoring and early warning system for open-pit mines, provided as an embodiment of this application.

[0051] In this embodiment, an AI-based intelligent open-pit mine safety monitoring and early warning system includes the following: The data acquisition unit is used to acquire multi-source data from the open-pit mine operation area and perform alignment processing on the multi-source data, specifically including: Deploy fixed video surveillance cameras, drone inspection image acquisition terminals, meteorological sensors, mine equipment operation status sensors, and personnel positioning and trajectory acquisition devices to acquire multi-source data within the open-pit mine operation area.

[0052] It should be noted that the fixed video surveillance camera should be installed at a height of 6 to 12 meters and cover the target work area. The drone inspection image acquisition terminal should collect data according to the flight path with an overlap of not less than 70% in the forward direction and not less than 60% in the lateral direction. The meteorological sensor and the mining equipment operation status sensor should complete the zero-point calibration and range calibration before being put into use. The personnel positioning and trajectory acquisition device should use GNSSRTK positioning module or UWB positioning module to meet the positioning needs of the large-scale and obstructed environment of the open-pit mine operation area.

[0053] The acquired multi-source data includes video surveillance data, meteorological environmental data, mining equipment operation data, and personnel positioning data. The video surveillance data includes image sequences of open-pit mine operation sites. The acquired meteorological environmental data includes parameters such as temperature, humidity, wind speed, wind direction, and precipitation. The acquired mining equipment operation data includes mining equipment temperature data, mining equipment vibration data, and mining equipment current data. The acquired personnel positioning data includes personnel location information and personnel trajectory information.

[0054] It should be noted that the video surveillance data includes image sequences of the open-pit mine operation site with a frame rate set to 15 to 25 frames per second. The meteorological environment data does not include image sequences and is recorded with a sampling period of 1 to 5 minutes. The sampling frequency of the mining equipment operation data is set to 100 Hz to 1000 Hz to capture vibration and transient changes in current. The update frequency of the personnel positioning data is set to 1 Hz to 10 Hz to reconstruct the continuity of personnel trajectories.

[0055] The acquired video surveillance data, meteorological environmental data, mining equipment operation data, and personnel positioning data are processed with unified timestamp synchronization and spatial coordinate mapping. Alignment is completed under a unified time reference and a unified coordinate reference, and the aligned multi-source data is output.

[0056] The environment construction unit is used to construct the dynamic environment tensor field of open-pit mining operations based on aligned multi-source data using a data fusion algorithm. Specifically, it includes: After alignment, the multi-source data undergoes quality assessment and noise suppression; video surveillance data undergoes frame deblurring and illumination normalization; meteorological and environmental data undergoes outlier removal; mining equipment operation data undergoes sensor drift correction; and personnel positioning data undergoes trajectory continuity verification.

[0057] It should be noted that the frame deblurring process adopts a frame-by-frame detection method based on the sharpness threshold, the illumination normalization process adopts the brightness histogram equalization method, the outlier removal process is completed by setting a physical threshold range, the sensor drift correction process is implemented by the zero-point offset correction method, and the trajectory continuity verification process is completed by the dual constraints of time interval and spatial distance.

[0058] Based on a unified time reference, time resampling is performed on video surveillance frame sequence data, meteorological environment time series data, mining equipment operation time series data, and personnel positioning time series data, and a unified sampling period is configured to generate a time index sequence.

[0059] It should be noted that the unified sampling period is set to a range of 1 to 5 seconds to ensure that data from different sources can be aligned and remain synchronized under a unified time reference.

[0060] A three-dimensional grid coordinate system for the open-pit mine operation area is established based on a unified coordinate benchmark. The grid resolution is configured, and the field of view of video surveillance data, meteorological and environmental observation locations, mine equipment installation locations, and personnel positioning locations are mapped to the three-dimensional grid coordinate system to generate a spatial index sequence.

[0061] It should be noted that the resolution of the three-dimensional grid coordinate system is set to 1 to 5 meters, which can reflect the spatial distribution characteristics of the open-pit mine operation area while ensuring computational efficiency.

[0062] Video feature vectors are formed by extracting the detection results, motion vectors, and regional occlusion probabilities of the work objects from video surveillance frame sequence data; meteorological feature vectors are formed by extracting temperature, humidity, wind speed, wind direction, and precipitation parameters from meteorological environmental time series data; equipment feature vectors are formed by extracting mine equipment temperature statistics, mine equipment vibration frequency band energy, and mine equipment current fluctuation amplitude from mine equipment operation time series data; and personnel feature vectors are formed by extracting personnel density, personnel speed, and personnel dwell time from personnel positioning time series data.

[0063] It should be noted that the detection results of the work objects are obtained by outputting bounding box information through the target detection network, the motion vector is obtained by calculating the displacement between consecutive frames, and the regional occlusion probability is obtained by calculating the pixel-level difference. Temperature, humidity, wind speed, wind direction and precipitation parameters are directly provided by the sensors. The mining equipment operation data are calculated into statistics by mean, variance or frequency domain energy. Personnel density, personnel speed and personnel stay time are calculated by location point set statistical methods.

[0064] Based on the time index sequence and spatial index sequence, the video feature vector, meteorological feature vector, equipment feature vector and personnel feature vector are subjected to spatiotemporal weighted fusion processing. The weighting factor is calculated based on the data alignment integrity score, sensor self-inspection status score and historical stability score to generate a dynamic environment tensor unit containing a multi-source feature channel set and a quality mask. The dynamic environment tensor units are stacked according to the time index sequence and spatial index sequence to form the dynamic environment tensor field of open-pit mining operations.

[0065] It should be noted that the spatiotemporal weighted fusion processing is implemented using a weighted average method. The weighting factors of data from different sources are allocated between 0 and 1. The quality mask of the dynamic environment tensor unit is used to mark the location of missing data or outliers to ensure the stability of subsequent calculations.

[0066] The risk estimation unit is used to extract risk factors and estimate risk weights based on the dynamic environmental tensor field of open-pit mining operations. When new risk events occur, it generates initial risk weights through a meta-learning unit and dynamically corrects the initial risk weights using an online gradient update mechanism to obtain the corrected risk weights. Specifically, it includes: Meteorological environmental feature factors, mining equipment operation feature factors, personnel distribution feature factors, and work scene image feature factors are extracted from the dynamic environment tensor field. The meteorological environmental feature factors, including temperature, humidity, wind speed, wind direction, and precipitation parameters, are normalized. The mining equipment operation feature factors, including temperature statistics, vibration frequency band energy, and current fluctuation amplitude, are standardized. The personnel distribution feature factors, including personnel density, personnel speed, and personnel dwell time, are statistically modeled. The work scene image feature factors, including the detection results of work objects, motion vectors, and regional occlusion probabilities, are convolutionally encoded. The processed data are then summarized to form a risk factor set.

[0067] It should be noted that the normalization and standardization methods adopt linear scaling, the statistical modeling method adopts distribution-fitting statistical modeling, and the convolutional feature encoding adopts multi-layer convolution and pooling operations to extract feature vectors. The above processing methods are all well-known in this technical field, and their general implementation methods will not be described in detail in this invention.

[0068] Risk weights are estimated using a set of risk factors. An initial risk weight is assigned to each risk factor set using a probability calculation method based on Bayesian inference, and the risk factor weight distribution is output.

[0069] It should be noted that the probability calculation method of Bayesian inference is based on the combination of prior probability and likelihood function, and the risk factors are weighted by the posterior probability distribution. The above probability calculation method is a well-known technical content, and its general implementation will not be described in this invention.

[0070] The meta-learning unit is used to adaptively model the input distribution of the risk factor set under conditions of distribution shift or small sample size, through task-level parameter extraction, fast parameter update and cross-task transfer learning mechanism, to generate initial risk weights that can reflect the characteristics of new risk events.

[0071] It should be noted that task-level parameter extraction is accomplished by sharing the underlying network parameters, fast parameter updates employ the few-step gradient descent method, and the cross-task transfer learning mechanism adopts a transfer method based on embedding features from similar tasks.

[0072] When a new type of risk event appears in the dynamic environment tensor field, the meta-learning unit is called to generate the initial risk weight of the new risk event. The meta-learning unit establishes the matching relationship between risk factors and risk levels under small sample conditions through a fast parameter adaptive algorithm, and outputs the initial risk weight corresponding to the new risk event. The new types of risk events include various extreme weather changes or sudden anomalies in various mining equipment.

[0073] The initial risk weights of new risk events are fused with the historical risk factor weight distributions. An online gradient update mechanism is used to iteratively correct the fusion result. The online gradient update mechanism calculates and corrects the gradient based on the risk event feedback error and gradually updates the risk weights under the control of the learning rate parameter, outputting the corrected risk weight vector.

[0074] It should be noted that the online gradient update mechanism uses mini-batch gradient descent or stochastic gradient descent, with the learning rate parameter set between 0.001 and 0.01 to ensure stable updates of risk weights.

[0075] A hierarchical recognition unit is used to construct a multi-task hierarchical AI recognition neural network, which includes a task type recognition sub-network and a risk factor recognition sub-network. The modified risk weights are embedded as input features into the cross-attention layer of the task type recognition sub-network and the risk factor recognition sub-network, outputting task type classification results and risk factor detection results, specifically including: A multi-task hierarchical AI recognition neural network is constructed, consisting of an input layer and a feature extraction layer. The input layer receives the corrected risk weight vector and the feature vector of the dynamic environment tensor field, while the feature extraction layer performs convolutional encoding and sequence modeling on the feature vector of the dynamic environment tensor field.

[0076] It should be noted that convolutional coding uses a combination of multiple convolutional kernels and pooling layers to extract local spatial features, while sequence modeling uses long short-term memory networks or gated recurrent units to model time series features to ensure that the temporal dependencies of features are preserved.

[0077] In the task type identification subnetwork, a feature extraction structure composed of convolutional neural units and recurrent neural units is set up to output the feature vector of candidate task types; in the risk factor identification subnetwork, a feature extraction structure composed of convolutional neural units and graph convolutional units is set up to output the feature vector of candidate risk factors.

[0078] It should be noted that graph convolutional units are used to process graph-structured data consisting of relationships between risk factors. Nodes represent risk factors, and edge weights represent the correlation between risk factors. Graph convolutional units propagate features through adjacency matrices to model the relationships between risk factors.

[0079] A cross-attention layer is introduced between the job type identification subnetwork and the risk factor identification subnetwork. The cross-attention layer is used to establish the correlation mapping between the candidate feature vectors of job types and the candidate feature vectors of risk factors, and to realize the feature cross constraint through the attention weight matrix.

[0080] It should be noted that the attention weight matrix of the cross-attention layer is calculated from the query vector, key vector, and value vector. The query vector comes from the candidate feature vector of the task type, and the key vector and value vector come from the candidate feature vector of the risk factor. The attention distribution is generated by dot product attention calculation to realize the interaction between different task features.

[0081] The corrected risk weights are embedded as input features into the attention weight matrix of the cross-attention layer. The cross-attention layer adjusts the weight distribution of the candidate feature vectors for job types and candidate feature vectors for risk factors according to the corrected risk weights, and generates a fused feature tensor.

[0082] It should be noted that the modified risk weights are embedded as weighting coefficients in the normalization stage of the attention distribution in the cross-attention layer. This is used to amplify the feature contribution of high-risk factors and suppress the interference of low-risk factors, thereby improving the discriminative ability of the fused feature tensor.

[0083] The fused feature tensor is fed into the classification output layer of the job type identification subnetwork and the detection output layer of the risk factor identification subnetwork, respectively. The job type identification subnetwork outputs the job type classification result, and the risk factor identification subnetwork outputs the risk factor detection result.

[0084] It should be noted that the classification output layer uses a combination of a fully connected layer and a Softmax function, while the detection output layer uses a combination of a fully connected layer and a Sigmoid function, in order to achieve classification of multiple job types and detection of multi-label risk factors.

[0085] The adaptive switching unit calculates the risk score change rate based on the job type classification results and risk factor detection results. When a job status changes abruptly, it runs the AI ​​recognition sub-models corresponding to the two job statuses in parallel, selects the effective AI recognition sub-model based on the risk score change rate, and outputs the risk recognition result after adaptive switching. Specifically, this includes: A job status score is generated based on the job type classification results. The job status score is obtained by calculating the classification probability of the dynamic environment tensor field feature vector through the job type recognition subnetwork.

[0086] It should be noted that the numerical range of the job status score is a decimal between 0 and 1. The closer the value is to 1, the higher the confidence level of the job type being identified, and the closer the value is to 0, the lower the confidence level of the job type being identified.

[0087] Risk factor scores are generated based on the risk factor detection results. These scores are obtained by calculating the confidence level of the risk factor set through a risk factor identification sub-network.

[0088] It should be noted that the risk factor scores are also decimals between 0 and 1, and correspond to different risk factors. The scores of different risk factors can be combined into a vector for subsequent weighted fusion.

[0089] The work status score and risk factor score are weighted and fused to configure a risk score sequence, and the rate of change of the risk score sequence is calculated based on continuous time segments.

[0090] It should be noted that the weighting ratio of the risk scoring sequence can be set according to the historical assessment accuracy. For example, if the weighting ratio of the operation status score is 0.4 and the weighting ratio of the risk factor score is 0.6, it can ensure that the risk factor detection has higher sensitivity in the overall assessment.

[0091] The risk score change rate is used to determine the stability of the operation status. When the risk score change rate exceeds the preset change rate threshold, it is determined that the open-pit mine operation status has changed abruptly, and the parallel operation mechanism of the AI ​​recognition sub-model is triggered.

[0092] It should be noted that the preset change rate threshold is set between 10% and 20%. When the change in risk score between two consecutive time segments exceeds the preset change rate threshold, it is determined to be a sudden change in the work status.

[0093] The AI ​​recognition sub-models corresponding to the two types of operation states are the normal operation state AI recognition sub-model and the emergency risk operation state AI recognition sub-model.

[0094] It should be noted that the AI ​​recognition sub-model for normal operation status is mainly built on a stable operation scenario dataset during the training phase, while the AI ​​recognition sub-model for sudden risk operation status is mainly built on a dataset of extreme weather and equipment anomaly scenarios during the training phase.

[0095] The AI ​​recognition sub-model for normal operating conditions adopts a hybrid architecture combining convolutional neural networks and recurrent neural networks, which is used for classifying operating types and detecting risk factors in stable environments.

[0096] The AI ​​identification sub-model for sudden risk operation status adopts a hybrid architecture combining convolutional neural networks and graph convolutional networks, which is used for the detection of highly sensitive risk factors and the determination of risk levels in extreme environments.

[0097] After a sudden change in the work status, the AI ​​identification sub-model for normal work status and the AI ​​identification sub-model for sudden risk work status run in parallel, respectively outputting work type classification results and risk factor detection results.

[0098] It should be noted that the parallel operation mechanism completes the inference by sharing input data and performing inference in two sub-models respectively, with the inference latency controlled within 1 second to ensure the real-time nature of risk identification.

[0099] By comparing the risk score change rate and feedback error distribution of the two AI recognition sub-models, the one with a better risk score change rate and feedback error distribution than the other sub-model is selected as the effective AI recognition sub-model, and the risk recognition result after adaptive switching is output.

[0100] It should be noted that the feedback error distribution is obtained by calculating the difference between the model output and the historical data of the system. The feedback process is a well-known technical content in this field, and the general implementation method will not be described in detail in this invention.

[0101] The instruction output unit is used to classify and rate the open-pit mine operation area based on the risk identification results after adaptive switching. Once the rating reaches a preset level, it pushes the corresponding preset instructions, specifically including: A comprehensive risk score is calculated based on the risk identification results after adaptive switching. The comprehensive risk score is obtained by fusing the job status score of the job type classification result and the risk factor score of the risk factor detection result according to a weighted coefficient.

[0102] It should be noted that the weighting coefficients are configured based on the historical risk prediction accuracy. If the weighting coefficient for the work status score is set to 0.3 and the weighting coefficient for the risk factor score is set to 0.7, the response sensitivity to the risk factor detection results will be enhanced.

[0103] Five risk levels are configured within the range of 0 to 100 based on the comprehensive risk score: the first risk level corresponds to 0–20, the second risk level corresponds to 21–40, the third risk level corresponds to 41–60, the fourth risk level corresponds to 61–80, and the fifth risk level corresponds to 81–100.

[0104] It should be noted that the comprehensive risk score is calculated from the normalized input data to ensure that data from different sources have a unified dimension.

[0105] The open-pit mine operation area is rated and divided according to the risk level corresponding to the comprehensive risk score, and the corresponding risk level label is output.

[0106] Preset instructions for each risk level are set: the first risk level pushes the operation instruction to maintain normal operation status; the second risk level pushes the local operation area inspection instruction; the third risk level pushes the operation instruction to reduce the intensity of operation and dispatch backup equipment; the fourth risk level pushes the control instruction to divert personnel, reduce equipment load and implement local control; and the fifth risk level pushes the linkage instruction to shut down the entire area, carry out emergency evacuation and emergency rescue.

[0107] It should be noted that the execution forms of the preset instructions include three methods: text message, voice alarm and automatic device control. Among them, automatic device control is achieved through interaction between control signals and the device control terminal. This type of control signal interaction is a well-known technical content in this field, and the general implementation method will not be described in detail in this invention.

[0108] When the risk level of an open-pit mine operation area reaches level three or above, a preset instruction push mechanism is automatically triggered, and corresponding preset instructions are issued through the mine dispatch center control terminal, personnel positioning terminal and equipment control terminal.

[0109] Thus, an AI-based intelligent mine safety monitoring and early warning system for open-pit mines has been completed.

[0110] It should be understood that in the various embodiments of this application, the order of the units or modules of the above processes does not mean the order of execution. The execution order of each unit or module should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0111] Those skilled in the art will recognize that the algorithms or steps described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0112] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A smart open-pit mine safety monitoring and early warning system based on AI recognition, characterized in that, Including the following: The data acquisition unit is used to acquire multi-source data from the open-pit mine operation area and perform alignment processing on the multi-source data. The environment construction unit is used to construct the dynamic environment tensor field of open-pit mining operations based on the aligned multi-source data through a data fusion algorithm. The risk estimation unit is used to extract risk factors and estimate risk weights based on the dynamic environment tensor field of open-pit mining operations. When a new risk event occurs, the initial risk weights are generated through the meta-learning unit and dynamically corrected by combining the online gradient update mechanism to obtain the corrected risk weights. The hierarchical recognition unit is used to construct a multi-task hierarchical AI recognition neural network, which includes a task type recognition sub-network and a risk factor recognition sub-network. The modified risk weights are embedded as input features into the cross-attention layer of the job type identification subnetwork and the risk factor identification subnetwork, and the job type classification results and risk factor detection results are output. The adaptive switching unit is used to calculate the risk score change rate based on the job type classification results and risk factor detection results. When the job status changes abruptly, it runs the AI ​​recognition sub-models corresponding to the two job statuses in parallel, selects the effective AI recognition sub-model based on the risk score change rate, and outputs the risk recognition result after adaptive switching. The instruction output unit is used to classify the open-pit mine operation area according to the risk identification results after adaptive switching, and push the corresponding preset instruction after the rating reaches the preset level.

2. The AI-based intelligent open-pit mine safety monitoring and early warning system as described in claim 1, characterized in that, Based on the aligned multi-source data, a dynamic environmental tensor field for open-pit mining operations is constructed using a data fusion algorithm, specifically including: Deploy fixed video surveillance cameras, drone inspection image acquisition terminals, meteorological sensors, mine equipment operation status sensors, and personnel positioning and trajectory acquisition devices to acquire multi-source data within the open-pit mine operation area; The acquired multi-source data includes video surveillance data, meteorological environmental data, mining equipment operation data, and personnel positioning data. The video surveillance data includes image sequences of open-pit mine operation sites; the acquired meteorological environmental data includes temperature, humidity, wind speed, wind direction, and precipitation parameters; the acquired mining equipment operation data includes mining equipment temperature data, mining equipment vibration data, and mining equipment current data; and the acquired personnel positioning data includes personnel location information and personnel trajectory information. The acquired video surveillance data, meteorological environmental data, mining equipment operation data, and personnel positioning data are processed with unified timestamp synchronization and spatial coordinate mapping. Alignment is completed under a unified time reference and a unified coordinate reference, and the aligned multi-source data is output.

3. The AI-based intelligent open-pit mine safety monitoring and early warning system as described in claim 2, characterized in that, Based on the aligned multi-source data, a dynamic environmental tensor field for open-pit mining operations is constructed using a data fusion algorithm, specifically including: After alignment, the multi-source data will undergo quality assessment and noise suppression; the video surveillance data will undergo frame deblurring and illumination normalization; the meteorological and environmental data will undergo outlier removal; the mining equipment operation data will undergo sensor drift correction; and the personnel positioning data will undergo trajectory continuity verification. Based on a unified time reference, time resampling is performed on video surveillance frame sequence data, meteorological environment time series data, mining equipment operation time series data and personnel positioning time series data, a unified sampling period is configured and a time index sequence is generated. A three-dimensional grid coordinate system for the open-pit mine operation area is established based on a unified coordinate benchmark. The grid resolution is configured, and the field of view of video surveillance data, meteorological and environmental observation locations, mine equipment installation locations, and personnel positioning locations are mapped to the three-dimensional grid coordinate system to generate a spatial index sequence. Video feature vectors are formed by extracting the detection results, motion vectors, and regional occlusion probabilities of the work objects from video surveillance frame sequence data; meteorological feature vectors are formed by extracting temperature, humidity, wind speed, wind direction, and precipitation parameters from meteorological environmental time series data; equipment feature vectors are formed by extracting temperature statistics, vibration frequency band energy, and current fluctuation amplitude of mining equipment from mining equipment operation time series data; and personnel feature vectors are formed by extracting personnel density, personnel speed, and personnel dwell time from personnel positioning time series data. Based on the time index sequence and spatial index sequence, the video feature vector, meteorological feature vector, equipment feature vector and personnel feature vector are subjected to spatiotemporal weighted fusion processing. The weighting factor is calculated based on the data alignment integrity score, sensor self-inspection status score and historical stability score to generate a dynamic environment tensor unit containing a multi-source feature channel set and a quality mask. The dynamic environment tensor units are stacked according to the time index sequence and spatial index sequence to form the dynamic environment tensor field of open-pit mining operations.

4. The AI-based intelligent open-pit mine safety monitoring and early warning system as described in claim 3, characterized in that, Risk factor extraction and risk weight estimation are performed based on the dynamic environmental tensor field of open-pit mining operations, specifically including: Meteorological environmental feature factors, mining equipment operation feature factors, personnel distribution feature factors, and work scene image feature factors are extracted from the dynamic environment tensor field. The meteorological environmental feature factors, including temperature, humidity, wind speed, wind direction, and precipitation parameters, are normalized. The mining equipment operation feature factors, including temperature statistics, vibration frequency band energy, and current fluctuation amplitude, are standardized. The personnel distribution feature factors, including personnel density, personnel speed, and personnel dwell time, are statistically modeled. The work scene image feature factors, including work object detection results, motion vectors, and regional occlusion probability, are convolutionally encoded. The processed data are summarized to form a risk factor set. Risk weights are estimated using a set of risk factors. An initial risk weight is assigned to each risk factor set using a probability calculation method based on Bayesian inference, and the risk factor weight distribution is output.

5. The AI-based intelligent open-pit mine safety monitoring and early warning system as described in claim 4, characterized in that, When a new type of risk event occurs, initial risk weights are generated through a meta-learning unit, and then dynamically corrected using an online gradient update mechanism to obtain corrected risk weights, specifically including: The meta-learning unit is used to adaptively model the input distribution of the risk factor set under the conditions of distribution shift or small sample size, through task-level parameter extraction, fast parameter update and cross-task transfer learning mechanism, and generate initial risk weights that can reflect the characteristics of new risk events. When a new type of risk event appears in the dynamic environment tensor field, the meta-learning unit is called to generate the initial risk weight of the new risk event. The meta-learning unit establishes the matching relationship between risk factors and risk levels under small sample conditions through a fast parameter adaptive algorithm, and outputs the initial risk weight corresponding to the new risk event. The initial risk weights of new risk events are fused with the historical risk factor weight distributions. An online gradient update mechanism is used to iteratively correct the fusion result. The online gradient update mechanism calculates and corrects the gradient based on the risk event feedback error and gradually updates the risk weights under the control of the learning rate parameter, outputting the corrected risk weight vector.

6. The AI-based intelligent open-pit mine safety monitoring and early warning system as described in claim 1, characterized in that, Construct a multi-task hierarchical AI recognition neural network, which includes a task type recognition sub-network and a risk factor recognition sub-network, specifically including: A multi-task hierarchical AI recognition neural network is constructed, consisting of an input layer and a feature extraction layer. The input layer receives the corrected risk weight vector and the feature vector of the dynamic environment tensor field, while the feature extraction layer performs convolutional encoding and sequence modeling on the feature vector of the dynamic environment tensor field. In the task type recognition subnetwork, a feature extraction structure composed of convolutional neural units and recurrent neural units is set up to output the feature vector of candidate task types; in the risk factor recognition subnetwork, a feature extraction structure composed of convolutional neural units and graph convolutional units is set up to output the feature vector of candidate risk factors. A cross-attention layer is introduced between the job type identification subnetwork and the risk factor identification subnetwork. The cross-attention layer is used to establish the correlation mapping between the candidate feature vectors of job types and the candidate feature vectors of risk factors, and to realize the feature cross constraint through the attention weight matrix.

7. The AI-based intelligent open-pit mine safety monitoring and early warning system as described in claim 6, characterized in that, The modified risk weights are embedded as input features into the cross-attention layer of the job type identification subnetwork and the risk factor identification subnetwork, outputting job type classification results and risk factor detection results, specifically including: The corrected risk weights are embedded as input features into the attention weight matrix of the cross-attention layer. The cross-attention layer adjusts the weight distribution of the candidate feature vectors of job type and candidate feature vectors of risk factor according to the corrected risk weights to generate a fused feature tensor. The fused feature tensor is fed into the classification output layer of the job type identification subnetwork and the detection output layer of the risk factor identification subnetwork, respectively. The job type identification subnetwork outputs the job type classification result, and the risk factor identification subnetwork outputs the risk factor detection result.

8. The AI-based intelligent open-pit mine safety monitoring and early warning system as described in claim 1, characterized in that, The risk score change rate is calculated based on the job type classification results and risk factor detection results, specifically including: A job status score is generated based on the job type classification results. The job status score is obtained by calculating the classification probability of the dynamic environment tensor field feature vector through the job type recognition subnetwork. Risk factor scores are generated based on the risk factor detection results. The risk factor scores are obtained by calculating the detection confidence of the risk factor set through a risk factor identification sub-network. The work status score and risk factor score are weighted and fused to configure a risk score sequence, and the rate of change of the risk score sequence is calculated based on continuous time segments. The risk score change rate is used to determine the stability of the operation status. When the risk score change rate exceeds the preset change rate threshold, it is determined that the open-pit mine operation status has changed abruptly, and the parallel operation mechanism of the AI ​​recognition sub-model is triggered.

9. The AI-based intelligent open-pit mine safety monitoring and early warning system as described in claim 8, characterized in that, When a job status changes abruptly, two AI recognition sub-models corresponding to different job statuses are run in parallel. The effective AI recognition sub-model is selected based on the rate of change of risk scores, and the risk recognition result after adaptive switching is output, specifically including: The AI ​​recognition sub-models corresponding to the two types of operation states are the normal operation state AI recognition sub-model and the sudden risk operation state AI recognition sub-model. The AI ​​recognition sub-model for normal operating status adopts a hybrid architecture combining convolutional neural networks and recurrent neural networks, which is used for operation type classification and risk factor detection in stable environments. The AI ​​identification sub-model for sudden risk operation status adopts a hybrid architecture combining convolutional neural networks and graph convolutional networks, which is used for the detection of highly sensitive risk factors and the determination of risk levels in extreme environments. After a sudden change in the work status, the AI ​​recognition sub-model for normal work status and the AI ​​recognition sub-model for sudden risk work status run in parallel, and output the work type classification results and risk factor detection results respectively. By comparing the risk score change rate and feedback error distribution of the two AI recognition sub-models, the one with a better risk score change rate and feedback error distribution than the other sub-model is selected as the effective AI recognition sub-model, and the risk recognition result after adaptive switching is output.

10. The AI-based intelligent open-pit mine safety monitoring and early warning system as described in claim 9, characterized in that, Based on the risk identification results after adaptive switching, the open-pit mine operation area is rated and divided. Once the rating reaches a preset level, the corresponding preset instructions are pushed out, including: A comprehensive risk score is calculated based on the risk identification results after adaptive switching. The comprehensive risk score is obtained by fusing the job status score of the job type classification results and the risk factor score of the risk factor detection results according to a weighted coefficient. Five risk levels are configured based on the comprehensive risk score, ranging from 0 to 100: the first risk level corresponds to 0–20, the second risk level corresponds to 21–40, the third risk level corresponds to 41–60, the fourth risk level corresponds to 61–80, and the fifth risk level corresponds to 81–100. The open-pit mine operation area is rated and divided according to the risk level corresponding to the comprehensive risk score, and the corresponding risk level label is output. Preset instructions for each risk level: Level 1 risk pushes operation instructions to maintain normal operation; Level 2 risk pushes inspection instructions for local operation areas; Level 3 risk pushes operation instructions to reduce operation intensity and dispatch backup equipment; Level 4 risk pushes control instructions for personnel diversion, equipment unloading and local lockdown; Level 5 risk pushes coordinated instructions for full area shutdown, emergency evacuation and emergency rescue. When the risk level of an open-pit mine operation area reaches level three or above, a preset instruction push mechanism is automatically triggered, and corresponding preset instructions are issued through the mine dispatch center control terminal, personnel positioning terminal and equipment control terminal.

Images