Metal mine production big data intelligent analysis method, equipment and medium

Abstract

The invention discloses a metal mine production big data intelligent analysis method and device and a medium, and relates to the technical field of intelligent mine information, and the device comprises an acquisition and modeling module, an analysis module, a prediction and optimization module and a visualization module; the data acquisition unit is provided with a vibration sensor, a temperature sensor and a geological pressure sensor, and is used for acquiring mine equipment operation parameters, mineral processing process indexes and geological environment parameters in real time to obtain multi-source data; the data preprocessing unit is used for performing timestamp alignment, missing value compensation, noise elimination and unit normalization processing on the multi-source data to obtain standardized multi-source data; the feature extraction unit adopts a multi-head Transform network, performs space-time modeling and semantic feature fusion on standardized multi-source data, and outputs a tensor embedding vector, the risk identification result is converted into an executable process adjustment strategy, the crossing from passive early warning to active optimization is realized, and the efficiency and the intelligent level are enhanced.

Description

Technical Field

[0001] This invention relates to the field of intelligent mine information technology, and in particular to an intelligent analysis method, equipment and medium for big data in metal mine production. Background Technology

[0002] With the continuous advancement of the Industrial Internet and smart mine construction, data acquisition and analysis technologies in the metal mine production process have developed rapidly. Traditional mine informatization mainly relies on SCADA (Supervisory Control and Data Acquisition) and MES (Manufacturing Execution System), which deploy various sensors to monitor equipment operating status, process parameters and environmental conditions in real time, and store the data in relational databases or time-series databases for subsequent querying and statistical analysis. The application of big data technology in the mining field is gradually deepening, and some advanced mines have built data platforms based on Hadoop or Spark architecture, realizing centralized management and offline analysis of multi-source heterogeneous data. Existing technologies still face key challenges in dealing with complex dynamic equipment in metal mines. Because mining production equipment involves the coupling of multiple physical fields such as geological environment, electromechanical equipment and mineral processing technology, the operating state exhibits strong nonlinearity, time-varying and spatial dependence characteristics. For example, analysis methods based on fixed topology diagrams cannot adapt to network structure changes caused by process flow adjustments or equipment start-up and shutdown, resulting in a lag in state representation. On the other hand, relying solely on time series models ignores the physical connection relationships between equipment and the upstream and downstream dependencies of the process, limiting the ability to accurately trace the source of equipment-level anomalies. Summary of the Invention

[0003] In view of the aforementioned existing problems, the present invention is proposed.

[0004] Therefore, this invention provides a smart analysis device for big data in metal mine production to solve the problems of lagging state representation and insufficient risk tracing capabilities caused by the coupling of multiple physical fields.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an intelligent analysis device for big data in metal mine production, comprising: a data acquisition and modeling module, an analysis module, a prediction and optimization module, and a visualization module; the data acquisition and modeling module includes a data acquisition unit, a data preprocessing unit, and a feature extraction unit; the data acquisition unit is equipped with vibration sensors, temperature sensors, and geological pressure sensors to collect real-time operating parameters of mining equipment, mineral processing indicators, and geological environment parameters to obtain multi-source data; the data preprocessing unit performs timestamp alignment, missing value compensation, noise removal, and unit normalization on the multi-source data to obtain standardized multi-source data; the feature extraction unit uses a multi-head Transformer network to perform spatiotemporal modeling and semantic feature fusion on the standardized multi-source data, and outputs tensor embedding vectors.

[0006] As a preferred embodiment of the intelligent analysis equipment for big data in metal mine production described in this invention, the equipment comprises: a data acquisition and modeling module, an analysis module, a prediction and optimization module, and a visualization module. The data acquisition and modeling module includes a data acquisition unit, a data preprocessing unit, and a feature extraction unit; The data acquisition unit includes a vibration sensor, a temperature sensor, and a geological pressure sensor, which are used to collect mining equipment operating parameters, mineral processing indicators, and geological environment parameters in real time to obtain multi-source data. The data preprocessing unit is used to perform timestamp alignment, missing value compensation, noise removal and unit normalization on multi-source data to obtain standardized multi-source data. The feature extraction unit employs a multi-head Transformer network to perform spatiotemporal modeling and semantic feature fusion on standardized multi-source data, and outputs tensor embedding vectors.

[0007] As a preferred embodiment of the intelligent analysis equipment for big data in metal mine production according to the present invention, the analysis module includes a dynamic graph construction unit, a temporal convolution unit, and a node update unit. The dynamic graph construction unit establishes a dynamic adjacency matrix based on the spatial correlation in the tensor embedding vector and the process flow to obtain a dynamic graph. The temporal convolutional unit performs temporal feature propagation and spatial convolution calculation based on the dynamic graph, extracts node association features across time slices, and obtains a temporal feature matrix; The node update unit is based on the temporal feature matrix. It fuses the node's neighborhood features and time weights through a feature aggregation function, updates the node's state vector, and outputs the node representation matrix.

[0008] As a preferred embodiment of the intelligent analysis equipment for big data in metal mine production described in this invention, the prediction and optimization module includes a deep prediction unit, a risk identification unit, and a reinforcement learning decision-making unit. The deep prediction unit performs health scoring and potential failure probability prediction on the node representation matrix based on the graph attention mechanism to obtain the node health score and failure probability. The risk identification unit generates multi-level risk labels based on the dynamic changing trends of node health scores and failure probabilities, and forms a risk vector. The reinforcement learning decision-making unit establishes a multi-objective reward function based on the risk vector and production status data, combined with a deep reinforcement learning algorithm, performs strategy optimization calculations on the risk vector, generates process parameter optimization instructions, and outputs the process parameter optimization instructions to the visualization module.

[0009] As a preferred embodiment of the intelligent big data analysis equipment for metal mine production described in this invention, the specific steps for obtaining standardized multi-source data are as follows: Real-time sampling is performed using vibration sensors, temperature sensors, flow meters, and geological pressure sensors deployed at mining site equipment and geological monitoring points to generate multi-source data. Based on multi-source data, the timestamps of different device acquisition channels are synchronized using the PTP protocol, and missing data points are filled in by bidirectional interpolation. Unit standardization conversion is performed to output standardized multi-source data.

[0010] As a preferred embodiment of the intelligent analysis equipment for big data in metal mine production according to the present invention, the specific steps for the output tensor embedding vector are as follows: Standardized multi-source data is organized according to the sampling area number, equipment number and sampling time to establish a three-dimensional tensor structure. The sliding time window mechanism is used to slice and stack the three-dimensional tensor structure within a continuous sampling period to form a temporal tensor block. The temporal tensor block is input into the encoding layer of the multi-head Transformer network to calculate the correlation weights between spatial nodes and temporal nodes. Multidimensional feature fusion and compression are then performed to output a deep semantic feature matrix, which is then mapped to a tensor embedding vector.

[0011] In a preferred embodiment of the intelligent big data analysis equipment for metal mine production described in this invention, the output node represents a matrix, and the specific steps are as follows: Equipment operating parameters, process flow characteristics and geological coupling indicators are extracted from tensor embedding vectors. A dynamic adjacency matrix is ​​constructed based on the spatial correlation and process dependence between features. Nodes represent mining equipment and edge weights characterize the physical connection strength and temporal correlation to form a dynamic graph. The dynamic graph is input into a temporal convolutional network to capture the cooperative change patterns between devices in different time slices. The spatial topological features and temporal evolution features are integrated through a residual fusion structure to output a temporal feature matrix. The temporal feature matrix is ​​subjected to node feature aggregation processing. The neighborhood set of each node is obtained based on the adjacency matrix of the dynamic graph. The weighted average coefficient is calculated based on the edge weight and the feature similarity between nodes. The temporal features of the adjacent nodes are aggregated to obtain the node feature vector set. The node feature vector set is input into the gated recurrent unit to dynamically model the temporal variation of node features. The historical state and current input features are fused through the update gate and reset gate mechanism to extract the temporal evolution pattern of the node at different times and output the node update feature matrix. The node update feature matrix is ​​normalized and fused with residuals. The original node features are then superimposed with the node update feature matrix through a residual connection structure to obtain the node representation matrix.

[0012] As a preferred embodiment of the intelligent big data analysis equipment for metal mine production described in this invention, the specific steps for generating process parameter optimization instructions are as follows: Based on the node representation matrix, the health score and failure probability of each node are calculated using a graph attention neural network model to generate a node health score sequence. Cluster analysis is performed based on the time change rate and process deviation characteristics in the node health score sequence to classify node risk levels and generate multi-level risk labels; the multi-level risk labels are then fused with the node failure probability to output a risk vector. Collect production status data, input the risk vector and production status data into a deep reinforcement learning model, use equipment operating efficiency, energy consumption level and safety risk as state variables, establish a multi-objective reward function and conduct strategy iterative training to obtain the optimal process parameter adjustment scheme; The optimal process parameter adjustment scheme is converted into instructions and subjected to safety constraints to generate process parameter optimization instructions, which are then output to the visualization module.

[0013] As a preferred embodiment of the intelligent big data analysis equipment for metal mine production described in this invention, the visualization module is used for three-dimensional visualization and dynamic simulation verification of process parameter optimization instructions. The specific steps are as follows: Input the process parameter optimization instructions into the digital twin rendering engine, and based on the three-dimensional spatial structure model of the mine and the topology relationship of the equipment, update the virtual equipment operating status, energy consumption curve and process flow path to generate dynamic visualization. The virtual simulation images are compared synchronously with the actual production monitoring data. The deviation detection algorithm is used to calculate the energy efficiency improvement rate and risk change trend after the process parameters are adjusted, and a dynamic performance evaluation result is formed.

[0014] Secondly, this invention provides a method for intelligent analysis of big data in metal mine production, including, Real-time acquisition, preprocessing, and spatiotemporal feature modeling of multi-source heterogeneous information from metal mine production sites are performed, and tensor embedding vectors are output. A dynamic graph is constructed based on the topological relationship of mining equipment and the mineral processing flow. The temporal graph convolution algorithm is used to perform feature propagation and neighborhood information aggregation on the node state, and outputs a node representation matrix. The prediction network based on graph attention mechanism calculates node health scores and failure probabilities, and generates process parameter optimization instructions. The process parameter optimization instructions are displayed in three dimensions and verified by dynamic simulation.

[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the intelligent analysis method for big data in metal mine production as described in the second aspect of the present invention.

[0016] The beneficial effects of this invention are as follows: by constructing a dynamic graph and combining it with temporal convolution and node update mechanisms, it accurately captures the spatiotemporal correlation and state evolution laws between equipment, improves the accuracy of equipment behavior representation under complex working conditions, transforms risk identification results into executable process adjustment strategies, realizes the leap from passive early warning to active optimization, and significantly enhances the safety, efficiency and intelligence level of mine operation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A schematic diagram of intelligent big data analysis equipment for metal mines.

[0019] Figure 2 This is a flowchart for the data acquisition and modeling module.

[0020] Figure 3 This is a flowchart of the analysis module.

[0021] Figure 4 The flowcharts are for the prediction and optimization module and the visualization module. Detailed Implementation

[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a big data intelligent analysis device for metal mine production, including the following steps: The acquisition and modeling module is used to acquire, preprocess, and model the spatiotemporal features of multi-source heterogeneous information from metal mine production sites in real time, and output tensor embedding vectors.

[0026] The acquisition and modeling module is connected to the analysis module via a high-speed data bus for real-time transmission of tensor embedding vectors; The data acquisition and modeling module includes a data acquisition unit, a data preprocessing unit, and a feature extraction unit.

[0027] The data acquisition unit is equipped with vibration sensors, temperature sensors and geological pressure sensors to collect mining equipment operating parameters, mineral processing indicators and geological environment parameters in real time, and obtain multi-source data. The data preprocessing unit is used to perform timestamp alignment, missing value compensation, noise removal and unit normalization on multi-source data to obtain standardized multi-source data; The feature extraction unit employs a multi-head Transformer network to perform spatiotemporal modeling and semantic feature fusion on standardized multi-source data, and outputs tensor embedding vectors.

[0028] Real-time sampling is performed using vibration sensors, temperature sensors, flow meters, and geological pressure sensors deployed at mining site equipment and geological monitoring points to generate multi-source data.

[0029] Specifically, vibration sensors are installed on mining equipment such as crushers, hoists, and conveying devices to collect acceleration signals and vibration spectra during equipment operation, reflecting the stability and wear status of the equipment. Temperature sensors and flow meters are installed on slurry pipelines and flotation units in the mineral processing stage to collect information on changes in slurry temperature, mineral processing medium temperature, and flow rate, in order to characterize the energy transfer efficiency and media flow characteristics of the mineral processing process. Geological pressure sensors are installed at geological monitoring points to collect parameters such as surrounding rock stress, pore water pressure, and formation deformation, in order to characterize the stability of the geological structure and the stress distribution. Various sensors are configured with sampling frequencies according to sampling requirements. For example, the sampling frequency of a vibration sensor can be set to an example value of 500Hz to capture high-frequency mechanical vibration signals; the sampling frequency of a temperature sensor can be set to an example value of 1Hz to continuously monitor temperature change trends; the sampling frequency of a flow meter can be set to an example value of 10Hz to track medium flow fluctuations; and the sampling frequency of a geological pressure sensor can be set to an example value of 5Hz to record dynamic changes in ground stress. The sampling signals of all sensors are uniformly connected to the industrial Ethernet bus through the field data acquisition terminal, and the sampling clock is uniformly calibrated through a time synchronization protocol. After sampling, the monitoring signals output by each sensor are encapsulated into multi-source data with timestamps and device numbers as indexes.

[0030] Based on multi-source data, the timestamps of different device acquisition channels are synchronized using the PTP protocol, and missing data points are filled in by bidirectional interpolation. Unit standardization conversion is performed to output standardized multi-source data.

[0031] Specifically, time-series signals from vibration sensors, temperature sensors, flow meters, and geological pressure sensors are read, and the original timestamps of each sampled signal are extracted. The time reference of each acquisition channel is uniformly calibrated using the PTP protocol. Time axis resampling processing is performed on the multi-source time series calibrated by the PTP protocol, and missing data points in the sampling interval are detected. Interpolation coefficients are calculated based on the time interval between known data points before and after the missing point. Trend correction values ​​are calculated based on the signal change rate of adjacent times. Interpolation values ​​are generated by averaging before and after the missing points. Physical quantities of different dimensions (such as temperature, flow rate, vibration amplitude, and geological pressure) are mapped to dimensionless intervals and standardized using the minimum-maximum normalization formula to obtain standardized multi-source data.

[0032] Standardized multi-source data is organized according to sampling area number, equipment number and sampling time to establish a three-dimensional tensor structure. The sliding time window mechanism is used to slice and stack the three-dimensional tensor structure within a continuous sampling period to form a time-series tensor block.

[0033] Specifically, a table of correspondence between mining areas and equipment is established using the mining area number as a spatial hierarchical index, and the equipment number is used as the next-level index. Standardized multi-source data of each equipment in the same mining area are arranged in the order of sampling time to form a structured multi-dimensional time series set. Within each mining area, a three-dimensional tensor structure is constructed based on the spatial coordinates and timestamp information of the equipment. The spatial dimension is used to describe the geographical distribution relationship between different mining areas and equipment, the temporal dimension is used to describe the continuity of sampling time, and the feature dimension is used to describe the multi-source monitoring features such as equipment operating parameters, process flow rate, geological pressure, and temperature. A sliding time window mechanism is used to slice and stack the three-dimensional tensor structure within a continuous sampling period. The sliding time window mechanism continuously slides on the time axis by setting a fixed-length time window (e.g., 30 minutes in the example) and an overlap step size (e.g., 5 minutes in the example). Each slide extracts tensor slices within the window and stacks them in chronological order to form a temporal tensor block.

[0034] The temporal tensor block is input into the encoding layer of the multi-head Transformer network to calculate the correlation weights between spatial nodes and time, and to perform multi-dimensional feature fusion and compression to output a deep semantic feature matrix. The deep semantic feature matrix is ​​then mapped to a tensor embedding vector.

[0035] Specifically, the dimensions of the temporal tensor block are adjusted from (number of spatial nodes, number of time steps, feature dimension) to (product of number of spatial nodes and number of time steps, ...). The feature dimension is formed by merging the time dimension and the spatial node dimension to form a serialized structure. The multi-dimensional feature vector of each spatial node is mapped to a unified high-dimensional feature space through the input projection layer of the multi-head Transformer network. The input projection layer is a learnable linear transformation matrix that maps the dimension of each feature vector from the original feature dimension to the hidden dimension specified by the encoding layer of the multi-head Transformer network. The multi-head self-attention mechanism calculates the correlation weights between all spatial nodes and time slices to obtain a weighted feature matrix. The weighted feature matrix is ​​then concatenated along the feature channel dimension, and the multi-head output is fused through linear transformation. The fused feature matrix is ​​input to the feedforward neural network layer. In the Transformer model, the feedforward neural network layer is an important component of each sub-layer of the encoder and decoder. After the self-attention mechanism, its main function is to perform further nonlinear transformation, feature enhancement, and dimension mapping on the feature representation extracted and fused by the self-attention layer. In the feedforward process, feature dimension compression and deep feature extraction are achieved through linear transformation and nonlinear activation functions to obtain a deep semantic feature matrix. The deep semantic feature matrix is ​​then mapped into a tensor embedding vector through a linear mapping layer.

[0036] It should be noted that the training of the multi-head Transformer network is based on historical production sample data, with time-series tensor blocks as input and historical production status classification results or key process indicator predictions as supervision signals. The network parameters are updated by minimizing the cross-entropy loss function or the mean square error loss function. The Adam optimizer is used, and a dynamic learning rate is set to balance convergence speed and stability. The multi-head Transformer network is trained alternately on samples from multiple mining areas to obtain the trained multi-head Transformer network.

[0037] The analysis module receives tensor embedding vectors, constructs a dynamic graph based on the topological relationships of mining equipment and the mineral processing flow, and uses a temporal graph convolution algorithm to perform feature propagation and neighborhood information aggregation on the node states, outputting a node representation matrix.

[0038] The analysis module includes a dynamic graph construction unit, a temporal convolution unit, and a node update unit; The dynamic graph construction unit establishes a dynamic adjacency matrix based on the spatial correlation in the tensor embedding vector and the process flow to obtain a dynamic graph. The temporal convolutional unit performs temporal feature propagation and spatial convolution calculation based on the dynamic graph, extracts node association features across time slices, and obtains a temporal feature matrix; The node update unit is based on the temporal feature matrix. It fuses the node's neighborhood features and time weights through a feature aggregation function, updates the node's state vector, and outputs the node representation matrix.

[0039] Equipment operating parameters, process flow characteristics, and geological coupling indicators are extracted from tensor embedding vectors. A dynamic adjacency matrix is ​​constructed based on the spatial correlation and process dependence between features. Mining equipment is represented by nodes, and the physical connection strength and temporal correlation are represented by edge weights to form a dynamic map.

[0040] Specifically, based on the structural dimension of the tensor embedding vector, the feature vectors containing equipment operating status, mineral processing parameters and geological environmental variables are grouped by equipment number, and the feature sequence of each equipment node in different time slices is extracted. By calculating the spatial correlation and temporal variation trend among different features in the tensor embedding vectors, a feature spatial similarity matrix is ​​obtained by extracting feature vectors representing different spatial nodes and calculating cosine similarity. The temporal variation trend is calculated by forming a time series from the tensor embedding vectors of each node along the time dimension and calculating the Pearson correlation coefficient between the time series of different nodes to obtain the synchronicity matrix of the temporal variation trend. This determines the dependencies between equipment nodes. Spatial correlation is obtained by calculating the Euclidean distance matrix between equipment in the three-dimensional coordinate system and the geological attribute similarity matrix. The upstream and downstream connection information in the process flow topology table is then used to correlate with equipment operation. The correlation coefficient matrix of parameters is comprehensively calculated to obtain the process dependence (such as the feeding dependence between the crusher and the ball mill, and the slurry conveying dependence between the ball mill and the classifier). Based on the spatial correlation matrix and the process dependence matrix, a weighted fusion strategy is used to generate a dynamic adjacency matrix. The physical connection strength is determined according to the connection of process pipelines, conveyor belts or material channels between equipment. The signal correlation is calculated based on the time series Pearson correlation coefficient of the operating parameters. The time correlation is calculated by measuring the similarity of the rate of change of features between nodes within a sliding window. With equipment nodes as vertices and edge weights as connection weights, a dynamic graph is formed on the basis of the dynamic adjacency matrix.

[0041] It should be noted that the process flow topology table is derived from the original design data of mine production, such as piping and instrumentation diagrams, process flow diagrams and equipment layout diagrams, which describe the physical connections between equipment, material flow direction and process sequence.

[0042] The dynamic graph is input into a temporal convolutional network to capture the collaborative change patterns between devices in different time slices. The spatial topological features and temporal evolution features are integrated through a residual fusion structure to output a temporal feature matrix.

[0043] Specifically, the topological relationships and spatial correlation features of equipment nodes, process nodes, and geological nodes are extracted from the dynamic graph generated by the dynamic graph construction unit. Based on the temporal convolution unit, a time sliding window mechanism is used to sample the node states in the dynamic graph at multiple time steps, for example, the window size is set to 5 time steps, forming a node temporal feature sequence. Through the spatiotemporal graph convolution algorithm, such as combining graph convolution operations and one-dimensional convolution in the time dimension, convolution calculation is performed on the node temporal feature sequence to fuse the node state changes in the time dimension and the neighborhood dependencies in the spatial dimension. Feature normalization processing is performed on the convolution output, such as using layer normalization methods, and nonlinear activation processing, such as using the ReLU activation function, to generate a temporal feature matrix.

[0044] It should be noted that the spatiotemporal graph convolution algorithm is a deep learning algorithm used to process graph-structured time series data. Its core is to capture the topological relationships between nodes in the spatial dimension through graph convolution, and at the same time capture the dynamic evolution pattern in the time dimension through temporal convolution, so as to achieve collaborative learning of spatiotemporal features.

[0045] The temporal feature matrix is ​​subjected to node feature aggregation processing. The neighborhood set of each node is obtained based on the adjacency matrix of the dynamic graph. The weighted average coefficient is calculated based on the edge weight and the feature similarity between nodes. The temporal features of the adjacent nodes are aggregated to obtain the node feature vector set.

[0046] Specifically, based on the adjacency matrix of the dynamic graph, the list of directly connected nodes of each node is extracted to form a neighborhood set; the feature similarity between a node and its neighboring nodes is calculated, for example, by comparing the feature vectors of corresponding nodes in the temporal feature matrix using the cosine similarity algorithm to obtain a similarity value, and combined with the edge weights in the adjacency matrix, the weighted average coefficient between each node and its neighboring nodes is calculated using the weighted average formula; the temporal feature vectors of the neighboring nodes are linearly weighted and summed using the weighted average coefficient to aggregate and generate the enhanced feature vector of each node, thus obtaining the node feature vector set.

[0047] It should be noted that the expression for calculating the weighted average coefficient is as follows: ; in, It is the weighted average coefficient, representing the coefficient from the neighboring nodes. Aggregate to central node The weighting ratio Indicates from neighboring nodes Aggregate to central node The edge weights, Indicates from neighboring nodes Aggregate to central node Feature vector similarity, Indicates the current central node being processed. Indicates the central node A certain neighboring node, This indicates the index variable used in the summation symbol. Indicates the central node The neighborhood set, Indicates the central node To the node The edge weights, Indicates the central node To the node Feature vector similarity.

[0048] The node feature vector set is input into the gated recurrent unit to dynamically model the temporal variation of node features. By fusing historical states and current input features through update gate and reset gate mechanisms, the temporal evolution pattern of nodes at different times is extracted, and the node update feature matrix is ​​output.

[0049] Specifically, the node feature vector set is input into the gated recurrent unit (GRU). The GRU calculates the combination of the current input features and the previous hidden state through a reset gate. The reset gate uses the sigmoid activation function to generate a reset gate value. For example, the sigmoid function maps the input to between 0 and 1. The update gate controls the fusion ratio between the current state and the historical state. The update gate also uses the sigmoid activation function to generate an update gate value, which is used to adjust the degree of retention of historical hidden states. Candidate hidden states are calculated by combining the product of the reset gate value and the previous hidden state with the current input features and generating candidate values ​​through the tanh activation function. Finally, the hidden states are updated by weighting the update gate value with the historical hidden states and then complementary weighting it with the candidate hidden states to generate new hidden states. This process extracts the temporal evolution pattern of node features at different time points and outputs the node update feature matrix.

[0050] The node update feature matrix is ​​normalized and fused with residuals. The original node features are then superimposed with the node update feature matrix through a residual connection structure to obtain the node representation matrix.

[0051] Specifically, the eigenvalues ​​of each dimension of the node update feature matrix are normalized. By calculating the mean and standard deviation of each feature dimension, the eigenvalues ​​are mapped to a unified numerical range, resulting in a normalized node update feature matrix. This normalized node update feature matrix is ​​then aligned and matched with the original node feature matrix output from the previous stage in terms of dimensions. The original node feature matrix and the node update feature matrix are then weighted and superimposed element-wise using a residual connection structure. Simultaneously, the patterns reflecting dynamic changes over time in the node update feature matrix are superimposed, so that the node feature representation contains both historical state information and temporal evolution characteristics, resulting in a fused feature matrix. The fused feature matrix is ​​then transformed using a nonlinear activation function to output the node representation matrix.

[0052] It should be noted that residual connection structures are used to solve the problems of gradient vanishing / exploding and network degradation in the training of deep neural networks. The core is to allow the original input information to "skip" one or more layers of the network and pass it to a deeper network without any transformation or with only a few transformations.

[0053] The prediction and optimization module receives the node representation matrix, calculates the node health score and failure probability through a prediction network based on a graph attention mechanism, and generates process parameter optimization instructions.

[0054] The prediction and optimization module includes a deep prediction unit, a risk identification unit, and a reinforcement learning decision-making unit; The deep prediction unit performs health scoring and potential failure probability prediction on the node representation matrix based on the graph attention mechanism to obtain the node health score and failure probability. The risk identification unit generates multi-level risk labels based on the dynamic changing trends of node health scores and failure probabilities, and forms a risk vector. The reinforcement learning decision-making unit establishes a multi-objective reward function based on the risk vector and production status data, combined with a deep reinforcement learning algorithm, performs strategy optimization calculations on the risk vector, generates process parameter optimization instructions, and outputs the process parameter optimization instructions to the visualization module.

[0055] Based on the node representation matrix, a graph attention neural network model is used to calculate the health score and failure probability of each node, generating a node health score sequence.

[0056] Specifically, the feature vector of each node in the node representation matrix is ​​input into the input layer of the graph attention neural network model. The feature correlation weight between the node and its neighboring nodes is calculated through a multi-head attention mechanism. The graph attention neural network model takes the neighborhood set of the node as the input range and performs weighted aggregation operation on the features of adjacent nodes. The aggregated node features are transformed by a non-linear activation function and then enter the fully connected layer. The health score value is output through the mapping parameters obtained by training. The graph attention neural network model calculates the prediction residual and the trend of time step change based on the health score. It performs difference analysis between the node's historical health score and the current prediction result, and normalizes the failure risk through the Sigmoid function to obtain a failure probability value between 0 and 1. The node health scores are arranged in time series order to form a node health score sequence.

[0057] It should be noted that the training of the graph attention neural network model is based on supervised learning using historical production operation data. During the training phase, the node representation matrix is ​​used as input, and the actual health status labels of the equipment and known fault records are used as supervision signals. The prediction error is calculated using the mean square error function or the cross-entropy loss function. The parameters of the graph attention neural network model are iteratively updated using the Adam optimization algorithm to obtain the trained graph attention neural network model.

[0058] Cluster analysis is performed based on the time change rate and process deviation characteristics in the node health score sequence to classify node risk levels and generate multi-level risk labels; the multi-level risk labels are then fused with the node failure probability to output a risk vector.

[0059] Specifically, time difference calculation is performed on the node health score sequence to obtain the rate of change information between adjacent time steps, including first-order difference (i.e., instantaneous change rate, calculated as the difference in health scores between adjacent time steps) and second-order difference (i.e., change acceleration, calculated as the difference in adjacent first-order difference values, used to characterize the intensification or deceleration of the change trend). At the same time, process deviation features are extracted from the process parameters associated with the node representation matrix, including indicators such as mineral processing flow deviation rate, energy consumption deviation rate, and temperature change amplitude, and these, together with the time change rate features, form a node state feature set. K-means algorithm or density-based DBSCAN algorithm is used to perform cluster analysis on the nodes. Based on the mean of the health score and the variance of the change trend of the cluster centers, risk levels are divided to generate multi-level risk labels. The multi-level risk labels corresponding to each node are fused with the node failure probability. Through weighted linear combination or principal component analysis, the time change, process deviation, and failure probability information are fused in the feature space to obtain the risk vector.

[0060] Production status data is collected, and the risk vector and production status data are input into a deep reinforcement learning model. With equipment operating efficiency, energy consumption level and safety risk as state variables, a multi-objective reward function is established and the strategy is iteratively trained to obtain the optimal process parameter adjustment scheme.

[0061] Specifically, production status data is collected in real time from key monitoring points in the mining production process. The production status data includes multi-dimensional operating characteristics such as equipment operating efficiency parameters, energy consumption level indicators, ore beneficiation yield parameters, and safety risk indicators. After time alignment and normalization, the risk vector and production status data are jointly input into the state input layer of the deep reinforcement learning model. In the deep reinforcement learning model, equipment operating efficiency, energy consumption level and safety risk are used as core state variables to establish a multi-objective reward function. The deep reinforcement learning model optimizes the process parameter adjustment scheme in the action space through multiple rounds of policy iteration. In each iteration, the policy network weights are updated according to the feedback of the reward function, gradually converging to the global optimal policy. The optimal process parameter adjustment scheme is obtained through the policy output layer.

[0062] It should be noted that the deep reinforcement learning model adopts a training mechanism based on the Actor-Critic structure. The Actor network is responsible for outputting process parameter adjustment actions, and the Critic network is responsible for evaluating the value function of the current action. During the training phase, historical production state data and risk vectors are used as training samples, and the goal is to maximize the cumulative reward. The network parameters are iteratively optimized through the gradient descent algorithm to obtain the trained deep reinforcement learning model.

[0063] The optimal process parameter adjustment scheme is converted into instructions and subjected to safety constraints to generate process parameter optimization instructions, which are then output to the visualization module.

[0064] Specifically, based on the optimal process parameter adjustment scheme, continuous optimization variables (such as equipment speed adjustment range, mineral processing flow distribution ratio, energy consumption control threshold, etc.) are discretized and encoded according to the control instruction format supported by mine automation control, and the discretized instruction parameters are obtained. Safety constraint processing is performed on the discretized instruction parameters. By constraining and verifying the adjustment scheme, including standard constraints such as equipment operating safety threshold, mineral processing boundary conditions, geological stress stability range, and energy consumption limit range, the process parameter adjustment scheme after safety constraint verification is numerically mapped to an instruction set that conforms to the execution standard, generating process parameter optimization instructions, and outputting the process parameter optimization instructions to the visualization module in DDS message format.

[0065] The visualization module is used to display and dynamically simulate process parameter optimization commands in three dimensions.

[0066] The visualization module includes an instruction parsing unit, a 3D rendering unit, and a performance evaluation unit; The instruction parsing unit receives process parameter optimization instructions and performs semantic parsing and parameter mapping. The 3D rendering unit dynamically updates the equipment status, process path, and geological environment elements in the mine virtual space based on process parameter optimization instructions; The performance evaluation unit synchronously compares the digital twin rendering results with the on-site monitoring data and generates dynamic performance evaluation results.

[0067] The process parameter optimization instructions are input into the digital twin rendering engine. Based on the three-dimensional spatial structure model of the mine and the topology relationship of the equipment, the virtual equipment operating status, energy consumption curves and process flow paths are updated to generate dynamic visualization screens.

[0068] Specifically, the digital twin rendering engine dynamically updates the operating status of the corresponding equipment based on the values ​​such as equipment rotation speed, ore dressing flow rate and energy consumption control threshold contained in the process parameter optimization instructions. The update process includes the synchronous adjustment of multi-dimensional parameters such as equipment rotation speed, conveyor belt carrying rate, slurry flow path and geological pressure distribution. The digital twin rendering engine performs linked calculations on the associated device nodes based on the device topology, calculates the trend of energy consumption curve changes and the trajectory of process flow changes, and generates a dynamic dataset reflecting the operating status under the current optimization strategy. The digital twin rendering engine redraws the device operation screen in virtual 3D space and generates dynamic visualization screens by combining lighting mapping and physical simulation effects.

[0069] It should be noted that the digital twin rendering engine is the core visualization component of the digital twin device. It is a complex software whose core task is to deeply integrate the real-time data and business logic of the physical entity with 3D computer graphics technology to generate an immersive, interactive, and dynamically evolving virtual scene that evolves in sync with the physical world.

[0070] The virtual simulation images are compared synchronously with the actual production monitoring data. The deviation detection algorithm is used to calculate the energy efficiency improvement rate and risk change trend after the process parameters are adjusted, and a dynamic performance evaluation result is formed.

[0071] Specifically, based on the timestamp alignment mechanism, the equipment operation data, energy consumption curves, and process flow paths in the virtual simulation screen generated by the digital twin rendering engine are synchronized in time and matched with the production monitoring data collected in real time by the on-site monitoring equipment. A one-to-one correspondence is established through equipment number, mining area number, and process node identifier, so that the virtual simulation data and the actual production monitoring data are consistent in spatial coordinates and time dimensions. The deviation detection algorithm is used to compare and analyze the key parameters (including energy consumption value, equipment speed, ore dressing flow rate, and slurry concentration, etc.) in the virtual simulation data and production monitoring data point by point. The change range of each parameter before and after the process adjustment is calculated, and the deviation distribution is statistically analyzed. The dynamic difference of time series data is evaluated based on the sliding window mechanism. The energy efficiency improvement rate after the process parameter adjustment is calculated by the difference rate function (a mathematical method used to quantify the degree or magnitude of difference between two sets of data, including absolute difference rate and relative difference rate). Combined with the risk level weight in the risk vector, the risk change trend in the production process after adjustment is analyzed. The results of energy efficiency improvement rate and risk change trend are summarized to form dynamic performance evaluation results.

[0072] It should be noted that the deviation detection algorithm is a methodology whose core objective is to quantify and evaluate the consistency between the "virtual simulation prediction value" and the "physical world measured value", thereby determining the effectiveness of the optimization strategy and the health status of the equipment operation.

[0073] This embodiment also provides a method for intelligent analysis of big data in metal mine production, including: Real-time acquisition, preprocessing, and spatiotemporal feature modeling of multi-source heterogeneous information from metal mine production sites are performed, and tensor embedding vectors are output. A dynamic graph is constructed based on the topological relationship of mining equipment and the mineral processing flow. The temporal graph convolution algorithm is used to perform feature propagation and neighborhood information aggregation on the node state, and outputs a node representation matrix. The prediction network based on graph attention mechanism calculates node health scores and failure probabilities, and generates process parameter optimization instructions. The process parameter optimization instructions are displayed in three dimensions and verified by dynamic simulation.

[0074] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the intelligent analysis method for big data in metal mine production as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0075] In summary, this invention, by constructing a dynamic graph and combining it with temporal convolution and node update mechanisms, accurately captures the spatiotemporal correlation and state evolution patterns between equipment, improves the accuracy of equipment behavior representation under complex working conditions, transforms risk identification results into executable process adjustment strategies, and achieves a leap from passive early warning to proactive optimization, significantly enhancing the safety, efficiency, and intelligence level of mine operations.

[0076] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A smart big data analysis device for metal mine production, characterized in that: It includes a data acquisition and modeling module, an analysis module, a prediction and optimization module, and a visualization module; The data acquisition and modeling module includes a data acquisition unit, a data preprocessing unit, and a feature extraction unit; The data acquisition unit includes a vibration sensor, a temperature sensor, and a geological pressure sensor, which are used to collect mining equipment operating parameters, mineral processing indicators, and geological environment parameters in real time to obtain multi-source data. The data preprocessing unit is used to perform timestamp alignment, missing value compensation, noise removal and unit normalization on multi-source data to obtain standardized multi-source data. The feature extraction unit employs a multi-head Transformer network to perform spatiotemporal modeling and semantic feature fusion on standardized multi-source data, and outputs tensor embedding vectors.

2. The intelligent analysis equipment for big data in metal mine production as described in claim 1, characterized in that: The analysis module includes a dynamic graph construction unit, a temporal convolution unit, and a node update unit; The dynamic graph construction unit establishes a dynamic adjacency matrix based on the spatial correlation in the tensor embedding vector and the process flow to obtain a dynamic graph. The temporal convolutional unit performs temporal feature propagation and spatial convolution calculation based on the dynamic graph, extracts node association features across time slices, and obtains a temporal feature matrix; The node update unit is based on the temporal feature matrix. It fuses the node's neighborhood features and time weights through a feature aggregation function, updates the node's state vector, and outputs the node representation matrix.

3. The intelligent analysis equipment for big data in metal mine production as described in claim 1, characterized in that: The prediction and optimization module includes a deep prediction unit, a risk identification unit, and a reinforcement learning decision-making unit. The deep prediction unit performs health scoring and potential failure probability prediction on the node representation matrix based on the graph attention mechanism to obtain the node health score and failure probability. The risk identification unit generates multi-level risk labels based on the dynamic changing trends of node health scores and failure probabilities, and forms a risk vector. The reinforcement learning decision-making unit establishes a multi-objective reward function based on the risk vector and production status data, combined with a deep reinforcement learning algorithm, performs strategy optimization calculations on the risk vector, generates process parameter optimization instructions, and outputs the process parameter optimization instructions to the visualization module.

4. The intelligent analysis equipment for big data in metal mine production as described in claim 1, characterized in that: The specific steps to obtain standardized multi-source data are as follows. Real-time sampling is performed using vibration sensors, temperature sensors, flow meters, and geological pressure sensors deployed at mining site equipment and geological monitoring points to generate multi-source data. Based on multi-source data, the timestamps of different device acquisition channels are synchronized using the PTP protocol, and missing data points are filled in by bidirectional interpolation. Unit standardization conversion is performed to output standardized multi-source data.

5. The intelligent analysis equipment for big data in metal mine production as described in claim 1, characterized in that: The specific steps for embedding the output tensor vector are as follows. Standardized multi-source data is organized according to the sampling area number, equipment number and sampling time to establish a three-dimensional tensor structure. The sliding time window mechanism is used to slice and stack the three-dimensional tensor structure within a continuous sampling period to form a temporal tensor block. The temporal tensor block is input into the encoding layer of the multi-head Transformer network to calculate the correlation weights between spatial nodes and temporal nodes. Multidimensional feature fusion and compression are then performed to output a deep semantic feature matrix, which is then mapped to a tensor embedding vector.

6. The intelligent analysis equipment for big data in metal mine production as described in claim 2, characterized in that: The output node represents a matrix, and the specific steps are as follows. Equipment operating parameters, process flow characteristics and geological coupling indicators are extracted from tensor embedding vectors. A dynamic adjacency matrix is ​​constructed based on the spatial correlation and process dependence between features. Nodes represent mining equipment and edge weights characterize the physical connection strength and temporal correlation to form a dynamic graph. The dynamic graph is input into a temporal convolutional network to capture the cooperative change patterns between devices in different time slices. The spatial topological features and temporal evolution features are integrated through a residual fusion structure to output a temporal feature matrix. The temporal feature matrix is ​​subjected to node feature aggregation processing. The neighborhood set of each node is obtained based on the adjacency matrix of the dynamic graph. The weighted average coefficient is calculated based on the edge weight and the feature similarity between nodes. The temporal features of the adjacent nodes are aggregated to obtain the node feature vector set. The node feature vector set is input into the gated recurrent unit to dynamically model the temporal variation of node features. The historical state and current input features are fused through the update gate and reset gate mechanism to extract the temporal evolution pattern of the node at different times and output the node update feature matrix. The node update feature matrix is ​​normalized and fused with residuals. The original node features are then superimposed with the node update feature matrix through a residual connection structure to obtain the node representation matrix.

7. The intelligent analysis equipment for big data in metal mine production as described in claim 3, characterized in that: The specific steps for generating the process parameter optimization instructions are as follows: Based on the node representation matrix, the health score and failure probability of each node are calculated using a graph attention neural network model to generate a node health score sequence. Cluster analysis is performed based on the time change rate and process deviation characteristics in the node health score sequence to classify node risk levels and generate multi-level risk labels; Multi-level risk labels and node failure probabilities are fused to output a risk vector. Collect production status data, input the risk vector and production status data into a deep reinforcement learning model, use equipment operating efficiency, energy consumption level and safety risk as state variables, establish a multi-objective reward function and conduct strategy iterative training to obtain the optimal process parameter adjustment scheme; The optimal process parameter adjustment scheme is converted into instructions and subjected to safety constraints to generate process parameter optimization instructions, which are then output to the visualization module.

8. The intelligent analysis equipment for big data in metal mine production as described in claim 7, characterized in that: The visualization module is used to perform three-dimensional visualization and dynamic simulation verification of process parameter optimization commands. The specific steps are as follows. Input the process parameter optimization instructions into the digital twin rendering engine, and based on the three-dimensional spatial structure model of the mine and the topology relationship of the equipment, update the virtual equipment operating status, energy consumption curve and process flow path to generate dynamic visualization. The virtual simulation images are compared synchronously with the actual production monitoring data. The deviation detection algorithm is used to calculate the energy efficiency improvement rate and risk change trend after the process parameters are adjusted, and a dynamic performance evaluation result is formed.

9. A method for intelligent analysis of big data in metal mine production, based on the intelligent analysis equipment for big data in metal mine production as described in any one of claims 1 to 8, characterized in that: include, Real-time acquisition, preprocessing, and spatiotemporal feature modeling of multi-source heterogeneous information from metal mine production sites are performed, and tensor embedding vectors are output. A dynamic graph is constructed based on the topological relationship of mining equipment and the mineral processing flow. The temporal graph convolution algorithm is used to perform feature propagation and neighborhood information aggregation on the node state, and outputs a node representation matrix. The prediction network based on graph attention mechanism calculates node health scores and failure probabilities, and generates process parameter optimization instructions. The process parameter optimization instructions are displayed in three dimensions and verified by dynamic simulation.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the intelligent analysis method for big data in metal mine production as described in claim 9.

Images