Mine environment data LED display method based on edge calculation

Abstract

The invention discloses a mine environment data LED display method based on edge calculation, and the method comprises the steps: 1, collecting multi-source environment data, and carrying out the preprocessing and index block coding; step 2, a gating fusion feature tensor is generated through an improved EdgeViT network; 3, executing risk reasoning and evaluation through a risk perception module, and outputting a risk evaluation result; 4, when a communication abnormal state is detected, risk prediction reasoning is executed based on the historical cache; 5, if the collaborative model is deployed, fusing the risk assessment results of the improved EdgeViT model and the collaborative model; sixthly, a visual control instruction is generated through an LED display control module, and risk information is dynamically displayed; and step 7, collecting feedback information, and performing incremental updating on the EdgeViT network and the risk sensing module based on the feedback information. According to the invention, the real-time performance of mine environment monitoring, the reliability of risk early warning and the intelligent response level of LED display are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial environment monitoring and edge intelligent control, and particularly relates to a mine environment data LED display method based on edge computing. BACKGROUND

[0002] With the increasing demand for intelligent construction of mines and safety supervision of underground operations, edge perception, risk assessment and visual warning control technology for multi-source environmental data have attracted widespread attention. Existing mine monitoring relies mainly on centralized deployment architecture, mainly through a central server to collect and analyze temperature, humidity, gas concentration and other indicators, and to control LED display terminals to display risk information in the form of static charts or fixed text. However, the following problems exist in the actual mine operating environment:

[0003] Environmental data has highly nonlinear and strong disturbance characteristics, and the sampling frequency and transmission delay of different types of sensors differ significantly, making it difficult to synchronize environmental data and accurately capture rapidly fluctuating abnormal trends. In the absence of efficient modeling capabilities at the edge node, traditional rule-driven or shallow models perform poorly in terms of risk identification accuracy, adaptability and robustness, and cannot effectively respond to dynamic changes in complex mine operating conditions. Existing LED display control strategies are mostly static coding, which cannot dynamically adjust color, frequency flashing and graphic content according to risk level, trend and confidence, resulting in poor readability and timeliness of warning information, making it difficult to meet the rapid response needs of high-risk environments. User feedback and operation log information cannot be used to update the model, causing model aging and response delay, and failing to achieve closed-loop autonomous optimization on the edge side.

[0004] Therefore, how to provide a mine environment data LED display method based on edge computing is a problem that those skilled in the art need to solve. SUMMARY

[0005] One object of the present application is to provide a mine environment data LED display method based on edge computing. The present application fully integrates edge computing, improved EdgeViT network, risk perception reasoning and LED intelligent control technology, and describes in detail the collection of multi-source environmental data, grouping modeling, dynamic channel gating, communication anomaly prediction reasoning, visualization control and feedback updating, which can efficiently identify and intelligently display mine environmental risk information, and has the advantages of high risk identification accuracy, strong display response timeliness and strong edge autonomy robustness.

[0006] According to the mine environment data LED display method based on edge computing of the present application, the following steps are included: Step 1: Collect multi-source environmental data and perform preprocessing and index grouping coding to generate a structured token embedding tensor; Step two: structured token embedding tensor is input into the improved EdgeViT network deployed on the edge node to perform group modeling, cross-group attention fusion and local dynamic channel gating operation, and generate a gated fusion feature tensor; Step three: the gated fusion feature tensor is input into the risk perception module deployed on the edge node to perform risk inference and evaluation, and output the risk evaluation result; Step four: when the edge node detects a communication abnormal state, perform risk prediction inference based on the historical cache structured token embedding tensor and the gated fusion feature tensor to generate a backup risk evaluation result; Step five: if the edge node is deployed with a collaborative model, the risk evaluation results of the improved EdgeViT model and the collaborative model are fused to generate a collaborative risk evaluation result; Step six: input the risk evaluation result, backup risk evaluation result or collaborative risk evaluation result into the LED display control module deployed on the edge node to generate visual control instructions and drive the LED display terminal to dynamically display risk information; Step seven: collect feedback information and perform incremental update of the EdgeViT network and the risk perception module based on the feedback information.

[0007] Optionally, the step one specifically includes: Collect multi-source environmental data through environmental sensors deployed in the mine operation area, including temperature, humidity, gas concentration, dust density, carbon monoxide content and wind speed; Preprocess the multi-source environmental data, specifically: Extract the timestamp information of the multi-source environmental data, and map the multi-source environmental data to a unified time axis based on a set time step to form a standard time series; Complete the missing items in the standard time series by linear interpolation, and remove the outliers in the standard time series using a local statistical threshold method based on a sliding window to form a standard cleaned sequence; Map the original values in the standard cleaned sequence to the interval [0,1] by the min-max normalization method to form a standardized data set; the standardized data set is a two-dimensional matrix, each column corresponds to a normalized numerical sequence of an environmental index category, and each row represents multi-dimensional environmental data at a unified time step; At each time step, divide the standardized data set into multiple channels according to the environmental index category, and convert the normalized numerical sequence of each channel to an index token of a set dimension by embedding mapping function to perform embedding mapping operation, and all index tokens constitute a structured token embedding sequence at the current time step; The structured token embedding sequences of all time steps are stacked to form a structured token embedding tensor.

[0008] Optionally, the step two specifically comprises: The structured token embedding tensor is input into an improved EdgeViT network deployed on the edge node, and the dimensions of the structured token embedding tensor include a time step dimension, an environmental indicator category dimension, and an embedding dimension of each token. The structured token embedding tensor is grouped and modeled in the environmental indicator category dimension, specifically: According to the similarity of physical properties, temperature and humidity are divided into a hot and humid environment group, gas concentration and carbon monoxide concentration are divided into a harmful gas group, dust density is divided into a particle aerosol group, and wind speed is divided into an air flow ventilation group. The token embedding sequences of each group are formed into independent token grouping sequences to form a set of four indicator groupings. The four sets of indicator groupings are respectively subjected to intra-group self-attention operation to extract intra-group dependent features and time sequence features, and four sets of self-attention indicator tensors are generated. The four sets of self-attention indicator tensors are respectively generated into four sets of intra-group representative vectors through weighted pooling. The four sets of intra-group representative vectors are stacked into a group indicator feature tensor, and normalized dot product attention operation is performed on the group indicator feature tensor to generate an inter-group attention matrix. Based on the inter-group attention matrix, weighted fusion is performed on the group indicator feature tensor to generate a group fusion feature tensor, and the group fusion feature matrix is subjected to dimension compression and nonlinear transformation through an MLP structure to generate a group fusion perception tensor. The fusion perception vectors of each indicator grouping in the group fusion perception tensor are respectively converted into channel fusion perception vectors of each environmental indicator category through channel mapping according to the environmental indicator categories contained therein, and are spliced into a channel fusion perception tensor according to the order of the environmental indicator categories. The channel fusion perception tensor is divided into several time segments according to the time steps, and the local volatility score of each environmental indicator category is calculated using the average absolute change rate within a set sliding window at each time step. The local volatility score of each environmental indicator category is attributed to the indicator grouping to which it belongs, and the mean and standard deviation of the local volatility scores of all environmental indicator categories within the group are calculated at each time step. Based on the local volatility score and the corresponding mean and standard deviation, the channel activation weight of each environmental indicator category at each time step is calculated. The channel fusion perception tensor is based on the channel fusion perception vector of each environment index category at the current time step, and the corresponding channel activation weight is multiplied to generate a weighted channel perception vector; The weighted channel perception vectors of all time steps are stacked in time sequence to form a gated fusion feature tensor.

[0009] Optionally, the step three specifically includes: The gated fusion feature tensor is subjected to average pooling and maximum pooling along the time step dimension, and is spliced to generate a time-aggregated feature vector; The gated fusion feature tensor is subjected to average pooling and maximum pooling along the environment index category dimension, and is spliced to generate an index-aggregated feature vector; The gated fusion feature tensor is respectively unfolded into a two-dimensional feature matrix along the time step dimension and the environment index category dimension, and is respectively input into independent linear mapping networks for projection compression to generate two projection vectors, and is jointly modeled through an MLP structure to generate a channel cross feature vector; The time-aggregated feature vector, the index-aggregated feature vector, and the channel cross feature vector are spliced to form a multi-scale fusion perception vector; A risk level guide vector is introduced, which is a risk level label set in a historical time window converted into a risk level embedding vector sequence through an embedding mapping function, and fused by weighted average; The multi-scale fusion perception vector and the risk level guide vector are fused by residual weighting to obtain a risk fusion vector; The risk fusion vector is subjected to risk reasoning and evaluation operations to output a risk evaluation result, which includes a risk level label and a classification result, a trend score vector, and a confidence score, specifically: The risk fusion vector is mapped to a three-dimensional risk probability vector through a fully connected layer and a Softmax function, and the index of the maximum probability value in the three-dimensional risk probability vector is obtained through an argmax function as the risk level label of the current time step; The risk level label and the classification result include 0: normal level, 1: early warning level, and 2: high-risk level; The risk fusion vector is subjected to linear transformation and activation operations through a fully connected layer and a Sigmoid function to generate a trend score vector, and the trend score vector has a dimension of K, indicating the risk level fluctuation trend in the next K time steps after the current time step; The risk fusion vector is output through a fully connected layer and a HardSigmoid activation function to output a confidence score, which is a real number in the interval [0, 1], indicating the confidence strength of the current risk level label.

[0010] Optionally, the step four specifically comprises: The communication abnormal state comprises network connection disconnection, data upload failure, no periodic communication confirmation signal received, or communication reconnection failure; Loading the structured token embedding tensor cached in the setting history time window and the gated fusion feature tensor; Inputting the historically cached structured token embedding tensor into the improved EdgeViT network to perform group modeling, cross-group attention fusion, and local dynamic channel gating operation, and generating the autonomous gated fusion feature tensor; Inputting the autonomous gated fusion feature tensor into the risk perception module to perform multi-scale fusion and residual weighting, and generating the autonomous risk fusion vector; Performing risk reasoning and evaluation on the autonomous risk fusion vector, and outputting the backup risk evaluation result, which comprises a risk level label, a trend score vector, and a confidence score.

[0011] Optionally, the step five specifically comprises: The improved EdgeViT model comprises an improved EdgeViT network and a risk perception module deployed on an edge node; Respectively extracting the confidence scores in the risk evaluation results of the improved EdgeViT model and the collaborative model; Dividing the confidence scores of the two models by the sum of the confidence scores to obtain corresponding confidence weighting coefficients; Based on the confidence weighting coefficients, respectively fuse the risk level labels, trend score vectors, and confidence scores of the improved EdgeViT model and the collaborative model to generate collaborative risk level labels, collaborative trend score vectors, and collaborative confidence scores, and constitute the collaborative risk evaluation result.

[0012] Optionally, the step six specifically comprises: The visualization control instruction comprises color coding control, stroboscopic mode adjustment, brightness level setting, and graphical content arrangement; The color coding control is based on the risk level label of the current time step: when the risk level label is 0, the LED display color is set to green; when the risk level label is 1, the LED display color is set to yellow; when the risk level label is 2, the LED display color is set to red; The stroboscopic mode adjustment is based on the mean value of the trend score vector to determine the stroboscopic frequency: when the mean value of the trend score vector is less than 0.3, set to no flicker mode; when the mean value of the trend score vector is between 0.3 and 0.6, set to low-frequency flicker mode; when the mean value of the trend score vector is higher than 0.6, set to high-frequency flicker mode; The luminance level setting is segmented mapping based on the confidence score value: when the confidence score is less than 0.4, the luminance is set to a low level; when the confidence score is between 0.4 and 0.7, the luminance is set to a medium level; and when the confidence score is higher than 0.7, the luminance is set to a high level. The graphic content arrangement includes two parts of text prompts and graphic identifiers: according to the risk level label, the text prompt information "normal", "warning" or "high risk" is displayed; according to the change trend direction of the trend score vector, the arrow graphic identifier is dynamically displayed, the arrow upwards indicates that the risk is rising, the arrow downwards indicates that the risk is weakening, and the arrow horizontally indicates that the risk is stable.

[0013] Optionally, the feedback information includes user interaction feedback information of the LED display terminal and mine operation event log.

[0014] The beneficial effects of the present application are: Firstly, the present application lays a high-quality data foundation for risk perception by deploying environmental sensors in the mine operation area, collecting multi-source environmental data and constructing structured token embedding tensors combined with embedded mapping functions. Secondly, by using the improved EdgeViT network deployed in the edge node, the index grouping strategy based on physical property division, intra-group self-attention and cross-group attention fusion are used to realize the grouping modeling and dynamic channel gating of the mine multi-dimensional environmental data, and the expression ability and local sensitivity of feature extraction are improved. In the risk perception module, the time aggregation feature, the index aggregation feature and the channel cross feature are fused, and the historical risk level guide vector is combined to generate a risk fusion vector through residual weighting, which further improves the comprehensive discrimination performance of the risk level label, the trend score and the confidence score.

[0015] In addition, under the edge node communication abnormal state, the present application realizes autonomous risk prediction reasoning based on the structured token embedding tensor and the gated fusion feature tensor of the historical cache, which can ensure the sustainable operation of the LED display control in the network disconnection state; in the scene with collaborative model deployment, the local and collaborative model output results are integrated through the confidence weighted fusion mechanism, which effectively improves the robustness and adaptability of the risk assessment result. In the LED display control process, combined with the risk level label, the trend score and the confidence score, the visual control instructions such as color coding, frequency flashing mode, luminance level and graphic content are dynamically generated, which improves the readability and response efficiency of the mine risk state. The present application also collects the user interaction feedback information of the LED display terminal and the mine operation event log to incrementally update the EdgeViT network and the risk perception module, forming an edge autonomous feedback loop.

[0016] In conclusion, the application realizes the edge closed-loop control process of mine environment monitoring, risk inference and LED display, improves the accuracy of mine environment risk identification, the intelligence of visual display and the autonomy of edge deployment, and can provide a feasible technical path for real-time safety monitoring in complex mine scenes. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and together with the description serve to explain the application, and do not limit the application. In the drawings:

[0018] Figure 1 is a schematic diagram of a mine environment data LED display method based on edge computing proposed by the application; Figure 2 is an improved EdgeViT modeling flowchart in a mine environment data LED display method based on edge computing proposed by the application; Figure 3 is a risk perception and inference structure diagram in a mine environment data LED display method based on edge computing proposed by the application; Figure 4 is a network outage prediction inference flowchart in a mine environment data LED display method based on edge computing proposed by the application. DETAILED DESCRIPTION

[0019] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and therefore only show the components related to the application.

[0020] Reference Figures 1-4 A mine environment data LED display method based on edge computing, comprising the following steps: Step one: collect multi-source environment data, and perform preprocessing and index grouping coding to generate a structured token embedding tensor; Step two: input the structured token embedding tensor into the improved EdgeViT network deployed on the edge node, perform grouping modeling, cross-group attention fusion and local dynamic channel gate operation, and generate a gated fusion feature tensor; Step three: input the gated fusion feature tensor into the risk perception module deployed on the edge node to perform risk inference and evaluation, and output the risk evaluation result; Step four: when the edge node detects a communication abnormal state, perform risk prediction inference based on the historical cache structured token embedding tensor and the gated fusion feature tensor to generate a backup risk evaluation result; Step five: if the edge node is deployed with a collaborative model, the improved EdgeViT model is fused with the risk assessment result of the collaborative model to generate a collaborative risk assessment result; Step six: input the risk assessment result, backup risk assessment result or collaborative risk assessment result into the LED display control module deployed in the edge node to generate visual control instructions and drive the LED display terminal to dynamically display risk information; Step seven: collect feedback information and perform incremental update on the EdgeViT network and the risk perception module based on the feedback information.

[0021] In this embodiment, the step one specifically includes: Through the environmental sensor deployed in the mine operation area, multi-source environmental data is collected, including temperature, humidity, gas concentration, dust density, carbon monoxide content and wind speed; The multi-source environmental data is preprocessed, specifically: The timestamp information of the multi-source environmental data is extracted, and based on the set time step, the multi-source environmental data is mapped to a unified time axis to form a standard time sequence; The missing items in the standard time sequence are completed by linear interpolation, and the abnormal values in the standard time sequence are removed by using a local statistical threshold method based on a sliding window to form a standard cleaning sequence; The original values in the standard cleaning sequence are mapped to the interval [0, 1] by the minimum-maximum normalization method to form a standardized data set; the standardized data set is a two-dimensional matrix, each column corresponds to a normalized numerical sequence of an environmental index category, and each row represents multi-dimensional environmental data at a unified time step; At each time step, the standardized data set is divided into multiple channels according to the environmental index category, and the normalized numerical sequence of each channel is subjected to embedding mapping operation by embedding mapping function to convert into index token of a set dimension, and all index tokens constitute a structured token embedding sequence at the current time step; The structured token embedding sequences of all time steps are stacked and spliced to form a structured token embedding tensor.

[0022] In the present application, the structured token embedding tensor can enhance the modeling capability of the improved EdgeViT network for the correlation and evolution trend between different environmental indicators, facilitate the perception and discrimination of the downstream risk inference module for the time sequence pattern and local anomaly, and improve the accuracy of the overall risk identification and the real-time performance of the system response.

[0023] In this embodiment, the step two specifically includes: The structured token embedding tensor is input to an improved EdgeViT network deployed at an edge node, and the dimension of the structured token embedding tensor includes a time step dimension, an environmental indicator category dimension, and an embedding dimension of each token; The structured token embedding tensor is grouped and modeled in the environmental indicator category dimension, specifically: According to the similarity of the physical properties of the environmental indicators, temperature and humidity are divided into a thermal and humid environment group, gas concentration and carbon monoxide concentration are divided into a harmful gas group, dust density is divided into a particle aerosol group, and wind speed is divided into an air flow ventilation group; The token embedding sequence of each group is formed into an independent token grouping sequence to form a set of four indicator groups; In the present application, the structured token embedding tensor is grouped and modeled in the environmental indicator category dimension, and is divided into a thermal and humid environment group, a harmful gas group, a particle aerosol group, and an air flow ventilation group according to the similarity of the environmental physical properties, which can realize exclusive modeling paths for similar environmental factors without increasing the complexity of the model. This grouping method significantly enhances the expression consistency and semantic relevance between indicators in the group, and avoids the modeling interference caused by directly mixing environmental indicators with different physical properties.

[0024] Intra-group self-attention operations are respectively performed on the four sets of indicator grouping sets to extract intra-group dependent features and time sequence features, and four sets of self-attention indicator tensors are generated; Four sets of intra-group representative vectors are respectively generated by weighted pooling from the four sets of self-attention indicator tensors; The four sets of intra-group representative vectors are stacked into a group indicator feature tensor, and a normalized dot product attention operation is performed on the group indicator feature tensor to generate an inter-group attention matrix The dimension of the inter-group attention matrix is 4x4: ; Wherein, represents a group feature tensor, is a normalization function, is the dimension of the intra-group representative vector; Based on the inter-group attention matrix, a group fusion feature tensor is generated by performing weighted fusion on the group indicator feature tensor, and the group fusion feature matrix is compressed in dimension and transformed nonlinearly through an MLP structure to generate a group fusion perception tensor; In this invention, introducing self-attention within a group fully uncovers the temporal dependencies and implicit correlations within the indicator group. A normalized dot-product attention matrix is ​​constructed using an inter-group weighted fusion strategy to weight the representative vectors of each indicator group, effectively achieving the aggregation and dynamic adjustment of global information across groups. This approach enhances the flexibility and expressiveness of inter-group feature fusion, and strengthens the overall model's ability to model potential correlations between multi-dimensional environmental features.

[0025] The fusion perception vectors of each index group in the group fusion perception tensor are converted into channel fusion perception vectors of each environmental index category according to the environmental index categories they contain through channel mapping, and then concatenated into a channel fusion perception tensor according to the order of environmental index categories. The channel fusion sensing tensor is divided into several time segments according to time steps, and the local volatility score of each environmental indicator category is calculated at each time step using the average absolute rate of change within a set sliding window: ; in, Indicates environmental indicator categories At time step Local volatility score on The set sliding window length, Indicates environmental indicator categories At time step Normalized values ​​on Indicates environmental indicator categories At time step Normalized values ​​on; The local volatility score of each environmental indicator category is assigned to its respective indicator group, and the mean and standard deviation of the local volatility scores of all environmental indicator categories within the group are calculated at each time step. Based on the local volatility score and its corresponding mean and standard deviation, calculate the channel activation weight for each environmental indicator category at each time step: ; in, It is the Sigmoid activation function. This is an adjustable scaling factor. Indicates the grouping of environmental indicators At time step The mean of the local volatility score, Indicates the grouping of environmental indicators At time step The standard deviation of the local volatility score on the graph. This represents a small constant to prevent the denominator from being zero; Based on the channel fusion perception tensor, the channel fusion perception vector of each environmental indicator category at the current time step is multiplied by the corresponding channel activation weight to generate a weighted channel perception vector. The weighted channel sensing vectors of all time steps are stacked in chronological order to form a gated fusion feature tensor.

[0026] In this invention, by calculating the normalized rate of change of each environmental indicator within a set sliding window, channel activation weights are dynamically generated, and a weighted dot product operation is performed on the channel fusion sensing tensor, enabling specific attention to abnormal fluctuations or abrupt changes. This method can effectively filter out stable but non-risk-indicating redundant features, while enhancing the responsiveness to key risk factors, thereby improving the sensing accuracy and robustness of the improved EdgeViT network in detecting subtle abnormal changes and determining risk levels.

[0027] In this embodiment, step three specifically includes: The gated fusion feature tensor is subjected to average pooling and max pooling along the time step dimension, and then concatenated to generate a temporal aggregated feature vector; The gated fusion feature tensor is subjected to average pooling and max pooling along the environmental indicator category dimension, and then concatenated to generate an indicator aggregate feature vector. The gated fusion feature tensor is expanded into two-dimensional feature matrices along the time step dimension and the environmental indicator category dimension, respectively, and then input into independent linear mapping networks to perform projection compression, generating two projection vectors. These vectors are then jointly modeled using an MLP structure to generate channel cross feature vectors. The time-aggregated feature vector, the index-aggregated feature vector, and the channel cross feature vector are concatenated to form a multi-scale fusion perception vector; A risk level guidance vector is introduced, which is generated by converting the risk level labels of a set historical time window into a risk level embedding vector sequence through an embedding mapping function, and then fusing the risk level embedding vector sequence through a weighted average. The risk fusion vector is obtained by fusing the multi-scale fusion perception vector and the risk level guidance vector through residual weighting. Perform risk inference and assessment operations on the risk fusion vector, and output the risk assessment results. The risk assessment results include risk level labels and classification results, trend score vectors, and confidence scores, specifically: The risk fusion vector is mapped to a three-dimensional risk probability vector through a fully connected layer and a Softmax function. The index of the maximum probability value in the three-dimensional risk probability vector is obtained by the argmax function and used as the risk level label for the current time step. The risk level label and classification result include 0: normal level, 1: warning level and 2: high risk level. The risk fusion vector is linearly transformed and activated by a fully connected layer and a sigmoid function to generate a trend score vector. The trend score vector has a dimension of K, which represents the risk level fluctuation trend within K consecutive time steps after the current time step. The value range of each vector element is [0,1]. The larger the value, the higher the probability of the risk level rising. The risk fusion vector is passed through a fully connected layer and a HardSigmoid activation function to output a confidence score. The confidence score is a real number in the interval [0,1], which represents the credibility of the current risk level label. The closer the confidence score is to 1, the more credible the current predicted risk level is.

[0028] In this embodiment, step four specifically includes: The abnormal communication states include network connection loss, data upload failure, failure to receive periodic communication acknowledgment signals, or communication reconnection failure. Load the structured token embedding tensor cached within the set historical time window and the gated fusion feature tensor; The structured tokens of the historical cache are embedded into a tensor input to the improved EdgeViT network, and group modeling, cross-group attention fusion and local dynamic channel gating operations are performed to generate an autonomous gating fusion feature tensor. The autonomous gated fusion feature tensor is input into the risk perception module to perform multi-scale fusion and residual weighting, generating an autonomous risk fusion vector; Perform risk inference and assessment on the autonomous risk fusion vector, and output the backup risk assessment results, which include risk level labels, trend score vectors and confidence scores.

[0029] In this invention, the above steps ensure the continuity and reliability of risk perception and LED display when communication anomalies are detected at edge nodes. By pre-caching structured token embedding tensors and gated fusion feature tensors within a historical time window as input sources for autonomous inference, the improved EdgeViT network can be quickly invoked to reuse validated group modeling, cross-group attention fusion, and local dynamic channel gating mechanisms to generate autonomous gated fusion feature tensors in abnormal situations such as network disconnection, data upload failure, or communication reconnection failure. Subsequently, this autonomous feature tensor is subjected to multi-scale fusion and residual weighting to obtain an autonomous risk fusion vector, and risk inference is performed to output a backup risk assessment result containing risk level labels, trend score vectors, and confidence scores. This step completes the "perception-inference-display" process locally, improving robustness and real-time response capabilities in unreliable communication environments, avoiding the loss of risk information or delayed alarms due to communication interruptions, and ensuring continuous safety visualization and emergency response capabilities in the mine operation area.

[0030] In this embodiment, step five specifically includes: The improved EdgeViT model includes an improved EdgeViT network and a risk awareness module deployed at edge nodes; Confidence scores were extracted from the risk assessment results of the improved EdgeViT model and the collaborative model, respectively. Divide the confidence scores of the two models by the sum of their confidence scores to obtain the corresponding confidence weighting coefficients. Based on the confidence weighting coefficient, the risk level labels, trend score vectors and confidence scores of the improved EdgeViT model and the collaborative model are fused to generate collaborative risk level labels, collaborative trend score vectors and collaborative confidence scores, which constitute the collaborative risk assessment results.

[0031] In this invention, by extracting the confidence scores from the outputs of the improved EdgeViT model and the collaborative model respectively, and calculating a weighting coefficient based on the confidence score ratio, adaptive fusion of the confidence levels of prediction results from different models can be achieved. This design fully leverages the advantages of the improved EdgeViT model in time-series feature modeling, while introducing the complementary capabilities of the collaborative model in identifying global trends or abnormal fluctuations, thereby generating a comprehensive evaluation result that includes collaborative risk level labels, collaborative trend score vectors, and collaborative confidence scores. This design effectively alleviates the overfitting and bias problems of single models in complex environments, and improves the decision stability and risk discrimination accuracy after multi-source model fusion.

[0032] In this embodiment, step six specifically includes: The visual control commands include color coding control, strobe mode adjustment, brightness level setting, and graphic content layout. The color coding control is based on the risk level label setting at the current time step: when the risk level label is 0, the LED display color is set to green; when the risk level label is 1, the LED display color is set to yellow; and when the risk level label is 2, the LED display color is set to red. The flicker mode adjustment is based on the mean of the trend score vector to determine the flicker frequency: when the mean of the trend score vector is less than 0.3, it is set to a flicker-free mode; when the mean of the trend score vector is between 0.3 and 0.6, it is set to a low-frequency flicker mode; when the mean of the trend score vector is greater than 0.6, it is set to a high-frequency flicker mode. The brightness level setting is based on a segmented mapping of confidence scores: when the confidence score is less than 0.4, the brightness is set to low level; when the confidence score is between 0.4 and 0.7, the brightness is set to medium level; and when the confidence score is higher than 0.7, the brightness is set to high level. The graphic content layout includes two parts: text prompts and graphic icons. According to the risk level label, the text prompts "normal", "warning" or "high risk" are displayed. According to the changing trend of the trend score vector, the arrow graphic icons are dynamically displayed. An upward arrow indicates an increased risk, a downward arrow indicates a decreased risk, and a horizontal arrow indicates a stable risk.

[0033] For example, in a specific data collection scenario, when the methane concentration reached 1.05%, the carbon monoxide concentration was 29 ppm, and the wind speed dropped to 0.28 m / s, the risk perception module deployed at the edge node determined it to be at the warning level. The LED panel displayed "Warning" on a yellow background, accompanied by mid-frequency flashing, with a confidence score of 0.81 and a brightness setting of medium. Subsequently, when the wind speed further decreased to 0.16 m / s and the dust concentration increased to 850 μg per cubic meter, the risk perception module determined it to be at the high-risk level. The LED immediately turned red and flashed at a high frequency, with a trend score vector mean of 0.74 and a confidence score rising to 0.88. This method effectively alerts workers to evacuate quickly and avoids potential poisoning and respiratory injuries.

[0034] In this embodiment, the feedback information includes user interaction feedback information from the LED display terminal and mine operation event logs.

[0035] Example 1: To verify the feasibility of this invention in practice, it was applied to a typical deep coal mine operation area. This mine has a high density of workers, large fluctuations in gas concentration, complex communication conditions, and limited lighting. Conventional risk warning and safety information display methods suffer from problems such as strong lag, poor prompting effect, and inability to perceive risks after network outage. There is an urgent need for an edge risk perception and visualization method that can perform real-time reasoning, autonomous display, and sustainable operation locally.

[0036] In this scenario, three main roadways at a depth exceeding 800 meters in the mine were selected, and six types of environmental sensors—temperature, humidity, methane concentration, carbon monoxide, dust, and wind speed—were deployed at a frequency of 1Hz. These sensors were connected to an edge computing unit via an industrial protocol gateway, and the method of this invention was deployed on an embedded edge node system equipped with an RK3588 chip. A 60-second time window was set. The embedded edge node system constructed a structured token embedding tensor in real time and synchronized the current risk level guidance vector. After fusing gating features, the vector was input into the risk perception module for evaluation and display.

[0037] Under normal communication conditions, the embedded edge node system can reliably identify risk levels and display them on an LED screen in the form of color, strobe, brightness, and graphics, indicating the current work environment's risk level. The LED display control module triggers color switching, strobe mode changes, and brightness level adjustments based on the risk level label, enabling operators to quickly identify the risk status. In communication interruption scenarios, the embedded edge node system caches historical token sequences and feature tensors, invokes autonomous inference during network outages, and can still output the risk level and LED control signals within 10ms, ensuring uninterrupted information display.

[0038] To further verify the effectiveness of this invention, the method was compared with a fixed threshold early warning method and an LSTM-based edge inference method. 1000 sets of real-world operational data were selected for experimental evaluation. Evaluation metrics included risk level identification accuracy, early warning rate, network outage tolerance rate, risk information readability score, LED response latency, and user response speed. The experimental results are shown in Table 1.

[0039] Table 1. Performance Comparison of the Invention Method and Comparative Methods in Mine Risk Perception Tasks

[0040] As shown in Table 1, the method of this invention outperforms the fixed threshold warning method and the LSTM-based edge inference method in multiple performance indicators. Regarding risk level identification accuracy, the method of this invention achieves 96.7%, significantly higher than the 79.2% of the fixed threshold warning method and the 88.6% of the LSTM-based edge inference method, demonstrating higher identification accuracy and significantly improving the accuracy of risk warning. In terms of early warning rate, the method of this invention achieves 91.4%, far exceeding the 63.5% and 74.2% of the comparative schemes, effectively enhancing the foresight and proactiveness of the warning. Regarding edge deployment stability, the method of this invention has 100% network outage tolerance, enabling autonomous inference under communication anomalies and ensuring continuous risk assessment output. In contrast, the fixed threshold warning method cannot provide any assessment under network outage conditions, and the LSTM-based edge inference method can only maintain a 41.7% fault tolerance rate, indicating significantly insufficient reliability.

[0041] Furthermore, in terms of LED information display effects, the method of this invention achieved a readability score of 9.3, higher than the fixed threshold warning method's 6.1 and the LSTM-based edge inference method's 7.2, indicating that the visual control commands generated by this invention are more recognizable and user-friendly in terms of color, flicker, and brightness adjustment. Simultaneously, the LED response latency is controlled within 25 milliseconds, far superior to other methods, ensuring high timeliness. Regarding the average user response time, this invention is only 3.4 seconds, significantly better than the fixed threshold warning method's 7.8 seconds and the LSTM-based edge inference method's 5.6 seconds, indicating that users can identify risk levels and react more quickly.

[0042] This invention demonstrates superior performance in edge perception, autonomous reasoning without network access, risk assessment accuracy, and LED visualization, effectively improving risk perception capabilities and early warning response efficiency in mining environments.

[0043] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A mine environment data LED display method based on edge computing, characterized in that, The method comprises the following steps: Step one: collect multi-source environmental data, and perform preprocessing and index grouping and coding to generate a structured token embedding tensor; Step two: input the structured token embedding tensor into an improved EdgeViT network deployed on the edge node to perform grouping modeling, cross-group attention fusion and local dynamic channel gating operation, and generate a gated fusion feature tensor; Step three: input the gated fusion feature tensor into a risk perception module deployed on the edge node to perform risk reasoning and evaluation, and output a risk evaluation result; Step four: when the edge node detects a communication abnormal state, perform risk prediction reasoning based on the historical cached structured token embedding tensor and gated fusion feature tensor to generate a backup risk evaluation result; Step five: if a collaborative model is deployed on the edge node, fuse the risk evaluation results of the improved EdgeViT model and the collaborative model to generate a collaborative risk evaluation result; Step six: input the risk evaluation result, backup risk evaluation result or collaborative risk evaluation result into an LED display control module deployed on the edge node to generate a visual control instruction and drive the LED display terminal to dynamically display risk information; Step seven: collect feedback information and perform incremental update on the EdgeViT network and the risk perception module based on the feedback information.

2. The mine environment data LED display method based on edge computing according to claim 1, characterized in that, The step one specifically comprises: Collect multi-source environmental data through environmental sensors deployed in the mine operation area, wherein the multi-source environmental data includes temperature, humidity, gas concentration, dust density, carbon monoxide content and wind speed; Preprocess the multi-source environmental data, specifically: Extract the timestamp information of the multi-source environmental data, and map the multi-source environmental data to a unified time axis based on a set time step to form a standard time series; Complete the missing items in the standard time series by linear interpolation, and remove the outliers in the standard time series by a local statistical threshold method based on a sliding window to form a standard cleaned sequence; Map the original values in the standard cleaned sequence to the interval [0, 1] by a min-max normalization method to form a standardized data set; the standardized data set is a two-dimensional matrix, each column corresponds to a normalized numerical sequence of an environmental index category, and each row represents multi-dimensional environmental data at a unified time step; At each time step, divide the standardized data set into multiple channels according to the environmental index category, and convert the normalized numerical sequence of each channel into an index token of a set dimension by an embedding mapping function to form a structured token embedding sequence at the current time step; Stack and splice the structured token embedding sequences of all time steps to form a structured token embedding tensor.

3. The mine environment data LED display method based on edge computing according to claim 1, characterized in that, The step two specifically comprises: Input the structured token embedding tensor into an improved EdgeViT network deployed on the edge node, wherein the dimensions of the structured token embedding tensor include time step dimension, environmental index category dimension and embedding dimension of each token; The structured token embedding tensor is grouped and modeled in the environment index category dimension, specifically: According to the physical property similarity of the environmental indicators, temperature and humidity are divided into the thermal and humid environment group, gas concentration and carbon monoxide concentration are divided into the harmful gas group, dust density is divided into the particle aerosol group, and wind speed is divided into the air flow ventilation group; The token embedding sequence of each group is constructed into an independent token grouping sequence to form four groups of index grouping sets; Intra-group self-attention operation is performed on the four groups of index grouping sets to extract intra-group dependent features and time sequence features, and four groups of self-attention index tensors are generated; Four groups of intra-group representative vectors are generated by weighted pooling of the four groups of self-attention index tensors; The four groups of intra-group representative vectors are stacked into a group index feature tensor, and normalized dot product attention operation is performed on the group index feature tensor to generate an inter-group attention matrix; Based on the inter-group attention matrix, weighted fusion is performed on the group index feature tensor to generate a group fusion feature tensor, and dimension compression and nonlinear transformation are performed on the group fusion feature matrix through an MLP structure to generate a group fusion perception tensor; The fusion perception vectors of each index grouping in the group fusion perception tensor are converted into channel fusion perception vectors of each environmental indicator category through channel mapping according to the environmental indicator categories contained, and are spliced into a channel fusion perception tensor in the order of environmental indicator categories; The channel fusion perception tensor is divided into several time segments according to the time steps, and the average value of the absolute change rate in the set sliding window is used to calculate the local volatility score of each environmental indicator category at each time step; The local volatility score of each environmental indicator category is attributed to the index grouping to which it belongs, and the mean and standard deviation of the local volatility scores of all environmental indicator categories in the group are calculated at each time step; Based on the local volatility score and the corresponding mean and standard deviation, the channel activation weight of each environmental indicator category at each time step is calculated; Based on the channel fusion perception tensor, the channel fusion perception vector of each environmental indicator category at the current time step is multiplied by the corresponding channel activation weight to generate a weighted channel perception vector; All time steps of the weighted channel perception vector are stacked in time order to form a gated fusion feature tensor.

4. The mine environment data LED display method based on edge computing according to claim 1, characterized in that, The step three specifically includes: Average pooling and maximum pooling are performed on the gated fusion feature tensor along the time step dimension, and the time aggregation feature vector is generated by splicing; Average pooling and maximum pooling are performed on the gated fusion feature tensor along the environmental indicator category dimension, and the index aggregation feature vector is generated by splicing; The gated fusion feature tensor is unfolded into a two-dimensional feature matrix along the time step dimension and the environmental indicator category dimension, respectively, and is input into independent linear mapping networks to perform projection compression, generating two projection vectors, and performing joint modeling through an MLP structure to generate a channel cross feature vector; The time aggregation feature vector, the index aggregation feature vector, and the channel cross feature vector are spliced to form a multi-scale fusion perception vector; introducing a risk level guide vector, the risk level guide vector being a risk level label of a set historical time window converted into a risk level embedding vector sequence through an embedding mapping function, and the risk level embedding vector sequence being fused to generate a risk fusion vector through weighted average; fusing the multi-scale fusion perception vector and the risk level guide vector through residual weighting to obtain a risk fusion vector; performing risk reasoning and evaluation operations on the risk fusion vector to output a risk evaluation result, the risk evaluation result including a risk level label and a classification result, a trend score vector, and a confidence score, specifically: mapping the risk fusion vector to a three-dimensional risk probability vector through a fully connected layer and a Softmax function, and obtaining the index of the maximum probability value in the three-dimensional risk probability vector through an argmax function as the risk level label of the current time step; the risk level label and the classification result include 0: normal level, 1: early warning level, and 2: high-risk level; performing linear transformation and activation operations on the risk fusion vector through a fully connected layer and a Sigmoid function to generate a trend score vector, the trend score vector having a dimension of K, representing the risk level fluctuation trend in the next K time steps after the current time step; outputting a confidence score through a fully connected layer and a HardSigmoid activation function, the confidence score being a real number in the interval [0, 1], representing the confidence strength of the current risk level label.

5. The edge computing-based mine environment data LED display method according to claim 1, characterized in that, The step four specifically includes: The communication abnormal state includes network connection disconnection, data upload failure, failure to receive a periodic communication confirmation signal, or communication reconnection failure; loading the structured token embedding tensor and the gated fusion feature tensor cached in the set historical time window; inputting the historically cached structured token embedding tensor into the improved EdgeViT network to perform grouping modeling, cross-group attention fusion, and local dynamic channel gating operations to generate an autonomous gated fusion feature tensor; inputting the autonomous gated fusion feature tensor into the risk perception module to perform multi-scale fusion and residual weighting to generate an autonomous risk fusion vector; performing risk reasoning and evaluation on the autonomous risk fusion vector to output a backup risk evaluation result, the backup risk evaluation result including a risk level label, a trend score vector, and a confidence score.

6. The edge computing-based mine environment data LED display method according to claim 1, characterized in that, The step five specifically includes: The improved EdgeViT model includes an improved EdgeViT network and a risk perception module deployed on an edge node; extracting the confidence scores in the risk evaluation results of the improved EdgeViT model and the collaborative model, respectively; dividing the confidence scores of the two models by the sum of the confidence scores to obtain corresponding confidence weighting coefficients, respectively; based on the confidence weighting coefficients, fusing the risk level labels, trend score vectors, and confidence scores of the improved EdgeViT model and the collaborative model, respectively, to generate a collaborative risk level label, a collaborative trend score vector, and a collaborative confidence score, and forming a collaborative risk evaluation result.

7. The edge computing-based mine environment data LED display method according to claim 1, characterized in that, The step six specifically includes: The visualization control instruction includes color coding control, stroboscopic mode adjustment, brightness level setting and graphic content arrangement. The color coding control is based on the risk level label of the current time step to set the LED display color: when the risk level label is 0, the LED display color is set to green; when the risk level label is 1, the LED display color is set to yellow; when the risk level label is 2, the LED display color is set to red. The stroboscopic mode adjustment is based on the mean value of the trend score vector to determine the stroboscopic frequency: when the mean value of the trend score vector is less than 0.3, it is set to the non-flashing mode; when the mean value of the trend score vector is between 0.3 and 0.6, it is set to the low-frequency flashing mode; when the mean value of the trend score vector is higher than 0.6, it is set to the high-frequency flashing mode. The brightness level setting is based on the segmented mapping of the confidence score value: when the confidence score is less than 0.4, the brightness is set to low level; when the confidence score is between 0.4 and 0.7, the brightness is set to medium level; when the confidence score is higher than 0.7, the brightness is set to high level. The graphic content arrangement includes two parts of text prompt and graphic identification: according to the risk level label, the text prompt information "normal", "warning" or "high risk" is displayed; according to the change trend direction of the trend score vector, the arrow graphic identification is dynamically displayed, the arrow pointing up indicates that the risk is rising, the arrow pointing down indicates that the risk is weakening, and the arrow horizontally indicates that the risk is stable. 8.The mine environment data LED display method based on edge computing according to claim 1, wherein, The feedback information includes user interaction feedback information of the LED display terminal and mine operation event log.

Images