Federal learning-based cross-region prediction method and system for mine dust concentration

Abstract

The application belongs to the technical field of mine safety monitoring and intelligent control, and specifically provides a mine dust concentration cross-region prediction method and system based on federated learning, which mainly comprises the following steps: obtaining vibration spectrum, wind speed data and temperature and humidity data of multiple mining areas, preprocessing the vibration spectrum, wind speed data and temperature and humidity data, and generating a federated standardized feature matrix; based on the federated standardized feature matrix, spatial features are extracted through three-dimensional convolution, high-risk area attention weighting is performed, and time series modeling processing is performed to generate a multi-mining area federated prediction model. The application can accurately capture the dust diffusion dynamics across regions, collaboratively schedule equipment to efficiently suppress dust, and help new mining areas quickly build a precise prediction and prevention and control system, effectively improving the scientificity, collaboration and safety of mine dust control.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of mine safety monitoring and intelligent control, and particularly relates to a mine dust concentration cross-region prediction method and system based on federated learning. BACKGROUND

[0002] In the field of mine dust control, the existing technology usually collects vibration, wind speed, temperature and humidity data in a single mining area with various sensors, and after simple preprocessing, uses traditional models to monitor and locally predict dust concentration. Some methods rely on single-point monitoring data to determine dust conditions, and dust removal equipment is often operated independently according to preset programs or manual experience.

[0003] However, the geological conditions of the mine are complex, and the dust diffusion is dynamically affected by many factors. Single-point monitoring cannot fully reflect the dust distribution and diffusion trend, and there is a lack of cross-region data fusion and coordination mechanism. Different mining area models cannot share knowledge, and independent operation of equipment cannot efficiently link to dust migration, resulting in limited accuracy of dust concentration prediction, difficulty in timely and accurate regulation of dust removal measures, and impact on dust control effect and mine operation safety. SUMMARY

[0004] The application provides a mine dust concentration cross-region prediction method and system based on federated learning, effectively solving the problem of dust prediction lacking cross-region coordination and equipment regulation difficulty in adapting to complex working conditions in the prior art, achieving accurate prediction of dust diffusion trend, coordinated scheduling of dust removal equipment, improving dust suppression efficiency, assisting new mining area rapid adaptation, and ensuring mine operation safety.

[0005] To achieve the above purpose, the application adopts the following technical solutions:

[0006] In a first aspect, the application provides a mine dust concentration cross-region prediction method based on federated learning, comprising:

[0007] Obtain vibration spectrum, wind speed data, and temperature and humidity data of multiple mining areas, preprocess the vibration spectrum, wind speed data, and temperature and humidity data, and generate a federated standardized feature matrix.

[0008] Based on the federated standardized feature matrix, spatial features are extracted by three-dimensional convolution, high-risk area attention weighting is performed, and time series modeling processing is performed to generate a multi-mining area federated prediction model.

[0009] Input the real-time standardized feature matrix into the multi-mining area federated prediction model, perform dust concentration prediction and particle migration simulation processing, and generate a dust migration linkage equipment coordinate set.

[0010] According to the dust migration linkage equipment coordinate set, spray device matching optimization and instruction generation processing are performed, and a dust suppression efficiency analysis report is formed; input the standardized feature matrix of the new mining area into the multi-mining area federal prediction model, perform residual compensation and risk visualization processing, and output a cross-regional visualization prediction report.

[0011] Further, the vibration spectrum, wind speed data, and temperature and humidity data are preprocessed to generate a federal standardized feature matrix, including:

[0012] The vibration spectrum, wind speed data, and temperature and humidity data are timestamp calibrated and feature confusion processed to generate a desensitized environmental feature dataset.

[0013] The desensitized environmental feature dataset is subjected to vibration anomaly filtering and wind direction discretization to generate a preprocessed feature set.

[0014] The preprocessed feature set is fused, and null value filling and normalization processing are performed to generate a standardized feature matrix.

[0015] Further, the preprocessed feature set is fused, and null value filling and normalization processing are performed to generate a standardized feature matrix, including:

[0016] The vibration direction preprocessing feature set and temperature and humidity parameters in the fused preprocessing feature set are subjected to multi-source data alignment fusion processing to generate a fused feature dataset.

[0017] The fused feature dataset is subjected to null matrix filling based on spatial and temporal proximity to generate a complete environmental feature matrix.

[0018] The complete environmental feature matrix is subjected to maximum and minimum normalization by channel to generate a standardized feature matrix.

[0019] Further, based on the federal standardized feature matrix, a multi-mining area federal prediction model is generated through three-dimensional convolution to extract spatial features, high-risk area attention weighting, and time series modeling processing, including:

[0020] The standardized feature matrix of each mining area is subjected to feature space alignment to generate a federal input dataset.

[0021] The federal input dataset is input into a three-dimensional convolution module to output a tunneling face spatial feature map.

[0022] The tunneling face spatial feature map is input into a high-risk area attention module to output a spatial weighted feature map.

[0023] The spatial weighted feature map is input into a long short-term memory network to output a spatiotemporal fusion feature vector.

[0024] The spatiotemporal fusion feature vectors of each mining area are aggregated to perform federal parameter fusion and model optimization to generate an optimized federal model.

[0025] Furthermore, the real-time standardized feature matrix is ​​input into the multi-mine area federated prediction model to perform dust concentration prediction and particle migration simulation processing, generating a coordinate set of dust migration linkage equipment, including:

[0026] The real-time standardized feature matrix is ​​input into the optimized federated model, which outputs a map of future dust concentration distribution.

[0027] The concentration gradient is calculated based on the dust concentration distribution map, and the dust diffusion direction field is generated.

[0028] Input the dust diffusion direction field into the particle migration simulation module, and output the dust trajectory prediction map;

[0029] Obtain the tunnel topology; analyze the areas exceeding the standard in the dust trajectory prediction map, and generate a coordinate set of linkage equipment based on the tunnel topology.

[0030] Furthermore, based on the coordinate set of the dust migration linkage equipment, the spray device matching optimization and command generation processing are performed to generate a dust suppression efficiency analysis report, including:

[0031] Parse the coordinate set of the linked equipment, perform spray device matching and coverage optimization, and generate a device startup priority queue.

[0032] Instructions are formulated based on the device priority queue to generate a set of device control instructions.

[0033] The device control command set is sent to the terminal, feedback verification and efficiency evaluation are performed, and a dust suppression efficiency report is obtained.

[0034] Furthermore, the standardized feature matrix of the new mining area is input into the multi-mining area federated prediction model, residual compensation and risk visualization processing are performed, and a cross-regional visualized prediction report is output, including:

[0035] The standardized feature matrix of the new mining area is input into the optimized federated model, and the initial dust prediction value is output.

[0036] Based on the initial predicted values, residual compensation and risk mapping are performed to generate a dust generation risk heat map.

[0037] Integrate risk heatmaps with historical accident databases to generate visual prediction reports.

[0038] Furthermore, this federated learning-based cross-regional prediction method for mine dust concentration also includes: integrating dust suppression efficiency analysis reports and visualization prediction reports, and updating the plug-and-play interface for coal mine dust through knowledge distillation and graph construction.

[0039] Furthermore, by integrating dust suppression efficiency analysis reports and visualization prediction reports, and through knowledge distillation and graph construction, the plug-and-play interface for coal mine dust is updated, including:

[0040] Collect dust suppression efficiency report and prediction report, perform error analysis and feature sorting, and generate core knowledge feature set.

[0041] Input the core knowledge feature set into the knowledge distillation module, and output the lightweight prediction model.

[0042] Based on the lightweight model, a dust knowledge graph is constructed, and a plug-and-play interface is generated.

[0043] In a second aspect, the application provides a mine dust concentration cross-region prediction system based on federated learning, which comprises:

[0044] The federated data preprocessing module: obtains the vibration spectrum, wind speed data and temperature and humidity data of multiple mining areas, preprocesses the vibration spectrum, wind speed data and temperature and humidity data, and generates a federated standardized feature matrix.

[0045] The space-time federated modeling module: based on the federated standardized feature matrix, the spatial features are extracted by three-dimensional convolution, the high-risk area attention weighting and time series modeling processing are performed, and a multi-mine area federated prediction model is generated.

[0046] The dust simulation linkage module: input the real-time standardized feature matrix into the multi-mine area federated prediction model, perform dust concentration prediction and particle migration simulation processing, and generate a dust migration linkage device coordinate set.

[0047] The cross-domain risk prediction module: according to the dust migration linkage device coordinate set, the spray device matching optimization and instruction generation processing are performed, and a dust suppression efficiency analysis report is formed; input the new mine area standardized feature matrix into the multi-mine area federated prediction model, perform residual compensation and risk visualization processing, and output a cross-region visual prediction report.

[0048] In a third aspect, the application provides a mine dust concentration cross-region prediction device based on federated learning, which comprises a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program to realize the steps of the mine dust concentration cross-region prediction method based on federated learning in the first aspect.

[0049] In a fourth aspect, the application provides a storage medium, which stores computer program instructions, and when the computer program instructions are read and run by a processor, the steps of the mine dust concentration cross-region prediction method based on federated learning in the first aspect are executed.

[0050] The beneficial effects of the application are:

[0051] The application effectively solves the problems of lack of cross-regional cooperation in dust prediction and difficulty in adapting to complex working conditions in device regulation and control in the prior art by collecting and preprocessing vibration spectrum, wind speed, temperature and humidity data of multiple mining areas, modeling and predicting through three-dimensional convolution and federated learning, combining particle migration simulation, device linkage optimization and new mining area residual compensation, realizing accurate capture of dust diffusion dynamics across regions, coordinated scheduling of dust suppression equipment, helping to quickly build a precise prediction and prevention and control system in new mining areas, and improving the scientificity, collaboration and safety of mine dust control.

[0052] Other features and advantages of the present application will be set forth in the following description, and in part will become apparent to those skilled in the art from the description, or can be learned by practice of the present application. The objects and other advantages of the present application can be realized and obtained by the structure indicated in the specification and drawings. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0054] Figure 1 The flowchart of the mine dust concentration cross-regional prediction method based on federated learning of the present application is shown;

[0055] Figure 2 The module diagram of the mine dust concentration cross-regional prediction system based on federated learning of the present application is shown. DETAILED DESCRIPTION

[0056] In order to solve the problems raised in the background art, the present application collects and preprocesses vibration spectrum, wind speed, temperature and humidity data of multiple mining areas, models and predicts through three-dimensional convolution and federated learning, combines particle migration simulation, device linkage optimization and new mining area residual compensation, realizes accurate prediction of dust diffusion trend, coordinated scheduling of dust removal equipment, improves dust suppression efficiency, helps new mining areas to quickly adapt, and ensures the safety of mine operation.

[0057] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be clearly and completely described in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0058] In some embodiments, asFigure 1 As shown, the present application provides a mine dust concentration cross-regional prediction method based on federated learning, comprising:

[0059] S1. Obtain the vibration spectrum, wind speed data, and temperature and humidity data of multiple mining areas, preprocess the vibration spectrum, wind speed data, and temperature and humidity data, and generate a federated standardized feature matrix.

[0060] S2. Based on the federated standardized feature matrix, spatial features are extracted through three-dimensional convolution, high-risk area attention weighting and time series modeling processing are performed, and a multi-mining area federated prediction model is generated.

[0061] S3. Input the real-time standardized feature matrix into the multi-mining area federated prediction model, perform dust concentration prediction and particle migration simulation processing, and generate a dust migration linkage device coordinate set.

[0062] S4. According to the dust migration linkage device coordinate set, perform spray device matching optimization and instruction generation processing, form a dust suppression efficiency analysis report; input the new mining area standardized feature matrix into the multi-mining area federated prediction model, perform residual compensation and risk visualization processing, and output a cross-regional visualization prediction report.

[0063] In some embodiments, the preprocessing of the vibration spectrum, wind speed data, and temperature and humidity data in S1 to generate the federated standardized feature matrix comprises:

[0064] S11. Time stamp calibration and feature confusion processing are performed on the vibration spectrum, wind speed data, and temperature and humidity data to generate a desensitization environment feature data set.

[0065] The vibration spectrum represents the frequency domain features of the tunneling machine cutting head vibration signal collected by the mine intrinsic safety acceleration sensor, and reflects the correlation between device operating state and dust production intensity.

[0066] The wind speed data represents the wind flow parameters monitored by the ultrasonic wind speed instrument in the mine tunnel in real time, and the dust migration path is characterized by the change of wind speed gradient.

[0067] The temperature and humidity data are collected by the mine temperature and humidity integrated sensor, and the temperature and humidity data affect the dust settlement through the water adsorption effect.

[0068] When performing time stamp calibration processing on the vibration spectrum, wind speed data, and temperature and humidity data of multiple mining areas, the atomic clock signal of the mine explosion-proof computer is taken as the reference, the time deviation of the original data of each sensor is corrected to within a specified error (such as ±50 milliseconds), and a time-aligned environment data set is generated.

[0069] Extract the device location identifier in the time-aligned environmental data set, convert it into a non-geographically meaningful code (such as "E128°N35°" to "LOC#9A3F") through an irreversible hash algorithm, while completely retaining the numerical values of the vibration spectrum, wind speed data, direction angle, temperature value, and humidity value.

[0070] Output the desensitized environmental feature data set, which has a data structure containing: vibration amplitude value, wind speed vector, temperature and humidity value field, and obfuscated location identifier.

[0071] S12. Perform vibration anomaly filtering and wind direction discretization on the desensitized environmental feature data set to generate a preprocessed feature set.

[0072] During vibration anomaly filtering, extract the vibration amplitude value for sliding window analysis. If a data point deviates from the window mean by more than a specified value (such as 3 times the standard deviation), it is determined to be an outlier and is removed and replaced with linear interpolation.

[0073] During wind direction discretization, the continuous wind direction angle in the wind speed vector is divided into a specified number of intervals (such as 16 intervals of 22.5°, 0°-22.5° is recorded as north wind, 22.5°-45° is recorded as northeast wind, and so on), to eliminate mechanical jittering errors of the wind direction sensor.

[0074] Output the preprocessed feature set, including the effective vibration feature set and the discrete wind direction feature set.

[0075] S13. Fuse the preprocessed feature set, perform null value filling and normalization processing, and generate a standardized feature matrix.

[0076] In some embodiments, S13 fuses the preprocessed feature set, performs null value filling and normalization processing, and generates a standardized feature matrix, including:

[0077] S131. Perform multi-source data alignment fusion processing on the vibration direction preprocessed feature set and the temperature and humidity parameters in the fused preprocessed feature set to generate a fused feature data set.

[0078] The vibration direction preprocessed feature set includes the vibration intensity sequence of the effective vibration feature set and the direction label of the discrete wind direction feature set.

[0079] The temperature and humidity parameters are the temperature and humidity value fields directly extracted from the desensitized environmental feature data set (such as temperature value 35.2°C, humidity value 28.0%RH).

[0080] The time stamp matching algorithm is adopted to merge the data at the same time point and the same spatial position of the roadway into a single record, and a fusion feature data set of three-dimensional tensor structure is output, and the dimension is: time point number x spatial coordinate number x 4 feature channels, and the 4 channels store vibration intensity, direction label, temperature value and humidity value in sequence.

[0081] S132. Based on the spatiotemporal proximity, the null matrix filling is performed on the fusion feature data set to generate a complete environmental feature matrix.

[0082] The three-dimensional tensor of the fusion feature data set is subjected to null value detection. When the feature channel at a certain spatial position and time appears null value (such as temperature and humidity sensor failure), a weighted average filling algorithm based on spatiotemporal proximity is adopted to process, thereby generating a complete environmental feature matrix.

[0083] S133. The maximum and minimum normalization is performed on the complete environmental feature matrix in a channel to generate a standardized feature matrix.

[0084] The maximum and minimum normalization is independently performed on the four feature channels of the complete environmental feature matrix, and a three-dimensional standardized feature matrix is output, and the numerical range is unified as [0, 1], and the dimension is the same as the input matrix.

[0085] In some embodiments, based on the federal standardized feature matrix in S2, spatial features, high-risk area attention weighting and time series modeling processing are extracted by three-dimensional convolution to generate a multi-mine area federal prediction model, including:

[0086] S21. The feature space alignment is performed on the standardized feature matrix of each mine area to generate a federal input data set.

[0087] The dynamic time warping algorithm is adopted to eliminate the difference in grid resolution of the roadway network in different mine areas, and the data of each mine area is uniformly mapped to a reference coordinate system through spatial interpolation, and a federal input data set is output, including the number of time frames, the number of standard space grids and the number of feature channels.

[0088] S22. The federal input data set is input into a three-dimensional convolution module to output a drivage face spatial feature map.

[0089] The federal input data set is input into a three-dimensional convolution module for processing, and a convolution kernel is designed first (such as a 3×3×4 kernel sliding along the spatial dimension), and a drivage face spatial feature map is generated by extracting the local spatial correlation of the drivage face (such as the relationship between the vibration high value area and the temperature gradient).

[0090] S23. The drivage face spatial feature map is input into a high-risk area attention module to output a spatial weighted feature map.

[0091] The spatial feature map of the tunneling face is input into a high-risk area attention module for processing. First, attention weight is generated, and high-risk coordinates are located based on a historical dust accident library (such as a roadway corner and a ventilation dead angle). Then, feature weighting is carried out, and the feature weight of the high-risk area is increased by a specified percentage (such as 300%), and the weight of the non-dangerous area is reduced to a specified percentage (such as 50%). Finally, a spatial weighted feature map is output.

[0092] S24. The spatial weighted feature map is input into a long short-term memory network to output a spatio-temporal fusion feature vector.

[0093] The spatial weighted feature map is input into a long short-term memory network (LSTM) for time step setting (such as allowing each spatial point to independently input a 30-minute sequence (30 steps)), and then feature fusion is carried out to output a spatio-temporal fusion feature vector.

[0094] S25. Aggregate the spatio-temporal fusion feature vectors of each mining area to perform federated parameter fusion and model optimization to generate an optimized federated model.

[0095] Aggregate the spatio-temporal fusion feature vectors of each mining area to perform federated learning. First, parameter aggregation is performed using the FedAvg algorithm to average the model parameters of each mining area. Then, back propagation is performed to update the global parameters based on the true value of the dust concentration. Then, multiple iterations are performed until the convergence condition is met (such as a decrease in the loss function < 0.001 / round). An optimized federated model is generated, and the model architecture is fixed and deployed to each mining area.

[0096] In some embodiments, the real-time standardized feature matrix in S3 is input into a multi-mine area federated prediction model to perform dust concentration prediction and particle migration simulation processing to generate a dust migration linkage device coordinate set, including:

[0097] S31. The real-time standardized feature matrix is input into the optimized federated model to output a future dust concentration distribution map.

[0098] The real-time standardized feature matrix (time frame x spatial point x 4 feature channels) is input into a multi-mine area federated prediction model. The model is trained by federated learning to fuse mine area data, and the model outputs a future dust concentration distribution map as a three-dimensional prediction tensor, including a time dimension (K future time slices), a spatial dimension (N fixed standard spatial points), and a data layer (dust concentration value).

[0099] The multi-mine area federated prediction model uses an Encoder-Decoder architecture to capture the spatio-temporal correlation of dust diffusion through spatio-temporal convolution layers, and is suitable for mines with different geological conditions.

[0100] S32. Calculate the concentration gradient based on the dust concentration distribution map to generate a dust diffusion direction field.

[0101] The concentration gradient is calculated based on the dust concentration distribution map, the gradient generation principle is to calculate the concentration difference of adjacent grids to generate a diffusion intensity scalar field, and the maximum concentration change direction is extracted and represented by an angle of 0°-360°.

[0102] The output dust diffusion direction field is a two-dimensional vector field, each spatial point contains gradient intensity and direction angle.

[0103] S33. The dust diffusion direction field is input into the particle migration simulation module, and the dust trajectory prediction map is output.

[0104] The dust diffusion direction field is input into the particle migration simulation module, and the dust trajectory prediction map is output.

[0105] The output dust trajectory prediction map is a space-time probability density distribution map (hot spot area probability value is 0-1); wherein, represents the particle position vector, represents the velocity field composed of gradient intensity and direction angle, represents the instantaneous velocity, t represents the time variable.

[0106] S34. Obtain the roadway topology; analyze the over-standard area of the dust trajectory prediction map, and generate the linkage device coordinate set combined with the roadway topology.

[0107] Analyze the over-standard area of the dust trajectory prediction map (such as concentration>10mg / m³), generate the linkage device coordinate set combined with the roadway topology, and the decision logic can be to identify the dust aggregation area with a probability greater than a specified value (such as 0.7), and locate the key control point along the dust diffusion path (such as from the source 、 , L represents the length of the diffusion path), and the output device coordinate set is a list of device deployment positions, each device deployment position contains a roadway partition ID and three-dimensional coordinates.

[0108] In some embodiments, S4 is performed according to the dust migration linkage device coordinate set, the spray device matching optimization and instruction generation process is performed, and a dust suppression efficiency analysis report is formed, including:

[0109] Sa41. Analyze the linkage device coordinate set, perform spray device matching and coverage optimization, and generate a device start priority queue.

[0110] When matching the spray device, construct a KD-Tree spatial index according to the three-dimensional coordinates of the roadway, and retrieve controllable spray devices within a specified distance (such as 10 meters) from the target point; when optimizing coverage, first determine the effective coverage radius R of the spray, and then solve the minimum coverage set: , the constraint condition is: , the output device starts a priority queue; wherein, n represents the total number of candidate spraying devices, representing the state of device activation, representing the coverage of the device, representing the coordinates of the spraying device, and the high-risk area is a set of coordinates where the dust exceeds the standard, such as a space area where the dust concentration is greater than a specified value (such as 10 mg / m³).

[0111] Sa42. Formulate instructions based on the device priority queue to generate a set of device control instructions.

[0112] Formulate control instructions based on the device priority queue, and then carry out parameter generation, such as spraying intensity I calculated according to the formula , and duration T calculated according to the formula , and finally output the set of device control instructions; wherein, representing the dust suppression coefficient, which can be 0.8, device priority quantization data (represented by a value in the interval 0-1 to reflect the priority level), representing the predicted concentration, L representing the length of the diffusion path, and v representing the wind speed.

[0113] Sa43. Send the set of device control instructions to the terminal, perform feedback verification and efficiency evaluation, and obtain a dust suppression efficiency report.

[0114] During feedback verification, use CRC32 verification to ensure the integrity of the instructions, and the terminal feeds back a response code.

[0115] During efficiency evaluation, calculate the real-time dust concentration drop rate and energy consumption efficiency.

[0116] The generated dust suppression efficiency analysis report contains the dust concentration change curve (comparing predicted and measured values), device response delay distribution histogram, and energy quantization index.

[0117] In some embodiments, S4 inputs the standardized feature matrix of the new mining area into the mining area federal prediction model to perform residual compensation and risk visualization processing, and outputs a cross-area visualization prediction report, including:

[0118] Sb41. Input the standardized feature matrix of the new mining area into the optimized federal model to output the initial dust prediction value.

[0119] The initial dust prediction value is a three-dimensional concentration distribution tensor, including time dimension (K future time slices), space dimension (N fixed standard space points), and data layer (dust concentration value).

[0120] Sb42. Perform residual compensation and risk mapping based on the initial prediction value to generate a dust production risk heat map.

[0121] Real-time data calibration, access new mine local dust sensor real-time monitoring value, calculate residual , Again, the space compensation field is constructed, and the residual is diffused to the whole roadway by Kriging interpolation to generate a residual correction matrix , , ; wherein, represents the predicted value, represents the measured value, represents the compensation coefficient of the new mine area, which is determined by fitting the residual distribution of the new mine area debugging database.

[0122] Finally, risk thermodynamic mapping is performed, and the compensated concentration is mapped to the thermodynamic map to obtain a dust risk thermodynamic map.

[0123] When mapping the compensated concentration to the thermodynamic map, a number of sequentially increasing dust mass concentration intervals can be set, which correspond to low risk, medium risk and high risk respectively, and are marked with colors.

[0124] For example, if the dust mass concentration is <10 mg / m³, it represents low risk and is marked green; if the dust mass concentration is 10-50 mg / m³, it represents medium risk and is marked yellow; and if the dust mass concentration is greater than 50 mg / m³, it represents high risk and is marked red.

[0125] Sb43. Integrate the risk thermodynamic map with the historical accident library to generate a visual prediction report.

[0126] Retrieve similar working condition accident cases in recent years (such as 5 years) (such as working condition accident cases with a concentration deviation of <8% and a geological type similarity of >75%), and calculate the risk index , the formula is: ; wherein, represents the concentration risk weight, represents the accident frequency weight, and are determined by multiple linear regression of historical accident data of the coal mine, represents the normalized value of the number of historical accidents in the specified range.

[0127] Perform three-dimensional visualization construction, overlay the thermodynamic map into a semi-transparent data layer by rendering the roadway stereoscopic model through WebGL, dynamically mark the expansion direction of the high-risk area with arrows pointing to the main path of dust diffusion, and finally output a cross-regional visual prediction report.

[0128] In some embodiments, the mine dust concentration cross-regional prediction method based on federated learning further comprises: S5. Integrate the dust suppression efficiency analysis report and the visual prediction report, update the coal mine dust plug-and-play interface through knowledge distillation and graph construction processing.

[0129] In some embodiments, the dust suppression efficiency analysis report and the visualized prediction report are integrated, and a coal mine dust plug-and-play interface is updated through knowledge distillation and graph construction processing, including:

[0130] S51. Collecting the dust suppression efficiency report and the prediction report, performing error analysis and feature sorting, and generating a core knowledge feature set.

[0131] The dust suppression efficiency analysis report includes fields such as actual dust concentration reduction rate, equipment response delay, and energy consumption statistics; the visualized prediction report provides spatial distribution of dust concentration prediction deviation and historical accident correlation data. Through random forest algorithm, features with contribution greater than a specified value (such as 0.3) are selected to generate a core knowledge feature set, which has a high-dimensional vector data structure, retains the original physical meaning, but eliminates redundant parameters.

[0132] S52. Input the core knowledge feature set into the knowledge distillation module, and output a lightweight prediction model.

[0133] The core knowledge feature set is input into the knowledge distillation module, and knowledge transfer is achieved using a teacher-student model architecture. The teacher model can be the original federal prediction model, and the student model can be a lightweight MobileNetV3 architecture. The distillation process minimizes the knowledge transfer loss function L.

[0134] Exemplary, where 0.7 represents the weight coefficient of KL divergence, represents the KL divergence loss term, which measures the difference between the probability distributions output by the teacher model and the student model, T represents the output distribution of the teacher model, S represents the output distribution of the student model, and 0.3 represents the weight coefficient of cross-entropy, represents the cross-entropy loss term, which measures the difference between the predicted value of the student model and the true label y.

[0135] KL divergence forces the student model to learn the probability distribution of the teacher model, and cross-entropy guarantees the basic prediction accuracy.

[0136] S53. Constructing a dust knowledge graph based on the lightweight model and generating a plug-and-play interface.

[0137] Based on the lightweight prediction model, the dust knowledge graph is constructed, and the model hidden layer features are mapped to graph nodes.

[0138] The graph nodes include three types of entities: risk source entities (such as "fault zone cutting area"), dust suppression equipment entities (such as "high-pressure spray device SPJ-205"), and accident case entities.

[0139] ​The inter-entity relationship is connected through feature attention weight, and the association weight between the node of "vibration intensity > 1.2G" and the node of "tooth wear accident" is 0.88.

[0140] Finally, a plug-and-play interface is generated, the input layer of which receives a four-dimensional tensor of standardized feature matrix, and the output layer provides a JSON format early warning instruction containing risk coordinates, concentration value and device action.

[0141] In some embodiments, as shown in Figure 2 The present application provides a mine dust concentration cross-region prediction system based on federated learning, which comprises:

[0142] The federated data preprocessing module acquires vibration spectrum, wind speed data and temperature and humidity data of multiple mining areas, pre-processes the vibration spectrum, wind speed data and temperature and humidity data, and generates a federated standardized feature matrix.

[0143] The spatio-temporal federated modeling module generates a multi-mining-area federated prediction model by extracting spatial features, high-risk area attention weighting and time series modeling processing based on the federated standardized feature matrix through three-dimensional convolution.

[0144] The dust simulation linkage module inputs the real-time standardized feature matrix into the multi-mining-area federated prediction model, performs dust concentration prediction and particle migration simulation processing, and generates a dust migration linkage device coordinate set.

[0145] The cross-domain risk prediction module performs spray device matching optimization and instruction generation processing according to the dust migration linkage device coordinate set, forms a dust suppression efficiency analysis report, inputs the standardized feature matrix of a new mining area into the multi-mining-area federated prediction model, performs residual compensation and risk visualization processing, and outputs a cross-region visual prediction report.

[0146] In some embodiments, the present application provides a mine dust concentration cross-region prediction device based on federated learning, which comprises a memory and a processor; the memory is used to store a computer program; and the processor is used to execute the computer program to realize the steps of the mine dust concentration cross-region prediction method based on federated learning.

[0147] In some embodiments, the present application provides a storage medium, which stores computer program instructions, and the computer program instructions are read and run by a processor to execute the steps of the mine dust concentration cross-region prediction method based on federated learning.

[0148] Any reference to storage, memory, database or other medium herein includes non-volatile and / or volatile storage. Non-volatile storage can include read-only memory (ROM), programmable ROM (PROM), electronically programmable ROM (EPROM), or electrically erasable programmable ROM (EEPROM). Volatile storage can include random-access memory (RAM), or external cache memory.

[0149] It should be noted that, in the present document, relational terms such as "first" and "second", and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. In other words, any of the elements listed in the description of the above embodiments can be combined with any of the other elements to form a new embodiment.

[0150] Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art will understand that modifications can be made to the foregoing embodiments, or other embodiments can be used, without departing from the spirit and scope of the inventive subject matter.

Claims

1. A method for cross-region prediction of mine dust concentration based on federated learning, characterized in that, The method comprises the following steps: Obtain the vibration spectrum, wind speed data, and temperature and humidity data of multiple mining areas, preprocess the vibration spectrum, wind speed data, and temperature and humidity data, and generate a federal standardized feature matrix; Based on the federal standardized feature matrix, spatial features, high-risk area attention weighting, and time series modeling processing are extracted through three-dimensional convolution to generate a multi-mining area federal prediction model; Input the real-time standardized feature matrix into the multi-mining area federal prediction model to perform dust concentration prediction and particle migration simulation processing, and generate a dust migration linkage device coordinate set; According to the dust migration linkage device coordinate set, spray device matching optimization and instruction generation processing are performed to form a dust suppression efficiency analysis report; input the new mining area standardized feature matrix into the multi-mining area federal prediction model to perform residual compensation and risk visualization processing, and output a cross-regional visual prediction report; Wherein, the preprocessing of the vibration spectrum, wind speed data, and temperature and humidity data to generate the federal standardized feature matrix comprises: Timestamp calibration and feature confusion processing are performed on the vibration spectrum, wind speed data, and temperature and humidity data to generate a desensitized environmental feature dataset; Vibration anomaly filtering and wind direction discretization are performed on the desensitized environmental feature dataset to generate a preprocessed feature set; The preprocessed feature set is fused to perform null value filling and normalization processing to generate a standardized feature matrix; Based on the federal standardized feature matrix, spatial features, high-risk area attention weighting, and time series modeling processing are extracted through three-dimensional convolution to generate a multi-mining area federal prediction model, which comprises: Feature space alignment is performed on the standardized feature matrix of each mining area to generate a federal input dataset; The federal input dataset is input into a three-dimensional convolution module to output a tunneling face spatial feature map; The tunneling face spatial feature map is input into a high-risk area attention module to output a spatial weighted feature map; The spatial weighted feature map is input into a long short-term memory network to output a spatiotemporal fusion feature vector; The spatiotemporal fusion feature vectors of each mining area are aggregated to perform federal parameter fusion and model optimization to generate an optimized federal model.

2. The method of claim 1, wherein the method is a method of predicting a mine dust concentration across regions based on federated learning. The preprocessed feature set is fused to perform null value filling and normalization processing to generate a standardized feature matrix, which comprises: Multi-source data alignment fusion processing is performed on the vibration direction preprocessed feature set and temperature and humidity parameters to generate a fused feature dataset; Based on the spatiotemporal proximity, the fused feature dataset is filled with null value matrix to generate a complete environmental feature matrix; The complete environmental feature matrix is divided into channels to perform maximum and minimum normalization to generate a standardized feature matrix.

3. The method of claim 1, wherein the method is a method of predicting a mine dust concentration across regions based on federated learning. The real-time standardized feature matrix is input into the multi-mining area federal prediction model to perform dust concentration prediction and particle migration simulation processing to generate a dust migration linkage device coordinate set, which comprises: The real-time standardized feature matrix is input into the optimized federal model to output a future dust concentration distribution map; Based on the dust concentration distribution map, a concentration gradient is calculated to generate a dust diffusion direction field; The dust diffusion direction field is input into a particle migration simulation module to output a dust trajectory prediction map; Obtain the roadway topology, analyze the exceeding standard area of the dust trajectory prediction map, and generate a linkage device coordinate set in combination with the roadway topology.

4. The method of claim 1, wherein the method is a method of predicting a mine dust concentration across areas based on a federal learning. According to the dust migration linkage device coordinate set, spray device matching optimization and instruction generation processing are performed to form a dust suppression efficiency analysis report, which comprises: Analyzing the linkage device coordinate set, performing spray device matching and coverage optimization, and generating a device start priority queue; Formulating instructions based on the device priority queue to generate a device control instruction set; Sending the device control instruction set to the terminal, performing feedback verification and efficiency evaluation, and obtaining a dust suppression efficiency report.

5. The cross-regional prediction method for mine dust concentration based on federated learning according to claim 1, characterized in that, Input the new mine standardized feature matrix into the multi-mine federal prediction model, perform residual compensation and risk visualization processing, and output the cross-regional visual prediction report, including: Input the new mine standardized feature matrix into the optimization federal model to output the initial dust prediction value; Based on the initial prediction value, perform residual compensation and risk mapping to generate a dust production risk heat map; Integrate the risk heat map and the historical accident library to generate a visual prediction report.

6. The method of claim 1, wherein the method is a method of predicting a mine dust concentration across regions based on federated learning. Also including: Integrate the dust suppression efficiency analysis report and the visual prediction report, update the coal mine dust plug-and-play interface through knowledge distillation and graph construction processing.

7. The method of claim 6, wherein the method further comprises: Integrate the dust suppression efficiency analysis report and the visual prediction report, update the coal mine dust plug-and-play interface through knowledge distillation and graph construction processing, including: Collecting dust suppression efficiency reports and prediction reports, performing error analysis and feature sorting, and generating a core knowledge feature set; Input the core knowledge feature set into the knowledge distillation module to output a lightweight prediction model; Based on the lightweight model, build a dust knowledge graph to generate a plug-and-play interface.

8. A mine dust concentration cross-region prediction system based on federated learning, characterized in that, It includes: Federal data preprocessing module: obtain vibration spectrum, wind speed data, and temperature and humidity data of multiple mines, preprocess the vibration spectrum, wind speed data, and temperature and humidity data, and generate a federal standardized feature matrix; Spacetime federal modeling module: based on the federal standardized feature matrix, extract spatial features through three-dimensional convolution, weight high-risk areas, and perform time series modeling processing to generate a multi-mine federal prediction model; Dust simulation linkage module: input real-time standardized feature matrix into the multi-mine federal prediction model to perform dust concentration prediction and particle migration simulation processing, and generate a dust migration linkage device coordinate set; Cross-domain risk prediction module: based on the dust migration linkage device coordinate set, perform spray device matching optimization and instruction generation processing to form a dust suppression efficiency analysis report; input the new mine standardized feature matrix into the multi-mine federal prediction model to perform residual compensation and risk visualization processing, and output a cross-regional visual prediction report; Wherein, preprocessing the vibration spectrum, wind speed data, and temperature and humidity data to generate a federal standardized feature matrix includes: Timestamp calibration and feature confusion processing on the vibration spectrum, wind speed data, and temperature and humidity data to generate a desensitized environmental feature dataset; Perform vibration anomaly filtering and wind direction discretization on the desensitized environmental feature dataset to generate a preprocessed feature set; Fuse the preprocessed feature set, perform null value filling and normalization processing, and generate a standardized feature matrix; Based on the federal standardized feature matrix, extract spatial features through three-dimensional convolution, weight high-risk areas, and perform time series modeling processing to generate a multi-mine federal prediction model, including: Perform feature space alignment on each mine standardized feature matrix to generate a federal input dataset; Input the federal input dataset into the three-dimensional convolution module to output a heading face spatial feature map; The spatial feature map of the tunneling face is input into the high-risk area attention module, and a spatial weighted feature map is output; The spatial weighted feature map is input into the long short-term memory network, and a spatio-temporal fusion feature vector is output; The spatio-temporal fusion feature vectors of each mining area are aggregated, and federal parameter fusion and model optimization are performed to generate an optimized federal model.

Images