Mine fully-mechanized coal mining equipment state identification and Wi-Fi transmission method based on deep learning

Abstract

The invention discloses a mine fully-mechanized coal mining equipment state identification and Wi-Fi transmission method based on deep learning, and relates to the technical field of mine equipment state monitoring, and the method comprises the steps: collecting equipment operation parameters and environment vibration data, and constructing a space-time-vibration combined database; analyzing a vibration multipath effect by combining a Rayleigh fading model and Monte Carlo simulation, and calculating a path loss value; constructing a space-state-multipath matrix by using a deep learning model, generating a multi-dimensional characteristic spectrum of the equipment, screening abnormal equipment based on health scores, and performing analysis to obtain abnormal scores; through comparison with a preset threshold value, three-level dynamic responses, namely a conventional mode and an enhanced recognition mode, are realized; for a high-concurrency scene, a sliding window statistics and priority flow limiting mechanism is adopted. Self-adaptively adjusting a threshold value and optimizing a multi-path compensation strategy by evaluating the recognition accuracy, the communication stability and the resource consumption; according to the method, the state recognition precision and the communication reliability of the underground equipment are remarkably improved, and meanwhile, the resource utilization efficiency is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of mine equipment status monitoring, and specifically to a mine fully mechanized mining equipment status identification and Wi-Fi6 transmission method based on deep learning. Background Art

[0002] The stability and efficiency of underground communication systems in coal mines are of great significance to ensuring safe production in mines and improving operational efficiency. However, due to the particularity of the underground environment in coal mines, such as narrow tunnels, complex geological structures, severe electromagnetic interference, and significant multipath effects, traditional wireless communication technologies are difficult to meet the needs of underground communications. Existing underground communication systems mostly rely on wired transmission or early wireless communication technologies. These technologies have obvious deficiencies in coverage, data transmission rate, and anti-interference capabilities, and cannot achieve real-time monitoring and efficient communication of underground equipment status. In addition, with the development of intelligent coal mines, the demand for status identification and concurrent data transmission of comprehensive mining equipment is increasing. Traditional communication methods also show obvious limitations in integrating program-controlled dispatching systems, public network telephones, and underground wireless calls.

[0003] Although some research has attempted to introduce Wi-Fi technology into underground coal mine communications, early Wi-Fi technology still struggles to adapt to the complex underground environment in terms of bandwidth, coverage, and multipath mitigation. This is particularly true in scenarios with high-density device access and large amounts of concurrent data transmission, further limiting the performance of existing technologies. To address this issue, a deep learning-based method for identifying the status of fully mechanized mining equipment and implementing Wi-Fi 6 transmission is proposed. Summary of the Invention

[0004] The purpose of this invention is to provide a mine fully mechanized mining equipment status recognition and Wi-Fi6 transmission method based on deep learning.

[0005] The purpose of the present invention can be achieved through the following technical solutions:

[0006] The deep learning-based method for identifying and transmitting the state of fully-mechanized mining equipment in a mine is described. The method involves collecting key operating parameters and environmental vibration data of fully-mechanized mining equipment in real time through cameras and multiple sensors deployed underground, forming a joint space-time-vibration database. Vibration multipath path loss values ​​are then obtained through multipath effect analysis. The data from the joint space-time-vibration database is input into a preset deep learning model to output a baseline for each device state. A space-state-multipath matrix incorporating vibration multipath characteristics is then constructed to extract a set of baseline feature vectors.

[0007] Based on the extracted baseline feature vector set, key features are activated through weighted nonlinear transformation and ReLU function. Iterative training of convolutional neural networks is introduced to generate a multidimensional feature map of the device. A device health score is obtained based on device status indicators, vibration multipath path loss values, and a preset upper limit on retransmission times. The multidimensional feature map of the device is then screened to obtain a multidimensional feature map of abnormal devices.

[0008] Mapping the multi-dimensional feature map of abnormal devices to anomaly scores and comparing them with the preset initial switching threshold to achieve anomaly classification. Matching corresponding recognition modes is performed based on the anomaly classification. This also includes the identification and control of concurrent anomaly reporting in the enhanced recognition mode.

[0009] A comprehensive evaluation is conducted on the actual recognition and multipath compensation effects under different recognition modes, including the status recognition accuracy, wireless communication stability, and resource consumption of each abnormal device. Based on the evaluation results, the initial switching threshold is automatically adjusted, and the device adaptability of the abnormal recognition and compensation strategy is intelligently optimized according to the multipath compensation performance in different modes.

[0010] As a further solution of the present invention: forming a joint spatiotemporal-vibration database, and obtaining a vibration multipath path loss value through multipath effect analysis, including: aligning key operating parameters and environmental vibration data according to timestamps and device spatial positions, constructing a spatiotemporal index for each data item, with each data item using the timestamp as the primary key and the device ID and spatial coordinates as auxiliary indexes, and obtaining a joint spatiotemporal-vibration feature through alignment and integration;

[0011] A time series database is used to store sensor data, a relational database is used to manage device metadata, and a spatiotemporal-vibration joint database is obtained by associating multi-source data with a unified timestamp.

[0012] Based on the space-time-vibration joint database, the vibration acceleration time series of each device at different time points and spatial positions are extracted; then, based on the geometric dimensions of the downhole environment and the relative position relationship between the sensor and the device, the number and distribution of typical multipath propagation paths are determined; using the preset Rayleigh fading statistical model, the envelope amplitude distribution of each path is calculated; combined with the small displacement changes caused by the actual measured vibration, the relative phase and amplitude attenuation factors of each path are adjusted, and the instantaneous composite signal of multiple paths is simulated by Monte Carlo; by statistically analyzing the simulation results, the additional loss distribution characteristics of the multipath path are obtained, and then the vibration multipath path loss value L is obtained. i , where i is the index of the mine fully-mechanized mining equipment and i=(1,2,...,n), n is the number of mine fully-mechanized mining equipment and n is a positive integer.

[0013] As a further solution of the present invention: the original multi-parameter data in the spatiotemporal-vibration joint database is input into a pre-trained deep learning model, and the state S of each device is obtained through the device detection branch output in the deep learning model. i ∈{0,1}, where 0 indicates the device is in a normal state and 1 indicates the device is in an abnormal state;

[0014] The baseline feature vector regression branch in the deep learning model is based on reading the actual coordinates (x i ,y i ,z i ); According to the vibration multipath path loss value L corresponding to the device i and the preset upper limit of retransmission times M i , keep a small dimension vector Δ for each device i =[L i ,M i ], and based on the pre-stored image of the entire mine operation area, a fixed-size sparse three-dimensional grid is established. Each grid unit corresponds to a certain spatial voxel. For the i-th device, its coordinates (x i ,y i ,z i ) is mapped to the nearest voxel unit as the position index dimension of the matrix; on the basis of the voxel unit, the state level dimension and the multipath impact dimension are assigned to each device to form a three-dimensional index, and the corresponding position (x, y, z, S i ), fill in the multi-radial quantity Δ corresponding to the device i , and obtain the space-state-multipath matrix.

[0015] As a further solution of the present invention: based on the space-state-multipath matrix, a baseline feature vector set is extracted, including: expanding the three-dimensional space-state-multipath matrix into a one-dimensional vector to obtain the baseline feature vector set at the current time t Among them, S i is the device state corresponding to device i, L i is the vibration multipath path loss value corresponding to device i, M i is the preset upper limit of retransmission times corresponding to device i, C i is the coding rate adjustment coefficient corresponding to device i, and T is the transposed symbol.

[0016] As a further solution of the present invention: generating a multi-dimensional feature map of the device, including: based on the baseline feature vector set of each device i at the current moment obtained in step 1, and the spatial coordinates (x i ,y i ,z i) and type identification; multiply the feature vectors in the baseline feature vector set by the preset first set of weight matrices through weighted nonlinear affine transformation, and add the corresponding preset bias vector. Apply the ReLU activation function element by element on the affine transformation result, and use the output after the previous step of activation as the input of the next layer of affine transformation again. Repeat the ReLU activation operation, and the output of the last layer of mapping is recorded as the intermediate feature vector Where l represents the lth layer currently being calculated and l=1,2,...,L-1, L represents the total number of layers included in the weighted nonlinear affine transformation design, Represents the output of the lth layer;

[0017] Intermediate representation of all devices According to the spatial coordinates (x i ,y i ,z i ) and type identification, mapped to a sparse three-dimensional grid, and spliced ​​to obtain the initial feature map F (0) ; with F (0) As input, it is continuously updated through convolutional neural network iteration. In each iteration, the corresponding vibration multipath path loss value L of each device is extracted i , generating channel-level gain ΔW i , represents the weight vector that needs to be dynamically adjusted on each feature channel of the i-th device, in the initial feature map F (0) Multiply the channel by 1+ΔW i Get the final device multi-dimensional feature map F (L) .

[0018] As a further solution of the present invention: the screening principle of the abnormal device multi-dimensional feature map is: in the device multi-dimensional feature map, for each device i, extract the device status indicator S from its corresponding position i , vibration multipath path loss value L i and the preset upper limit of retransmission times M i , the device health score SP corresponding to each device is obtained by weighted summation i , and the device health score SP corresponding to each device i Compare with the preset device health threshold and calculate the device health score SP i The device index i that is less than the preset device health threshold is collected into the abnormal device set S abn , for each i∈S abn In the device multi-dimensional feature map F (L) The spatial coordinates (x i ,y i ,z i ); Taking the spatial coordinate as the center, intercept the small tensor formed by its neighborhood as the multi-dimensional feature map of the abnormal device of the device Wherein, j is the index of the abnormal device and j=(1,2,...m), and m is the number of abnormal devices screened from the device multi-dimensional feature map.

[0019] As a further solution of the present invention: matching the corresponding recognition pattern according to the abnormal classification includes: for each abnormal device multi-dimensional feature map Through the global average pooling operation, it is compressed into a one-dimensional feature vector V j ;

[0020] The one-dimensional feature vector V j Input to an independent fully connected layer, and generate a scalar anomaly score AS through linear transformation and nonlinear activation function j ; The linear transformation is based on the formula: z j =σ·V j +u; where z j is the intermediate output result of the linear transformation, σ and u are the independent preset weights and preset biases of the fully connected layer; the nonlinear activation is based on the formula: AS j =sigma(z j ); where sigma is the Sigmoid function, which limits the score to the interval [0,1];

[0021] Get the score AS for each abnormal device j Then, compare it with the preset initial switching threshold T low and T high Make comparisons;

[0022] When AS j ≤T low When it is detected, it is marked as a low-level anomaly and the regular recognition mode is maintained;

[0023] When T low <AS j ≤T high When it is detected, it is marked as a medium-level abnormality, and a light compensation is performed while maintaining the normal recognition mode;

[0024] When AS j >T high When it is detected, it is marked as a high-level abnormality and automatically switches to enhanced recognition mode.

[0025] As a further solution of the present invention: a sliding time window ΔT of fixed length win The total number of devices that are judged as high-level abnormalities in the window is counted as the sliding window device statistics value m abn (t), according to the formula: m abn (t)=|{j|AS j (t')≥Thigh ,t'∈[t-ΔT win ,t]}|; Among them, AS j (t') represents the abnormal score of the jth device at time t', t'∈[t-ΔT win ,t] represents all moments t' from the current moment t back to t-ΔT win The states within this interval are all included in the statistics;

[0026] Set the sliding window device statistics value m abn (t) and the preset concurrent reporting threshold m th For comparison, the preset concurrent reporting threshold m th Indicates that at any ΔT win The maximum number of high-level abnormal devices allowed to enter enhanced mode within the window;

[0027] When the sliding window device statistics value m abn (t) Satisfy the preset concurrent reporting threshold m th This means that it is determined to be a high-concurrency exception reporting scenario. From then on until the end of the next sliding window, the system enters a high-concurrency current limiting state.

[0028] As a further solution of the present invention: Strengthen the control of concurrent abnormality reporting in the identification mode, including: abnormality score AS based on abnormal equipment j and Device Health Score SP j , the priority YP of each abnormal device is obtained by weighted summation j ; According to the priority YP of each abnormal device j Sort all abnormal devices in descending order to generate a priority queue;

[0029] Based on the generated priority queue, take the first m abn The abnormal devices are recorded as the enhanced identification device set of this sliding window and enter the enhanced identification mode, and the remaining devices are postponed to the next sliding window for re-evaluation.

[0030] As a further solution of the present invention, the initial switching threshold is automatically adjusted, including: calculating the state recognition accuracy, wireless communication stability, and resource consumption score for each abnormal device j within a preset period; for each abnormal device, subtracting the resource consumption and stability penalties from the recognition accuracy in the corresponding recognition mode to obtain a comprehensive performance score; averaging the conventional recognition mode scores and enhanced recognition mode scores of all devices to obtain two global averages, and comparing the difference between the two;

[0031] When the global average score in the enhanced recognition mode is greater than that in the conventional recognition mode, the enhanced recognition mode is judged to be excellent; the switching threshold T originally used for graded switching is automatically lowered by the preset downward adjustment value.low and T high ;

[0032] When the global average score in the enhanced recognition mode is less than or equal to the normal recognition mode, the switching threshold T is automatically increased by the preset upward value. low and T high , reducing the trigger range of enhanced recognition mode.

[0033] Beneficial effects of the present invention:

[0034] (1) This invention, through deep integration of vibration multipath effect analysis and equipment status recognition, innovatively quantifies the multipath path loss value caused by downhole equipment vibration, combines the Rayleigh fading model with Monte Carlo simulation to accurately characterize the dynamic characteristics of the channel; constructs a space-state-multipath three-dimensional matrix, associates equipment coordinates, state labels and multi-radial quantities, and provides structured input for deep learning. Based on weighted nonlinear affine transformation and ReLU activation, it generates a multi-dimensional feature map of the equipment, and dynamically optimizes the feature weights through channel-level gain, significantly improving the model's expression ability; at the same time, it designs a dynamic classification mechanism: screens abnormal equipment based on the equipment health score, triggers a three-level recognition mode based on the abnormal score, and introduces sliding window concurrency control and priority scheduling to achieve on-demand resource allocation; through comprehensive evaluation of recognition accuracy, communication stability and resource overhead, it automatically optimizes the switching threshold to form an adaptive closed loop; the high bandwidth and low latency characteristics of Wi-Fi6 support real-time backhaul of multi-source data, ensuring efficient system collaboration. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The present invention will be further described below with reference to the accompanying drawings.

[0036] Figure 1 This is a schematic diagram of the overall method flow of the mine fully mechanized mining equipment status recognition and Wi-Fi6 transmission method based on deep learning of the present invention;

[0037] Figure 2 This is a schematic diagram of the method for screening the multi-dimensional feature map of abnormal devices in step 2 of the present invention;

[0038] Figure 3 Schematic diagram of the method for matching the recognition pattern in step three of the present invention;

[0039] Figure 4 It is a schematic diagram of the method for automatically adjusting the initial switching threshold in step four of the present invention. DETAILED DESCRIPTION

[0040] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts shall fall within the scope of protection of the present invention.

[0041] Example 1

[0042] See also Figure 1 As shown, the present invention is a method for identifying the state of fully mechanized mining equipment in a mine and transmitting it via Wi-Fi 6 based on deep learning, comprising the following steps:

[0043] Step 1: Through cameras and multi-type sensors deployed underground, key operating parameters and environmental vibration data of mine fully-mechanized mining equipment are collected in real time to form a space-time-vibration joint database. Vibration multipath path loss values ​​are obtained through multipath effect analysis. The original multi-parameter data is input into a preset deep learning model to output a sub-equipment state baseline, and a space-state-multipath matrix integrating vibration multipath characteristics is constructed to extract a baseline feature vector set. Specifically, the mine fully-mechanized mining equipment includes coal mining machines, hydraulic supports, etc., and each mine fully-mechanized mining equipment is assigned a unique equipment number, namely, an equipment ID, and basic information such as equipment type and location is registered in the system. Key operating parameters include temperature, current, and pressure; environmental vibration data are vibration acceleration and equipment displacement amplitude.

[0044] All data is aligned by timestamp and device spatial location, and a spatiotemporal index is constructed for each data item. Each data item uses the timestamp as the primary key and the device ID and spatial coordinates as auxiliary indexes. Through alignment, the above multi-source data is integrated into a spatiotemporal-vibration joint feature.

[0045] A time series database is used to store sensor data, a relational database is used to manage device metadata, and a spatiotemporal-vibration joint database is obtained by associating multi-source data with a unified timestamp.

[0046] Based on the obtained space-time-vibration joint database, the vibration acceleration time series of each device at different time points and spatial positions are extracted; then, according to the geometric dimensions of the downhole environment and the relative position relationship between the sensor and the equipment, the number and distribution of typical multipath propagation paths are determined; then, using the preset Rayleigh fading statistical model, the Rayleigh fading statistical model calculates the envelope amplitude distribution of each path on the basis of assuming that the signal amplitude on each reflection or scattering path obeys a complex random variable with zero mean and Gaussian distribution; on this basis, combined with the small displacement changes caused by the actual measured vibration, the relative phase and amplitude attenuation factors of each path are adjusted, and the instantaneous synthetic signals of multiple paths are simulated by Monte Carlo; finally, by statistically analyzing the simulation results, the additional loss distribution characteristics of the multipath path caused by vibration are obtained, and then the vibration multipath path loss value L is obtained. i , where i is the index of the mine fully-mechanized mining equipment and i=(1,2,...,n), n is the number of mine fully-mechanized mining equipment and n is a positive integer;

[0047] It should be noted that the Rayleigh fading statistical model is an existing mature technology and will not be described in detail in this embodiment.

[0048] The original multi-parameter data in the spatiotemporal-vibration joint database is input into the pre-trained deep learning model (convolutional network shown in the figure), and the state S of each device is obtained through the device detection branch output in the deep learning model. i ∈{0,1}, where 0 indicates the device is in a normal state and 1 indicates the device is in an abnormal state;

[0049] Furthermore, the baseline feature vector regression branch in the deep learning model reads the actual coordinates (x i ,y i ,z i ); According to the vibration multipath path loss value L corresponding to the device i and the preset upper limit of retransmission times M i , keep a small dimension vector Δ for each device i =[L i ,M i ], used to characterize the vibration driving effect of the device on the wireless channel at the current moment; and based on the pre-stored image of the entire mine operation area, a fixed-size sparse three-dimensional grid is established, each grid unit corresponds to a certain spatial voxel, and for the i-th device, its coordinates (x i ,y i ,z i) is mapped to the nearest voxel unit as the position index dimension of the matrix; on the basis of the voxel unit, the state level dimension and the multipath impact dimension are assigned to each device, forming a three-dimensional index in total, and the multi-radial quantity Δ corresponding to the device is filled in the corresponding position (x, y, z, S) of the three-dimensional matrix i , and obtain the three-dimensional space-state-multipath matrix;

[0050] Furthermore, the above three-dimensional space-state-multipath matrix is ​​expanded into a one-dimensional vector to obtain the baseline feature vector set at the current time t Among them, S i is the device state corresponding to device i, L i is the vibration multipath path loss value corresponding to device i, M i is the preset upper limit of retransmission times corresponding to device i, C i is the coding rate adjustment coefficient corresponding to device i, and T is the transposed symbol.

[0051] See also Figure 2 As shown, step 2: Based on the baseline feature vector set generated in step 1, key features are activated through weighted nonlinear transformation and ReLU function; convolutional neural network iterative training is introduced to generate a multi-dimensional feature map of the device, and the device health score is obtained based on the device status indicator, vibration multipath path loss value and the preset retransmission limit. The multi-dimensional feature map of the device is then screened to obtain a multi-dimensional feature map of abnormal devices;

[0052] Based on the baseline feature vector set of each device i at the current moment obtained in step 1, and the spatial coordinates of the device (x i ,y i ,z i ) and type identification; multiply the feature vectors in the baseline feature vector set by the preset first set of weight matrices through weighted nonlinear affine transformation, and add the corresponding preset bias vector. Apply the ReLU activation function element by element on the affine transformation result, and use the output after the previous step of activation as the input of the next layer of affine transformation again. Repeat the weight product, bias addition and ReLU activation operations. Each layer of ReLU activation clears some unnecessary feature components to zero, making the hidden layer representation naturally sparse. The output of the last layer mapping is recorded as the intermediate feature vector Where l represents the lth layer currently being calculated and l=1,2,...,L-1, L represents the total number of layers included in the weighted nonlinear affine transformation design, Represents the output of the lth layer;

[0053] Intermediate representation of all devices According to the spatial coordinates (x i ,y i ,z i) and type identification, mapped to a sparse three-dimensional grid, and pieced together into an initial feature map F with a shape of A×H×D×E (0) ; Where A is the number of voxels in the grid in the X direction, H is the number of voxels in the grid in the Y direction, D is the number of voxels in the grid in the Z direction, and E is the number of channels carried by each voxel; F (0) As input, it is continuously updated through convolutional neural network iteration. In each iteration, the corresponding vibration multipath path loss value L of each device is extracted i , generating channel-level gain ΔW i , represents the weight vector that needs to be dynamically adjusted on each feature channel of the i-th device, and then in the initial feature map F (0) Multiply the channel by 1+ΔW i Get the final device multi-dimensional feature map F (L) ;

[0054] Furthermore, in the multi-dimensional feature map of the device, for each device i, the device status indicator S is first extracted from its corresponding position i , vibration multipath path loss value L i and the preset upper limit of retransmission times M i , the device health score SP corresponding to each device is obtained by weighted summation i , and the device health score SP corresponding to each device i Compare with the preset device health threshold and calculate the device health score SP i The device index i that is less than the preset device health threshold is collected into the abnormal device set S abn , for each i∈S abn In the device multi-dimensional feature map F (L) The spatial coordinates (x i ,y i ,z i ); Taking the spatial coordinate as the center, intercept the small tensor formed by its neighborhood as the multi-dimensional feature map of the abnormal device of the device Wherein, j is the index of the abnormal device and j=(1,2,...m), and m is the number of abnormal devices screened from the device multi-dimensional feature map.

[0055] The multi-dimensional feature map of abnormal equipment obtained in this embodiment is constructed based on the spatiotemporal-vibration joint database generated in step one and the equipment health score extracted in step two, providing multi-dimensional feature support for abnormality classification.

[0056] See also Figure 3As shown, step three: map the multi-dimensional feature map of the abnormal device to the abnormal score, and compare it with the preset initial switching threshold to achieve abnormal classification, match the corresponding recognition mode according to the abnormal classification, and also include the identification and control of concurrent abnormal reporting in the enhanced recognition mode;

[0057] S1: Map the abnormal device multi-dimensional feature map to the abnormal score and compare it with the preset initial switching threshold to achieve abnormal classification, including: For each abnormal device multi-dimensional feature map Through the Global Average Pooling (GAP) operation, it is compressed into a one-dimensional feature vector V j ;

[0058] The one-dimensional feature vector V j Input to a fully connected layer, the one-dimensional feature vector V j Input to an independent fully connected layer, and generate a scalar anomaly score AS through linear transformation and nonlinear activation function j ; The linear transformation is based on the formula: z j =σ·V j +u; where z j is the intermediate output result of the linear transformation, σ and u are the independent preset weights and preset biases of the fully connected layer, and have no shared relationship with the weight matrix of step 2; the nonlinear activation is based on the formula: AS j =sigma(z j ); where sigma is the Sigmoid function, which limits the score to the interval [0,1];

[0059] Get the score AS for each abnormal device j Then, compare it with the preset initial switching threshold T low and T high Make comparisons;

[0060] When AS j ≤T low When it is detected, it is marked as a low-level anomaly and the regular recognition mode is maintained;

[0061] When T low <AS j ≤T high When the error is detected, it is marked as a medium-level abnormality. On the basis of maintaining the normal recognition mode, a lightweight dynamic compensation strategy is added based on the preset adjustment range, such as a small adjustment of the upper limit of the number of retransmissions.

[0062] When AS j >T highWhen the device is marked as a high-level abnormality, it will automatically switch to the enhanced recognition mode; the enhanced recognition mode includes training round improvement, dynamic adjustment of learning rate, batch size and other hyper parameters according to the vibration multipath loss value L of the device. i and real-time communication quality feedback to dynamically adjust compensation parameters;

[0063] S2: When the system automatically switches to enhanced recognition mode, it also includes the identification and control of high-concurrency exception reporting, including:

[0064] The identification conditions for high concurrency exception reporting are: a sliding time window of fixed length ΔT win The total number of devices that are judged as high-level abnormalities in the window is counted as the sliding window device statistics value m abn (t), according to the formula: m abn (t)=|{j|AS j (t')≥T high ,t'∈[t-ΔT win ,t]}|; Among them, AS j (t') represents the abnormal score of the jth device at time t', t'∈[t-ΔT win ,t] represents all moments t' from the current moment t back to t-ΔT win The states within this interval are all included in the statistics;

[0065] Set the sliding window device statistics value m abn (t) and the preset concurrent reporting threshold m th For comparison, the preset concurrent reporting threshold m th Indicates that at any ΔT win The maximum number of high-level abnormal devices allowed to enter enhanced mode within the window;

[0066] When the sliding window device statistics value m abn (t) Satisfy the preset concurrent reporting threshold m th This means that the scenario is determined to be a high-concurrency exception reporting scenario. From then on, until the end of the next sliding window, the system enters a high-concurrency current limiting state;

[0067] Abnormal score AS based on abnormal devices j and Device Health Score SP j , the priority YP of each abnormal device is obtained by weighted summation j ; According to the priority YP of each abnormal device j Sort all abnormal devices in descending order to generate a priority queue;

[0068] Based on the generated priority queue, take the first m abnThe abnormal devices are recorded as the enhanced identification device set of this sliding window, and the remaining devices are postponed to the next sliding window for re-evaluation;

[0069] The abnormal reporting control process is that during the remaining time of this sliding window, only the devices in the enhanced recognition device set can enter the enhanced recognition mode, increase the training rounds and compensation intensity, and initiate high-priority reporting to the monitoring platform. After the window ends, the enhanced recognition device set is re-counted and updated according to the same process.

[0070] See also Figure 4 As shown, step 4: Comprehensively evaluate the actual recognition and multipath compensation effects under different recognition modes, including the status recognition accuracy, wireless communication stability, and resource consumption of each abnormal device. Based on the evaluation results, the initial switching threshold is automatically adjusted, and the device adaptability of the abnormal recognition and compensation strategy is intelligently optimized according to the multipath compensation performance in different modes. This includes:

[0071] During the entire identification and transmission process, the key operating parameters, identification status, multipath compensation parameters, and corresponding wireless communication quality indicators such as signal strength RSSI, packet loss rate, latency, and throughput of each mine fully mechanized mining equipment are transmitted back to the ground monitoring platform in real time via the Wi-Fi 6 network. Each piece of data contains a unique device ID, timestamp, spatial location, current identification mode, status judgment result and its confidence level, current multipath compensation parameters including path loss, number of retransmissions, compensation gain, and resource usage including CPU, memory, and bandwidth consumption.

[0072] For each abnormal device j, the state recognition accuracy, wireless communication stability, and resource expenditure score are calculated within a preset period. The state recognition accuracy is the number of correct judgments divided by the total number of judgments. The wireless communication stability is a stability score that combines the retransmission rate, packet loss rate, and RSSI fluctuations. The resource expenditure score is a weighted calculation based on the proportion of resource units used by the device in this period and the average transmit power.

[0073] Mark each abnormal device's status recognition accuracy, wireless communication stability, and resource overhead score in different modes and store them in the evaluation database;

[0074] For each abnormal device, a comprehensive performance score is obtained by weighting the recognition accuracy in the corresponding recognition mode by subtracting the resource consumption and stability penalties. The scores of all devices in the conventional recognition mode and the enhanced recognition mode are averaged to obtain two global averages, and the difference between the two is compared.

[0075] It should be noted that the conventional recognition mode includes low-level anomalies and medium-level anomalies;

[0076] When the global average score in the enhanced recognition mode is greater than that in the conventional recognition mode, the enhanced recognition mode is judged to be excellent; the switching threshold T originally used for graded switching is automatically lowered by the preset downward adjustment value. low and T high , so that more devices can enter enhanced mode;

[0077] When the global average score in the enhanced recognition mode is less than or equal to the normal recognition mode, the switching threshold T is automatically increased by the preset upward value. low and T high , reduce the trigger range of the enhanced recognition mode;

[0078] When adjusting, always ensure that the threshold is between the minimum and maximum boundaries set by the system;

[0079] Furthermore, a comprehensive evaluation of multipath compensation effectiveness includes: extracting the compensation parameters used in both modes and the corresponding communication stability and resource consumption scores for a small number of devices with the worst performance during the current cycle; using an optimization algorithm, finding a new set of compensation parameters that improves the stability scores of this group of abnormal devices while maintaining resource consumption within the original level; clustering and averaging these new parameters by device type or geographic location to derive the optimal compensation parameters for each abnormal device group; and returning the optimized parameters to the identification pattern matching in step three, replacing the old parameters.

[0080] Through the Wi-Fi 6 wireless communication group, the device status identification and multipath compensation results with device numbers are transmitted back to the ground monitoring platform in real time, and the vibration-related multipath compensation process is recorded in the database according to the device ID; the monitoring platform regularly performs statistical analysis on the identification results and communication quality of each device; and based on the adaptive learning algorithm, it automatically adjusts the deep learning model parameters and multipath compensation mechanism, and continuously optimizes the status identification and overall wireless communication transmission performance of each device according to the vibration and data changes.

[0081] In this embodiment, a joint space-time-vibration database is constructed by collecting equipment operating parameters and environmental vibration data; the Rayleigh fading model and Monte Carlo simulation are combined to analyze the vibration multipath effect and calculate the path loss value; a deep learning model is used to construct a space-state-multipath matrix and generate a multidimensional feature map of the equipment. Abnormal equipment is screened based on the health score and analyzed to obtain an abnormal score; a three-level dynamic response is achieved by comparing with the preset threshold: normal mode and enhanced recognition mode; for high-concurrency scenarios, a sliding window statistics and priority current limiting mechanism are adopted. By evaluating the recognition accuracy, communication stability and resource consumption, the threshold is adaptively adjusted and the multipath compensation strategy is optimized; this method significantly improves the accuracy of downhole equipment status recognition and communication reliability, while ensuring resource utilization efficiency.

[0082] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters and thresholds in the formulas are set by technicians in this field according to actual conditions.

[0083] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A method for identifying the state of fully mechanized mining equipment in a mine and transmitting it via Wi-Fi 6 based on deep learning, characterized in that: include: Through cameras and multi-type sensors deployed underground, key operating parameters and environmental vibration data of mine fully mechanized mining equipment are collected in real time to form a joint space-time-vibration database. After multipath effect analysis, the vibration multipath path loss value is obtained. The data from the joint space-time-vibration database is input into a preset deep learning model, which outputs the status baseline of each device. A space-state-multipath matrix integrating the vibration multipath characteristics is constructed to extract a set of baseline feature vectors. Based on the extracted baseline feature vector set, key features are activated through weighted nonlinear transformation and ReLU function. Iterative training of convolutional neural networks is introduced to generate a multidimensional feature map of the device. A device health score is obtained based on device status indicators, vibration multipath path loss values, and a preset upper limit on retransmission times. The multidimensional feature map of the device is then screened to obtain a multidimensional feature map of abnormal devices. Mapping the multi-dimensional feature map of abnormal devices to anomaly scores and comparing them with the preset initial switching threshold to achieve anomaly classification. Matching corresponding recognition modes is performed based on the anomaly classification. This also includes the identification and control of concurrent anomaly reporting in the enhanced recognition mode. A comprehensive evaluation is conducted on the actual recognition and multipath compensation effects under different recognition modes, including the status recognition accuracy, wireless communication stability, and resource consumption of each abnormal device. Based on the evaluation results, the initial switching threshold is automatically adjusted, and the device adaptability of the abnormal recognition and compensation strategy is intelligently optimized according to the multipath compensation performance in different modes.

2. The method for identifying the state of fully mechanized mining equipment and Wi-Fi 6 transmission based on deep learning according to claim 1 is characterized in that: A joint spatiotemporal-vibration database is formed. After multipath effect analysis, the vibration multipath path loss value is obtained. This includes: aligning key operating parameters and environmental vibration data by timestamp and device spatial location, constructing a spatiotemporal index for each data item. Each data item uses the timestamp as the primary key and the device ID and spatial coordinates as auxiliary indexes. The joint spatiotemporal-vibration features are obtained through alignment and integration. A time series database is used to store sensor data, a relational database is used to manage device metadata, and a spatiotemporal-vibration joint database is obtained by associating multi-source data with a unified timestamp. Based on the space-time-vibration joint database, the vibration acceleration time series of each device at different time points and spatial positions are extracted; then, based on the geometric dimensions of the downhole environment and the relative position relationship between the sensor and the device, the number and distribution of typical multipath propagation paths are determined; using the preset Rayleigh fading statistical model, the envelope amplitude distribution of each path is calculated; combined with the small displacement changes caused by the actual measured vibration, the relative phase and amplitude attenuation factors of each path are adjusted, and the instantaneous composite signal of multiple paths is simulated by Monte Carlo; by statistically analyzing the simulation results, the additional loss distribution characteristics of the multipath path are obtained, and then the vibration multipath path loss value L is obtained. i , where i is the index of the mine fully-mechanized mining equipment and i=(1,2,...,n), n is the number of mine fully-mechanized mining equipment and n is a positive integer.

3. The method for identifying the state of fully mechanized mining equipment and Wi-Fi 6 transmission based on deep learning according to claim 1 is characterized in that: Constructing a space-state-multipath matrix that integrates vibration multipath characteristics includes: inputting the original multi-parameter data in the space-time-vibration joint database into a pre-trained deep learning model, and obtaining the state S of each device through the device detection branch output in the deep learning model. i ∈{0,1}, where 0 indicates the device is in a normal state and 1 indicates the device is in an abnormal state; The baseline feature vector regression branch in the deep learning model is based on reading the actual coordinates (x i ,y i ,z i ); According to the vibration multipath path loss value L corresponding to the device i and the preset upper limit of retransmission times M i , keep a small dimension vector Δ for each device i =[L i ,M i ], and based on the pre-stored image of the entire mine operation area, a fixed-size sparse three-dimensional grid is established. Each grid unit corresponds to a certain spatial voxel. For the i-th device, its coordinates (x i ,y i ,z i ) is mapped to the nearest voxel unit as the position index dimension of the matrix; on the basis of the voxel unit, the state level dimension and the multipath impact dimension are assigned to each device to form a three-dimensional index, and the corresponding position (x, y, z, S i ), fill in the multi-radial quantity Δ corresponding to the device i , and obtain the space-state-multipath matrix.

4. The method for identifying the state of fully mechanized mining equipment and Wi-Fi 6 transmission based on deep learning according to claim 3 is characterized in that: Based on the space-state-multipath matrix, a baseline feature vector set is extracted, including: expanding the three-dimensional space-state-multipath matrix into a one-dimensional vector to obtain the baseline feature vector set at the current time t Among them, S i is the device state corresponding to device i, L i is the vibration multipath path loss value corresponding to device i, M i is the preset upper limit of retransmission times corresponding to device i, C i is the coding rate adjustment coefficient corresponding to device i, and T is the transposed symbol.

5. The method for identifying the state of fully mechanized mining equipment and Wi-Fi 6 transmission based on deep learning according to claim 1 is characterized in that: Generate a multi-dimensional feature map of the device, including: the baseline feature vector set of each device i at the current moment obtained in step 1, and the spatial coordinates of the device (x i ,y i ,z i ) and type identification; multiply the feature vectors in the baseline feature vector set by the preset first set of weight matrices through weighted nonlinear affine transformation, and add the corresponding preset bias vector. Apply the ReLU activation function element by element on the affine transformation result, and use the output after the previous step of activation as the input of the next layer of affine transformation again. Repeat the ReLU activation operation, and the output of the last layer of mapping is recorded as the intermediate feature vector Where l represents the lth layer currently being calculated and l=1,2,...,L-1, L represents the total number of layers included in the weighted nonlinear affine transformation design, Represents the output of the lth layer; Intermediate representation of all devices According to the spatial coordinates (x i ,y i ,z i ) and type identification, mapped to a sparse three-dimensional grid, and spliced ​​to obtain the initial feature map F (0) ; with F (0) As input, it is continuously updated through convolutional neural network iteration. In each iteration, the corresponding vibration multipath path loss value L of each device is extracted i , generating channel-level gain ΔW i , represents the weight vector that needs to be dynamically adjusted on each feature channel of the i-th device, in the initial feature map F (0) Multiply the channel by 1+ΔW i Get the final device multi-dimensional feature map F (L) .

6. The method for identifying the state of fully mechanized mining equipment and Wi-Fi 6 transmission based on deep learning according to claim 5 is characterized in that: The screening principle of the abnormal device multi-dimensional feature map is as follows: for each device i in the device multi-dimensional feature map, extract the device status indicator S from its corresponding position. i , vibration multipath path loss value L i and the preset upper limit of retransmission times M i , the device health score SP corresponding to each device is obtained by weighted summation i , and the device health score SP corresponding to each device i Compare with the preset device health threshold and calculate the device health score SP i The device index i that is less than the preset device health threshold is collected into the abnormal device set S abn , for each i∈S abn In the device multi-dimensional feature map F (L) The spatial coordinates (x i ,y i ,z i ); Taking the spatial coordinate as the center, intercept the small tensor formed by its neighborhood as the multi-dimensional feature map of the abnormal device of the device Wherein, j is the index of the abnormal device and j=(1,2,...m), and m is the number of abnormal devices screened from the device multi-dimensional feature map.

7. The method for identifying the state of fully mechanized mining equipment and Wi-Fi 6 transmission based on deep learning according to claim 1 is characterized in that: Matching the corresponding identification pattern according to the abnormal classification includes: multi-dimensional feature map of abnormal devices for each abnormal device Through the global average pooling operation, it is compressed into a one-dimensional feature vector V j ; The one-dimensional feature vector V j Input to an independent fully connected layer, and generate a scalar anomaly score AS through linear transformation and nonlinear activation function j ; The linear transformation is based on the formula: z j =σ·V j +u; where z j is the intermediate output result of the linear transformation, σ and u are the independent preset weights and preset biases of the fully connected layer; the nonlinear activation is based on the formula: AS j =sigma(z j ); where sigma is the Sigmoid function, which limits the score to the interval [0,1]; Get the score AS for each abnormal device j Then, compare it with the preset initial switching threshold T low and T high Make comparisons; When AS j ≤T low When it is detected, it is marked as a low-level anomaly and the regular recognition mode is maintained; When T low <AS j ≤T high When it is detected, it is marked as a medium-level abnormality, and a light compensation is performed while maintaining the normal recognition mode; When AS j >T high When it is detected, it is marked as a high-level abnormality and automatically switches to enhanced recognition mode.

8. The method for identifying the state of fully mechanized mining equipment and Wi-Fi 6 transmission based on deep learning according to claim 7 is characterized in that: Strengthen the identification conditions for concurrent exception reporting in identification mode, including: a sliding time window of fixed length ΔT win The total number of devices that are judged as high-level abnormalities in the window is counted as the sliding window device statistics value m abn (t), according to the formula: m abn (t)=|{j|AS j (t')≥T high ,t'∈[t-ΔT win ,t]}|; Among them, AS j (t') represents the abnormal score of the jth device at time t', t'∈[t-ΔT win ,t] represents all moments t' from the current moment t back to t-ΔT win The states within this interval are all included in the statistics; Set the sliding window device statistics value m abn (t) and the preset concurrent reporting threshold m th For comparison, the preset concurrent reporting threshold m th Indicates that at any ΔT win The maximum number of high-level abnormal devices allowed to enter enhanced mode within the window; When the sliding window device statistics value m abn (t) Satisfy the preset concurrent reporting threshold m th This means that it is determined to be a high-concurrency exception reporting scenario. From then on until the end of the next sliding window, the system enters a high-concurrency current limiting state.

9. The method for identifying the state of fully mechanized mining equipment and Wi-Fi 6 transmission based on deep learning according to claim 8 is characterized in that: Strengthen the control of concurrent anomaly reporting in identification mode, including: anomaly score AS based on abnormal devices j and Device Health Score SP j , the priority YP of each abnormal device is obtained by weighted summation j ; According to the priority YP of each abnormal device j Sort all abnormal devices in descending order to generate a priority queue; Based on the generated priority queue, take the first m abn The abnormal devices are recorded as the enhanced identification device set of this sliding window and enter the enhanced identification mode, and the remaining devices are postponed to the next sliding window for re-evaluation.

10. The method for state identification and Wi-Fi 6 transmission of mine fully mechanized mining equipment based on deep learning according to claim 1 is characterized in that: Automatically adjust the initial switching threshold, including: calculating the state recognition accuracy, wireless communication stability, and resource consumption scores for each abnormal device j within a preset period; for each abnormal device, subtracting the resource consumption and stability penalties from the recognition accuracy weighted by the corresponding recognition mode to obtain a comprehensive performance score; averaging the conventional recognition mode scores and enhanced recognition mode scores of all devices to obtain two global averages, and comparing the difference between the two; When the global average score in the enhanced recognition mode is greater than that in the conventional recognition mode, the enhanced recognition mode is judged to be excellent; the switching threshold T originally used for graded switching is automatically lowered by the preset downward adjustment value. low and T high ; When the global average score in the enhanced recognition mode is less than or equal to the normal recognition mode, the switching threshold T is automatically increased by the preset upward value. low and T high , reducing the trigger range of enhanced recognition mode.

Images