A double-in double-out coal mill coal quality measurement polling method and system

Abstract

The application provides a double-in double-out coal mill coal quality measurement polling method and system, relating to the field of industrial equipment state monitoring and wireless communication. The method comprises: receiving the rotation angle of the wireless sensor node relative to the stationary base station, calling the channel quality mapping table to predict the expected channel quality of each angle interval in the next rotation period; dividing the rotation period into fixed time slots, matching the expected signal-to-noise ratio with the priority weight, and dynamically adjusting the coding scheme and transmission power of the node; receiving the three-axis vibration signal sent by the node in the fixed time slot, generating a feature vector through the lightweight one-dimensional convolutional neural network in the node and uploading; constructing a state matrix based on the feature vectors of all nodes and inputting it into the fully connected neural network layer for material level inference. The application solves the problems of low wireless communication reliability and unsynchronized time of multi-node data in a rotating metal sealed environment, and realizes continuous and reliable monitoring of the material level of the coal mill and accurate fusion of multi-source data.

Description

Technical Field

[0001] This application relates to the field of industrial equipment condition monitoring and wireless communication, and in particular to a method and system for polling coal quality measurement in a dual-inlet dual-outlet coal mill. Background Technology

[0002] The double-inlet, double-outlet coal mill is a core piece of equipment in the pulverizing system of a thermal power plant. Real-time monitoring of the material level (coal-to-ball ratio) inside the mill is crucial for controlling the coal feed rate and preventing accidents such as dry milling or full milling. Currently, the mainstream non-contact vibration telemetry method typically involves installing wireless vibration sensors on the surface of the rotating mill cylinder to collect vibration signals during the cylinder's rotation and transmitting them wirelessly to a stationary base station for analysis.

[0003] However, in practical applications, due to the continuous rotation of the coal mill cylinder and its large metal structure, the relative positions of the transmitting and receiving antennas change periodically as the sensor rotates with the cylinder. This causes the wireless channel quality to fluctuate strongly with the angle. When the sensor rotates to the side facing away from the base station, the signal strength drops sharply, and the packet loss rate increases. Furthermore, to improve monitoring accuracy, multiple sensor nodes are typically arranged along the cylinder's axis. However, because each node collects data independently and lacks a unified time reference, the data from multiple nodes is not synchronized, making effective multi-source feature fusion analysis impossible. Summary of the Invention

[0004] To address the issues of low wireless communication reliability and asynchronous data time among multiple nodes in existing technologies under rotating metal enclosed environments, this application proposes a polling method and system for coal quality measurement in a dual-input, dual-output coal mill. This method utilizes rotation period prediction and channel quality mapping to achieve deterministic communication scheduling, and combines edge computing feature compression and time synchronization mechanisms to achieve continuous and reliable monitoring of coal mill material level and accurate fusion of multi-source data.

[0005] To achieve the above objectives, this application adopts the following technical solution: In a first aspect, a method for polling coal quality measurement in a dual-inlet, dual-outlet coal mill includes: installing angle sensors on wireless sensor nodes on the rotating cylinder of the dual-inlet, dual-outlet coal mill to collect the rotation angle of the rotating cylinder relative to a stationary base station in real time, and transmitting the rotation angle to the stationary base station wirelessly; wherein the angle sensor is a magnetoresistive angle sensor or an incremental encoder, with a collection accuracy of ±0.5°, a sampling frequency ≥100Hz, and the wireless signal must penetrate the metal wall of the rotating cylinder; the stationary base station predicts the channel quality parameters corresponding to each angle interval in the next rotation cycle based on the rotation angle and a pre-generated channel quality mapping table, wherein the channel quality mapping table includes parameters that allow the wireless signal to penetrate. The attenuation behind the metal wall is corrected by a coefficient; the stationary base station divides the rotation period of the rotating cylinder into M fixed time slots, where M = K × N, N is the total number of vibration sensor nodes installed on the surface of the rotating cylinder, and K is the number of time slots allocated to each sensor node per rotation period, with K ranging from 2 to 4; based on the channel prediction results, the stationary base station allocates the K time slots with the highest expected signal-to-noise ratio in each rotation period to the sensor node with the highest current priority, and sends the time slot allocation results to each sensor node before the start of the next rotation period via broadcast signaling; the stationary base station dynamically adjusts the modulation and coding scheme and transmission power of each time slot according to the expected signal-to-noise ratio of each time slot.

[0006] Each vibration sensor node acquires the triaxial vibration acceleration signal of the rotating cylinder at a sampling rate of 1kHz, and each 1024 time points constitute a raw data block. The lightweight one-dimensional convolutional neural network (1D-CNN) built into each node performs forward inference calculation on the raw data block and outputs a 16-dimensional floating-point feature vector. The data size of the raw data block is 6144 bytes, the data size of the feature vector is 64 bytes, and the compression ratio is 96:1. The network structure of the 1D-CNN is as follows: input layer → first convolutional layer → first pooling layer → second convolutional layer → second pooling layer → flattening layer → first fully connected layer → second fully connected layer. The input layer size is 1024×3. The first convolutional layer contains 32 1×3 convolutional kernels. The first pooling layer is a max pooling with a window size of 4 and a stride of 4. The second convolutional layer contains 64 1×1 convolutional kernels. The second pooling layer is a max pooling with a window size of 4 and a stride of 4. The first fully connected layer is 4096→128, and the second fully connected layer is 128→16.

[0007] Each sensor node caches the feature vector generated in the previous cycle and calculates the difference between the current feature vector and the Euclidean distance. If the difference is less than a preset threshold, only a 1-byte "hold" flag is sent; otherwise, the complete feature vector is sent. After receiving the feature vectors from each sensor node, the stationary base station constructs a state matrix. The stationary base station inputs the aggregated state matrix into the fully connected neural network layer at the base station end to complete the material level inference and outputs the inference result to the distributed control system (DCS) for coal feed control and coal mill operation status monitoring.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the process of constructing the channel quality mapping table specifically includes: receiving offline calibration data and constructing a basic mapping table. ,in For angle Signal strength at that location For angle Signal-to-noise ratio at the location. For angle Bit error rate at the receiving angle; Signal arrival time distribution and corresponding power attenuation coefficient Construct a multipath feature matrix Where m is the number of resolvable multipaths; obtain the attenuation coefficient of the wireless signal penetrating the metal wall of the rotating cylinder, and for the angle The signal strength at the specified location is corrected to obtain the corrected signal strength. The basic mapping table, multipath feature matrix, and corrected signal strength are combined. Construct a three-layer channel quality mapping table based on angle index association. ,in: The first layer is , used for real-time channel quality query; The second layer is , used for multipath equalization parameter calculation; The third layer is = , The rate of change of channel quality, It is the acceleration of channel quality changes, used to predict the channel quality change trend at the next moment.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the online update process of the channel quality mapping table specifically includes: receiving data packets from the receiving node, extracting the actual receiving angle and the actual measured channel quality parameters, including the actual received signal strength, actual signal-to-noise ratio, and actual bit error rate; calculating an adaptive forgetting factor based on the channel quality change rate, updating the mapping table value for the corresponding angle, and performing smoothing processing on adjacent angles; calculating the root mean square error (RMSE) of the updated mapping table, confirming the update as valid and storing the update result when the RMSE is less than a first threshold, triggering local recalibration when the RMSE is between the first and second thresholds, and triggering a global recalibration warning when the RMSE is greater than the second threshold; obtaining the history of rotation period changes, using linear regression to predict the channel quality change rate for the next period, and using this prediction for time slot allocation pre-scheduling.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, the process of dynamically adjusting the coding scheme and transmit power of a node specifically includes: dividing the expected signal-to-noise ratio (SNR) into continuous intervals, constructing a mapping relationship between the continuous intervals and the modulation and coding schemes, wherein the continuous intervals include low SNR intervals, medium SNR intervals, high SNR intervals, and ultra-high SNR intervals; calculating the effective data rate based on the modulation and coding scheme, and synchronously adjusting the transmit power of the node based on the continuous intervals; activating a hysteresis comparator when the expected SNR fluctuates at the boundaries of adjacent intervals to prevent frequent switching of the coding scheme; and increasing the transmit power of the node according to the interval where the current expected SNR is located, and using a lower-order modulation and coding scheme for retransmission when a data packet transmission fails.

[0011] In conjunction with the first aspect mentioned above, one possible implementation method for training and deploying a lightweight one-dimensional convolutional neural network specifically includes: In the offline phase, a high-performance server is used as the teacher model, and triaxial vibration data under various operating conditions (empty mill, half-mill, full mill, steel ball loading deviation, and coal grindability variation) is used as the training set to train the lightweight one-dimensional convolutional neural network as the student model through knowledge distillation; the weights of the trained student model are converted from floating-point numbers to fixed-point numbers, and the forward inference computation graph is solidified into the read-only memory of the node's low-power microcontroller unit; in the online inference phase, the node only performs forward inference computation, inputting the real-time collected triaxial vibration data and directly outputting the feature vector.

[0012] In conjunction with the first aspect mentioned above, in one possible implementation, the process of constructing the state matrix specifically includes: obtaining the geographical location or installation angle identifiers of all nodes, and sorting and interpolating the feature vectors from different nodes at the same time according to the corresponding rotation angle to form a two-dimensional state matrix, which is used to reflect the circumferential state distribution of the cylinder; obtaining the two-dimensional state matrices at consecutive times and stacking them to form a three-dimensional spatiotemporal state tensor; and inputting the three-dimensional spatiotemporal state tensor into a fully connected neural network layer to infer the overall material level and coal quality distribution uniformity within the current cylinder.

[0013] In conjunction with the first aspect mentioned above, in one possible implementation, the priority of each node is determined based on the state evaluation value. The priority mapping relationship includes: calculating the state evaluation value of the corresponding node based on the feature vector, sorting the state evaluation values ​​of all nodes from high to low to obtain a priority weight sequence, and determining the baseline number of fixed time slots allocated to each node in each rotation cycle based on the priority weight sequence; sorting the expected signal-to-noise ratio (SNR) of each node in the current rotation cycle from high to low according to the channel quality mapping table to obtain an expected SNR sequence; performing a weighted matching of the expected SNR sequence and the priority weight sequence to allocate more fixed time slots to nodes with high priority and in angle intervals with high expected SNR; marking the corresponding node as an emergency priority node when the state evaluation value exceeds a preset emergency threshold; for emergency priority nodes, forcibly allocating additional fixed time slots exceeding the baseline number, and dynamically preempting fixed time slots from the fixed time slots already allocated to nodes at the end of the priority weight sequence as needed; reducing the number of fixed time slots reserved by the preempted node to a preset minimum guaranteed value, which is adaptively determined based on the total number of nodes in the system and the cycle duration.

[0014] In conjunction with the first aspect mentioned above, one possible implementation also includes a time slot synchronization and timestamp calibration module: at the start of the rotation cycle, a time slot synchronization beacon is broadcast, which contains the current cycle number and a global time reference; when a node sends a triaxial vibration signal, it embeds a local timestamp and converts it into global time using the global time reference; the time deviation of nodes at the same angle in adjacent cycles is calculated, and the frequency division coefficient of the node's local sampling clock is dynamically adjusted according to the time deviation to achieve sub-millisecond synchronization with the rotation cycle of the rotating cylinder.

[0015] Secondly, a coal quality measurement polling system for a dual-inlet, dual-outlet coal mill is provided, comprising: a communication unit for receiving rotation angle, broadcast time slot synchronization beacons, and data packets uploaded by receiving nodes; a processing unit for calling a channel quality mapping table to perform channel prediction, allocating fixed time slots, adjusting the coding scheme and transmission power, constructing a state matrix, and performing material level inference; a storage unit for storing the channel quality mapping table, the fixed weights of a lightweight one-dimensional convolutional neural network, and the model parameters of the fully connected neural network layers; and a wireless sensor node integrating a triaxial vibration sensor, a low-power microcontroller unit, and a wireless transceiver.

[0016] In conjunction with the second aspect above, in one possible implementation, the read-only memory of the low-power microcontroller unit contains a forward inference computation graph of a lightweight one-dimensional convolutional neural network; and a wireless transceiver is used to transmit data packets containing feature vectors within allocated fixed time slots.

[0017] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this application do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a method for polling coal quality measurement in a dual-inlet, dual-outlet coal mill, provided in this application embodiment; Figure 2 The flowchart for constructing and updating the channel quality mapping table in a polling method for coal quality measurement of a dual-input dual-output coal mill provided in this application embodiment; Figure 3 A flowchart of fixed time slot allocation and dynamic scheduling in a coal quality measurement polling method for a dual-inlet dual-outlet coal mill provided in this application embodiment; Figure 4 A lightweight one-dimensional convolutional neural network structure diagram is provided in the polling method for coal quality measurement of a dual-inlet dual-outlet coal mill according to an embodiment of this application. Figure 5The flowchart of knowledge distillation training and model deployment in a coal quality measurement polling method for a dual-inlet dual-outlet coal mill provided in this application embodiment; Figure 6 The flowchart of state matrix construction and material level inference in a coal quality measurement polling method for a dual-inlet dual-outlet coal mill provided in this application embodiment; Figure 7 A flowchart of node priority scheduling and time slot allocation in a polling method for coal quality measurement of a dual-inlet dual-outlet coal mill provided in this application embodiment; Figure 8 The flowchart of time slot synchronization and timestamp calibration in a coal quality measurement polling method for a dual-inlet dual-outlet coal mill provided in this application embodiment is shown. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0020] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention.

[0021] Example 1: like Figure 1 As shown, this embodiment provides a polling method for coal quality measurement in a dual-inlet, dual-outlet coal mill. This method deeply integrates wireless communication technology with the physical laws of rotating machinery, solving the technical challenges of unreliable communication and data asynchrony among multiple nodes in a sealed rotating metal environment. The method specifically includes the following steps: Step 100: Receive the rotation angle of the wireless sensor node installed on the rotating cylinder of the double-inlet double-outlet coal mill relative to the stationary base station. The wireless signal needs to penetrate the metal wall of the rotating cylinder.

[0022] The wireless sensor node refers to the vibration monitoring equipment installed on the outer wall of the rotating cylinder of the double-inlet, double-outlet coal mill. It integrates a triaxial accelerometer, a low-power microcontroller unit, and a wireless transceiver. The rotation angle refers to the circumferential position of the sensor node relative to the stationary base station during the cylinder's rotation, usually expressed as an angle value ranging from 0° to 360°. Its acquisition is a fundamental variable for channel quality prediction and time slot scheduling. The angle information is collected in real time by a non-contact magnetoresistive angle sensor or incremental encoder installed at the end of the cylinder. The stationary base station is a data receiving and processing control platform fixedly installed on the ground or support near the coal mill. The rotating cylinder is a large metal structure, and its metal walls have a significant penetration attenuation effect on wireless signals; signals must pass through these metal walls to be transmitted from the sensor node to the base station.

[0023] In some implementations, the wireless sensor node can be directly installed on the outer wall of the cylinder before use. This allows the cylinder of the dual-inlet, dual-outlet coal mill to continuously rotate during operation (typically between 15 and 20 revolutions per minute). A non-contact magnetoresistive angle sensor installed at the end of the cylinder collects rotation angle data in real time at a frequency of at least 100Hz. This angle data is then transmitted to a stationary base station via a separate low-rate wireless channel. Upon receiving the current angle, the base station uses it as a core input variable for subsequent channel prediction and time slot allocation.

[0024] Meanwhile, the base station pre-stores an offline channel quality mapping table. This table is generated through an offline calibration process: with the coal mill stopped, a standard signal source is set up at the sensor installation location on the cylinder surface. The signal strength, signal-to-noise ratio, and bit error rate received by the base station are recorded angle by angle in 1° increments, and corrections are made for metal wall penetration loss, thus establishing the correspondence between angle and channel quality. When the system is running online, the base station can call this mapping table based on the current angle to predict the expected channel quality for each angle interval in the next rotation cycle, thereby providing a basis for subsequent fixed time slot allocation, coding scheme, and transmit power adjustment.

[0025] It should be noted that, due to the large metal structure of the cylinder, the wireless signal must penetrate the metal wall of the cylinder to be transmitted from the node to the base station. This results in an extremely complex signal transmission path, which is not only obstructed by line-of-sight propagation but also accompanied by severe multipath effects and penetration attenuation. Therefore, in this step, a non-contact magnetoresistive angle sensor or incremental encoder (with an accuracy of ±0.5 degrees and an update frequency of no less than 100Hz) installed at the end of the cylinder is used to obtain the current rotation angle of the node in real time. This serves as the basic variable for all subsequent prediction and scheduling operations to ensure accurate tracking of the cylinder's rotation position.

[0026] Step 200: Call the channel quality mapping table to predict the rotation angle and obtain the prediction result.

[0027] The channel quality mapping table is a pre-constructed data structure indexed by rotation angle, recording the channel quality parameters of wireless signals penetrating the metal wall of the rotating cylinder at different angles. Expected channel quality refers to predicting the communication channel state for each angle interval within the next rotation cycle based on the current rotation angle, typically using signal-to-noise ratio (SNR) as the primary quantification indicator. The prediction result is the expected SNR sequence for each angle interval in the next cycle, generated by querying the channel quality mapping table and combining it with the current angle information. This sequence is used for subsequent time slot allocation and coding / modulation decisions. The rotation cycle refers to the time required for the rotating cylinder of the dual-inlet, dual-outlet coal mill to complete one 360° rotation; its value is determined by the cylinder's rotational speed.

[0028] In some implementations, a stationary base station receives the current rotation angle from an angle sensor and uses this angle as an index to call a pre-generated channel quality mapping table. It can then look up the channel quality parameters of the corresponding angle and its adjacent angles in the mapping table based on the current angle. Combining the channel quality change rate and change acceleration stored in the mapping table, a linear extrapolation method is used to predict the expected signal-to-noise ratio of each angle interval (e.g., every 30° interval) in the next rotation cycle, and output it in the form of an expected signal-to-noise ratio sequence for subsequent time slot allocation and modulation and coding scheme adjustment.

[0029] It should be noted that due to the rotation of the cylinder, the relative position between the node and the base station changes periodically, causing the wireless channel quality to fluctuate strongly with angle. For example, when the node rotates to the side facing away from the base station, the signal strength may drop sharply by 20 to 40 dB. Therefore, this embodiment constructs an offline calibration process. While the coal mill is stopped, a standard signal source is set at the sensor installation position on the cylinder surface. The signal strength, signal-to-noise ratio, and bit error rate received by the base station are recorded angle by angle in 1° increments. Correction is made for metal wall penetration loss, thereby establishing a channel quality mapping table that includes the correspondence between angle and channel quality, allowing it to record channel parameters at different rotation angles. The additional loss caused by signal penetration through the metal wall is also considered, thus introducing a penetration attenuation correction coefficient to correct the theoretical model or offline measurement values, thereby more accurately reflecting the real channel environment. During use, this mapping table can be directly consulted to predict the channel quality of each angle interval in the next rotation cycle based on the current angle, achieving a shift from passive adaptation to active avoidance.

[0030] Step 300: Divide the rotation period of the rotating cylinder into fixed time slots, match the expected signal-to-noise ratio sequence and priority weight sequence of the prediction results, and dynamically adjust the coding scheme and transmission power of the nodes.

[0031] In this system, a fixed time slot is a time window that divides the rotation period equally on the time axis. Each time slot is allocated to a specific wireless sensor node for data transmission. The priority weight sequence is a list of nodes ranked by their transmission urgency, ordered from highest to lowest based on their state evaluation values. The coding scheme refers to the combination of modulation and channel coding methods used during wireless signal transmission, directly affecting data transmission rate and interference resistance. Transmit power refers to the radio frequency output power level of the wireless sensor node when transmitting signals.

[0032] In some implementations, this embodiment employs a Time Division Multiple Access (TDMA) protocol, dividing a complete rotation cycle into several fixed time slots. For example, for a coal mill with a rotation speed of 18 revolutions per minute, the rotation cycle is approximately 3.33 seconds, which can be divided into 12 time slots, each approximately 278 milliseconds. Based on the channel quality predicted in step S200 and combined with the data priority of each node, high-quality time slots are allocated to high-priority nodes. Simultaneously, the modulation and coding scheme (MCS) and transmit power of the nodes are dynamically adjusted according to the expected signal-to-noise ratio (SNR) of different time slots. For example, 16QAM high-order modulation is used in time slots with high SNR to improve the transmission rate, while BPSK low-order modulation is used in time slots with low SNR to ensure robustness. This prediction-based deterministic scheduling avoids the blind retransmissions that occur in traditional Carrier Sense Multiple Access (CSMA) protocols during channel fading, significantly reducing the packet loss rate.

[0033] Step 400: Receive the triaxial vibration signal sent by the receiving node within a fixed time slot, generate a feature vector representing the coal quality or material level state through a lightweight one-dimensional convolutional neural network deployed in the node, and upload the feature vector as part of the data packet.

[0034] Among them, the triaxial vibration signal refers to the vibration acceleration data of the rotating cylinder in the three orthogonal directions of X, Y, and Z, collected by the triaxial MEMS accelerometer built into the wireless sensor node. The lightweight one-dimensional convolutional neural network refers to a deep learning model that can run on a low-power microcontroller unit, specifically designed to extract features from one-dimensional time-series signals. The feature vector is the low-dimensional numerical sequence output by the lightweight one-dimensional convolutional neural network after performing forward inference calculations on the input triaxial vibration data, used to characterize the current coal quality or material level state. The data packet is a wireless transmission frame in which the node encapsulates the feature vector, local timestamp, and other state information according to a predetermined format.

[0035] In some implementations, the wireless sensor node collects triaxial vibration acceleration signals at a sampling rate of 1kHz at 1024 time points within a fixed time slot allocated to it, forming a raw data block of size 1024×3. Simultaneously, its internal low-power microcontroller unit reads a pre-programmed lightweight one-dimensional convolutional neural network forward inference computation graph from read-only memory, and performs convolution, pooling, and fully connected operations layer by layer using the raw data block as input.

[0036] At this point, the specific structure of the network is as follows: the input layer receives 1024×3 data and performs the following operations: The first convolutional layer uses 32 1×3 convolutional kernels to extract local features, and then outputs a 1024×32 feature map after ReLU activation. The first pooling layer uses max pooling with a window size of 4 and a stride of 4 to compress the feature map size to 256×32. The second convolutional layer uses 64 1×1 convolutional kernels to perform feature transformation, and then performs ReLU activation to output a 256×64 feature map. The second pooling layer also uses max pooling with a window size of 4 and a stride of 4 to compress the feature map to 64×64. The flattening layer expands the 64×64 feature map into a 4096-dimensional vector; The first fully connected layer maps 4096 dimensions to 128 dimensions and activates them using ReLU; The second fully connected layer maps 128 dimensions to a 16-dimensional output feature vector; Finally, the node can encapsulate the 16-dimensional floating-point feature vector (64 bytes in total) with the local timestamp into a data packet, and upload it to the stationary base station via a wireless transceiver within a fixed time slot.

[0037] It should be noted that this embodiment deploys a lightweight one-dimensional convolutional neural network (1D-CNN) within the sensor node, and trains and quantizes this network offline, embedding it into the node's low-power microcontroller unit (MCU). This allows the triaxial vibration signals (e.g., 1024 sampling points) collected by the node to be directly input into the network for forward inference, outputting a low-dimensional feature vector (e.g., 16-dimensional) locally. This achieves feature extraction and data compression at the edge, with a compression ratio of up to 96:1, significantly alleviating the pressure on wireless transmission.

[0038] It should be understood that although this embodiment uses a one-dimensional convolutional neural network as an example, other lightweight neural network models, such as recurrent neural networks (RNNs) or their variants, can also be used in other embodiments, as long as they can achieve the mapping from vibration signals to feature vectors.

[0039] Step 500: Based on the feature vectors of all nodes, construct a state matrix and input it into a fully connected neural network layer to infer the material level and obtain the inference result.

[0040] The state matrix is ​​a two-dimensional array formed by sorting and interpolating feature vectors from different wireless sensor nodes at the same time according to the installation angle of each node on the rotating cylinder. It reflects the material level and coal quality distribution status around the cylinder. The fully connected neural network layer is a feedforward neural network structure deployed on a stationary base station or in a host computer. Each neuron in this layer is connected to all neurons in the previous layer, and it is used to learn the nonlinear mapping relationship between feature vectors and material level. Material level inference refers to the calculation process by which the fully connected neural network layer outputs the overall material level percentage and coal quality distribution uniformity index within the current coal mill cylinder based on the input state matrix or three-dimensional spatiotemporal state tensor. The inference results are used for coal feed control and coal mill operation status monitoring.

[0041] In some implementations, the stationary base station first receives the feature vectors uploaded by all wireless sensor nodes within their respective fixed time slots and extracts the installation angle identifier corresponding to each node. This allows the feature vectors from different nodes at the same time to be directly sorted according to their installation angles from 0° to 360°. For the angle interval between adjacent nodes (e.g., a 60° interval between two nodes), linear interpolation or spline interpolation methods are used to generate feature vectors for the virtual positions, thereby constructing a two-dimensional state matrix of dimension 360×D, where 360 ​​represents the circumferential angular resolution and D represents the dimension of the feature vectors.

[0042] The base station simultaneously acquires two-dimensional state matrices for T consecutive time periods (T is set to 10, corresponding to approximately 10 rotation cycles), stacks them in the time dimension to form a three-dimensional spatiotemporal state tensor with dimensions T×360×D. This three-dimensional spatiotemporal state tensor is then flattened into a one-dimensional vector and input into a pre-trained fully connected neural network layer. This network contains an input layer, two hidden layers (containing 128 and 64 neurons respectively, with ReLU activation function), and an output layer (containing two neurons, outputting the overall material level percentage and coal quality distribution uniformity respectively). The final inference result is obtained through forward propagation calculation. The base station outputs the inference result to the distributed control system for coal feed rate control and coal mill operation status monitoring.

[0043] Based on the above technical solution, by constructing a complete technical chain of rotation angle prediction, channel mapping, TDMA scheduling, edge inference, and state fusion, the channel quality for each angle interval can be predicted during its use by receiving the rotation angle in real time and calling a channel quality mapping table containing metal wall penetration correction coefficients, thus solving the problem of unpredictable channel state. This is combined with techniques such as dividing the rotation period into fixed time slots, matching the expected signal-to-noise ratio with node priority weights, and dynamically adjusting the coding scheme and transmit power. This not only actively avoids channel degradation intervals and improves transmission robustness to reduce packet loss rate, but also effectively corrects signal attenuation in metal wall penetration scenarios. Simultaneously, a lightweight one-dimensional convolutional neural network is deployed on the node side to compress the three-axis vibration signal into feature vectors, alleviating line bandwidth pressure. Finally, on the base station side, a state matrix is ​​constructed based on the multi-node feature vectors and input into a fully connected neural network to complete material level inference. The synchronization benchmark naturally provided by the TDMA time slot structure solves the problem of multi-node data time misalignment. The aforementioned methods, from channel prediction, deterministic scheduling, edge compression to synchronous fusion, are interconnected, enabling continuous and reliable monitoring and precise fusion of multi-source data in a rotating metal sealed environment.

[0044] Example 2: like Figure 2 As shown, this embodiment, based on step S200 of embodiment 1, provides a detailed explanation of the construction and updating process of the channel quality mapping table. This mapping table is not a simple static lookup table, but a dynamic model that integrates physical environment characteristics and the time dimension. Its construction process specifically includes the following steps: Step 201: Receive offline calibration data and construct a basic mapping table. ,in For angle Signal strength at that location For angle Signal-to-noise ratio at the location. For angle The bit error rate at the location.

[0045] During the shutdown and maintenance of the coal mill, technicians fixed the base station transmitting antenna to the actual installation position and set up a standard signal source at the sensor installation position on the surface of the cylinder.

[0046] As the control cylinder rotates slowly (in 1-degree increments), the signal parameters received by the base station can be recorded angle by angle. This allows the acquisition of raw channel data (which does not yet take into account the impact of metal wall penetration).

[0047] It should be understood that the step angle for offline calibration is not limited to 1 degree. It can be adjusted to 0.5 degrees or 2 degrees according to the actual accuracy requirements, as long as it can cover the range of one revolution of the cylinder.

[0048] Step 202, Receiving Angle Signal arrival time distribution and corresponding power attenuation coefficient Construct a multipath feature matrix , where m is the number of resolvable multipaths.

[0049] Because the coal mill cylinder is a large metal structure, wireless signals undergo complex reflection, diffraction, and scattering on its surface, resulting in multipath propagation. Moreover, signals from different paths arrive at different times and each carries different levels of power attenuation.

[0050] Therefore, this step extracts multipath features at each angle through high-frequency sampling and channel impulse response analysis. For example, at a certain angle, the direct path may be blocked by the cylinder, while the path reflected from the cylinder surface becomes the main component. In this case, the signal arrival time distribution τ(θ) records the arrival time delay of the reflected path, and the power attenuation coefficient β(θ) records the corresponding attenuation. This provides key parameters for subsequent multipath equalization at the receiver.

[0051] Step 203: Obtain the attenuation coefficient of the wireless signal penetrating the metal wall of the rotating cylinder, and adjust the angle. The signal strength at the specified location is corrected to obtain the corrected signal strength. The basic mapping table, multipath feature matrix, and corrected signal strength are combined. Construct a three-layer channel quality mapping table based on angle index association. .

[0052] This is a key calibration step in this embodiment for the specific scenario of a dual-inlet, dual-outlet coal mill. Since its wireless signal is transmitted from the sensor nodes inside the mill body to the external base station, it must penetrate the metal wall. The material, thickness, and openings of the metal wall all cause significant signal attenuation, and this attenuation is often non-linear, making it difficult to calculate using simple theoretical formulas.

[0053] Therefore, this embodiment obtains the penetration attenuation coefficient in advance through actual measurement or simulation. And the correction formula is as follows: Signal strength at angle θ To perform correction. A positive value indicates penetration loss. By introducing this correction coefficient, the signal strength in the mapping table can accurately reflect the level received by the external base station, thus avoiding prediction errors caused by ignoring penetration loss.

[0054] Its three-layer mapping table serves different functional requirements in each layer: The first layer is , used for real-time channel quality query; The second layer is , used for multipath equalization parameter calculation; The third layer is = , The rate of change of channel quality, It is the acceleration of channel quality changes, used to predict the channel quality change trend at the next moment.

[0055] For example, when When the value is negative and large in absolute terms, it indicates that the channel quality is deteriorating rapidly, allowing for early power boosting or modulation scheme switching, rather than waiting for a complete channel outage before responding. This hierarchical design, from static parameters to dynamic trends, and from single indicators to multi-dimensional features, constructs a comprehensive channel profile, greatly improving the accuracy and robustness of predictions.

[0056] After constructing the above mapping table, the system must also have the ability to update online adaptively to cope with changes in the mass distribution of the mill cylinder caused by liner wear and coal quality variations during long-term operation, which in turn leads to channel characteristic drift. The online update process specifically includes: Step 204: Receive data packets from the receiving node and extract the actual receiving angle and the actual measured channel quality parameters, including the actual received signal strength, the actual signal-to-noise ratio, and the actual bit error rate.

[0057] In some implementations, during normal operation, each time a data packet is successfully received from a receiving node, the physical layer information is parsed to obtain the measured channel parameters at the current angle. This real-time data is used as the basis for updating the mapping table.

[0058] Step 205: Calculate the adaptive forgetting factor based on the channel quality change rate, update the mapping table value of the corresponding angle, and perform smoothing processing on adjacent angles.

[0059] Among them, the channel quality change rate refers to the instantaneous derivative of the channel quality parameter with respect to the angle at the current angle, used to reflect the trend of channel condition deterioration or improvement. The adaptive forgetting factor is a dynamically adjusted weighting coefficient used to balance the contribution ratio of historical measurements and current real-time measurements when updating the channel quality mapping table. Adjacent angle smoothing refers to smoothing adjacent angles (such as...) after updating the mapping table value of the current angle. The mapping table values ​​are also adjusted appropriately to prevent discontinuous jumps in the mapping table surface.

[0060] In some implementations, the current angle is extracted from the successfully received data packet. Corresponding actual channel quality parameters Among them, actual channel quality parameters This includes the actual received signal strength, actual signal-to-noise ratio, and actual bit error rate. By simultaneously querying the predicted value for that angle in the mapping table, the channel quality change rate at the current angle can be directly calculated. That is, it is obtained by dividing the difference in signal-to-noise ratio between two consecutive updates by the change in angle.

[0061] And an adaptive forgetting factor is introduced. Its calculation logic is related to the rate of change of channel quality: When the channel changes drastically (e.g.) Larger), smaller The value of gives greater weight to real-time measurements, enabling the mapping table to quickly track changes; When the channel is relatively stable, increase The value of is used to retain more historical information and suppress the influence of transient noise. The update formula can be expressed as: ,in For angle The updated channel quality mapping table is located here. For angle The original channel quality mapping table, For angle The actual channel quality parameters at that location.

[0062] In addition, to avoid data jumps at a single angle point, the update operation will also update adjacent angles (e.g., ...). ±1° The adjacent angles (±2°) are subjected to weighted smoothing, for example, Gaussian kernel weights are used to correct the mapping values ​​of adjacent angles to ensure the continuity of the mapping surface.

[0063] Step 206: Calculate the root mean square error of the updated mapping table. When the root mean square error is less than the first threshold, confirm the update is valid and store the update result. When the root mean square error is between the first threshold and the second threshold, trigger local recalibration. When the root mean square error is greater than the second threshold, trigger a global recalibration warning.

[0064] This embodiment establishes a hierarchical error handling mechanism to ensure the security of updates. The first and second thresholds within it are preset limits based on the system's fault tolerance capability.

[0065] If the updated error is very small, it means that the measured value matches the predicted value, and the update is effective; If the error is in the middle range, it indicates that there may be an anomaly in a local area (such as local coal accumulation in the cylinder). Automatic local recalibration is triggered to perform intensive measurement and correction in this angle range. If the error is too large and exceeds the second threshold, it may mean that a major structural change or hardware failure has occurred. At this time, a global recalibration warning is triggered, prompting maintenance personnel to intervene and check.

[0066] The aforementioned hierarchical processing mechanism not only ensures the system's adaptability but also prevents erroneous data from polluting the mapping table, demonstrating strong robustness.

[0067] Step 207: Obtain the history of rotation cycle changes, and use linear regression to predict the channel quality change rate of the next cycle for time slot allocation pre-scheduling.

[0068] The system also records the historical sequence of channel quality change rates and uses a simple linear regression model to predict the trend of change in the next cycle. For example, if a continuous downward trend in channel quality is detected recently, the system can reserve more power margin or allocate better time slot resources in advance to achieve pre-schedule of resources.

[0069] This step transforms passive response into proactive planning, further enhancing the determinism of communication.

[0070] In summary, the construction and updating process described above makes the channel quality mapping table in this embodiment no longer a static data table, but a dynamic model that can adapt to environmental drift and has multi-dimensional features and trend prediction capabilities, providing accurate decision-making basis for subsequent time slot scheduling and power control.

[0071] Example 3: like Figure 3 As shown, this embodiment, based on step S300 of embodiment 1, provides a detailed explanation of the specific process for dynamically adjusting the node's coding scheme and transmission power. This process is a core element in ensuring the reliability of wireless communication in rotating scenarios, and specifically includes the following steps: Step 301: Divide the expected signal-to-noise ratio into continuous intervals and construct the mapping relationship between the continuous intervals and the modulation and coding scheme. The continuous intervals include low signal-to-noise ratio interval, medium signal-to-noise ratio interval, high signal-to-noise ratio interval and ultra-high signal-to-noise ratio interval.

[0072] Based on the predicted signal-to-noise ratio (SNR) from the channel quality mapping table, the channel state is divided into four levels. As a specific example, the classification criteria are as follows: When the expected SNR is less than 5dB, it is defined as the low signal-to-noise ratio range, corresponding to a poor channel. When the expected SNR is between 5dB and 10dB, it is defined as the medium signal-to-noise ratio range, corresponding to a general channel. When the expected SNR is between 10dB and 20dB, it is defined as a high signal-to-noise ratio range, corresponding to a good channel. When the expected SNR is greater than 20dB, it is defined as the ultra-high signal-to-noise ratio range, corresponding to a high-quality channel.

[0073] This allows for the direct pre-setting of different modulation and coding schemes (MCS) for these four intervals. For example: In the low signal-to-noise ratio range, BPSK (Binary Phase Shift Keying) modulation is used, with a coding rate of 1 / 2. It has the strongest anti-interference capability but the lowest transmission rate. The signal-to-noise ratio range uses QPSK (Quadrature Phase Shift Keying) modulation with a coding rate of 3 / 4; The high signal-to-noise ratio range uses 16QAM (quadrature amplitude modulation) modulation with a coding rate of 1 / 2; the ultra-high signal-to-noise ratio range uses 16QAM modulation with a coding rate of 3 / 4 and the highest transmission rate.

[0074] It should be understood that the above-mentioned specific expected signal-to-noise ratio (SNR) threshold and modulation and coding scheme (MCS) mapping relationship is only a preferred configuration in this embodiment. In practical applications, it can be adjusted according to factors such as the specific metal material of the coal mill cylinder and the upper limit of the sensor's transmission power, as long as the channel quality and transmission rate can be matched.

[0075] Step 302: Calculate the effective data rate based on the modulation and coding scheme, and adjust the transmit power of the node based on the continuous interval synchronization.

[0076] The effective data rate refers to the number of effective information bits that can be successfully transmitted per unit time under a specific modulation and coding scheme. Its value depends on the modulation order, coding rate, and symbol rate.

[0077] Once the modulation and coding scheme (MCS) is determined, the effective data rate for the current time slot can be calculated. Simultaneously, to further optimize energy consumption and communication quality, the system will also adjust the node's transmit power. For example: In the high signal-to-noise ratio range, due to the excellent channel conditions, the transmit power can be appropriately reduced (e.g., from +20dBm to +10dBm) to save battery energy of the sensor node. In the low signal-to-noise ratio range, the command node increases the transmit power to the maximum value (e.g., +20dBm) to counteract signal attenuation.

[0078] Therefore, the optimal balance between communication reliability and energy consumption can be achieved directly through a joint adaptive strategy of power and rate.

[0079] Step 303: When the expected signal-to-noise ratio fluctuates at the boundary of adjacent intervals, activate the hysteresis comparator to prevent frequent switching of the coding scheme.

[0080] In actual operation, fluctuations in the rotational speed of the coal mill cylinder or external interference may cause the expected signal-to-noise ratio (SNR) to fluctuate around the boundary values ​​of two intervals. For example: The expected signal-to-noise ratio (SNR) fluctuates between 9.8 dB and 10.2 dB. If the threshold in step S301 is used directly, the modulation and coding scheme (MCS) will frequently switch between QPSK and 16QAM. This "ping-pong effect" will lead to a significant increase in system control signaling overhead and may even cause the modem to lose synchronization. Therefore, a hysteresis comparator mechanism is directly introduced. Specifically, this involves setting a hysteresis threshold H (e.g., H = 1 dB) at the interval boundaries. When the expected signal-to-noise ratio (SNR) crosses the boundary from low to high, it needs to exceed the boundary value +H (e.g., from 9dB to 11dB) before it can be determined to enter the next higher level range. Conversely, when the expected signal-to-noise ratio (SNR) crosses the boundary from high to low, it needs to be below the boundary value -H (e.g., from 11dB to 9dB) to be considered as entering the next lower level range.

[0081] This bidirectional threshold determination allows the current modulation and coding scheme (MCS) to remain unchanged during boundary fluctuations, effectively suppressing system oscillations caused by frequent switching and ensuring the stability of the communication link.

[0082] Step 304: When data packet transmission fails, increase the node's transmit power according to the current expected signal-to-noise ratio range, and retransmit using a low-order modulation and coding scheme.

[0083] Despite the system's predictive and adaptive adjustments, transmission failures may still occur under extreme conditions. In such cases, this embodiment employs a down-order retransmission strategy. If the initial transmission fails, the system first determines the current SNR interval: If the current signal-to-noise ratio range is already in the low range and BPSK is being used, the transmission power will not be downgraded, but will be directly increased to the maximum value allowed by the hardware (such as +23dBm) for retransmission. If the current signal-to-noise ratio is in the medium or higher range, the modulation and coding scheme (MCS) will be reduced by one level (e.g., from 16QAM to QPSK), while the transmit power will be increased by a preset step (e.g., +3dB), and then retransmission will be scheduled in the next available time slot.

[0084] This "downgraded order + increased power" strategy improves the retransmission success rate. Compared to traditional simple retransmission mechanisms, the strategy in this embodiment can proactively adapt to channel degradation, avoiding resource waste and latency accumulation caused by blind retransmission under poor channel conditions, and ensuring that critical monitoring data can ultimately be successfully transmitted to the base station.

[0085] Based on the above technical solution, by dividing the expected signal-to-noise ratio into continuous intervals and constructing a mapping relationship with the modulation and coding scheme, the problem of unreliable transmission caused by the strong periodic fluctuation of wireless channel quality with angle in a rotating metal confined environment is solved. This achieves a fundamental shift in wireless communication in rotating scenarios from passive response to active avoidance, providing deterministic communication assurance for continuous and reliable monitoring of coal mill material levels.

[0086] Example 4: like Figure 4-5 As shown, this embodiment, based on step S400 of Embodiment 1, provides a detailed explanation of the training and deployment method of a lightweight one-dimensional convolutional neural network. In traditional coal mill monitoring systems, vibration data analysis typically relies on a host computer or cloud server, requiring the transmission of large amounts of raw waveform data, which places extremely high demands on wireless bandwidth. This embodiment, however, achieves feature extraction and data compression at the edge side by deploying a lightweight neural network locally on the sensor nodes. The specific implementation process is as follows: Step 401: In the offline stage, a high-performance server is used as the teacher model, and triaxial vibration data including the working conditions of empty mill, half-mill, full mill, steel ball loading deviation and coal grindability change are used as the training set to train a lightweight one-dimensional convolutional neural network as the student model for knowledge distillation.

[0087] The teacher model refers to a deep neural network model trained on a high-performance server with strong feature extraction capabilities. The student model refers to a lightweight one-dimensional convolutional neural network deployed on a low-power microcontroller unit, with a much smaller number of parameters than the teacher model. Knowledge distillation training involves comparing the output probability distribution of the teacher model as a soft label with the prediction results of the student model on the same input data, and updating the weight parameters of the student model by minimizing the difference between the two.

[0088] The internal operating conditions of a coal mill are extremely complex. Different material levels, coal qualities, and steel ball loading rates can all lead to significant differences in the spectral characteristics of vibration signals. To ensure the lightweight model has sufficient generalization ability, this embodiment collects triaxial vibration data of the coal mill under various typical operating conditions, including but not limited to: empty mill state (no coal or steel balls in the mill), half-mill state (normal working material level), full mill state (material level reaches the upper limit), steel ball loading deviation state (steel ball wear or excessive replenishment), and coal grindability variation state (hardness differences caused by changes in coal type). In other words, the above data constitute a rich and representative sample space.

[0089] Due to limited computing resources at the sensor node, complex deep neural networks cannot be directly run. Therefore, this embodiment employs knowledge distillation technology: First, a complex and highly accurate teacher model is trained on a high-performance server. This model can accurately extract deep features from vibration signals and output the class probability distribution of each sample as a soft label. Simultaneously, a simplified student model is constructed—the lightweight one-dimensional convolutional neural network in this embodiment—which outputs its own predicted distribution. Finally, the KL divergence (quantifying the "distance" between the probability distributions output by the teacher and student models) is calculated as the distillation loss. Simultaneously, the student model also calculates cross-entropy loss with the true labels, and the weighted sum of the distillation loss and cross-entropy loss is used as the total loss function. Finally, the weight parameters of the student model are iteratively updated using the backpropagation algorithm, enabling the student model's output distribution to approximate the teacher model with a significantly reduced number of parameters, thus completing the knowledge distillation training. During training, the student model not only learns the real labels, but also the output probability distribution of the teacher model, thus "transferring" the knowledge of the teacher model to the student model.

[0090] This method allows student models to maintain recognition accuracy close to that of teacher models even with a significant reduction in the number of parameters, solving the technical challenge of running complex AI models on low-power MCUs.

[0091] Step 402: Convert the weights of the trained student model from floating-point numbers to fixed-point numbers, and solidify the forward inference computation graph into the read-only memory of the low-power microcontroller unit of the node.

[0092] Floating-point numbers refer to a data format used in computers to represent real numbers. They are typically stored using 32-bit single-precision or 64-bit double-precision, and include a sign bit, exponent bit, and mantissa bit. Fixed-point numbers refer to a numerical representation where the decimal point is fixed at a specific position. They are typically stored using 8-bit or 16-bit integers, and conversion between floating-point and integer values ​​is achieved through scaling factors. A forward inference computation graph refers to the data flow diagram of a neural network from the input layer to the output layer. It includes the computational nodes of each layer (convolution, pooling, activation, fully connected, etc.) and their connections, and is stored in the form of a static data structure. Since deep learning frameworks typically use 32-bit floating-point numbers for computation, this places a significant burden on the storage space and computing units of an MCU. Therefore, this embodiment employs model quantization technology to convert model weights and activation values ​​from 32-bit floating-point quantization to 8-bit fixed-point quantization. The quantization process is not a simple numerical truncation; instead, it establishes a mapping relationship between floating-point and fixed-point values ​​by statistically analyzing the distribution range of the weights, and introduces inverse quantization to compensate for precision loss. Specifically, this includes: statistically analyzing the numerical distribution range of all weights in the trained student model, calculating the maximum absolute value as the quantization benchmark, and determining the scaling factor as the maximum value divided by 127 (for 8-bit signed integers). Then, a quantization operation is performed on each floating-point weight, i.e., the weight is multiplied by the scaling factor, rounded to the nearest integer, and clamped to the range of -128 to 127 to obtain the corresponding 8-bit fixed-point integer value. This reduces the size of the quantized model to one-quarter of its original size, and the execution efficiency of fixed-point operations on the MCU is far higher than that of floating-point operations.

[0093] Furthermore, to eliminate reliance on the operating system and dynamic memory management, this embodiment permanently stores the quantized model weights and the computation graph structure required for forward inference in the MCU's read-only memory (ROM) as static data. Specifically, this involves arranging the quantized weight matrix sequentially, along with the bias parameters and activation function types for each layer, and packaging them together into a binary forward inference computation graph. This binary computation graph is then embedded as a constant array into a C language source file and linked to the target code segment by the compiler. Finally, during firmware flashing, the code segment containing the computation graph is written to the starting address of the microcontroller's ROM, achieving permanent storage.

[0094] This means that after the node is powered on and running, there is no need to load model files or perform dynamic memory allocation; the weights can be read directly from ROM and executed according to the preset computation graph. This greatly improves the system's startup speed and operational stability, enabling it to adapt to the harsh electromagnetic environment and power supply conditions in industrial settings.

[0095] Step 403: During the online inference phase, the node only performs forward inference calculations, inputting the real-time collected triaxial vibration data and directly outputting the feature vector.

[0096] The online inference phase refers to the continuous working state during normal operation of the coal mill, where wireless sensor nodes collect vibration data in real time and perform neural network calculations.

[0097] In some implementations, when the coal mill is running normally, the microcontroller reads the pre-programmed forward inference computation graph and quantization weights from the read-only memory and loads them into the computation register. At this time, the sensor nodes continuously acquire triaxial vibration signals at a preset sampling rate (e.g., 1 kHz). Whenever a data block is fully acquired (e.g., 1024 sampling points, corresponding to a data length of approximately 1 second), the MCU triggers an inference task.

[0098] Since the model is already fixed and quantized, the MCU only needs to perform simple convolution, pooling, activation, and fully connected operations, without the need for backpropagation or gradient updates. This yields a low-dimensional feature vector (e.g., a 16-dimensional floating-point vector) as the inference result. This vector highly condenses the key information related to material level and coal quality in the original vibration signal.

[0099] Compared to transmitting raw waveform data (approximately 6KB), the feature vector data size is only 64 bytes, achieving a compression ratio as high as 96:1. This extremely high compression ratio directly alleviates the pressure on wireless transmission, enabling reliable transmission of monitoring data even in harsh channel environments where metal cylinders cause severe penetration and attenuation.

[0100] It should be understood that although this embodiment uses a one-dimensional convolutional neural network as an example, in other embodiments, other lightweight network structures, such as depth-separable convolutional networks or spiking neural networks, can be adopted according to the specific computing power and storage resources of the MCU, as long as they can achieve effective mapping from vibration signals to feature vectors.

[0101] In summary, addressing the technical challenges of limited wireless bandwidth and congestion / packet loss caused by large raw data transmission volumes in multi-node vibration monitoring of a dual-inlet, dual-outlet coal mill's rotating cylinder, this solution employs knowledge distillation technology in the offline phase. A high-performance server acts as the teacher model, guiding a lightweight one-dimensional convolutional neural network student model, significantly reducing the number of parameters while retaining high-precision feature extraction capabilities. Simultaneously, the trained floating-point weights are quantized into fixed-point numbers, and the forward inference computation graph is embedded in the read-only memory of the node's microcontroller unit, eliminating the dynamic overhead of loading the model at runtime. Furthermore, in the online inference phase, the nodes only perform forward inference computation, directly mapping the real-time collected triaxial vibration data into low-dimensional feature vectors. This allows for seamless integration of technical methods, compressing the raw vibration data from approximately 6KB per node per cycle to 64 bytes, achieving a compression ratio of 96:1. This fundamentally alleviates wireless bandwidth pressure, avoids packet loss due to data congestion, and reduces node computational load and power consumption by embedding the inference computation graph. Ultimately, this achieves lightweight and reliable transmission of vibration monitoring data in a sealed rotating metal environment.

[0102] Example 5: like Figure 6-7As shown, this embodiment, based on step S500 of embodiment 1, provides a detailed explanation of the process of constructing the state matrix and the priority scheduling mechanism based on this matrix. By constructing a state matrix that reflects the spatiotemporal distribution, the system can perceive the internal state of the coal mill from a global perspective and dynamically adjust the allocation of communication resources accordingly, achieving data-driven intelligent decision-making. Step 501: Obtain the geographical location or installation angle identifier of all nodes, and sort and interpolate the feature vectors from different nodes at the same time according to the corresponding rotation angle to form a two-dimensional state matrix, which is used to reflect the circumferential state distribution of the cylinder.

[0103] Multiple sensor nodes are typically installed on the coal mill cylinder. These nodes are distributed at different locations along the axial direction (such as the drive end, middle, and non-drive end) and at different installation angles along the circumference. Due to the limited number and discrete distribution of nodes, directly concatenated feature vectors cannot fully describe the continuous state of the cylinder's circumference.

[0104] This embodiment first extracts the feature vectors uploaded by all nodes at the same time and sorts them according to their installation angle identifiers (such as 0 degrees, 120 degrees, 240 degrees). Then, an interpolation algorithm (such as linear interpolation or spline interpolation) is used to fill in the angle intervals between adjacent nodes. For example, if there is a 60-degree angle interval between node A and node B, a virtual "feature vector" within this interval can be generated by interpolation, thereby constructing a two-dimensional state matrix with dimensions [360, D], where 360 ​​represents the circumferential angle resolution and D represents the dimension of the feature vector.

[0105] At this point, the matrix intuitively reflects the material level and coal quality distribution at various positions around the cylinder at a certain moment, providing complete spatial field data for subsequent analysis.

[0106] Step 502: Obtain the two-dimensional state matrix at consecutive time points and stack it to form a three-dimensional spatiotemporal state tensor.

[0107] The change in material level inside a coal mill is a dynamic process, and its movement trend is difficult to capture by a single moment's spatial slice.

[0108] Therefore, in this embodiment, the two-dimensional state matrices of T consecutive time moments are stacked in the time dimension to form a three-dimensional spatiotemporal state tensor with dimensions [T, 360, D]. At this time, the tensor not only contains the spatial distribution information, but also records the evolution of these distributions over time, such as the fluctuation of material level and the flow of pulverized coal.

[0109] It should be understood that the value of T can be set according to actual monitoring needs. For example, data from the most recent 10 rotation cycles can be taken to cover a time window of about 30 to 40 seconds, which is sufficient to capture the main dynamic characteristics inside the coal mill.

[0110] Step 503: Input the three-dimensional spatiotemporal state tensor into the fully connected neural network layer to infer the overall material level and coal quality distribution uniformity in the current cylinder.

[0111] After constructing the three-dimensional spatiotemporal state tensor, it is flattened or high-level features are extracted through convolutional layers, and then input into a fully connected neural network layer deployed on the base station side or host computer.

[0112] This network, trained offline, learns a nonlinear mapping between material level height and feature distribution from complex spatiotemporal data. Its output includes two key metrics: First, the overall material level, used to guide coal feed rate control; Second, the uniformity of coal quality distribution is used to assess whether there is local blockage or uneven grinding inside the coal mill.

[0113] This spatiotemporal fusion inference method effectively filters out random noise from single-point measurements, improving the robustness and accuracy of monitoring results.

[0114] After obtaining the status information of each node, this embodiment further performs priority scheduling based on the status evaluation value to ensure that critical data can be transmitted first. The specific process is as follows: Step 504: Calculate the state evaluation value of the corresponding node based on the feature vector, sort the state evaluation values ​​of all nodes from high to low to obtain the priority weight sequence, and determine the base number of fixed time slots allocated to each node in each rotation cycle based on the priority weight sequence.

[0115] The status evaluation value is a numerical indicator used to quantify the degree of anomaly in the area monitored by each wireless sensor node. A higher value indicates a greater deviation from normal conditions and a stronger urgency for data transmission. The priority weight sequence is a list of nodes ordered by sorting their status evaluation values ​​from highest to lowest, reflecting the priority level of each node in obtaining communication resources within the current rotation cycle. The baseline quantity refers to the number of fixed time slots initially allocated to each node according to the priority weight sequence, serving as the basis for subsequent dynamic scheduling. In some implementations, based on the feature vector uploaded by each node, through... Calculate the state evaluation value This is used to quantify the degree of anomaly in the area monitored by the node. For the current feature vector, , and The priority weights are determined using historical data. This is the reference vector under normal operating conditions. This represents the anomaly probability output by the node. This represents the characteristic rate of change. Its state evaluation value... The higher the value, the more abnormal the working conditions at that location (such as a sharp rise in material level or the presence of foreign objects), and the greater the urgency of data transmission.

[0116] Therefore, the calculated state evaluation value can be directly used. All nodes are sorted according to their numerical values, and a baseline number of time slots are initially allocated. For example, the top 30% of nodes are allocated 3 time slots, the middle 40% of nodes are allocated 2 time slots, and the bottom 30% of nodes are allocated 1 time slot.

[0117] Step 505: According to the channel quality mapping table, sort the expected signal-to-noise ratios of each node in the current rotation period from high to low to obtain the expected signal-to-noise ratio sequence.

[0118] In some implementations, the channel quality map constructed in Example 2 is used to predict the channel quality of each node at its angle in the next cycle. The expected signal-to-noise ratio (SNR) is sorted from high to low to obtain the channel quality sequence. This step ensures that high-priority data can be transmitted within the time window with the best channel conditions possible.

[0119] Step 506: Perform weighted matching of the expected signal-to-noise ratio sequence and the priority weight sequence to allocate more fixed time slots to nodes with high priority that are in the high expected signal-to-noise ratio angle range.

[0120] A weighted algorithm (such as weighted round-robin or weighted maximum signal-to-noise ratio algorithm) is used to comprehensively consider priority and channel quality. Ideally, high-priority nodes should be matched with high signal-to-noise ratio time slots.

[0121] If a node has extremely high priority but poor channel quality at the current angle, the system will allocate a relatively better available time slot to it and may compensate for channel loss by increasing the transmit power.

[0122] The aforementioned dual matching mechanism achieves a resource scheduling strategy of "using the best resources where they are most needed," maximizing the system's effective throughput.

[0123] Step 507: When the status evaluation value exceeds the preset emergency threshold, mark the corresponding node as an emergency priority node.

[0124] In some implementations, an emergency threshold is set to handle unexpected operating conditions (such as coal shortage or full mill operation). This threshold is determined when the state evaluation value of a certain node... If the threshold is exceeded, an emergency event is determined to have occurred at that location, and the node is immediately marked as an emergency priority node.

[0125] Step 508: For emergency priority nodes, force them to be allocated an additional fixed time slot exceeding the baseline number, and dynamically preempt the fixed time slots already allocated to nodes at the end of the priority weight sequence as needed.

[0126] In some implementations, emergency priority nodes enjoy the highest level of resource preemption rights. The system will forcibly allocate additional time slots to them to meet their high-frequency data upload needs.

[0127] These extra time slots are not generated out of thin air, but are "preempted" from the lowest priority (i.e., the most normal) nodes. For example, if node A is marked as urgent, the system may allocate some time slots from node B (which has the lowest status evaluation value) to node A. This dynamic preemption mechanism ensures that abnormal operating condition data can be given priority protection when the total system bandwidth is limited.

[0128] Step 509: The number of fixed time slots reserved by the preempted node is reduced to a preset minimum guarantee value, which is adaptively determined based on the total number of nodes in the system and the cycle duration.

[0129] In some implementations, to prevent the communication resources of low-priority nodes from being completely deprived, which would cause the system to lose its ability to monitor these areas, this embodiment introduces a "minimum guarantee value" mechanism. Even if a node's time slot is preempted, the number of time slots it retains will not be lower than the minimum guarantee value.

[0130] The minimum guaranteed value is not a fixed constant, but rather depends on the current total number of nodes N in the system and the rotation cycle duration. The result is obtained through adaptive calculation, for example, by setting it to... This ensures that each node has at least one opportunity to upload heartbeat packets or basic data within each cycle. This balances the urgency of data transmission with the completeness of system monitoring, reflecting a balance between fairness and efficiency.

[0131] Based on the above technical solution, a two-dimensional state matrix is ​​constructed by sorting and interpolating the feature vectors of each node according to the installation angle, and then stacking continuous time points to form a three-dimensional spatiotemporal state tensor. This solves the technical problem in existing technologies where the discrete distribution of nodes and lack of spatiotemporal correlation lead to an inability to fully perceive the circumferential direction of the cylinder and the dynamic material level distribution. Simultaneously, a priority weight sequence is generated based on the state evaluation values ​​and weighted and matched with the expected signal-to-noise ratio sequence. More time slots are allocated to high-priority nodes located in high-quality channel intervals, avoiding mismatch between channel resources and data importance, which could lead to the loss or delay of critical data. When the state evaluation value exceeds the emergency threshold, additional time slots are forcibly allocated to emergency nodes, and time slots are preempted from low-priority nodes as needed, while a minimum guarantee value is maintained to prevent monitoring blind spots. This effectively solves the defects of disordered competition for communication resources among multiple nodes and untimely response to abnormal operating conditions in rotating enclosed environments, realizing a shift from passive equal allocation to active on-demand scheduling, and improving the reliability of data transmission and the overall monitoring integrity of the system under channel degradation and sudden operating conditions.

[0132] Example 6: like Figure 8 As shown, this embodiment, based on Embodiment 1, provides a detailed explanation of the specific implementation process of time slot synchronization and timestamp calibration. In a multi-node collaborative monitoring scenario, each node operates independently. If a unified time reference is lacking, the collected data will not be aligned in the time dimension, leading to the failure of subsequent state matrix construction and feature fusion. This embodiment achieves sub-millisecond time synchronization through the following steps: Step 601: At the start of the rotation cycle, broadcast the time slot synchronization beacon, which contains the current cycle number and the global time reference.

[0133] This embodiment utilizes the inherent time slot structure of the TDMA protocol as the carrier for time synchronization. The base station, acting as the master clock source, broadcasts a high-priority synchronization beacon frame to all nodes at the beginning of each rotation cycle (e.g., when the zero-position marker on the cylinder passes a fixed position). This beacon frame not only contains the sequence number of the current cycle but also carries a precise global timestamp.

[0134] Since the wireless signal transmission delay is in the microsecond range and relatively fixed over short distances (usually less than 10 meters), each node can assume that its local time is aligned with the global time reference the moment it receives the beacon.

[0135] This broadcast and receive mechanism eliminates the need for complex handshake interactions between nodes, greatly simplifying the synchronization process and reducing communication overhead.

[0136] It should be understood that the timing of the synchronization beacon broadcast is not limited to the start of the period, but can also be carried out in a specific time slot within the period, as long as it can be ensured that all nodes receive it at the same physical moment.

[0137] Step 602: When the node sends a triaxial vibration signal, it embeds a local timestamp and converts it into global time using a global time reference.

[0138] After a node completes feature vector extraction locally, it reads the current value of the local counter as a local timestamp and embeds it into the protocol header of the data packet.

[0139] When the base station receives the data packet, it can convert the local timestamp into a unified global time based on the time slot number where the data packet is located and the global time reference established in step S601.

[0140] Its conversion formula can be expressed as:

[0141] in, For a unified global time, This is the global time at the start of this time slot. The local timestamp carried in the data packet. This is the local reference time recorded by the node during its last synchronization. The time slot length, The nominal period of the node's local clock.

[0142] This transformation enables the base station to map all data onto a unified time axis, even if there are slight differences in the local clocks of each node, providing an accurate time reference for the subsequent construction of the three-dimensional spatiotemporal state tensor.

[0143] Step 603: Calculate the time deviation of nodes at the same angle within adjacent cycles, and dynamically adjust the frequency division coefficient of the local sampling clock of the node according to the time deviation to achieve sub-millisecond synchronization with the rotation cycle of the rotating cylinder.

[0144] Because low-cost crystal oscillators typically have frequency errors at the ppm (parts per million) level, long-term operation can lead to the accumulation of time deviations. For example, a 50ppm crystal oscillator may accumulate 180ms of error after one hour of operation, which is unacceptable for monitoring systems that require analysis of the phase characteristics of vibration signals.

[0145] Therefore, this embodiment introduces closed-loop control logic to compensate for this drift. When the base station continuously monitors each node, if the time deviation Δt at the same angular position in adjacent cycles shows a systematic increase (e.g., a fixed increase of 0.1ms per cycle), it indicates that the node's local clock is either too fast or too slow.

[0146] At this point, the base station will send a frequency division coefficient adjustment command to the node via the downlink.

[0147] After receiving the instruction, the node can fine-tune the parameters of its sampling clock divider. For example, if the node clock is found to be too fast, the division factor is increased appropriately to reduce the sampling frequency; conversely, the division factor is decreased. This dynamic adjustment allows the node's sampling frequency to be "pulled" into strict synchronization with the cylinder's rotation frequency, thus fully meeting the stringent time alignment requirements of multi-source data fusion.

[0148] Based on the above technical solution, a unified global time reference is provided for each node through broadcast time slot synchronization beacons to solve the technical problem of misaligned data acquisition time and inability to perform effective multi-source feature fusion caused by independent operation of multiple nodes and lack of a unified clock. Furthermore, when nodes send feature vectors, they embed local timestamps and convert them to global time through the global time reference, enabling the data reported by each node to be mapped onto the same time axis. This provides a precise time reference for the subsequent construction of the three-dimensional spatiotemporal state tensor, eliminating state matrix misalignment and fusion failure caused by time deviation. Simultaneously, the base station calculates the time deviation of nodes at the same angle within adjacent periods and dynamically adjusts the frequency division coefficient of the node's local sampling clock, suppressing long-term crystal oscillator drift to the sub-millisecond level and avoiding long-term out-of-synchronization due to the accumulation of frequency errors in low-cost crystal oscillators. This collectively solves the defect of multi-node data time asynchrony in a rotating confined environment, realizing the transformation from independent asynchronous acquisition to unified synchronous fusion, improving the accuracy of multi-source data fusion and the reliability of state monitoring.

[0149] Example 7: This embodiment provides a coal quality measurement and polling system for a dual-inlet, dual-outlet coal mill, such as... Figure 1-8 As shown, this system serves as the hardware carrier for the methods described in Embodiments 1 to 6 above. Through the coordinated operation of its various functional units, it achieves reliable communication and intelligent monitoring in a rotating metal confined environment. This measurement polling system specifically includes a communication unit, a processing unit, a storage unit, and wireless sensor nodes.

[0150] The communication unit serves as the interface between the system and the outside world, and its hardware typically manifests as a wireless transceiver module and signal processing circuit integrated on the base station side. This unit is configured to receive rotation angle data and triaxial vibration signal data packets transmitted by wireless sensor nodes installed on the rotating cylinder of a dual-inlet, dual-outlet coal mill, as well as broadcast time-slot synchronization beacons to each node.

[0151] In practical deployments, the communication unit can employ an industrial-grade wireless module supporting the 2.4GHz band, coupled with a high-gain antenna, to enhance signal reception after penetrating metal walls. It should be understood that the communication unit is not limited to a single frequency band; depending on the electromagnetic environment at the site, a lower frequency band such as 433MHz or Sub-1GHz can be selected to obtain better diffraction performance, as long as the signal penetration and data transmission requirements are met.

[0152] The processing unit is the core control hub of the system, typically handled by a high-performance embedded processor or industrial control computer. This unit is configured to perform channel prediction by calling the channel quality mapping table in the storage unit, allocate fixed time slots based on the prediction results, dynamically adjust the node's coding scheme and transmit power, construct the state matrix, and perform computationally intensive tasks such as material level inference.

[0153] When in use, the processing unit connects at high speed to the communication unit and storage unit via an internal bus to obtain wireless channel status information in real time and run scheduling algorithms to generate control commands. For example, after receiving rotation angle information, the processing unit immediately triggers a lookup table prediction logic to calculate the expected signal-to-noise ratio for each angle interval in the next cycle, and generates a time slot allocation table accordingly, which is then distributed to each node through the communication unit.

[0154] Meanwhile, the storage unit is used to persist the key data models and parameters required for the operation of the storage system.

[0155] Specifically, the storage unit stores a channel quality mapping table, the fixed weights of a lightweight one-dimensional convolutional neural network, and the model parameters of the fully connected neural network layer. The storage unit can be non-volatile memory, such as flash memory or a solid-state drive, ensuring data is not lost after a power outage and restart. The channel quality mapping table stores the correspondence between angles and channel quality, allowing the processing unit to query it in real time; the model parameters of the fully connected neural network layer are used for material level inference calculations on the base station side, mapping the feature vectors uploaded by the nodes to the final material level value.

[0156] The wireless sensor node is a front-end sensing device deployed on the rotating cylinder of the coal mill, featuring extremely high hardware integration. Specifically, the wireless sensor node integrates a triaxial vibration sensor, a low-power microcontroller unit, and a wireless transceiver. Triaxial vibration sensors typically employ microelectromechanical systems (MEMS) accelerometers to collect vibration signals during the cylinder's rotation. Low-power microcontroller units (MCUs) act as the local brain of a node, responsible for data preprocessing and logic control; The wireless transceiver is responsible for sending the processed data to the base station.

[0157] In a preferred embodiment, the MCU can be an STM32L4 series low-power processor with an operating frequency of up to 80MHz and a floating-point unit (FPU) to meet the needs of lightweight neural network inference.

[0158] Furthermore, to enable edge computing at resource-constrained nodes, this embodiment features a specially designed storage structure for the MCU. The read-only memory of its low-power microcontroller unit contains the forward inference computation graph of a lightweight one-dimensional convolutional neural network.

[0159] Specifically, the forward inference computation graph is a binary file that has been trained offline, quantized (e.g., quantized from 32-bit floating-point weights to 8-bit fixed-point numbers), and compiled. By embedding the computation graph in ROM, the node does not need to load the model file upon power-up; it can directly read instructions from ROM to execute inference, greatly saving RAM resources and improving startup speed. This design enables low-cost MCUs to run AI models, thereby completing the compressed conversion from raw vibration signals to feature vectors locally on the node, significantly reducing the amount of data transmitted wirelessly.

[0160] Meanwhile, its wireless transceiver is configured to transmit data packets containing feature vectors within allocated fixed time slots.

[0161] Specifically, the operating mode of the wireless transceiver is controlled by the MCU. After receiving the time slot synchronization beacon broadcast by the base station, the node resolves the time slot number and start time assigned to it. The MCU then sets the timer accordingly. During non-allocated time slots, the wireless transceiver is in sleep or low-power listening mode to conserve battery power. When entering the assigned fixed time slot, the MCU wakes up the wireless transceiver, encapsulates the feature vector generated by local inference into a data packet, and completes the transmission within that time slot window.

[0162] This strict Time Division Multiple Access (TDMA) operating mode avoids communication conflicts between multiple nodes and ensures the determinism and reliability of data transmission in rotating scenarios with poor channel quality.

[0163] Through the aforementioned hardware architecture design, the system in this embodiment fully supports all the technical features described in the aforementioned method embodiments. From the underlying sensor data acquisition and lightweight inference at the edge, to the TDMA scheduling on the transmission side, and then to the data fusion and inference on the base station side, a closed-loop monitoring system is formed. In particular, by embedding the forward inference computation graph in the MCU's ROM and controlling the wireless transceiver to operate within fixed time slots, the contradiction between the limited wireless communication bandwidth and scarce computing resources on rotating components is resolved, providing a low-cost, high-reliability hardware solution for monitoring coal quality and level in dual-inlet, dual-outlet coal mills.

[0164] Example 8: To verify the practical application effect of the coal quality measurement and polling method and system for a dual-inlet, dual-outlet coal mill provided by this invention, this embodiment combines the actual operating conditions of a thermal power plant and sets specific coal mill parameters for detailed explanation. It should be understood that the specific values ​​given in this embodiment are merely illustrative examples, intended to intuitively demonstrate the beneficial effects of the technical solution, and do not constitute a limitation on the scope of protection of this invention.

[0165] Specifically, this embodiment selects a BBD4060 dual-inlet, dual-outlet coal mill as the monitoring object. The mill's cylinder diameter is 3.8 meters, its length is 6 meters, and its normal operating speed is set to 18 revolutions per minute. Based on this speed, the time period T_cycle required for the cylinder to complete one revolution is approximately 3.333 seconds. Four wireless sensor nodes are evenly arranged axially on the cylinder surface, located at the drive end, the left side of the center, the right side of the center, and the non-drive end, respectively. Each node integrates a three-axis MEMS accelerometer with a sampling rate set to 1 kHz, used to collect vibration signals during the cylinder's rotation. The base station is fixedly installed on a stationary platform approximately 5 meters from the side of the cylinder, with the antenna height level with the cylinder's central axis.

[0166] Based on the above hardware configuration, the system operates according to the method flow of Embodiments 1 to 6, and the specific process is as follows: First, rotation angle sensing and channel prediction are performed: a non-contact magnetoresistive angle sensor installed at the end of the cylinder outputs an angle signal in real time with an accuracy of 0.1 degrees.

[0167] When the base station receives the angle signal, it invokes a pre-built channel quality mapping table. Based on the current angle θ(t), it predicts the expected channel quality for each angle interval in the next cycle. For example, when it is predicted that the node is about to turn to the 180-degree to 200-degree interval away from the base station, the expected signal-to-noise ratio will drop sharply from the normal 15dB to about 2dB, and the system determines this interval as a "low signal-to-noise ratio interval".

[0168] The channel quality mapping table, calibrated offline, records the received signal strength (RSSI), signal-to-noise ratio (SNR), and correction coefficients after penetrating a metal wall for each degree angle from 0 to 360 degrees. Considering the cylinder is made of approximately 50 mm thick steel plate, the attenuation of the wireless signal during penetration is severe and non-linear; the correction coefficient K(θ) fluctuates by 10 to 15 dB at different angles.

[0169] Then, TDMA time slot allocation and dynamic scheduling are performed: The base station divides the 3.333-second rotation period into 12 fixed time slots, each with a length of approximately 278 milliseconds. Based on the prediction results, the time slot with the highest expected signal-to-noise ratio (e.g., corresponding to the 0-30 degree range) is allocated to the node with the highest current priority. Assuming that the current state evaluation value of the node on the left side of the middle is relatively high (indicating that the material level in this area fluctuates greatly), it is marked as high priority and allocated 3 high-quality time slots. At the same time, for nodes that are about to enter the low signal-to-noise ratio range, their modulation and coding scheme (MCS) is dynamically adjusted through downlink signaling, switching from the conventional 16QAM to BPSK, and the transmit power is increased from +5dBm to +20dBm to combat penetration attenuation and multipath fading.

[0170] Simultaneously, edge-side feature extraction is performed at the nodes: each node collects 1024 triaxial vibration data points (corresponding to approximately 1 second) within its assigned time slot. The low-power microcontroller unit (MCU) inside each node calls a lightweight one-dimensional convolutional neural network (1D-CNN) embedded in read-only memory (ROM) for forward inference. At this point, the network, trained through knowledge distillation, is able to extract a 16-dimensional feature vector from the complex vibration waveform. The original approximately 6KB of raw data is compressed into a 64-byte feature vector, achieving a compression ratio of 96:1. The node encapsulates this feature vector into a data packet, embeds a local timestamp, and transmits it to the base station via a wireless transceiver at the start of a specified fixed time slot.

[0171] After receiving data, the base station performs timestamp calibration and state fusion: The base station uses the broadcast time-slot synchronization beacon as a global time reference, converting the local timestamps uploaded by each node into a unified global time. By dynamically adjusting the frequency division coefficient of the sampling clocks of each node, long-term crystal oscillator drift is suppressed to within 10 μs / hour, achieving sub-millisecond time synchronization. Subsequently, the base station sorts and interpolates the feature vectors from four nodes at the same time according to their installation angles, constructing a two-dimensional state matrix reflecting the circumferential state distribution of the cylinder. Further, matrices for 10 consecutive cycles are stacked to form a three-dimensional spatiotemporal state tensor. This tensor is input into a fully connected neural network layer, inferring that the current overall material level is 65% (within the normal half-grinding range) and the coal quality distribution uniformity is 92% (no significant uneven grinding).

[0172] In summary: Traditional solutions typically employ the CSMA / CA protocol, lacking channel prediction capabilities and transmitting raw waveform data. When the coal mill rotates to the side facing away from the base station, the packet loss rate is high, leading to frequent interruptions in the level display. Furthermore, due to the lack of a unified time reference, the time alignment error of multi-node data fluctuates randomly, making it impossible to effectively train the multi-source fusion model, resulting in low level inference accuracy. However, by adopting the solution of this invention, the packet loss rate is steadily reduced due to actively avoiding channel fading intervals and employing adaptive modulation and coding, achieving continuous and reliable monitoring. Time alignment errors can also be controlled through TDMA time slot synchronization and drift compensation, meeting the requirements of phase analysis. Finally, thanks to edge-side feature compression and spatiotemporal state fusion, the level inference accuracy is improved.

[0173] The comparative data above demonstrates that this invention, by deeply integrating the physical laws of rotation with communication protocols and edge computing, constructs an integrated closed loop encompassing sensing, prediction, scheduling, compression, and fusion. This effectively solves the technical challenges of unreliable communication, data asynchrony, and bandwidth limitations in a sealed, rotating metal environment, achieving unexpected technical results. This solution not only significantly improves the reliability and accuracy of the monitoring system but also substantially reduces the power consumption and congestion of wireless transmission, extending the maintenance cycle of sensor nodes, and possesses extremely high engineering practical value.

[0174] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention, such as replacing the one-dimensional convolutional neural network with other lightweight network models with feature extraction capabilities, or applying the time-division multiple access scheduling strategy to other monitoring scenarios of rotating machinery, should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for polling coal quality in a dual-inlet, dual-outlet coal mill, characterized in that, include: The wireless sensor node installed on the rotating cylinder of the double-inlet double-outlet coal mill receives the rotation angle relative to the stationary base station, and the wireless signal needs to penetrate the metal wall of the rotating cylinder; The channel quality mapping table is invoked to predict the rotation angle and obtain the prediction result. The prediction result is the expected channel quality corresponding to each angle interval in the next rotation cycle. The channel quality mapping table contains coefficients to correct the attenuation of the wireless signal after penetrating the metal wall. The rotation period of the rotating cylinder is divided into fixed time slots. The expected signal-to-noise ratio sequence and priority weight sequence of the prediction results are matched, and the coding scheme and transmission power of the node are dynamically adjusted. The system receives the triaxial vibration signal sent by the node within the fixed time slot, generates a feature vector representing the coal quality or material level state through a lightweight one-dimensional convolutional neural network deployed in the node, and uploads the feature vector as part of the data packet. Based on the feature vectors of all nodes, a state matrix is ​​constructed and input into a fully connected neural network layer for material level inference. The inference results are used for coal feed control and coal mill operation status monitoring. The state matrix is ​​used to reflect the spatiotemporal distribution of the internal state of the coal mill.

2. The method for coal quality measurement and polling in a dual-inlet, dual-outlet coal mill according to claim 1, characterized in that, The process of constructing the channel quality mapping table specifically includes: Receive offline calibration data and construct a basic mapping table. , wherein For angle Signal strength at that location For angle Signal-to-noise ratio at the location For angle Bit error rate at the location; Receiving angle Signal arrival time distribution and corresponding power attenuation coefficient Construct a multipath feature matrix , where m is the number of resolvable multipaths; Obtain the attenuation coefficient of the wireless signal penetrating the metal wall of the rotating cylinder, and adjust the angle accordingly. The signal strength at the specified location is corrected to obtain the corrected signal strength. The basic mapping table, multipath feature matrix, and corrected signal strength are then combined. Construct a three-layer channel quality mapping table based on angle index association. ,in: The first layer is , used for real-time channel quality query; The second layer is , used for multipath equalization parameter calculation; The third layer is = , The rate of change of channel quality, It is the acceleration of channel quality changes, used to predict the channel quality change trend at the next moment.

3. The method for coal quality measurement and polling in a dual-inlet, dual-outlet coal mill according to claim 1, characterized in that, The online update process of the channel quality mapping table specifically includes: The data packets from the node are received, and the actual receiving angle and the actual measured channel quality parameters are extracted. The channel quality parameters include the actual received signal strength, the actual signal-to-noise ratio, and the actual bit error rate. The adaptive forgetting factor is calculated based on the channel quality change rate, the mapping table value for the corresponding angle is updated, and adjacent angles are smoothed. Calculate the root mean square error of the updated mapping table. When the root mean square error is less than the first threshold, confirm the update is valid and store the update result. When the root mean square error is between the first threshold and the second threshold, trigger local recalibration. When the root mean square error is greater than the second threshold, trigger a global recalibration warning. The history of rotation cycle changes is obtained, and linear regression is used to predict the channel quality change rate for the next cycle, which is then used for time slot allocation pre-scheduling.

4. The method for coal quality measurement and polling in a dual-inlet, dual-outlet coal mill according to claim 1, characterized in that, The process of dynamically adjusting the coding scheme and transmission power of the node specifically includes: The expected signal-to-noise ratio is divided into continuous intervals, and a mapping relationship between the continuous intervals and the modulation and coding scheme is constructed. The continuous intervals include low signal-to-noise ratio intervals, medium signal-to-noise ratio intervals, high signal-to-noise ratio intervals, and ultra-high signal-to-noise ratio intervals. The effective data rate is calculated based on the modulation and coding scheme, and the transmit power of the node is adjusted synchronously based on the continuous interval. When the expected signal-to-noise ratio fluctuates at the boundaries of adjacent intervals, the hysteresis comparator is activated to prevent frequent switching of the coding scheme. When a data packet transmission fails, the transmit power of the node is increased according to the expected signal-to-noise ratio range, and a low-order modulation and coding scheme is used for retransmission.

5. The method for coal quality measurement and polling in a dual-inlet, dual-outlet coal mill according to claim 1, characterized in that, The training and deployment method for the lightweight one-dimensional convolutional neural network specifically includes: In the offline phase, a high-performance server is used as the teacher model, and triaxial vibration data including empty mill, half-mill, full mill, steel ball loading deviation and coal grindability change are used as the training set to train the lightweight one-dimensional convolutional neural network as the student model for knowledge distillation. The weights of the trained student model are converted from floating-point numbers to fixed-point numbers, and the forward inference computation graph is stored in the read-only memory of the low-power microcontroller unit of the node. During the online inference phase, the node only performs the forward inference calculation, inputting the real-time collected triaxial vibration data and directly outputting the feature vector.

6. The method for coal quality measurement and polling in a dual-inlet, dual-outlet coal mill according to claim 1, characterized in that, The process of constructing the state matrix specifically includes: Obtain the geographical location or installation angle identifier of all the nodes, and sort and interpolate the feature vectors from different nodes at the same time according to the corresponding rotation angle to form a two-dimensional state matrix, which is used to reflect the circumferential state distribution of the cylinder. The two-dimensional state matrices at consecutive time points are obtained and stacked to form a three-dimensional spatiotemporal state tensor; The three-dimensional spatiotemporal state tensor is input into the fully connected neural network layer to infer the overall material level and coal quality distribution uniformity within the cylinder.

7. The method for coal quality measurement and polling in a dual-inlet, dual-outlet coal mill according to claim 6, characterized in that, The priority of each node is determined based on the state evaluation value, and the priority mapping relationship includes: The state evaluation value of the corresponding node is calculated based on the feature vector, and the state evaluation values ​​of all the nodes are sorted from high to low to obtain the priority weight sequence. The reference number of fixed time slots allocated to each node in each rotation cycle is determined according to the priority weight sequence. According to the channel quality mapping table, the expected signal-to-noise ratios of each node in the current rotation period are sorted from high to low to obtain the expected signal-to-noise ratio sequence. The expected signal-to-noise ratio sequence and the priority weight sequence are weighted and matched to allocate more fixed time slots to nodes with high priority that are in the high expected signal-to-noise ratio angle range. When the status evaluation value exceeds the preset emergency threshold, the corresponding node will be marked as an emergency priority node; For the emergency priority node, it is forcibly allocated an additional fixed time slot exceeding the baseline number, and dynamically preempts the fixed time slots already allocated to the node at the end of the priority weight sequence as needed; The number of fixed time slots reserved by the preempted node is reduced to a preset minimum guarantee value, which is adaptively determined based on the total number of nodes in the system and the cycle duration.

8. The method for coal quality measurement and polling in a dual-inlet, dual-outlet coal mill according to claim 1, characterized in that, It also includes a slot synchronization and timestamp calibration module: At the start of the rotation cycle, a time slot synchronization beacon is broadcast, the time slot synchronization beacon containing the current cycle number and a global time reference; When the node sends the triaxial vibration signal, it embeds a local timestamp and converts it into global time using the global time reference. The time deviation of the nodes at the same angle within adjacent cycles is calculated, and the frequency division coefficient of the local sampling clock of the node is dynamically adjusted according to the time deviation to achieve sub-millisecond synchronization with the rotation cycle of the rotating cylinder.

9. A coal quality measurement and polling system for a dual-inlet, dual-outlet coal mill, characterized in that, The method for measuring coal quality in a dual-inlet, dual-outlet coal mill, as described in any one of claims 1-8, specifically includes: The communication unit is used to receive rotation angles, broadcast time slot synchronization beacons, and data packets uploaded by the receiving node; The processing unit is used to call the channel quality mapping table to perform channel prediction, allocate fixed time slots, adjust the coding scheme and transmit power, construct the state matrix, and perform material level inference. The storage unit is used to store the channel quality mapping table, the fixed weights of the lightweight one-dimensional convolutional neural network, and the model parameters of the fully connected neural network layer. The wireless sensor node integrates a triaxial vibration sensor, a low-power microcontroller unit, and a wireless transceiver.

10. A coal quality measurement and polling system for a dual-inlet, dual-outlet coal mill according to claim 9, characterized in that, The read-only memory of the low-power microcontroller unit contains the forward inference computation graph of the lightweight one-dimensional convolutional neural network; the wireless transceiver is used to transmit data packets containing the feature vector within the allocated fixed time slots.

Images