Online identification and fault warning method for slack coal sorting state by fusing attention mechanism and knowledge distillation

Abstract

The application provides a pebble coal separation state online identification and fault early warning method fusing an attention mechanism and knowledge distillation. The method comprises the following steps: collecting multi-source sensor time series data of key parts of a pebble coal separation system, and pre-processing and working condition labeling the data to construct a multi-class labeled data set; constructing a high-precision teacher model based on a one-dimensional convolutional neural network and a Transformer encoder hybrid architecture, and training the teacher model using the data set; based on the knowledge distillation technology, using the soft labels output by the teacher model and the hard labels of the labeled data set as the supervision signals, training a student model and deploying the student model to an edge computing node of a distributed control system; using the deployed model to perform online inference on the sensor time series data, outputting the separation state category and the corresponding probability, and triggering corresponding warning or protection actions. The method fuses the attention mechanism and the knowledge distillation technology, realizes high-precision online identification of the separation state, and realizes lightweight deployment of the model.

Description

Technical Field

[0001] This application relates to the field of stone and coal sorting technology, and in particular to a method for online identification of stone and coal sorting status and fault early warning that integrates attention mechanism and knowledge distillation. Background Technology

[0002] Currently, coal and stone separation is a crucial component of the pulverizing system in coal-fired power plants, and its operational status directly impacts pulverized coal recovery efficiency and equipment safety. In related technologies, a basic process monitoring system has been constructed by using threshold monitoring of sensor parameters such as fan current, pipeline pressure, and material level height, along with univariate statistical methods.

[0003] However, in the relevant technologies, the stone and coal sorting monitoring schemes directly use fixed thresholds or single time point feature inputs, without fully exploring the long-range dependence and cross-channel coupling relationship of multi-sensor time series data. This may lead to serious delays in early warning, making it difficult to identify hidden faults such as decreased sorting efficiency and coal dust escape in advance, or causing frequent false alarms due to coal quality fluctuations, thereby affecting the operators' trust in alarm information and the accuracy of fault location. Summary of the Invention

[0004] This application aims to at least partially address one of the technical problems in the related art.

[0005] Therefore, the first objective of this application is to propose an online identification and fault early warning method for stone and coal sorting status that integrates attention mechanism and knowledge distillation.

[0006] The second objective of this application is to propose an electronic device.

[0007] The third objective of this application is to provide a computer-readable storage medium.

[0008] To achieve the above objectives, the first aspect of this application is to propose an online identification and fault early warning method for stone and coal sorting status that integrates attention mechanisms and knowledge distillation, comprising the following steps:

[0009] Collect time-series data from multiple sources of sensors at key parts of the stone and coal sorting system, and preprocess and label the time-series data of the multiple sources of sensors to construct a multi-category labeled dataset that includes normal state, low sorting efficiency state, material blockage state and coal powder escape state. A high-precision teacher model based on a hybrid architecture of a one-dimensional convolutional neural network and a Transformer encoder is constructed. The teacher model is trained using the labeled dataset so that the teacher model outputs the predicted probability distribution of various sorting states. Based on knowledge distillation technology, a lightweight student model is trained using the soft labels output by the teacher model and the hard labels of the labeled dataset as supervision signals, and the trained student model is deployed to the edge computing nodes of the distributed control system. The deployed student model performs online inference on the real-time sensor time-series data, outputs the sorting status category and corresponding probability at the current moment, and triggers the corresponding level of early warning or linkage protection action according to the preset hierarchical early warning rules.

[0010] To achieve the above objectives, a second aspect of this application also proposes an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the online identification and fault warning method for stone and coal sorting status fused with attention mechanism and knowledge distillation as described in any one of the first aspects above.

[0011] To achieve the above objectives, the third aspect of this application also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the online identification and fault warning method for stone and coal sorting status that integrates attention mechanism and knowledge distillation as described in any of the first aspects above.

[0012] The technical solution provided by the embodiments of this application brings at least the following beneficial effects: This application adopts a hybrid deep learning architecture combining a one-dimensional convolutional neural network and a Transformer encoder, supplemented by a multi-head attention mechanism and knowledge distillation technology to achieve model lightweighting. Through multi-source time-series data acquisition and working condition annotation, high-precision teacher model construction, knowledge distillation and lightweight model deployment, and online identification and hierarchical early warning, high-precision online identification of the state of stone and coal sorting and lightweight model deployment are achieved, which improves the accuracy, reliability and timeliness of online identification of the state of stone and coal sorting, and is easy to implement in practical applications.

[0013] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0014] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of an online identification and fault warning method for stone and coal sorting status that integrates attention mechanism and knowledge distillation, as proposed in an embodiment of this application. Figure 2This is a flowchart illustrating a specific online identification and fault warning method for stone and coal sorting status that integrates attention mechanism and knowledge distillation, as proposed in an embodiment of this application. Figure 3 This is a schematic diagram of the structure of an online identification and fault early warning system for stone and coal sorting that integrates attention mechanism and knowledge distillation, as proposed in an embodiment of this application. Detailed Implementation

[0015] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0016] The following description, with reference to the accompanying drawings, describes an online identification and fault warning method for stone and coal sorting status that integrates attention mechanisms and knowledge distillation, as proposed in an embodiment of this application.

[0017] Example 1 Figure 1 This is a flowchart of an online identification and fault warning method for stone and coal sorting status that integrates attention mechanism and knowledge distillation, as proposed in an embodiment of this application. Figure 1 As shown, the method includes the following steps: Step S101: Collect time-series data from multiple sources of sensors at key locations in the stone and coal sorting system, and preprocess and label the time-series data of the multiple sources of sensors to construct a multi-category labeled dataset that includes normal state, low sorting efficiency state, material blockage state, and coal powder escape state.

[0018] Specifically, this step acquires multi-source sensor time-series data reflecting the operating status of the stone and coal sorting system. This data originates from sensors deployed at various key locations within the system and covers a variety of physical quantity signals reflecting equipment operating conditions and sorting effectiveness. Preprocessing operations are performed on the acquired time-series data, including time alignment, missing value imputation, and outlier removal, to eliminate inconsistencies and noise interference during data acquisition.

[0019] Subsequently, based on domain expertise and operational records, the preprocessed time-series data is labeled with operating conditions, and the data samples within each time window are mapped to preset sorting state categories. This category set includes at least normal state, low sorting efficiency state, material blockage state, and coal powder escape state, thereby constructing a multi-category labeled dataset that covers all operating conditions and has clear state labels.

[0020] As one implementation method, data collection can cover a continuous operating cycle of no less than six months to ensure that the dataset includes typical operating conditions such as full load, reduced load, start-up and shutdown, and sudden changes in coal quality. The annotation process can adopt a double-blind cross-validation mechanism, which is completed and compared independently by multiple annotators. Inconsistent samples are arbitrated by senior experts. At the same time, for transitional samples with ambiguous boundaries, a state confidence level can be added to improve the annotation quality and the robustness of model training.

[0021] Therefore, the multi-class labeled dataset constructed in this step provides high-quality and comprehensive supervision signals for the training of subsequent deep learning models, ensuring that the models can accurately learn multi-source temporal feature patterns under different sorting states, thus laying a reliable data foundation for high-precision state identification and fault early warning.

[0022] Step S102: Construct a high-precision teacher model based on a hybrid architecture of a one-dimensional convolutional neural network and a Transformer encoder. Train the teacher model using a labeled dataset so that the teacher model outputs the predicted probability distribution of various sorting states.

[0023] Specifically, the teacher model constructed in this step adopts a hierarchical structure design. Its core components include a convolutional feature extraction module for extracting local temporal features of multi-sensor time-series data, a Transformer encoder module for capturing long-range dependencies between different sensor channels and between different time steps, an attention pooling module for compressing sequence features into a fixed-dimensional global feature vector, and a classification output module for outputting classification probabilities.

[0024] The convolutional feature extraction module performs local pattern mining on the input multi-channel temporal data through at least one layer of one-dimensional convolutional operations. Batch normalization and non-linear activation functions are applied after each convolutional layer to enhance feature representation capabilities, and pooling operations are used to progressively reduce temporal resolution. The Transformer encoder module introduces positional encoding into the feature sequence output by the convolutional module to preserve temporal order information. It utilizes a multi-head self-attention mechanism to model complex relationships across channels and time steps, and residual connections and layer normalization ensure training stability. The attention pooling module uses a learnable query vector to weight and aggregate the encoder output sequence, generating a feature representation with global discriminative capabilities. The classification output module maps global features to the probability distribution of the corresponding sorting state category.

[0025] During training, the cross-entropy loss function is used as the optimization objective, and the model parameters are iteratively updated through the backpropagation algorithm, so that the teacher model can accurately output the four-dimensional classification probability vectors of normal state, low sorting efficiency state, material blockage state, and coal powder escape state.

[0026] As one implementation, the teacher model can employ three cascaded one-dimensional convolutional blocks as the convolutional feature extraction module. Each convolutional block consists of two convolutional layers, a batch normalization layer, a ReLU activation function, and a max pooling layer, with the number of convolutional kernels increasing progressively. The Transformer encoder module can be configured as a four-layer encoder, each layer containing an eight-head self-attention mechanism, with a hidden layer dimension of 512. The attention pooling module can be configured with four learnable query vectors, outputting 1024-dimensional global features. The classification module consists of two fully connected layers and a Softmax activation function. During training, the Adam optimizer is used, combined with learning rate decay and early stopping strategies to prevent overfitting.

[0027] Therefore, by constructing a high-precision teacher model, this step can fully explore the local temporal features and cross-channel long-range dependencies in multi-sensor time-series data, providing high-quality soft-label knowledge for subsequent knowledge distillation, thus laying the foundation for lightweighting the model while ensuring recognition accuracy.

[0028] Step S103: Based on knowledge distillation technology, a lightweight student model is trained using the soft labels output by the teacher model and the hard labels of the labeled dataset as supervision signals, and the trained student model is deployed to the edge computing nodes of the distributed control system.

[0029] Specifically, this step first utilizes a pre-trained high-precision teacher model as a knowledge source. This teacher model can output predicted probability distributions for various sorting states, i.e., soft labels. Simultaneously, it combines these with manually labeled real state categories from the labeled dataset, i.e., hard labels, to form a dual supervision signal for training the student model. The student model employs a simplified network architecture, with significantly fewer parameters and lower computational complexity than the teacher model, to adapt to the limited resource constraints of edge computing nodes in distributed control systems.

[0030] During the distillation training process, the student model simultaneously learns the true class discrimination boundary corresponding to the hard labels and the inter-class similarity information contained in the soft labels. The model is optimized through a weighted combination loss function, which includes the hard label cross-entropy loss and the soft label KL divergence loss. The soft label loss softens the probability distribution of the teacher model by introducing a temperature parameter, enabling the student model to more fully transfer the decision knowledge of the teacher model.

[0031] After distillation training is completed, the lightweight student model is exported in a cross-platform compatible format, and model optimization is performed to reduce inference computation overhead. Then, it is integrated into the edge computing node of the distributed control system, and the corresponding data reading, preprocessing, inference and result parsing interfaces are configured to realize online deployment and real-time inference of the model.

[0032] As one implementation method, the student model can employ a simplified one-dimensional convolutional neural network architecture, retaining only a small number of convolutional layers and global average pooling layers, with its parameter count controlled to within one-tenth of that of the teacher model. In the distillation loss function, the weights of the hard-label cross-entropy loss and the soft-label KL divergence loss can be set to 0.3 and 0.7, respectively, and the temperature parameter can be set within the range of 4 to 10. Model export can utilize an open neural network exchange format, simultaneously performing operator fusion and constant folding optimization. Finally, the model is loaded through a dedicated algorithm function block in the distributed control system, providing a standardized inference interface.

[0033] Therefore, this step achieves knowledge transfer from a high-precision teacher model to a lightweight student model through knowledge distillation technology. While maintaining high recognition accuracy, it significantly reduces the computational complexity and storage requirements of the model, enabling the deep learning model to be successfully deployed on the edge computing nodes of resource-constrained distributed control systems, thus meeting the deployment requirements for real-time online identification and early warning in industrial sites.

[0034] Step S104: The deployed student model is used to perform online reasoning on the real-time collected sensor time series data, outputting the sorting status category and corresponding probability at the current moment, and triggering the corresponding level of early warning or linkage protection action according to the preset hierarchical early warning rules.

[0035] Specifically, the deployed student model continuously reads real-time data from multi-source sensors at key locations on-site at a preset sampling frequency. Each inference is based on a time-series data segment within a sliding time window to determine the state and outputs the sorted state category at the current moment and the probability value corresponding to each category.

[0036] The sorting status categories include at least normal status, low sorting efficiency status, material blockage status, and coal powder escape status. The system has built-in hierarchical early warning rules. These rules automatically match the corresponding early warning level and trigger the corresponding response action based on the status category and its probability value output by the model, combined with conditions such as the frequency or duration of the status's occurrence. The early warning levels include at least three levels: alert level, alarm level, and linkage protection level. Different levels correspond to different triggering conditions and execution actions: the alert level is used to remind operators to pay attention to potential anomalies; the alarm level is used to immediately notify on-site personnel and display fault information; and the linkage protection level automatically executes a preset sequence of protection actions when the alarm continues and preset conditions are met, such as closing branch valves, adjusting fan speed, or initiating a blockage clearing procedure.

[0037] As one implementation method, a prompt-level warning can be triggered when the model outputs "low sorting efficiency" multiple times consecutively and the probability exceeds a first threshold; an alarm-level warning can be triggered when the model outputs "material blockage" or "coal powder escape" and the probability exceeds a second threshold; and a linkage protection-level warning can be triggered when an alarm-level warning lasts for more than a preset duration and the probability of the state continues to exceed a second threshold. All parameters of the above warning rules can be configured and adjusted online according to the on-site working conditions. The system also supports the querying and statistics of historical warning records for continuous optimization of the warning strategy.

[0038] Therefore, this step achieves real-time online inference and hierarchical early warning through the deployed lightweight student model. It can meet the resource constraints of edge computing nodes while ensuring high recognition accuracy, realize early identification of the status of stone and coal sorting and hidden fault signs, and effectively balance the timeliness and accuracy of early warning through the hierarchical response mechanism, reduce the false alarm rate, and provide automated linkage means for equipment protection, significantly improving the reliability and security of system operation.

[0039] The following describes in detail the specific implementation process of each step in the online identification and fault warning method for stone and coal sorting based on the fusion attention mechanism and knowledge distillation of this application, using several specific embodiments.

[0040] Example 2 Based on the above embodiments, this embodiment provides a detailed description of the specific implementation of step S101, "collecting multi-source sensor time-series data from key parts of the stone and coal sorting system, preprocessing and labeling the multi-source sensor time-series data, and constructing a multi-category labeled dataset including normal state, low sorting efficiency state, material blockage state, and coal powder escape state".

[0041] In this embodiment, multi-source sensor time-series data for at least six months is first continuously collected from the historical database of the distributed control system (DCS) of the coal-fired power plant to ensure that the dataset covers all operating conditions, including full load, load reduction, start-up and shutdown, and sudden changes in coal quality. The collected sensor data specifically includes pressure and current signals before and after the blower, pressure signals at various sections of the conveying pipeline, material level signals in the sorting bin and gravel / coal bunker, carbon content detection data for gravel / coal, and opening status signals of the air supply valve and exhaust valve. Data acquisition uses a sampling frequency synchronized with on-site production. After acquisition, the raw multi-source sensor time-series data is obtained, which serves as input for subsequent steps.

[0042] Then, the acquired raw multi-source sensor time series data is time-aligned to ensure that each sensor channel is on the same time base. Subsequently, missing value and outlier cleaning is performed, such as filling missing values ​​with interpolation and removing outliers based on statistical thresholds, to obtain preprocessed time series data, which serves as the input for subsequent steps.

[0043] Finally, experienced professionals, combining DCS operation logs, equipment maintenance records, and periodic inspection reports, annotated the preprocessed time-series data with operational conditions. Specifically, two annotators independently labeled each time-series data sample with a unique state label, which could be one of the following: normal state, low sorting efficiency state, material blockage state, or coal powder escape state. After annotation, the annotation results of the two annotators were compared. If the annotation results were consistent, the label was adopted as the final label; if the annotation results were inconsistent, the data was submitted to senior experts for arbitration, who determined the final label based on on-site records and professional knowledge. For transitional samples with ambiguous boundaries, the annotators simultaneously labeled their state confidence level. This confidence level was used for sample weighting during subsequent model training; that is, samples with higher confidence levels were given higher weights in the loss function to improve the model's learning effect on clear samples. Through the above double-blind cross-validation mechanism, a multi-class labeled dataset containing normal state, low sorting efficiency state, material blockage state, and coal powder escape state was finally obtained, with each time-series data sample corresponding to a unique state label and an optional confidence level.

[0044] This embodiment effectively reduces the subjective bias of manual annotation through double-blind cross-validation and expert arbitration mechanism. At the same time, it introduces state confidence level for sample weighting, which significantly improves the accuracy and reliability of the labeled dataset. This provides a high-quality data foundation for the subsequent training of high-precision teacher models and lightweight student models, thereby ensuring the overall performance of the online identification and fault early warning method for stone and coal sorting.

[0045] Example 3 Based on the above embodiments, this embodiment provides a detailed description of the specific implementation of step S102, "constructing a high-precision teacher model based on a hybrid architecture of a one-dimensional convolutional neural network and a Transformer encoder, and training the teacher model using the labeled dataset so that the teacher model outputs the predicted probability distribution of various sorting states".

[0046] In this embodiment, the specific implementation process of step S102 is as follows. First, the input layer of the teacher model is constructed. This input layer receives time-series data from N sensor channels within a preset time window, constructing a two-dimensional feature matrix of T×N, where T is the time step and N is the number of sensor channels. As an example, T is set to 60, corresponding to 1 minute of time-series data, and N is no less than 10 sensor channels. The input data is first subjected to Z-score standardization to make the mean of each channel data 0 and the standard deviation 1. The standardization formula is: ,in Let be the raw acquisition value of the j-th sensor channel at the i-th time step. The values ​​are standardized. and Let be the mean and standard deviation of the j-th sensor channel on the training dataset, respectively. , The standardized two-dimensional feature matrix serves as the input for subsequent modules.

[0047] Next, a one-dimensional convolutional neural network local feature extraction layer for the teacher model is constructed. This layer consists of three cascaded one-dimensional convolutional blocks. Each convolutional block has the following structure: Conv1D layer, batch normalization layer, ReLU activation function, Conv1D layer, batch normalization layer, ReLU activation function, and MaxPool1D layer. Specifically, the first convolutional block has 64 kernels in its Conv1D layer with a kernel size of 3 and a stride of 1, and a pooling kernel size of 2. The second convolutional block has 128 kernels in its Conv1D layer, with the remaining parameters identical to the first block. The third convolutional block has 256 kernels in its Conv1D layer, with the remaining parameters identical to the first block. These three convolutional blocks sequentially halve the temporal resolution, resulting in a final output feature shape of (7, 256). The formula for calculating the convolution operation is as follows: ,in, Let be the output feature value of the l-th convolutional layer at time step t. The kernel size is the convolution kernel size. The number of channels for the input feature. As weight, This is a bias term.

[0048] Next, a Transformer encoder layer for the teacher model is constructed. This layer introduces sine and cosine positional encoding into the feature sequence output by the one-dimensional convolutional neural network, which is then element-wise added to the input features before being input to the encoder. The encoder consists of four Transformer encoder layers, each with eight self-attention layers, a hidden layer dimension of 512, and a feedforward network layer dimension of 2048, employing a Pre-LN structure to improve training stability. The calculation formula for the multi-head self-attention mechanism is as follows: , ,in , , These are query, key, and value matrices, respectively. Let be the dimension of the key vector. For the number of attention heads, The output is a linearly transformed weight matrix. This mechanism captures the long-range dependencies between different sensor channels and different time steps, and uses residual connections and layer normalization structures to ensure training stability.

[0049] Subsequently, a multi-head attention pooling layer for the teacher model is constructed. This layer sets four learnable query vectors with a dimension of 512. The scaling dot product attention weights of each query vector and the output sequence of the Transformer encoder layer are calculated. The sequences are then weighted and summed based on the weights to obtain four pooled features of 256 dimensions. After concatenation, a global feature vector of 1024 dimensions is obtained.

[0050] Next, construct the classification output layer of the teacher model. Global features are sequentially passed through a fully connected layer (1024→512), a ReLU activation function, a Dropout layer (with a deactivation rate of 0.3), a fully connected layer (512→4), and a Softmax activation function, finally outputting a four-dimensional classification probability vector corresponding to the four sorting states.

[0051] Finally, using the cross-entropy loss function as the optimization objective, the Adam optimizer was employed for end-to-end training of the teacher model. During training, a cosine annealing learning rate decay strategy and an early stopping mechanism were used to prevent overfitting. The initial learning rate was set to 1e-4, the weight decay coefficient to 1e-5, the batch size to 64, and the maximum number of training epochs to 200. Training was terminated when the validation set loss did not decrease for 20 consecutive epochs. The total number of parameters in the teacher model was approximately 12 million.

[0052] Therefore, through the above specific implementation methods, the teacher model can fully extract local temporal features and long-range dependencies from multi-source time-series data, output high-precision sorting state probability distribution, provide high-quality soft labels for subsequent knowledge distillation, and thus ensure the recognition accuracy of the lightweight student model.

[0053] Example 4 Based on the above embodiments, this embodiment provides a detailed description of the specific implementation of step S103 above, which involves "training a lightweight student model based on knowledge distillation technology, using the soft labels output by the teacher model and the hard labels of the labeled dataset as supervision signals, and deploying the trained student model to the edge computing nodes of the distributed control system".

[0054] In this embodiment, step S103 is further refined into three sequentially connected sub-steps. First, in the first sub-step, a lightweight student model is constructed. The input of this model is time-series data after Z-score normalization, with a shape of (T, N), where T = 60 time steps and N ≥ 10 sensor channels. The student model adopts a simplified one-dimensional convolutional neural network architecture, retaining only two Conv1D layers. The first layer has 32 convolutional kernels with a kernel size of 3, and the second layer has 64 convolutional kernels with a kernel size of 3. A global average pooling layer is then applied after the convolutional layers to compress the time-series features into a 64-dimensional one-dimensional feature vector. The classification module sequentially maps the 64-dimensional features to 32 dimensions through fully connected layers, then through the ReLU activation function, and finally through fully connected layers to map the 32-dimensional features to 4 dimensions. Finally, a Softmax activation function is applied to output a 4-dimensional classification probability vector, corresponding to the normal state, low sorting efficiency state, material blockage state, and coal powder escape state, respectively. The student model has approximately 80,000 parameters, which is only 0.7% of the parameters in the teacher model.

[0055] Subsequently, in the second sub-step, the student model is trained using the soft labels output by the teacher model and the hard labels from the labeled dataset as supervision signals, employing a distillation loss function. The distillation loss function is a weighted combination of the hard label cross-entropy loss and the soft label KL divergence loss, and its formula is: ,in, , .

[0056] In this embodiment, the hard label loss weighting coefficient Set to 0.3, the soft label loss weight coefficient Set to 0.7, distillation temperature coefficient The value of is between 4 and 10. By increasing the weight of soft labels and softening the predicted probability distribution of the teacher model, the student model can fully learn the similarity information between categories. The maximum number of training epochs for the student model is set to 100. It adopts the same early stopping mechanism as the teacher model, and judges the convergence status of the model based on the validation set loss. It usually converges after 50 to 80 training epochs. After training, the student model can output a classification probability distribution that is highly similar to that of the teacher model.

[0057] Finally, in the third sub-step, the trained student model is exported in the Open Neural Network Exchange Format (ONNX) to ensure cross-platform compatibility. During the export process, operator fusion and constant folding are performed simultaneously to optimize the model and reduce computational overhead for inference. Furthermore, the model is quantized using INT8, compressing the 32-bit floating-point parameters into 8-bit integers to reduce memory usage and computational load. Simultaneously, multi-core CPU parallel computing is enabled to allocate inference computation tasks, and an inference caching mechanism is configured to avoid redundant data computation within the sliding window. After these optimizations, the single-step inference latency can be controlled within 50ms, meeting the real-time requirement of a 1Hz sampling frequency. Subsequently, a dedicated algorithm function block is developed in the DCS system. This function block can load ONNX format models and provide a standardized inference interface. The input of the function block is a list of real-time sensor data, and the output is the sorted state category and its corresponding probability value. The algorithm function block is mounted to the corresponding measurement point using the DCS configuration tool, and supporting interfaces for real-time data reading, data preprocessing, model inference, and result parsing are configured to complete the deployment of the model on the DCS edge computing node.

[0058] This implementation method constructs an extremely lightweight student model and uses knowledge distillation training to compress the number of parameters to 0.7% of the teacher model while ensuring high recognition accuracy. Combined with INT8 quantization and multi-core parallel optimization, the model inference latency is controlled within 50ms, thereby realizing real-time, high-precision online identification of stone and coal sorting status and fault early warning on DCS edge computing nodes.

[0059] Example 5 Based on the above embodiments, this embodiment provides a detailed description of the specific implementation of step S104, which involves "using the deployed student model to perform online reasoning on the real-time collected sensor time-series data, outputting the current sorting status category and corresponding probability, and triggering the corresponding level of early warning or linkage protection action according to the preset hierarchical early warning rules".

[0060] In this embodiment, step S104 is implemented as follows. First, during the online inference phase, a lightweight student model deployed on the edge computing node of the distributed control system continuously reads real-time sensor data at a sampling frequency of 1Hz. Each inference is based on a sliding window of data from the most recent T time steps for state discrimination, where T is 60, corresponding to a 1-minute time-series data window. Specifically, the system maintains a circular buffer of length 60. Each time new data arrives, the buffer content is updated, and the 60×N dimensional feature matrix (N being the number of sensor channels) within the current window is input into the student model. The model outputs the sorting state category at the current moment and the probability value corresponding to each category. The state categories include four types: normal state, low sorting efficiency state, material blockage state, and coal powder escape state. This output serves as the input basis for subsequent graded early warning judgment.

[0061] Subsequently, in the tiered early warning phase, the system determines the state category and probability of the model output level by level according to the preset tiered early warning rules. When the model outputs a low sorting efficiency state M times consecutively and the state probability exceeds the preset threshold P1, a prompt-level early warning is triggered. In this embodiment, M is set to 3 and P1 is set to 0.6. That is, when the inference results are low sorting efficiency for 3 consecutive times and the probability exceeds 0.6 each time, a prompt message containing the state category, confidence level, and suggested operation is pushed to the operator's terminal.

[0062] When the model identifies a blockage or pulverized coal escape state and the predicted probability of the corresponding state exceeds the preset threshold P2, an alarm-level warning is triggered. In this embodiment, P2 is set to 0.7. After the alarm is triggered, an on-site audible and visual alarm is immediately activated, and the fault type, occurrence time, and confidence level are displayed on the DCS operation interface. When the alarm-level warning lasts for N seconds and the state probability continues to exceed the threshold P2, a linkage protection-level warning is triggered, and preset protection actions are automatically executed. In this embodiment, N is set to 30, that is, when the alarm state lasts for 30 seconds and the probability always exceeds 0.7, the corresponding branch valve is automatically closed, the fan speed is adjusted, or the backflushing and unblocking program is started, and the fault event is recorded and pushed to maintenance personnel simultaneously.

[0063] In addition, the parameters M, N, P1, and P2 in the early warning rules can be configured and adjusted online through the human-machine interface. The system synchronously records the triggering conditions and handling results of each early warning, supports statistical analysis based on historical data, and automatically generates early warning parameter optimization suggestions, thereby realizing continuous optimization of the early warning strategy.

[0064] The above implementation method closely integrates online reasoning with a hierarchical early warning mechanism, which enables real-time and accurate identification of the state of stone and coal sorting and progressive early warning response. This effectively avoids the lag and false alarm problems of single threshold alarms, while improving the system's adaptability through configurable parameters and historical data feedback mechanisms.

[0065] Example 6 This embodiment will describe in detail the complete implementation of the online identification and fault early warning method for stone and coal sorting status based on the fusion of attention mechanism and knowledge distillation proposed in this application. The specific process of the method used in this embodiment is as follows: Figure 2 As shown, this method integrates attention mechanism and knowledge distillation technology, aiming to solve the technical problems of traditional threshold alarm lag, high false alarm rate, and difficulty in deploying existing machine learning models on edge computing nodes of power plant distributed control systems.

[0066] In this embodiment, firstly, multi-source time-series data acquisition and operating condition labeling are performed. Operating data for at least six months is continuously collected from the historical database of the distributed control system of the coal-fired power plant to ensure the dataset covers all operating conditions, including full load, reduced load, start-up and shutdown, and sudden changes in coal quality. The collected multi-source sensor data specifically includes: pressure and current signals before and after the blower, pressure signals at various sections of the conveying pipeline, material level signals in the sorting bin and cobblestone bin, carbon content detection data of cobblestone, and opening status signals of the air supply valve and exhaust valve. Data acquisition adopts a sampling frequency synchronized with on-site production. After acquisition, the data undergoes time alignment and cleaning of missing and outlier values. For example, missing values ​​can be filled using linear interpolation, and outliers can be identified and removed using the 3σ principle. Subsequently, the preprocessed dataset is labeled with operating conditions. Professionals with on-site operating experience manually label the historical time-series data using the distributed control system operation logs, equipment maintenance records, and periodic inspection reports. The labeled target sorting states are categorized into four types: normal state, low sorting efficiency state, material blockage state, and coal powder escape state. Each time-series data sample corresponds to a unique state label. The labeling process employs a double-blind cross-validation mechanism, where two labelers independently complete the labeling and then compare the results. Samples with inconsistent labeling are arbitrated by senior experts. For transitional samples with ambiguous boundaries, their state confidence level is simultaneously labeled. This confidence level will be used for sample weighting during subsequent model training to improve the model's ability to identify ambiguous samples.

[0067] Next, a high-precision teacher model is constructed and trained. This teacher model employs a hybrid architecture combining a one-dimensional convolutional neural network and a Transformer encoder, aiming for optimal recognition accuracy without limiting the number of model parameters or inference computation. The core structure of the teacher model consists of: an input layer, a one-dimensional convolutional neural network local feature extraction layer, a Transformer encoder layer, a multi-head attention pooling layer, and a classification output layer. Specifically, the input layer receives time-series data from N sensor channels within a preset time window, constructing a T×N two-dimensional feature matrix, where T is the number of time steps and N is the number of sensor channels. In this embodiment, T is set to 60, corresponding to 1 minute of time-series data, and N is greater than or equal to 10 sensor channels. The input data is first subjected to Z-score standardization, ensuring that the mean of each channel's data is 0 and the standard deviation is 1. The standardization formula is: ,in, It is the raw acquisition value of the j-th sensor channel at the i-th time step. These are standardized values. It is the mean of the j-th sensor channel on the training dataset. It is the standard deviation of the j-th sensor channel on the training dataset. , The local feature extraction layer of the one-dimensional convolutional neural network consists of three cascaded one-dimensional convolutional blocks. Each block has the following structure: Conv1D layer, batch normalization layer, ReLU activation function, Conv1D layer, batch normalization layer, ReLU activation function, and MaxPool1D layer. Specifically, the first convolutional block has 64 kernels in its Conv1D layer with a kernel size of 3 and a stride of 1, and a pooling kernel size of 2. The second convolutional block has 128 kernels in its Conv1D layer, with the remaining parameters identical to the first block. The third convolutional block has 256 kernels in its Conv1D layer, with the remaining parameters identical to the first block. These three convolutional blocks sequentially halve the temporal resolution, resulting in a final output feature shape of (7, 256). The formula for this convolution operation is as follows: ,in, This represents the output feature value of the l-th convolutional layer at the t-th time step, where ReLU is the rectified linear activation function and K is the kernel size. The number of channels for the input feature. It is the weight of the l-th convolutional layer, the c-th input channel, and the k-th convolutional kernel. It is the input feature of the (l-1)th layer, the c-th channel, and the (t+k-1)-th time step. This is the bias term of the l-th convolutional layer. The Transformer encoder layer contains 4 Transformer encoder layers, each with 8 self-attention layers. The hidden layer dimension is 512, and the feedforward network layer dimension is 2048. A Pre-LN structure is used to improve training stability. Position encoding uses sine and cosine position encoding, which is added element-wise with the input features before being input to the encoder. The scaled dot product attention formula is: The formula for multi-head self-attention is:

[0068] Where Q, K, and V are the query, key, and value matrices, respectively, obtained by linear transformation of the Transformer input features. Let h be the dimension of the key vector, and h be the number of attention heads. The output linear transformation weight matrix is ​​used after concatenating the multi-head attention layers. The multi-head attention pooling module sets four learnable query vectors of 512 dimensions. The scaling dot product attention weights of each query vector and the Transformer output sequence are calculated. The sequences are then weighted and summed to obtain four 256-dimensional pooled features, which are concatenated to obtain a 1024-dimensional global feature vector. In the classification module, the global features are sequentially passed through a fully connected layer (1024→512), a ReLU activation function, a Dropout layer (with a 0.3% inactivation rate), a fully connected layer (512→4), and a Softmax activation function, ultimately outputting a 4-dimensional classification probability vector. In this embodiment, the total number of parameters in the teacher model is approximately 12 million. The teacher model uses the cross-entropy loss function as the optimization objective and employs the Adam optimizer for end-to-end training. The dataset is divided into training (70%), validation (15%), and test (15%) sets according to time sequence, using temporal partitioning rather than random partitioning to avoid data leakage. In the training configuration, the initial learning rate was set to 1e-4, the weight decay coefficient was set to 1e-5, the training batch size was 64, and the maximum number of training epochs was 200. An early stopping mechanism was employed, terminating training when the validation set loss did not decrease for 20 consecutive epochs. The learning rate was dynamically adjusted using a cosine annealing strategy. During training, data augmentation operations such as random pruning, random magnitude scaling (0.9-1.1 times), and random channel dropping (5% probability) were performed on the time-series data to improve the model's generalization ability.

[0069] Then, a lightweight student model is trained and deployed using knowledge distillation technology. The student model employs a simplified one-dimensional convolutional neural network architecture, retaining only two Conv1D layers. The first layer has 32 kernels with a kernel size of 3, and the second layer has 64 kernels with a kernel size of 3. A global average pooling layer follows the convolutional layers, compressing the temporal features into a 64-dimensional one-dimensional feature vector. The classification module sequentially passes through a fully connected layer (64→32), a ReLU activation function, a fully connected layer (32→4), and a Softmax activation function, outputting a 4-dimensional classification probability vector. In this embodiment, the total number of parameters in the student model is approximately 80,000, only 0.7% of the number of parameters in the teacher model. During distillation training, the student model simultaneously receives two types of supervision signals: one is manually labeled hard labels, i.e., the true state categories labeled by experts; the other is the soft labels output by the teacher model, i.e., the predicted probability distributions of each category output by the teacher model. The distillation loss function is defined as a weighted combination of the hard label cross-entropy loss and the soft label KL divergence loss, as shown in the formula:

[0070] in, , . The weighting coefficients for hard label loss are... The weighting coefficients for soft label loss are... This represents the distillation temperature coefficient. In this embodiment, Set to 0.3, Set to 0.7, The value range is 4 to 10. The maximum number of training rounds for the student model is set to 100. An early stopping mechanism consistent with the teacher model is adopted, and the model convergence status is judged based on the validation set loss. Convergence is typically achieved after 50-80 training rounds. After distillation training, the lightweight student model is exported in an open neural network exchange format to ensure cross-platform compatibility. During the export process, operator fusion and constant folding are performed simultaneously to optimize the model and reduce the computational overhead of model inference. Furthermore, considering the computational resource limitations of the distributed control system controller, three optimizations are performed on the model: INT8 quantization is used to compress the model's 32-bit floating-point parameters into 8-bit integers, reducing memory usage and computational load; multi-core CPU parallel computing is enabled to allocate inference computation tasks; and an inference caching mechanism is configured to avoid redundant data computation within the sliding window. After optimization, the model's single-step inference latency can be controlled within 50ms. A dedicated algorithm function block is developed in the distributed control system. This function block can load the open neural network exchange format model and provide a standardized inference interface. The input of the function block is a list of real-time sensor data, and the output is the sorted state category and its corresponding probability value. By using the distributed control system configuration tool, algorithm function blocks are mounted to the corresponding measurement points, achieving seamless integration with the existing distributed control system monitoring screen.

[0071] Finally, online identification and graded early warning are performed. After the lightweight model is deployed, real-time sensor data is continuously read at a sampling frequency of 1Hz. Each inference is based on the sliding window data of the most recent 60 time steps to determine the state, and outputs the sorting state category and the probability value corresponding to each category at the current moment. The system has a built-in three-level graded early warning rule. The trigger condition for the prompt-level early warning is that the following three conditions are met simultaneously: the model output state is "low sorting efficiency"; this state occurs continuously M times, where M is 3; and the predicted probability of this state exceeds the threshold P1, where P1 is 0.6. After the early warning is triggered, a prompt message containing the state category, confidence level, and suggested operation is pushed to the operator's terminal. The trigger condition for the alarm-level early warning is that the model output state is "material blockage" or "coal powder escape", and the predicted probability of the corresponding state exceeds the threshold P2, where P2 is 0.7. After the alarm is triggered, the on-site audible and visual alarm is immediately activated, and the fault type, occurrence time, confidence level, and corresponding on-site screen link are displayed on the distributed control system operation interface. The triggering conditions for the linkage protection level are that the following three conditions are met simultaneously: the alarm level warning for the corresponding fault has been triggered; the alarm state lasts for more than T seconds, where T is 30 seconds; and the predicted probability of the fault state continuously exceeds the threshold P2. Upon triggering, a preset sequence of protection actions is automatically executed, including closing the corresponding branch valve, adjusting the fan speed, and initiating the backflushing and unblocking procedure. The execution process and fault events are simultaneously recorded in the historical database for subsequent retrospective analysis. The parameters M, T, P1, and P2 in the above warning rules can all be configured and adjusted online through the human-machine interface. The system synchronously records the triggering conditions and handling results for each warning, supports statistical analysis based on historical data, and automatically generates warning parameter optimization suggestions.

[0072] In summary, this embodiment constructs a high-precision teacher model and uses knowledge distillation technology to obtain a lightweight student model, achieving real-time deployment on the edge computing nodes of a distributed control system. It can accurately identify four states in the stone and coal sorting process: normal, low sorting efficiency, material blockage, and coal powder escape. Based on the graded early warning rules, it can detect hidden fault signs in advance, effectively solving the technical problems of delayed early warning, high false alarm rate, and difficult model deployment in traditional methods.

[0073] To achieve the above embodiments, this application also proposes an online identification and fault early warning system for stone and coal sorting status that integrates attention mechanisms and knowledge distillation. Figure 3 This is a schematic diagram of the structure of an online identification and fault early warning system for stone and coal sorting that integrates attention mechanism and knowledge distillation, as proposed in an embodiment of this application. Figure 3 As shown, the system includes: The acquisition module 100 is used to collect time-series data from multiple sources of sensors in key parts of the stone and coal sorting system, and to preprocess and label the time-series data of multiple sources of sensors to construct a multi-category labeled dataset that includes normal state, low sorting efficiency state, material blockage state and coal powder escape state.

[0074] The construction module 200 is used to construct a high-precision teacher model based on a hybrid architecture of a one-dimensional convolutional neural network and a Transformer encoder. The teacher model is trained using a labeled dataset so that the teacher model outputs the predicted probability distribution of various sorting states.

[0075] Training module 300 is used to train a lightweight student model based on knowledge distillation technology, using soft labels output by the teacher model and hard labels of the labeled dataset as supervision signals, and then deploy the trained student model to the edge computing nodes of the distributed control system.

[0076] The output module 400 is used to perform online inference on the real-time sensor time-series data using the deployed student model, output the sorting status category and corresponding probability at the current moment, and trigger the corresponding level of early warning or linkage protection action according to the preset hierarchical early warning rules.

[0077] It should be noted that the foregoing explanation of the embodiment of the dynamic prevention and control method for landslides based on hierarchical management also applies to the system of this embodiment, and will not be repeated here.

[0078] In summary, the online identification and fault warning system for stone and coal sorting status that integrates attention mechanism and knowledge distillation in this application embodiment achieves high-precision online identification of sorting status and lightweight model deployment by integrating attention mechanism and knowledge distillation technology.

[0079] To implement the above embodiments, this application also proposes an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the online identification and fault warning method for stone and coal sorting status that integrates attention mechanism and knowledge distillation as described in any of the first aspect embodiments above.

[0080] To implement the above embodiments, this application also proposes a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the online identification and fault warning method for stone and coal sorting status by fusing attention mechanism and knowledge distillation as described in any one of the first aspect embodiments above.

[0081] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0082] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0083] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0084] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0085] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0086] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0087] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0088] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for online identification and fault early warning of stone and coal sorting status that integrates attention mechanisms and knowledge distillation, characterized in that, Includes the following steps: Collect time-series data from multiple sources of sensors at key parts of the stone and coal sorting system, and preprocess and label the time-series data of the multiple sources of sensors to construct a multi-category labeled dataset that includes normal state, low sorting efficiency state, material blockage state and coal powder escape state. A high-precision teacher model based on a hybrid architecture of a one-dimensional convolutional neural network and a Transformer encoder is constructed. The teacher model is trained using the labeled dataset so that the teacher model outputs the predicted probability distribution of various sorting states. Based on knowledge distillation technology, a lightweight student model is trained using the soft labels output by the teacher model and the hard labels of the labeled dataset as supervision signals, and the trained student model is deployed to the edge computing nodes of the distributed control system. The deployed student model performs online inference on the real-time sensor time-series data, outputs the sorting status category and corresponding probability at the current moment, and triggers the corresponding level of early warning or linkage protection action according to the preset hierarchical early warning rules.

2. The method according to claim 1, characterized in that, The constructed multi-class labeled dataset includes normal state, low sorting efficiency state, material blockage state, and coal powder escape state, including: Multi-source sensor time-series data are continuously collected from the historical database of the distributed control system within a preset time period. The multi-source sensor time-series data includes pressure and current signals before and after the fan, pressure signals of each section of the conveying pipeline, material level signals of the sorting bin and the stone and coal bin, carbon content detection data of the stone and coal, and opening status signals of the air supply valve and the exhaust valve. The time-series data from the multi-source sensors are time-aligned and cleaned of missing and outlier values ​​to obtain pre-processed time-series data. The preprocessed time series data is labeled with operating conditions, and a double-blind cross-validation mechanism is used for label comparison. The state confidence level is simultaneously labeled for transitional samples with blurred boundaries to obtain the multi-class labeled dataset.

3. The method according to claim 1, characterized in that, The construction of a high-precision teacher model based on a hybrid architecture of a one-dimensional convolutional neural network and a Transformer encoder, and the training of the teacher model using the labeled dataset, includes: The input layer of the teacher model is constructed, wherein the input layer receives time-series data from N sensor channels within a preset time window to construct a two-dimensional feature matrix, and performs standardization processing on the input data; A one-dimensional convolutional neural network local feature extraction layer is constructed for the teacher model. The local temporal features of each sensor channel are extracted through multiple one-dimensional convolution operations. After convolution, a batch normalization layer and a ReLU activation function are connected, and the temporal resolution is reduced through a pooling layer. The Transformer encoder layer of the teacher model is constructed to introduce learnable positional encoding into the feature sequence output by the one-dimensional convolutional neural network. Long-range dependencies are captured through a multi-head self-attention mechanism, and residual connections and layer normalization structures are adopted. Construct a multi-head attention pooling layer for the teacher model, set multiple learnable query vectors, calculate the scaling dot product attention weight of each query vector with the output sequence of the Transformer encoder layer, and perform weighted summation on the sequence based on the weights to obtain a global feature vector of fixed dimensions. The classification output layer of the teacher model is constructed, and a four-dimensional classification probability vector corresponding to the four sorting states is output through a fully connected layer and a Softmax activation function. Using the cross-entropy loss function as the optimization objective, the Adam optimizer is used to train the teacher model end-to-end. During the training process, learning rate decay and early stopping strategies are employed to prevent overfitting.

4. The method according to claim 3, characterized in that, The Transformer encoder layer that constructs the teacher model includes: Construct a multi-layer Transformer encoder, set the number of self-attention layers in each Transformer encoder layer, and set the corresponding dimensions of the hidden layers and feedforward network layers; Sine and cosine position codes are used as position codes, and are added element-wise to the input features before being input to the encoder; The calculation formula for the multi-head self-attention mechanism is set as follows: in, , , in, , , These are the query matrix, key matrix, and value matrix, respectively. Let be the dimension of the key vector. For the number of attention heads, To output the linear transformation weight matrix, This indicates a vector concatenation operation.

5. The method according to claim 1, characterized in that, The process of training a lightweight student model and deploying the trained student model to the edge computing nodes of a distributed control system includes: Based on the aforementioned supervision signal, the student model is trained using a distillation loss function, which is a weighted combination of hard-label cross-entropy loss and soft-label KL divergence loss, expressed by the following formula: in, , , The weighting coefficients for hard label loss are... The weighting coefficients for soft label loss are... This is the distillation temperature coefficient. The logits value output by the fully connected layer of the student model. The logits value is the output of the fully connected layer of the teacher model. For including temperature coefficient function, Let KL divergence loss function be used. p q These are the soft label probability distributions output by the teacher model and the student model, respectively. C Total number of categories; The trained student model is exported as an open neural network exchange format, and model optimization operations such as operator fusion and constant folding are performed. The model is loaded by the algorithm function block in the distributed control system, and the student model is configured with corresponding interfaces for real-time data reading, data preprocessing, model inference and result parsing.

6. The method according to claim 5, characterized in that, After exporting the trained student model to the Open Neural Network Exchange format, the following is also included: INT8 quantization is used to compress the floating-point parameters of the trained student model; The model is enabled to perform parallel computing on a multi-core CPU, and inference computing tasks are assigned to the multi-core CPU. Configure the model with an inference caching mechanism to prevent duplicate data calculations within the sliding window.

7. The method according to claim 1, characterized in that, The deployed student model performs online inference on the real-time collected sensor time-series data, outputs the sorting state category and corresponding probability at the current moment, and triggers corresponding level of early warning or linkage protection actions according to preset hierarchical early warning rules, including: The system continuously reads real-time sensor data at a preset sampling frequency. During each inference process, it performs state discrimination based on sliding window data from multiple time steps closest to the current time and outputs the sorting state category and the probability value corresponding to each category at the current time. Based on the preset hierarchical early warning rules, when the student model outputs a low sorting efficiency state M times in a row and the state probability exceeds the preset threshold P1, a prompt-level early warning is triggered. When the student model identifies a blockage or coal powder escape state and the state probability exceeds the preset threshold P2, an alarm-level warning is triggered. If the alarm-level warning lasts for N seconds and the state probability continues to exceed the threshold P2, a linkage protection-level warning is triggered, and preset protection actions are automatically executed, where M, P1, P2 and N are preset positive integers.

8. The method according to claim 7, characterized in that, Before triggering the corresponding level of early warning or linkage protection action according to the preset hierarchical early warning rules, the method further includes: The M, P1, P2, and N components in the hierarchical early warning rules can be configured and adjusted online via a human-machine interface. After triggering the corresponding level of early warning or linkage protection action according to the preset hierarchical early warning rules, the following is also included: Record the triggering conditions and handling results of each warning, and perform statistical analysis based on the recorded historical data to automatically generate suggestions for optimizing warning parameters.

9. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the online identification and fault warning method for stone and coal sorting status based on the fusion of attention mechanism and knowledge distillation as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the online identification and fault warning method for stone and coal sorting status based on the fusion attention mechanism and knowledge distillation as described in any one of claims 1-8.

Images