Early warning and prevention method and device based on dust big data tensor decomposition and learning

Abstract

The application provides a kind of early warning prevention and control method and device based on dust big data tensor decomposition and learning, it is related to intelligent control technical field, the method includes: collecting the field data of crushing and transportation site;Based on the modal information of field data, the feature alignment and mapping of each modal data of field data are carried out, and a multi-modal feature set is generated;Based on the multi-modal feature set and the core tensor, factor matrix pre-constructed, construct coupling tensor model;Based on the processing of multi-modal feature set, the corresponding high-order fusion feature vector is generated based on coupling tensor model;Combine high-order fusion feature vector and the physical layout of crushing and transportation site to construct heterogeneous graph;Call pre-constructed heterogeneous graph contrast learning model to process heterogeneous graph, and in the case of fault node, generate compensation strategy;Call pre-constructed strategy decoder to generate control instruction based on compensation strategy, and the control parameter of all nodes except fault node is included in control instruction.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of intelligent control technology for complex systems, and in particular to an early warning and prevention method and device based on tensor decomposition and learning of dust big data. Background Technology

[0002] In the operational environment of open-pit mine crushing and belt conveyor systems, the effectiveness of dust control systems is often limited by the fragmentation of environmental perception and the mechanization of control logic. Dust pollution not only harms the health of workers but also accelerates equipment wear and affects operational safety. Therefore, early warning and prevention based on tensor decomposition and learning of dust big data is a crucial aspect of safe production in open-pit mines. Although existing control methods have introduced sensing devices such as visual monitoring, noise monitoring, and physical sensors, this heterogeneous data is usually processed independently or simply pieced together, lacking the ability to mine high-order potential correlations between data and failing to fully leverage the collaborative sensing value of multi-source data. For example, abnormal impact sounds from the jaw plates of a crusher often precede a surge in dust concentration and are accompanied by specific visual texture changes. This cross-modal inherent logic is difficult to capture in traditional linear models, leading to a lag in control strategies.

[0003] While existing technologies utilize tensor decomposition for fault diagnosis, such as in smart substations where anomalies are detected by constructing tensor models of time and data types, these methods are typically only suitable for processing homogeneous scalar time-series data, such as current, voltage, or temperature values. Their core logic involves passively discovering faults by calculating and reconstructing residuals. In open-pit mine scenarios, however, data exhibits strong multimodal heterogeneity, encompassing unstructured image streams, frequency-domain acoustic signature streams, and time-series physical streams. Conventional tensor construction and standard decomposition methods cannot handle this heterogeneous data where dimensions and physical properties are completely mismatched. More critically, existing methods, upon detecting a fault, typically employ single-point repair strategies such as switching to backup equipment or adjusting clocks, lacking the ability for collaborative compensation based on the entire process topology. When critical nodes physically fail, existing systems cannot utilize other modal data to generate virtual features to maintain operation, nor can they schedule upstream and downstream related equipment for functional compensation, easily leading to localized faults escalating into global control paralysis. Summary of the Invention

[0004] To address the aforementioned technical problems, embodiments of the present invention provide an early warning and prevention method based on tensor decomposition and learning of dust big data, comprising: Collect on-site data from the crushing and transportation sites; the on-site data is multimodal data. Based on the modal information of the field data, feature alignment and mapping are performed on each modal data of the field data to generate a multimodal feature set; Based on the multimodal feature set and the pre-constructed core tensor and factor matrix, a coupled tensor model is constructed. The core tensor compresses and stores the fixed interaction rules between various modal data to realize the reverse derivation and completion of virtual features when modal data is missing. There are multiple factor matrices, each corresponding to a modal data. The multimodal feature set is processed based on the coupled tensor model to generate the corresponding high-order fusion feature vector; A heterogeneous graph is constructed by combining the high-order fusion feature vector and the physical layout of the crushing and transportation site. The heterogeneous graph includes the current operating state of each node. The heterogeneous graph is processed by a pre-built heterogeneous graph contrastive learning model, and a compensation strategy is generated in the case of faulty nodes. The compensation strategy includes a compensation path and a compensation intensity. The pre-built policy decoder is invoked to generate control instructions based on the compensation policy. The control instructions include control parameters for all nodes except the faulty node.

[0005] In one embodiment, the field data includes video data, audio data, and environmental physical parameters; The modal information based on the field data is used to perform feature alignment and mapping on each modal data of the field data to generate a multimodal feature set, including: Video features are obtained by extracting features from video data using a 3D convolutional neural network. Audio features are obtained by extracting features from audio data based on short-time Fourier transform and one-dimensional convolutional neural network; The environmental physical parameters are extracted based on a gated recurrent unit network to obtain environmental physical features. The video features, audio features, and environmental physical features are aligned, and a deep neural network extractor is used to map the aligned three types of features to the same latent feature space to obtain the multimodal feature set.

[0006] In one embodiment, constructing the core tensor and factor matrix includes: Acquire historical field data; The historical field data is used to construct a historical high-order observation tensor through the outer product operation; Perform coupled Tucker decomposition on the historical high-order observation tensor to obtain the historical high-order observation tensor relationship containing the core tensor and factor matrix to be solved; The historical field data and the alternating least squares method are used to iteratively optimize the historical high-order observation tensor relationship. The core tensor and factor matrix were obtained based on the iterative optimization results.

[0007] In one embodiment, constructing a coupled tensor model based on the multimodal feature set and pre-built core tensors and factor matrices includes: Pre-construct the initial model of the coupled tensor; Substituting the multimodal feature set, core tensor, and factor matrix into the initial model of the coupled tensor, we obtain the coupled tensor model. The method further includes: When the first modality data is missing in the field data, the virtual feature vector of the first modality data is generated by reverse derivation through the modular product operation based on the core tensor and factor matrix.

[0008] In one embodiment, constructing a heterogeneous graph by combining the higher-order fused feature vector and the physical layout of the crushing and transportation site includes: Based on the physical layout of the crushing and transportation site, a node topology diagram is constructed. The nodes are the various equipment nodes in the physical layout, including execution nodes, sensing nodes, and logic control nodes. The edges in the node topology diagram are heterogeneous edges, including material flow edges, spatial flow edges, and collaborative edges. The spatial flow edges are the edges between nodes with a distance less than a preset distance or between nodes in the downwind influence area of ​​the prevailing wind direction, as determined by the Gaussian plume diffusion model. The collaborative edges represent the logical association between various nodes under the same control loop. The initial feature vector of the sensing node is determined based on the higher-order fusion feature vector; The initial feature vectors of the execution nodes and logic control nodes are determined based on the real-time operating status of the execution nodes and logic control nodes in the field data. The heterogeneous graph is formed based on the node topology graph embedded with the initial feature vector, and the message passing mechanism between the nodes in the heterogeneous graph is the meta-path message passing mechanism.

[0009] In one embodiment, constructing the heterogeneous graph contrastive learning model includes: A dual-view comparison architecture is constructed, including a graph input layer, a view generation layer, a feature encoding layer, a comparison learning layer, and a self-healing inference layer; A global ideal view and an image mask view are constructed based on a historical heterogeneous graph for training. The global ideal view and the image mask view are used to simulate the state of each node under normal and abnormal operating conditions. The heterogeneous graph contrast learning model is formed by training the dual-view contrast architecture using the InfoNCE loss function and the global ideal view and image mask view used for training.

[0010] In one embodiment, the step of invoking a pre-built heterogeneous graph contrastive learning model to process the heterogeneous graph and generating a compensation strategy in the event of a faulty node includes: The pre-built heterogeneous graph contrastive learning model is invoked to process the heterogeneous graph and determine whether there are faulty nodes. If a faulty node exists, a compensation path that can compensate for the function provided by the faulty node is automatically calculated based on the heterogeneous graph and the pre-trained topological logic. The compensation intensity is determined by adjusting the attention weights. The compensation path includes nodes that have received fault information and whose feature vectors have drifted.

[0011] In one embodiment, the policy decoder is constructed from a multilayer perceptron including a control branch and a prediction branch; The invocation of the pre-built policy decoder to generate control instructions based on the compensation policy includes: The compensation strategy is input into the strategy decoder to process the compensation strategy based on the control branch of the strategy encoder, generate the control command, and simultaneously process the compensation strategy based on the prediction branch to predict the dust concentration change in the future time step and calculate the regression loss for parameter tuning of the heterogeneous graph contrastive learning model and the core tensor.

[0012] In one embodiment, the collection of on-site data from the crushing and transportation site includes: The system collects video footage of dust clouds, equipment operating noise, and environmental physical parameters at the crushing and transportation site, including time-series data on dust concentration, wind speed, air pressure, and material flow rate.

[0013] Another embodiment of the present invention also provides an early warning and prevention device based on tensor decomposition and learning of dust big data, comprising: The data acquisition module is used to collect on-site data from the crushing and transportation sites, and the on-site data is multimodal data. The feature alignment and mapping module is used to perform feature alignment and mapping on each modal data of the field data based on the modal information of the field data, and generate a multimodal feature set; The first construction module is used to construct a coupled tensor model based on the multimodal feature set and the pre-constructed core tensor and factor matrix. The core tensor compresses and stores the fixed interaction rules between various modal data to realize the reverse derivation and completion of virtual features when modal data is missing. There are multiple factor matrices, each corresponding to a modal data. The first processing module is used to process the multimodal feature set based on the coupled tensor model to generate the corresponding high-order fusion feature vector; The second construction module is used to construct a heterogeneous graph by combining the high-order fusion feature vector and the physical layout of the crushing and transportation site. The heterogeneous graph includes the current operating status of each node. The second processing module is used to call a pre-built heterogeneous graph contrastive learning model to process the heterogeneous graph, and generate a compensation strategy in the case of faulty nodes. The compensation strategy includes a compensation path and a compensation intensity. The generation module is used to call a pre-built policy decoder to generate control instructions based on the compensation policy. The control instructions include control parameters for all nodes except the faulty node.

[0014] Other features and advantages of this application will be set forth in the following description. The objectives and other advantages of this application can be realized and obtained through the structures particularly pointed out in the written description and drawings.

[0015] The technical solution of this application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating the early warning and control method based on tensor decomposition and learning of dust big data in an embodiment of the present invention.

[0018] Figure 2 This is a flowchart illustrating an early warning and control method based on tensor decomposition and learning of dust big data, according to another embodiment of the present invention.

[0019] Figure 3 This is an architecture diagram of an early warning and prevention system based on tensor decomposition and learning of dust big data in an embodiment of the present invention.

[0020] Figure 4 This is a structural block diagram of the early warning and control device based on dust big data tensor decomposition and learning in an embodiment of the present invention. Detailed Implementation

[0021] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but these are not intended to limit the scope of the invention.

[0022] It should be understood that various modifications can be made to the embodiments disclosed herein. Therefore, the following description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope of this disclosure will be apparent to those skilled in the art.

[0023] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present disclosure and, together with the general description of the disclosure given above and the detailed description of the embodiments given below, serve to explain the principles of the disclosure.

[0024] These and other features of the invention will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.

[0025] It should also be understood that although the invention has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of the invention.

[0026] The above and other aspects, features and advantages of this disclosure will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.

[0027] Specific embodiments of the present disclosure are described thereafter with reference to the accompanying drawings; however, it should be understood that the disclosed embodiments are merely examples of the present disclosure and can be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure the present disclosure. Therefore, the specific structural and functional details disclosed herein are not intended to be limiting, but are merely representative of the present disclosure and are intended to teach those skilled in the art to use the present disclosure in a variety of substantially any suitable detailed structures.

[0028] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in other embodiments,” all of which may refer to one or more of the same or different embodiments according to this disclosure.

[0029] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0030] like Figure 1 As shown, this embodiment of the invention provides an early warning and prevention method based on tensor decomposition and learning of dust big data, including: S1: Collect on-site data from the crushing and transportation site, wherein the on-site data is multimodal data; S2: Based on the modal information of the field data, perform feature alignment and mapping on each modal data of the field data to generate a multimodal feature set; S3: Based on the multimodal feature set and the pre-constructed core tensor and factor matrix, construct a coupled tensor model. The core tensor compresses and stores the fixed interaction rules between various modal data to realize the reverse derivation and completion of virtual features when modal data is missing. There are multiple factor matrices, each corresponding to a modal data. S4: Process the multimodal feature set based on the coupled tensor model to generate the corresponding high-order fusion feature vector; S5: Construct a heterogeneous graph by combining the high-order fusion feature vector and the physical layout of the crushing and transportation site. The heterogeneous graph includes the current operating status of each node. S6: Call the pre-built heterogeneous graph contrastive learning model to process the heterogeneous graph, and generate a compensation strategy in the case of faulty nodes. The compensation strategy includes a compensation path and a compensation intensity. S7: Invoke the pre-built policy decoder to generate control instructions based on the compensation policy. The control instructions include control parameters for all nodes except the faulty node.

[0031] Based on the above, the method in this embodiment is applicable to open-pit mine dust environments, such as crushing and transportation sites. The core of this method lies in achieving deep fusion of multi-source heterogeneous data through multimodal coupled tensor decomposition and utilizing heterogeneous graph comparative learning to achieve fault adaptive self-healing. As described above, the overall process of this embodiment follows a progressive logic of data acquisition, feature pre-mapping, tensor reconstruction, graph construction, self-healing training, and strategy execution. Each step relies on clear operating principles and parameter settings to ensure the accuracy, robustness, and engineering feasibility of the control strategy, providing complete methodological support for system implementation.

[0032] Specifically, in one embodiment, the collection of on-site data from the crushing and transportation site includes: S101: Collect dust cloud video, equipment operating noise, and environmental physical parameters at the crushing and transportation site. The environmental physical parameters include time series data of dust concentration, wind speed, air pressure, and material flow rate.

[0033] The field data in this embodiment is multimodal heterogeneous data. The acquisition and feature pre-mapping of multimodal heterogeneous data is the fundamental data support link for the entire control method. Multi-source data acquisition is completed using industrial explosion-proof cameras, high-sensitivity microphones, and integrated sensor arrays. Feature pre-mapping of each modality's data is achieved using three-dimensional convolutional neural networks, one-dimensional convolutional networks, and gated recurrent unit networks. Specifically, industrial explosion-proof cameras are deployed above the jaw crusher and at key nodes along the conveyor belt in the crushing station to collect dust cloud video streams; high-sensitivity microphones are deployed outside the crusher to collect operating noise from the crusher and fan; and integrated sensor arrays (including PM2.5 / 10, wind speed, and temperature / humidity probes) are arranged every 10 meters at the discharge port and within the conveyor belt chute to collect time-series data on dust concentration, wind speed, air pressure, and material flow rate. To address the issue of inconsistent sampling frequencies among multi-source data during the acquisition process, the system employs a strict spatiotemporal alignment strategy at the edge computing gateway: First, a high-precision time synchronization module is used to assign a unified UTC timestamp to data frames from all acquisition channels, ensuring consistent time references. Second, a sliding window algorithm is used to map data of different frequencies to a unified time step. For physical sensor data with lower sampling rates, cubic spline interpolation is used to fill time gaps, while a keyframe extraction strategy is employed for high frame rate video data. Finally, the multi-source data is aligned in the temporal dimension into a standard tensor slice sequence to prevent logical association failures during subsequent tensor decomposition due to data misalignment.

[0034] As described above, the field data in this embodiment includes video data, audio data, and environmental physical parameters. The step of performing feature alignment and mapping on the various modal data of the field data based on the modal information of the field data to generate a multimodal feature set includes: S201: Extracting video features from video data using a 3D convolutional neural network; S202: Audio features are obtained by extracting features from audio data based on short-time Fourier transform and one-dimensional convolutional neural network; S203: Based on a gated recurrent unit network, feature extraction is performed on the environmental physical parameters to obtain environmental physical features; S204: Align the video features, audio features, and environmental physical features, and use a deep neural network extractor to map the aligned three types of features to the same latent feature space to obtain the multimodal feature set.

[0035] For example, for visual image data, a three-dimensional convolutional neural network can be used as a feature extractor. This network structure includes four convolutional layers, two pooling layers, and one fully connected layer. The convolutional kernel sizes are set to 3×3×3, 3×3×3, 2×2×2, and 2×2×2, respectively. The pooling layer uses max pooling with a stride of 2. By performing convolution operations simultaneously in the temporal and spatial dimensions, the volume change rate and motion vector of dust clouds are captured, ultimately outputting a visual feature vector with a dimension of 1×256. For audio signal data, a spectrogram is first generated through short-time Fourier transform with a window length of 256ms and an overlap rate of 50%. Then, a one-dimensional convolutional network can be used to extract anomalous energy distribution features in the frequency domain. This one-dimensional convolutional network includes three convolutional layers and one global average pooling layer with convolutional kernel sizes of 3, 5, and 7, and a stride of 1 for each layer. This is used to capture audio signals of device malfunctions, outputting an acoustic feature vector with a dimension of 1×256. For time-series data such as environmental physical quantities, a gated recurrent unit network (GRU) can be used for processing. This network contains two hidden layers, each with 128 neurons and a dropout probability of 0.3. It captures the long-term dependence trend of parameters through a gating mechanism and outputs a 1×256-dimensional physical feature vector. These three types of feature vectors together constitute a multimodal feature set, laying the foundation for subsequent three-modal coupled tensor decomposition processing.

[0036] In one embodiment, constructing the core tensor and factor matrix includes: S8: Obtain historical on-site data; S9: The historical field data is constructed into a historical high-order observation tensor through the outer product operation; S10: Perform coupled Tucker decomposition on the historical high-order observation tensor to obtain the historical high-order observation tensor relationship containing the core tensor and factor matrix to be solved; S11: Iteratively optimize the historical high-order observation tensor relationship using the historical field data and the alternating least squares method; S12: The core tensor and factor matrix were obtained based on the iterative optimization results.

[0037] For example, this embodiment demonstrates that multimodal coupled tensor decomposition based on manifold mapping is a core component in ensuring the intrinsic robustness of the sensing system under extreme environments. Unlike standard tensor decomposition, which directly stacks raw data, this method first uses a deep neural network extractor to map three heterogeneous data types—visual, acoustic, and physical—to a unified latent feature space, and then constructs a coupled tensor model. This model aims to uncover the deep logical relationships between multimodal features, using a core tensor as logical memory to achieve reverse derivation and reconstruction of missing modalities, thereby effectively solving the problem of sensing interruption caused by a single modality when the sensor is iced or malfunctions.

[0038] The core principle is to first construct a high-order observation tensor from the mapped visual, acoustic, and physical feature vectors through an outer product operation, and then perform coupled Tucker decomposition based on this tensor:

[0039] in For the third-order observation tensor, I=J=K=256, This represents the outer product operation of vectors; This is the core tensor. In this process, the core tensor is no longer merely a tool for compressing and storing data, but is trained to be a mathematical entity that describes the invariant interaction logic between different modalities. These are factor matrices representing visual, acoustic, and physical modalities, respectively. Representing tensors and matrices Modular product is defined as .

[0040] The decomposition process employs a dual strategy of offline pre-training and online rapid inference to balance computational accuracy and real-time requirements. In the offline phase, all historical normal data is collected, and iterative optimization is performed using alternating least squares to minimize the reconstruction error.

[0041] The constraints are The number of iterations is set to 100, and the convergence threshold is set to 1e-5. During this process, the core tensor... Trained as a high-dimensional logical lookup table, it mathematically compresses and stores the invariant interaction rules between visual, acoustic, and physical modalities; for example, increased high-frequency noise and visual darkness necessarily correspond to high dust concentration. This logical memory characteristic allows the model to perform subsequent inferences even when a certain modal input is missing, by relying on other modalities. The approximate solution for the missing information is obtained by looking up the projection position in the table.

[0042] Furthermore, the construction of the coupled tensor model based on the multimodal feature set and the pre-constructed core tensor and factor matrix includes: S301: Pre-constructed initial model of coupled tensors; S302: Substitute the multimodal feature set, core tensor, and factor matrix into the initial model of the coupled tensor to obtain the coupled tensor model; The method further includes: S303: When the first modality data is missing in the field data, the virtual feature vector of the first modality data is generated by reverse derivation through the modular product operation based on the core tensor and factor matrix.

[0043] For example, during the online simulation phase, the system no longer performs time-consuming full least squares iterations, but instead uses fixed core tensors trained offline. Along with visual and acoustic factor matrices, real-time acquired features are mapped to the latent space using only fast tensor modulus multiplication operations. When data for a particular modality is lost due to sensor failure in high-altitude, cold-weather environments, the core tensor is used as the basis for... The factor matrices of the other two valid modes are used to inversely derive the virtual feature vector of the missing mode through modular multiplication. Assuming that physical mode data is missing due to sensor icing, the observation tensor degenerates into... Then the estimation formula for the physical mode factor matrix is:

[0044] in This represents the Moore-Penrose pseudo-inverse operation. This represents the generalized inverse of the core tensor. After reconstruction, the core tensor will contain multimodal high-order interaction logic. The pattern is unfolded and flattened, and then dimensionality is reduced by a fully connected layer to a fused feature vector of dimension 1×1024. This vector is then used as the initial feature embedding input for subsequent sensing nodes in the heterogeneous graph comparison learning process.

[0045] In one embodiment, constructing a heterogeneous graph by combining the higher-order fused feature vector and the physical layout of the crushing and transportation site includes: S401: Construct a node topology diagram based on the physical layout of the crushing and transportation site. The nodes are the various equipment nodes in the physical layout, including execution nodes, sensing nodes and logic control nodes. The edges in the node topology diagram are heterogeneous edges, including material flow edges, spatial flow edges and collaborative edges. The spatial flow edges are the edges between nodes with a distance less than a preset distance or between nodes in the downwind influence area of ​​the prevailing wind direction, as determined by the Gaussian plume diffusion model. The collaborative edges represent the logical association of each node under the same control loop. S402: Determine the initial feature vector of the sensing node based on the higher-order fusion feature vector; S403: Determine the initial feature vectors of the execution nodes and logic control nodes based on the real-time operating status of the execution nodes and logic control nodes in the field data; S404: The heterogeneous graph is formed based on the node topology graph embedded with the initial feature vector, and the message passing mechanism between the nodes in the heterogeneous graph is the meta-path message passing mechanism.

[0046] For example, the heterogeneous graph construction and meta-path message passing in this embodiment are the prerequisite foundational steps for heterogeneous graph comparative learning. The core lies in constructing a heterogeneous graph based on the physical layout of open-pit mine crushing and transportation, and realizing meta-path message passing through a relational graph convolutional network. Based on the physical layout and equipment relationships of open-pit mine crushing and transportation, a heterogeneous graph G=(V,E) is constructed, where the node set V contains three types of nodes: execution nodes, sensing nodes, and logic control nodes. Execution nodes correspond to various dust control actuators, specifically including crusher spray pumps, belt dust collector fans, and sealing cover motors; sensing nodes correspond to sensing modules and virtual sensing points output by multimodal coupling tensor decomposition; logic control nodes correspond to control units such as PLC substations. To ensure data alignment in the graph neural network input layer, the initial features of each type of node are uniformly constructed to a 1×1024 dimension: sensing nodes directly use 1×1024-dimensional embedding vectors reconstructed from multimodal coupling tensor decomposition; execution nodes and logic control nodes extract their real-time operating status feature vectors and upscale them to 1×1024 through their respective linear mapping layers. The edge set E contains three types of heterogeneous edges: material flow edges, spatial flow edges, and collaborative edges. Material flow edges represent the physical path of dust sources being transported from the crusher to the conveyor belt along with the ore. Spatial flow edges are calculated by introducing a Gaussian plume diffusion model to calculate the dust influence domain between equipment. When the physical distance between two equipment nodes is less than a set threshold or is within the downwind influence zone of the prevailing wind direction, a spatial flow edge is established to represent the neighborhood nonlinear influence of the wind field on dust diffusion. Collaborative edges represent the logical connections between nodes under the same control loop.

[0047] Metapath message passing is implemented using a relational graph convolutional network, and its theoretical formula is:

[0048] in For the l-th layer node i The characteristic is that R is the set of edge relation types. For nodes under relation r i Adjacent sets, Let r be the transformation matrix. Let σ be the node's own feature transformation matrix, and σ be the ReLU activation function. This message passing mechanism can effectively distinguish the impact of different types of edge relationships on node features, enabling it to identify the differences in the impact of material transport and spatial diffusion on dust, thereby achieving refined information aggregation and providing high-quality graph feature support for subsequent heterogeneous graph comparative learning.

[0049] Constructing the heterogeneous graph contrastive learning model includes: S13: Construct a dual-view contrast architecture, including a graph input layer, a view generation layer, a feature encoding layer, a contrast learning layer, and a self-healing inference layer; S14: Construct a global ideal view and an image mask view for training based on the historical heterogeneous graph. The global ideal view and the image mask view are used to simulate the state of each node under normal and abnormal working conditions. S15: The dual-view contrast architecture is trained using the InfoNCE loss function, as well as the global ideal view and the image mask view used for training, to form the heterogeneous graph contrast learning model.

[0050] Specifically, based on the aforementioned heterogeneous graph topology, the core of the Heterogeneous Graph Comparative Learning (HGCL) framework is to achieve logical self-healing of the system under abnormal operating conditions such as equipment failure and communication interruption. Its core logic lies in mining the functional substitutability and topological redundancy between devices through comparative learning, thereby achieving adaptive fault compensation. As the core support for the self-healing control link, this heterogeneous graph comparative learning framework adopts a dual-view comparative architecture, which includes, from top to bottom, a graph input layer, a view generation layer, a feature encoding layer, a comparative learning layer, and a self-healing inference layer. Each layer works together to complete the training and inference of the self-healing model.

[0051] The core operating principle of the framework revolves around the generation and comparison of dual views. It constructs a global ideal view and a fault mask view to simulate the graph states under different operating conditions, providing data support for comparative learning. The global ideal view... Based on the generation of a complete heterogeneous graph, it accurately represents the ideal operating condition where all devices are running normally and sensor data is complete; fault mask view. Abnormal working conditions are simulated through random masking operations. Specifically, during training, 20% of the node features are randomly set to zero, or 10% of the graph edge connections are cut. The masking operation is defined as follows:

[0052] The mask ratio mentioned above can be dynamically adjusted according to the actual working conditions of the open-pit mine. The adjustment range is set to 10%-30% to adapt to different levels of abnormal scenarios.

[0053] The self-healing performance of heterogeneous graph contrastive learning is achieved through training using the InfoNCE loss function, the theoretical expression of which is as follows:

[0054] In the formula, , Representing nodes respectively i Final feature representation under the global ideal view and the fault mask view; This is the cosine similarity function, used to quantify the similarity of node features; is the temperature coefficient used to adjust the similarity weight; N is the total number of nodes in the heterogeneous graph; M is the number of negative samples, which is set to N-1 in this invention, meaning that all other nodes except the target node are treated as negative samples.

[0055] The core optimization objective of the InfoNCE loss function is to minimize the feature difference of the same node in the ideal view and the masked view, while maximizing the feature difference between different nodes. This constraint forces the neural network to learn the node feature recovery ability under abnormal conditions. To ensure training effectiveness and engineering feasibility, specific parameter configurations include a two-layer multilayer perceptron for the feature encoding layer, with 1024 and 512 neurons respectively, and a dropout probability of 0.3 to prevent overfitting; the number of contrastive learning iterations is set to 200, the initial learning rate is set to 0.0005, and the Adam optimizer is used to achieve fast convergence of the loss function.

[0056] Furthermore, the step of calling a pre-built heterogeneous graph contrastive learning model to process the heterogeneous graph, and generating a compensation strategy in the case of faulty nodes, includes: S601: Call the pre-built heterogeneous graph contrastive learning model to process the heterogeneous graph and determine whether there are faulty nodes; S602: If a faulty node exists, a compensation path that can compensate for the function provided by the faulty node is automatically calculated based on the heterogeneous graph and the pre-trained topological logic, and the compensation intensity is determined by adjusting the attention weight. The compensation path includes nodes that have received node fault information and whose feature vectors have drifted.

[0057] For example, based on the training mechanism described above, the heterogeneous graph contrastive learning framework's workflow consists of two stages: training and inference. In the training stage, the model continuously compares and learns the graph features under ideal and simulated fault conditions, gradually mastering the functional substitutability and topological redundancy logic between devices. In the inference stage, when the perception layer confirms that a node has physically failed and cannot be recovered (e.g., the crusher's spray pump is completely damaged), the method in this embodiment does not attempt to repair the faulty node or simply switch to a backup machine. Instead, it initiates a topology-level collaborative compensation mechanism. Specifically, based on the topological logic obtained during training, the model automatically calculates the associated device paths that can compensate for the function of the faulty node and adjusts the compensation intensity through attention weights. The calculation expression is as follows: In the formula This represents a vector concatenation operation. and The attention parameters are learnable during model training, ultimately enabling the system to adapt and self-heal from faults. For example, the system automatically increases the negative pressure value of the downstream conveyor belt seal and the pressure value of the dry fog dust suppression at the transfer point, using enhanced cross-link management to offset the impact of source control failure. This collaborative compensation of dissimilar equipment based on graph reasoning enables the system to functionally self-heal under conditions of physical hardware damage.

[0058] In one embodiment, the policy decoder is constructed from a multilayer perceptron including a control branch and a prediction branch; the step of invoking the pre-built policy decoder to generate control instructions based on the compensation policy includes: S701: The compensation strategy is input into the strategy decoder to process the compensation strategy based on the control branch of the strategy encoder, generate the control command, and at the same time process the compensation strategy based on the prediction branch to predict the dust concentration change in the future time step, and calculate the mean square error regression loss. This loss is combined with the heterogeneous graph contrast learning loss to form a multi-task joint loss function for model parameter tuning.

[0059] Specifically, in this embodiment, policy decoding and parameter tuning are the core follow-up steps of heterogeneous graph contrastive learning. The core function is to map the self-healing trained and optimized node features into specific executable control commands, and to perform model parameter tuning through a multi-task joint loss function, ensuring the accuracy of control commands and the stability of model operation. This step uses a multilayer perceptron to construct the policy decoder. The fundamental reason for its self-healing compensation lies in the message passing mechanism of the graph contrastive network. When a node in the system, such as the crusher spray, malfunctions, the abnormal state features of that node are aggregated to its neighboring nodes through the edge relationships of the heterogeneous graph, such as the embedding vectors of the downstream belt seal and the vacuum fan. Because the model has undergone robust training through contrastive learning, the feature vectors of neighboring nodes will adaptively drift after receiving this fault information. To balance environmental perception and adaptive control, this step uses a multilayer perceptron with a dual-branch structure to construct the policy decoder. After adaptive drift, the node features are simultaneously input to both the prediction and control branches during the decoding phase. The prediction branch uses the node features to predict dust concentration changes over future time steps to calculate regression loss and supervise the model's perceptual representation capabilities. The control branch, used for action compensation, comprises three fully connected layers with 512, 256, and 64 neurons respectively. The output layer employs a sigmoid activation function to adapt to the normalization requirements of the control parameters. Based on the drifted node features, the control branch outputs enhanced compensation control signals, such as automatically increasing the inverter frequency from 30Hz to 45Hz and the spray pressure from 0.5MPa to 0.8MPa. This dual-branch architecture constitutes the physical implementation path for the system's adaptive self-healing mechanism.

[0060] Furthermore, in this embodiment, the model parameter tuning stage employs a multi-task joint loss function to achieve precise optimization, balancing dust control accuracy with system self-healing performance. Its expression is: In the formula The mean square error regression loss between the predicted dust concentration output from the above prediction branch and the actual sensor readings is used to strictly ensure the model's accuracy in predicting environmental evolution trends. The loss is for contrastive learning of heterogeneous graphs; , These are weighting coefficients, set to 0.6 and 0.4 respectively, to balance the priority of control accuracy and self-healing performance optimization.

[0061] In this embodiment, parameter tuning employs a five-fold cross-validation method to ensure the universality and stability of the parameter combination. In addition to the explicitly defined parameters, three key parameters are adjusted: the rank of the core tensor (32-128), the number of layers in the relational graph convolutional network (2-4), and the temperature coefficient of heterogeneous graph contrastive learning (0.05-0.2). When the self-healing recovery rate on the validation set reaches over 90% and the dust prediction error is less than 0.01, this set of parameters is determined as the optimal parameter combination, providing stable support for the model's practical engineering applications.

[0062] At the system level corresponding to the method of this embodiment, such as Figure 2 As shown, the overall system adopts a layered architecture design, which includes a multimodal heterogeneous perception module, a multimodal coupled tensor calculation module, a heterogeneous graph comparison self-healing decision module, and a closed-loop execution and feedback optimization module. Each module has a clear function and is closely connected, forming a closed-loop control of perception-processing-decision-execution-feedback to adapt to the high-altitude and extreme working conditions of open-pit mines and solve the technical bottlenecks of existing systems.

[0063] For example, the multimodal heterogeneous sensing module, as the front-end sensing core and data source of the entire control system, focuses on the acquisition and tensor preprocessing of multi-source heterogeneous data throughout the entire process of open-pit mine dust control, providing high-quality raw data support for the efficient operation of all subsequent modules. An explosion-proof and cold-resistant heterogeneous sensor array, combined with industrial 5G transmission and edge computing technology, completes data acquisition, spatiotemporal alignment, and preliminary feature extraction. Industrial explosion-proof cameras are deployed above the jaw crusher and at key nodes along the conveyor belt in the crushing station to collect dust cloud video streams; high-sensitivity microphones are deployed outside the crushing chamber to collect operating noise from the crusher and fan; integrated micro weather stations are arranged every 10m at the discharge port and within the conveyor belt chute to collect time-series data of environmental physical quantities such as dust concentration, wind speed, air pressure, temperature, humidity, and material flow rate. All sensing devices are designed to be explosion-proof and cold-resistant, with an operating temperature range of -40℃ to 40℃, making them suitable for the extreme working conditions of open-pit mines at high altitudes and cold conditions. They establish communication connections with the edge computing gateway via industrial 5G modules, with transmission latency controlled within 50ms to ensure real-time and stable data transmission. The edge computing gateway incorporates feature extraction algorithms such as 3D convolutional neural networks, short-time Fourier transforms, and gated recurrent unit networks. It performs denoising, alignment, and feature mapping on the collected unstructured video streams, audio streams, and time-series data, generating standardized visual, acoustic, and physical feature vectors, which are then output to the multimodal coupled tensor computation module.

[0064] The Multimodal Coupled Tensor Computation Module, relying on the core algorithm of Multimodal Coupled Tensor Decomposition, undertakes the tasks of preprocessing, deep fusion, and missing feature reconstruction of multimodal heterogeneous features. It focuses on solving the problem of perception interruption caused by single-modal failure in high-altitude and cold environments, providing complete and reliable node embedding vectors for subsequent decision-making modules. Deployed on a central control server, this module integrates a multimodal coupled tensor decomposition network and various dedicated feature extractors, strictly adhering to the parameter settings in the methodology, and receiving raw data and preliminary feature vectors transmitted from the multimodal heterogeneous perception modules. First, the data is standardized and preprocessed to normalize it to the [0,1] interval to eliminate noise interference. Then, a dedicated feature extractor further optimizes the features of each modality to meet the requirements of tensor operations. Next, a multi-order observation tensor is constructed using outer product operations, and Tucker decomposition is performed using alternating least squares (ALS) to iteratively optimize the core tensor and the factor matrices of each modality, uncovering higher-order potential correlations between multimodal data. When a data loss or anomaly is detected in a physical channel (e.g., sensor icing causing no change in value for 10 consecutive seconds), a missing feature reconstruction mechanism is automatically triggered. Using the factor matrices of the core tensor and the other two effective modalities, a virtual feature vector of the missing modality is generated through pseudo-inverse and modular multiplication operations. Finally, the reconstructed fused features are processed by a fully connected layer, flattened into a 1×1024-dimensional node embedding vector, and output to the heterogeneous graph comparison self-healing decision module.

[0065] The heterogeneous graph comparison self-healing decision module, serving as the core decision-making hub of the entire control system, is supported by a heterogeneous graph comparison learning mechanism. It coordinates the construction of heterogeneous graph topology, training of the self-healing model, and generation of control strategies, enabling adaptive self-healing after equipment failures and communication interruptions, and outputting precise control commands to the execution module. This module is also deployed on the central control server, incorporating a heterogeneous graph comparison learning network, a relational graph convolutional network, and a multilayer perceptron policy decoding unit. These components correspond one-to-one with the heterogeneous graph topology construction, heterogeneous graph comparison learning, and policy decoding stages in the methodology, adhering to predetermined theoretical formulas and parameter configurations. After receiving the node embedding vectors output by the multimodal coupled tensor computation module, the system first constructs a heterogeneous graph based on the physical layout of open-pit mine crushing and transportation. This graph includes execution nodes, sensing nodes, logic control nodes, and three types of heterogeneous edges, maintaining a dynamically updated equipment topology map. Then, a relational graph convolutional network is used to implement meta-path message passing, finely aggregating node features corresponding to different edge relationships. Next, a dual-view comparison architecture generates a global ideal view and a fault mask view. An InfoNCE loss function is used to train a self-healing model, uncovering functional substitutability and topological redundancy between devices. When a device fault or abnormal node feature is detected, the optimal compensation path is automatically derived based on the trained topological logic and attention weights, generating a collaborative control strategy. Finally, a multilayer perceptron strategy decoding unit maps node features to specific actuator operating parameter instructions, which are then sent to the closed-loop execution and feedback optimization module. Subsequently, the system re-collects the post-execution environment and equipment status through the multimodal heterogeneous sensing module, forming a closed-loop data stream for global control, driving the next round of graph updates and dynamic strategy adjustments.

[0066] The closed-loop execution and feedback optimization module, as the core carrier for the implementation and operational status feedback of the control strategy, constructs a complete closed loop of decision-making, execution, feedback, and optimization. It must accurately execute control commands issued by the upstream decision-making module to ensure dust control effectiveness, and also provide real-time operational data support for continuous optimization of model parameters. This module consists of various dust control actuators, corresponding to the strategy execution stage in the method composition. These mainly include crusher spray pumps, belt dust collector fans, dry fog dust suppression devices, and sealing cover motors, distributed in key locations such as open-pit mine crushing stations, belt conveyor lines, and transfer points, forming a one-to-one topological association with sensor nodes and control nodes. Connected to the field PLC system via industrial Ethernet, it receives control commands from the heterogeneous graph comparison self-healing decision-making module, converting normalized control parameters (0-100%) into specific physical parameters such as inverter frequency, solenoid valve opening, and spray pressure, precisely adjusting the operating status of each actuator with parameter adjustment accuracy controlled within ±1%. Simultaneously, the system collects real-time operational feedback from each actuator (such as actual motor speed and valve position signals) and changes in environmental parameters after treatment. This feedback information is used as new status inputs and transmitted back to the multimodal heterogeneous sensing module, providing real-time data support for subsequent feature extraction, tensor reconstruction, and model parameter optimization. The system also periodically uses feedback data to fine-tune and update the core tensors and graph network parameters to ensure the long-term stability and control accuracy of the system.

[0067] In practical applications, such as Figure 3 As shown, the deployment and operation of a typical open-pit mine crushing station and belt conveyor system are illustrated, and the practical application of the system and methods is further explained in conjunction with a specific high-altitude and cold-weather operation scenario: An explosion-proof high-definition industrial camera is installed above the jaw crusher in the crushing station to collect real-time video streams with a resolution of 1920×1080, capturing the dynamics of dust clouds above the crushing inlet and conveyor belt. Piezoelectric vibration sensors and directional microphones are deployed outside the crushing chamber to collect vibration and noise signals from the equipment. An integrated micro-weather station is arranged every 10 meters at the discharge port and within the conveyor belt chute to collect real-time dust concentration, wind speed, and temperature and humidity parameters. To address the issue of inconsistent sampling frequencies for multi-source data, this embodiment implements a strict spatiotemporal alignment strategy at the edge computing gateway. First, a unified UTC timestamp is applied to all channel data using a BeiDou time synchronization module. Then, a 256-millisecond sliding window is used to map heterogeneous data to a unified time step. For physical sensor data with low sampling rates, the system automatically performs cubic spline interpolation to fill time gaps; for high-frequency video data, keyframes are extracted. Finally, a standardized multimodal tensor slice sequence is formed and transmitted back to the central server via an industrial 5G module, ensuring the consistency of the logical benchmark for subsequent tensor calculations.

[0068] A dual computational strategy of offline pre-training and online rapid inference is adopted. In offline pre-training, three months of full-scale normal operation data are collected. Then, on a cloud server, the multimodal coupled tensor is trained 100 times using alternating least squares method to calculate a stable core tensor. The core tensor, serving as a high-dimensional logical lookup table, solidifies the interaction patterns between visual darkness, high-frequency noise, and high dust concentration. During online operation, the multimodal coupled tensor calculation module no longer performs time-consuming full iterations but instead loads the pre-trained core tensor for real-time projection. When the dust sensor data at the crusher outlet remains unchanged for 10 consecutive seconds, the system immediately triggers a reverse reconstruction mechanism, utilizing real-time acquired visual and acoustic features, combined with the core tensor... According to the formula Virtual physical feature vectors are generated in reverse within milliseconds, thus ensuring that the node embedding vectors input to the decision layer still contain complete environmental semantics even in the absence of physical perception.

[0069] The heterogeneous graph comparison self-healing decision module constructs an equipment topology graph containing over 50 nodes based on the physical layout of the open-pit mine. When constructing edge relationships, the system not only establishes material flow edges based on the belt conveyor links but also introduces a Gaussian plume diffusion model to calculate the dust impact domain, establishing spatial flow edges between equipment nodes with a physical distance of less than 10 meters or within the prevailing wind direction's influence zone. The system runs the heterogeneous graph comparison learning model in real time. When a mechanical failure occurs in the crusher's spray pump, the feature vector of that node becomes abnormal. At this time, based on the message passing mechanism of the graph comparison network, the negative state features of the faulty node are rapidly aggregated to the dry fog dust suppression and belt sealing nodes at the downstream transfer point through the material flow edges, causing adaptive drift in the feature vectors of these neighboring nodes. After capturing this feature drift, the strategy decoder (MLP) with a dual-branch structure simultaneously executes prediction and control tasks: on the one hand, the prediction branch anticipates the evolution trend of the upcoming increase in dust concentration at the downstream transfer point based on the drift feature, in order to calculate the regression loss and maintain the system's forward-looking perception; on the other hand, the control branch interprets the drift feature as a control signal for enhanced compensation and automatically generates self-healing instructions: increasing the pressure of the dry fog dust suppression device at the downstream transfer point from 0.4MPa to 0.8MPa, and increasing the speed of the negative pressure fan of the belt sealing cover in this section by 20%.

[0070] The aforementioned self-healing control commands are transmitted to the field PLC system via industrial Ethernet, driving the precise actions of each actuator. Through real-time monitoring, despite the failure of the spray device at the source crusher, thanks to proactive compensation and enhanced control measures in the downstream process, the dust concentration at the transfer point and along the subsequent conveyor belt remained within the standard range, preventing pollution spread. Simultaneously, the actuators transmit feedback indicating increased pressure and reached the required speed back to the sensing module, serving as new input for the next round of tensor calculation and graph update. This achieves a complete closed-loop management system, from sensing failure reconstruction to self-healing of the control logic.

[0071] like Figure 4 As shown, another embodiment of the present invention also provides an early warning and control device based on dust big data tensor decomposition and learning, comprising: The data acquisition module is used to collect on-site data from the crushing and transportation sites, and the on-site data is multimodal data. The feature alignment and mapping module is used to perform feature alignment and mapping on each modal data of the field data based on the modal information of the field data, and generate a multimodal feature set; The first construction module is used to construct a coupled tensor model based on the multimodal feature set and the pre-constructed core tensor and factor matrix. The core tensor compresses and stores the fixed interaction rules between various modal data to realize the reverse derivation and completion of virtual features when modal data is missing. There are multiple factor matrices, each corresponding to a modal data. The first processing module is used to process the multimodal feature set based on the coupled tensor model to generate the corresponding high-order fusion feature vector; The second construction module is used to construct a heterogeneous graph by combining the high-order fusion feature vector and the physical layout of the crushing and transportation site. The heterogeneous graph includes the current operating status of each node. The second processing module is used to call a pre-built heterogeneous graph contrastive learning model to process the heterogeneous graph, and generate a compensation strategy in the case of faulty nodes. The compensation strategy includes a compensation path and a compensation intensity. The generation module is used to call a pre-built policy decoder to generate control instructions based on the compensation policy. The control instructions include control parameters for all nodes except the faulty node.

[0072] In one embodiment, the field data includes video data, audio data, and environmental physical parameters; The modal information based on the field data is used to perform feature alignment and mapping on each modal data of the field data to generate a multimodal feature set, including: Video features are obtained by extracting features from video data using a 3D convolutional neural network. Audio features are obtained by extracting features from audio data based on short-time Fourier transform and one-dimensional convolutional neural network; The environmental physical parameters are extracted based on a gated recurrent unit network to obtain environmental physical features. The video features, audio features, and environmental physical features are aligned, and a deep neural network extractor is used to map the aligned three types of features to the same latent feature space to obtain the multimodal feature set.

[0073] In one embodiment, constructing the core tensor and factor matrix includes: Acquire historical field data; The historical field data is used to construct a historical high-order observation tensor through the outer product operation; Perform coupled Tucker decomposition on the historical high-order observation tensor to obtain the historical high-order observation tensor relationship containing the core tensor and factor matrix to be solved; The historical field data and the alternating least squares method are used to iteratively optimize the historical high-order observation tensor relationship. The core tensor and factor matrix were obtained based on the iterative optimization results.

[0074] In one embodiment, constructing a coupled tensor model based on the multimodal feature set and pre-built core tensors and factor matrices includes: Pre-construct the initial model of the coupled tensor; Substituting the multimodal feature set, core tensor, and factor matrix into the initial model of the coupled tensor, we obtain the coupled tensor model. The device further includes: The derivation module is used to generate a virtual feature vector of the first modality data by reverse derivation through modular multiplication operation based on the core tensor and factor matrix when the first modality data is missing in the field data.

[0075] In one embodiment, constructing a heterogeneous graph by combining the higher-order fused feature vector and the physical layout of the crushing and transportation site includes: Based on the physical layout of the crushing and transportation site, a node topology diagram is constructed. The nodes are the various equipment nodes in the physical layout, including execution nodes, sensing nodes, and logic control nodes. The edges in the node topology diagram are heterogeneous edges, including material flow edges, spatial flow edges, and collaborative edges. The spatial flow edges are the edges between nodes with a distance less than a preset distance or between nodes in the downwind influence area of ​​the prevailing wind direction, as determined by the Gaussian plume diffusion model. The collaborative edges represent the logical association between various nodes under the same control loop. The initial feature vector of the sensing node is determined based on the higher-order fusion feature vector; The initial feature vectors of the execution nodes and logic control nodes are determined based on the real-time operating status of the execution nodes and logic control nodes in the field data. The heterogeneous graph is formed based on the node topology graph embedded with the initial feature vector, and the message passing mechanism between the nodes in the heterogeneous graph is the meta-path message passing mechanism.

[0076] In one embodiment, constructing the heterogeneous graph contrastive learning model includes: A dual-view comparison architecture is constructed, including a graph input layer, a view generation layer, a feature encoding layer, a comparison learning layer, and a self-healing inference layer; A global ideal view and an image mask view are constructed based on a historical heterogeneous graph for training. The global ideal view and the image mask view are used to simulate the state of each node under normal and abnormal operating conditions. The heterogeneous graph contrast learning model is formed by training the dual-view contrast architecture using the InfoNCE loss function and the global ideal view and image mask view used for training.

[0077] In one embodiment, the step of invoking a pre-built heterogeneous graph contrastive learning model to process the heterogeneous graph and generating a compensation strategy in the event of a faulty node includes: The pre-built heterogeneous graph contrastive learning model is invoked to process the heterogeneous graph and determine whether there are faulty nodes. If a faulty node exists, a compensation path that can compensate for the function provided by the faulty node is automatically calculated based on the heterogeneous graph and the pre-trained topological logic. The compensation intensity is determined by adjusting the attention weights. The compensation path includes nodes that have received fault information and whose feature vectors have drifted.

[0078] In one embodiment, the policy decoder is constructed from a multilayer perceptron including a control branch and a prediction branch; The invocation of the pre-built policy decoder to generate control instructions based on the compensation policy includes: The compensation strategy is input into the policy decoder, which processes the compensation strategy based on the control branch of the policy encoder to generate the control command. At the same time, the compensation strategy is processed based on the prediction branch to predict the dust concentration change in the future time step and calculate the mean square error regression loss. This loss is combined with the heterogeneous graph contrast learning loss to form a multi-task joint loss function for model parameter tuning.

[0079] In one embodiment, the collection of on-site data from the crushing and transportation site includes: The system collects video footage of dust clouds, equipment operating noise, and environmental physical parameters at the crushing and transportation site, including time-series data on dust concentration, wind speed, air pressure, and material flow rate.

[0080] Another embodiment of the present invention also provides an electronic device, comprising: One or more processors; Memory, configured to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the early warning and prevention method based on dust big data tensor decomposition and learning as described above.

[0081] Furthermore, one embodiment of the present invention also provides a storage medium storing a computer program thereon. When executed by a processor, this program implements the early warning and control method based on tensor decomposition and learning of dust big data as described above. It should be understood that the various solutions in this embodiment have the corresponding technical effects in the above-described method embodiments, and will not be repeated here.

[0082] Furthermore, embodiments of the present invention also provide a computer program product, which is tangibly stored on a computer-readable medium and includes computer-readable instructions, which, when executed, cause at least one processor to perform an early warning and prevention method based on dust big data tensor decomposition and learning, as described in the embodiments above.

[0083] It should be noted that the computer storage medium of the present invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, system, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, a random access storage medium (RAM), a read-only storage medium (ROM), an erasable programmable read-only storage medium (EPROM or flash memory), an optical fiber, a portable compact disk read-only storage medium (CD-ROM), an optical storage medium, a magnetic storage medium, or any suitable combination thereof. In the present invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. In the present invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program configured for use by or in connection with an instruction execution system, system, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, antenna, optical fiber, RF, etc., or any suitable combination thereof.

[0084] Furthermore, those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.

[0085] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A system that specifies functions in one or more boxes.

[0086] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0087] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of protection of this application is limited to these examples; within the framework of this application, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of one or more embodiments of this application as described above, which are not provided in detail for the sake of brevity.

Claims

1. A method for early warning and prevention based on tensor decomposition and learning of dust big data, characterized in that, include: Collect on-site data from the crushing and transportation sites; the on-site data is multimodal data. Based on the modal information of the field data, feature alignment and mapping are performed on each modal data of the field data to generate a multimodal feature set; Based on the multimodal feature set and the pre-constructed core tensor and factor matrix, a coupled tensor model is constructed. The core tensor compresses and stores the fixed interaction rules between various modal data to realize the reverse derivation and completion of virtual features when modal data is missing. There are multiple factor matrices, each corresponding to a modal data. The multimodal feature set is processed based on the coupled tensor model to generate the corresponding high-order fusion feature vector; A heterogeneous graph is constructed by combining the high-order fusion feature vector and the physical layout of the crushing and transportation site. The heterogeneous graph includes the current operating state of each node. The heterogeneous graph is processed by a pre-built heterogeneous graph contrastive learning model, and a compensation strategy is generated in the case of faulty nodes. The compensation strategy includes a compensation path and a compensation intensity. The pre-built policy decoder is invoked to generate control instructions based on the compensation policy. The control instructions include control parameters for all nodes except the faulty node, so as to achieve fault adaptive topology self-healing. The data collected from the crushing and transportation sites includes: Collect video of dust clouds, equipment operating noise, and environmental physical parameters at the crushing and transportation site. The environmental physical parameters include time series data of dust concentration, wind speed, air pressure, and material flow rate.

2. The early warning and control method based on tensor decomposition and learning of dust big data according to claim 1, characterized in that, The on-site data includes video data, audio data, and environmental physical parameters; The modal information based on the field data is used to perform feature alignment and mapping on each modal data of the field data to generate a multimodal feature set, including: Video features are obtained by extracting features from video data using a 3D convolutional neural network. Audio features are obtained by extracting features from audio data based on short-time Fourier transform and one-dimensional convolutional neural network; The environmental physical parameters are extracted based on a gated recurrent unit network to obtain environmental physical features. The video features, audio features, and environmental physical features are aligned, and a deep neural network extractor is used to map the aligned three types of features to the same latent feature space to obtain the multimodal feature set.

3. The early warning and control method based on tensor decomposition and learning of dust big data according to claim 1, characterized in that, Constructing the core tensor and factor matrix includes: Acquire historical field data; The historical field data is used to construct a historical high-order observation tensor through the outer product operation; Perform coupled Tucker decomposition on the historical high-order observation tensor to obtain the historical high-order observation tensor relationship containing the core tensor and factor matrix to be solved; The historical field data and the alternating least squares method are used to iteratively optimize the historical high-order observation tensor relationship. The core tensor and factor matrix were obtained based on the iterative optimization results.

4. The early warning and control method based on tensor decomposition and learning of dust big data according to claim 1, characterized in that, The construction of the coupled tensor model based on the multimodal feature set and the pre-constructed core tensor and factor matrix includes: Pre-construct the initial model of the coupled tensor; Substituting the multimodal feature set, core tensor, and factor matrix into the initial model of the coupled tensor, we obtain the coupled tensor model. The method further includes: When the first modality data is missing in the field data, the virtual feature vector of the first modality data is generated by reverse derivation through the modular product operation based on the core tensor and factor matrix.

5. The early warning and control method based on tensor decomposition and learning of dust big data according to claim 1, characterized in that, The construction of the heterogeneous graph by combining the higher-order fused feature vectors and the physical layout of the crushing and transportation sites includes: Based on the physical layout of the crushing and transportation site, a node topology diagram is constructed. The nodes are the various equipment nodes in the physical layout, including execution nodes, sensing nodes, and logic control nodes. The edges in the node topology diagram are heterogeneous edges, including material flow edges, spatial flow edges, and collaborative edges. The spatial flow edges are the edges between nodes with a distance less than a preset distance or between nodes in the downwind influence area of ​​the prevailing wind direction, as determined by the Gaussian plume diffusion model. The collaborative edges represent the logical association between various nodes under the same control loop. The initial feature vector of the sensing node is determined based on the higher-order fusion feature vector; The initial feature vectors of the execution nodes and logic control nodes are determined based on the real-time operating status of the execution nodes and logic control nodes in the field data. The heterogeneous graph is formed based on the node topology graph embedded with the initial feature vector, and the message passing mechanism between the nodes in the heterogeneous graph is the meta-path message passing mechanism.

6. The early warning and control method based on tensor decomposition and learning of dust big data according to claim 1, characterized in that, Constructing the heterogeneous graph contrastive learning model includes: A dual-view comparison architecture is constructed, including a graph input layer, a view generation layer, a feature encoding layer, a comparison learning layer, and a self-healing inference layer; A global ideal view and an image mask view are constructed based on a historical heterogeneous graph for training. The global ideal view and the image mask view are used to simulate the state of each node under normal and abnormal operating conditions. The heterogeneous graph contrast learning model is formed by training the dual-view contrast architecture using the InfoNCE loss function and the global ideal view and image mask view used for training.

7. The early warning and control method based on tensor decomposition and learning of dust big data according to claim 1, characterized in that, The process involves calling a pre-built heterogeneous graph contrastive learning model to process the heterogeneous graph and generating a compensation strategy in the event of a faulty node, including: The pre-built heterogeneous graph contrastive learning model is invoked to process the heterogeneous graph and determine whether there are faulty nodes. If a faulty node exists, a compensation path that can compensate for the function provided by the faulty node is automatically calculated based on the heterogeneous graph and the pre-trained topological logic. The compensation intensity is determined by adjusting the attention weights. The compensation path includes nodes that have received fault information and whose feature vectors have drifted.

8. The early warning and control method based on tensor decomposition and learning of dust big data according to claim 1, characterized in that, The policy decoder is constructed from a multilayer perceptron that includes a control branch and a prediction branch; The invocation of the pre-built policy decoder to generate control instructions based on the compensation policy includes: The compensation strategy is input into the policy decoder, and the compensation strategy is processed based on the control branch of the policy decoder to generate the control command. At the same time, the compensation strategy is processed based on the prediction branch to predict the dust concentration change in the future time step and calculate the mean squared error regression loss. This loss is combined with the heterogeneous graph contrast learning loss to form a multi-task joint loss function for model parameter tuning.

9. An early warning and control device based on tensor decomposition and learning of dust big data, characterized in that, include: The data acquisition module is used to collect on-site data from the crushing and transportation sites, and the on-site data is multimodal data. The feature alignment and mapping module is used to perform feature alignment and mapping on each modal data of the field data based on the modal information of the field data, and generate a multimodal feature set; The first construction module is used to construct a coupled tensor model based on the multimodal feature set and the pre-constructed core tensor and factor matrix. The core tensor compresses and stores the fixed interaction rules between various modal data to realize the reverse derivation and completion of virtual features when modal data is missing. There are multiple factor matrices, each corresponding to a modal data. The first processing module is used to process the multimodal feature set based on the coupled tensor model to generate the corresponding high-order fusion feature vector; The second construction module is used to construct a heterogeneous graph by combining the high-order fusion feature vector and the physical layout of the crushing and transportation site. The heterogeneous graph includes the current operating status of each node. The second processing module is used to call a pre-built heterogeneous graph contrastive learning model to process the heterogeneous graph, and generate a compensation strategy in the case of faulty nodes. The compensation strategy includes a compensation path and a compensation intensity. The generation module is used to call a pre-built strategy decoder to generate control instructions based on the compensation strategy. The control instructions include control parameters for all nodes except the faulty node. The data collected from the crushing and transportation sites includes: Collect video of dust clouds, equipment operating noise, and environmental physical parameters at the crushing and transportation site. The environmental physical parameters include time series data of dust concentration, wind speed, air pressure, and material flow rate.

Images