Coal mining machine transmission system multi-modal fault diagnosis method based on physical topology constraint

Abstract

The present application relates to a kind of coal winning machine transmission system multimode fault diagnosis method based on physical topology constraint, belong to the technical field of fault diagnosis. Including: the physical main topology graph of coal winning machine transmission system is constructed;Multi-modal time series signal is obtained, and is mapped into the corresponding component node of physical main topology graph, the time series signal of different modalities of each component node is carried out time series feature coding, obtains the deep state feature of different modalities of each component node, and carries out node internal state consistency fusion, obtains the fusion node state representation of each component node;Propagation and aggregation are carried out in the adjacent relationship defined in physical main topology graph, obtain the updated node state representation of each component node, then carry out graph-level readout operation and node-level discrimination operation, respectively output system-level fault class and component-level fault position, fault probability or abnormal degree.The present application is more in line with actual mechanical transmission mechanism, can improve the integrity of fault feature expression and diagnostic stability under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and in particular to a multimodal fault diagnosis method for a coal mining machine transmission system based on physical topology constraints. Background Technology

[0002] Intelligent fault diagnosis is one of the key technologies for ensuring the safe and efficient operation of mining equipment. The transmission system of a coal mining machine, as the core power transmission unit, typically includes key transmission components such as a motor, drive shaft (input shaft), multi-stage gears, bearings, couplings, and output shaft. It is characterized by complex structure, numerous transmission stages, strong component coupling, severe load fluctuations, and harsh operating environments. Once key transmission components experience wear, cracks, loosening, abnormal meshing, bearing damage, or transmission instability, it can easily lead to downtime, decreased production efficiency, and even serious safety risks. With the continuous improvement of the intelligence, digitalization, and unmanned operation of industrial equipment, condition monitoring and fault diagnosis technologies based on multi-sensor monitoring data have become an important research direction in the field of high-end equipment operation and maintenance.

[0003] Faults in the transmission system of a coal mining machine typically do not occur in isolation. For example, abnormal gear meshing may cause increased shaft vibration, changes in bearing support conditions may affect gear meshing stability, and motor load fluctuations may be transmitted to the output end via the drive shaft and gear set. Therefore, faults in the transmission system of a coal mining machine often manifest as a complex process involving multiple components, multiple modes, and multiple paths of coupled propagation. Existing fault diagnosis methods can be broadly categorized as follows: (1) Methods based on artificial features and traditional machine learning: These methods typically perform time-domain, frequency-domain, or time-frequency-domain analysis on vibration signals to extract artificial features such as root mean square, kurtosis, peak factor, spectral peak factor, envelope spectrum features, and wavelet energy features. Then, models such as support vector machines, random forests, and K-nearest neighbors are used for classification and recognition. These methods are simple to implement, but they rely heavily on artificial feature engineering, have limited generalization ability, and are insufficient for fusing multi-source heterogeneous information under complex working conditions.

[0004] (2) Methods based on single-signal deep learning: These methods typically input vibration signals into one-dimensional convolutional networks, recurrent neural networks, attention networks, or other deep models to automatically extract fault features from the original time series. Compared with traditional manual feature methods, deep learning has stronger automatic representation capabilities, but most methods still focus on single-modal inputs and are difficult to fully utilize auxiliary state information such as current, speed, torque, and operating conditions.

[0005] (3) Deep learning methods based on multimodal fusion: Some existing solutions have begun to jointly utilize multimodal data such as vibration, current, acoustics, temperature, rotational speed or operating condition information, and improve fault identification performance through feature splicing, attention weighting, gating fusion and other methods. Although this type of method has improved the limitations of single-modal methods to some extent, most of them only perform simple fusion at the feature level, do not fully consider the correspondence between different modes and specific transmission components, and lack the expression of the fault propagation process along the actual transmission chain of the coal mining machine.

[0006] (4) Mechanical Fault Diagnosis Methods Based on Graph Neural Networks: In recent years, methods for fault diagnosis of rotating machinery, bearings, gearboxes, and complex equipment using graph neural networks have emerged. These methods typically abstract sensor measurement points, signal channels, sample features, or components into graph nodes, and then construct graph edges based on geometric distance, statistical correlation, K-nearest neighbor relationships, signal similarity, or empirical rules. Node information aggregation and classification are achieved through graph convolutional networks, graph attention networks, graph Transformers, and other methods. Although these methods can utilize graph structures for feature aggregation, their graph structures mainly originate from data statistical relationships, sensor arrangement relationships, or sample similarities, and may not correspond to the actual power transmission paths, gear meshing relationships, and bearing support relationships in the coal mining machine transmission system. Therefore, these methods struggle to accurately express the physical process of fault transmission, coupling, and diffusion along the coal mining machine transmission chain, and also find it difficult to directly correlate diagnostic results with specific transmission components. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a multimodal fault diagnosis method for a coal mining machine transmission system based on physical topology constraints. The technical solution of this invention is as follows: A multimodal fault diagnosis method for a coal mining machine transmission system based on physical topology constraints, comprising: S1. Based on the mechanical structure, component composition, assembly relationship and power transmission path of the coal mining machine transmission system, construct the physical main topology diagram of the coal mining machine transmission system; S2, acquire the multimodal timing signals within a preset time period during the operation of the coal mining machine transmission system; S3 maps multimodal time-series signals to the corresponding component nodes of the physical master topology graph, forming a set of multi-source state observations of the nodes; S4. Perform time-series feature encoding on the time-series signals of different modes in the multi-source state observation set of each component node to obtain the deep state features of different modes of each component node. S5, input the deep state features of different modes of each component node into the node fusion unit to perform intra-node state consistency fusion, and obtain the fused node state representation of each component node; S6. Within the adjacency relationship defined by the physical main topology graph, the fused node state representation of each component node is propagated and aggregated based on the fault impact information between adjacent component nodes according to edge direction, edge type and / or edge weight, so as to obtain the updated node state representation of each component node. S7 performs graph-level readout and node-level discrimination operations based on the updated node status representation of each component node, and outputs the system-level fault category and component-level fault location, fault probability or abnormality degree respectively.

[0008] Preferably, S1 includes: abstracting key transmission components in the coal mining machine transmission system as graph nodes, and abstracting the torque transmission relationship, gear meshing relationship, bearing support relationship, shaft connection relationship and mechanical constraint relationship between the graph nodes as graph edges, generating a physical main topology graph of the coal mining machine transmission system that includes component nodes, mechanical connection edges, edge direction, edge type and edge weight; The key transmission components include one or more of the following: motor, drive shaft, primary gear, secondary gear, tertiary gear, bearing, coupling, and output shaft; The graph edges are assigned one or more of the following attributes: Edge type attribute is used to distinguish torque transmission relationship, gear meshing relationship, bearing support relationship, shaft connection relationship and mechanical constraint relationship; The edge direction attribute is used to characterize the direction of power flow, the direction of fault impact transmission, or the priority direction of information propagation; Edge weights are used to characterize the coupling strength, contact strength, degree of transmission influence, or condition-related propagation strength between key transmission components.

[0009] Preferably, S2 includes: S21, acquire the original vibration signal, original motor current signal, original speed signal, original torque signal and original load signal within a preset time period during the operation of the coal mining machine transmission system, and form the original multimodal time sequence signal; S22, the original multimodal time series signal is denoised, outlier removed, normalized, time aligned and segmented by a sliding window to obtain a multimodal time series signal. The multimodal time series signal includes at least vibration signal, motor current signal, speed signal, torque signal and load signal.

[0010] Preferably, S3 includes: Based on the sensor installation location, signal physical meaning, component affiliation, and power transmission path, multimodal time-series signals are mapped to corresponding component nodes in the physical master topology diagram. Specifically: For vibration sensors installed on critical transmission components, the vibration signals in the multimodal time-series signals are directly mapped to the corresponding component nodes; For sensors installed between adjacent critical transmission components or near support locations, multimodal timing signals are mapped to one or more related component nodes based on the measurement point location, support relationship, or power transmission path. For motor current signals, speed signals, torque signals and load signals, they are broadcast to motor nodes, drive shaft nodes, gear nodes or other relevant transmission component nodes according to their physical objects and power transmission paths; For component nodes without direct sensors, a multi-source state observation set for the node is constructed using the status of adjacent nodes, global operating condition signals, and missing identifiers.

[0011] Preferably, the multi-source state observation set of the target component node includes: local vibration signal of the component corresponding to the target component node; motor current signal or local current response signal related to the target component node; speed signal, torque signal or load signal related to the transmission chain segment where the target component node is located; operating background signal related to the operating state of the target component node; and a missing identifier characterizing whether any mode of the target component node is missing.

[0012] Preferably, in step S5, the fusion mechanism of the fusion unit within the node adopts a cross-modal attention mechanism, a gated fusion mechanism, a dynamic weighted fusion mechanism, a splicing fusion mechanism, or a tensor fusion mechanism.

[0013] Preferably, in step S6, for the target component node, its updated node state representation is jointly determined by its own state and the states of its physical adjacent nodes. The contribution of the adjacent node states is jointly determined by the edge direction, edge type, and edge weight. By using a graph attention network or graph Transformer network that integrates edge attributes and direction constraints, the influence weight of adjacent nodes on the target component node is calculated based on the edge direction, edge type, and edge weight. The states of adjacent nodes are then weighted and aggregated to obtain the updated node state representation of the target component node.

[0014] Preferably, S7 includes: S71, perform graph-level readout operation on the node state representation after all component nodes are updated to obtain a global state representation that characterizes the overall operating state of the coal mining machine transmission system, and perform system-level diagnosis based on the global state representation to output the overall machine health status or fault category judgment result. S72 performs node-level diagnosis for each component node in the physical master topology graph based on its updated node state representation, and outputs the health probability, failure probability, failure type or abnormality degree of the corresponding key transmission component.

[0015] Preferably, the method further includes: S8 constructs modality-level missing, node-level missing, and time-slice-level missing samples during the model training phase, and combines masking reconstruction and consistency constraints for robust training.

[0016] Preferably, the method further includes: S9 outputs diagnostic interpretation results corresponding to the physical structure of the coal mining machine transmission system. The diagnostic interpretation results include key abnormal components, key contribution modes, key propagation edges, and fault impact propagation paths.

[0017] All of the above-mentioned optional technical solutions can be combined arbitrarily, and the present invention will not provide a detailed description of the structure after each combination.

[0018] By means of the above solution, the beneficial effects of the present invention are as follows: By acquiring multimodal time-series signals during the operation of the coal mining machine's transmission system and mapping them to the corresponding component nodes in the physical master topology graph, a multi-source state observation set of nodes is formed. This enables the creation of node state representations with clear component semantics, improving the completeness of fault feature expression and diagnostic stability under complex working conditions. Within the adjacency relationships defined by the physical master topology graph, fault impact information between adjacent component nodes is propagated and aggregated based on edge direction, edge type, and / or edge weight, resulting in updated node state representations for each component node. This makes the fault information propagation process more consistent with the actual mechanical transmission mechanism, and the component-level fault location results more physically plausible. By performing graph-level readout and node-level discrimination operations based on the updated node state representations of each component node, system-level fault categories and component-level fault locations, fault probabilities, or anomaly degrees are output, enabling both assessment of the overall machine health status and identification of specific faulty components.

[0019] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0020] Figure 1 A flowchart of a multimodal fault diagnosis method for a coal mining machine transmission system based on physical topology constraints provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the physical topology of the coal mining machine transmission system in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the process of converting multimodal time-series signals into deep state features of different modes in an embodiment of the present invention; Figure 4 This is a schematic diagram of intra-node state consistency fusion performed by the intra-node fusion unit in an embodiment of the present invention; Figure 5This is a schematic diagram of graph learning propagation under physical topological constraints in an embodiment of the present invention; Figure 6 This is a schematic diagram of the system-level and component-level diagnostic outputs in an embodiment of the present invention; Figure 7 This is a schematic diagram of the robust training mechanism for missing conditions in an embodiment of the present invention; Figure 8 This is a schematic diagram of interpretability analysis based on propagation weights and modal contributions in an embodiment of the present invention. Detailed Implementation

[0021] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0022] like Figure 1 As shown in the figure, the multimodal fault diagnosis method for a coal mining machine transmission system based on physical topology constraints provided in this embodiment of the invention includes: S1. Based on the mechanical structure, component composition, assembly relationship and power transmission path of the coal mining machine transmission system, construct the physical master topology diagram of the coal mining machine transmission system.

[0023] In one specific embodiment, S1 includes: abstracting key transmission components in the coal mining machine transmission system as graph nodes, and abstracting the torque transmission relationships, gear meshing relationships, bearing support relationships, shaft connection relationships, and mechanical constraint relationships between graph nodes as graph edges, thereby generating a physical main topology graph of the coal mining machine transmission system that includes component nodes, mechanical connection edges, edge directions, edge types, and edge weights.

[0024] Specifically, key transmission components with clear physical boundaries and functional attributes in the main transmission structure of the coal mining machine's transmission system, such as motors, drive shafts, primary gears, secondary gears, tertiary gears, bearings, couplings, and output shafts, are abstracted as graph nodes. The torque transmission relationships, gear meshing relationships, bearing support relationships, shaft connection relationships, and mechanical constraint relationships between nodes are abstracted as graph edges to construct a physical main topology graph reflecting the internal structure and fault propagation paths of the coal mining machine's transmission system. For example... Figure 2 As shown, it is a schematic diagram of the physical topology of the coal mining machine transmission system in an embodiment of the present invention.

[0025] The physical master topology graph in this embodiment of the invention can be represented as: G=(V, E, A, R, W); where: V={v1, v2, ..., v N} represents the set of component nodes in the transmission system of the coal mining machine, where N is the number of critical transmission components; E represents the set of mechanical connection edges between nodes; A∈R N×N Represents the adjacency matrix, when the critical transmission component v i With vj When there are torque transmission, gear meshing, bearing support, shaft connection, or mechanical constraint relationships between them, let A ij =1, otherwise let A =1. ij =0; R represents the set of edge types, for example, the edge between the motor and the input shaft is a torque transmission edge, the edge between the first-stage gear and the second-stage gear is a gear meshing edge, the edge between the bearing and the gear shaft or output shaft is a bearing support edge, and the edge between the coupling and the output shaft is a shaft connection edge; W represents the set of edge weights, used to characterize the coupling strength, transmission influence degree, or fault influence propagation strength between adjacent key transmission components. The edge weights can be uniformly initialized to 1, or they can be set according to the transmission ratio, meshing stiffness, support stiffness, shaft connection strength, coupling relationship between components, or expert experience. During model training, the edge weights can be adaptively corrected, but only within the existing physical topology edge set, and no new non-physical edges unrelated to the real mechanical structure are added.

[0026] The physical master topology graph is not automatically generated based on statistical correlations, sample similarities, spatial distances, or K-nearest neighbor relationships between sensor signals. Instead, it is pre-determined based on the actual mechanical structural relationships existing in the coal mining machine's transmission system. Specifically, the nodes in the physical master topology graph correspond one-to-one with the actual transmission components of the coal mining machine's transmission system, or are functionally correlated. These key transmission components include one or more of the following: motor, drive shaft, primary gear, secondary gear, tertiary gear, bearing, coupling, and output shaft. The edges in the physical master topology graph correspond to actual torque transmission relationships, gear meshing relationships, bearing support relationships, shaft connection relationships, or mechanical constraint relationships. This gives the model's graph structure a clear mechanical mechanism, rather than being a simple data correlation graph.

[0027] Preferably, the graph edges are set as directed edges to represent the direction of power and energy transmission in the transmission chain, or the preferred propagation direction of fault effects in the transmission system. Furthermore, self-loop edges can be set on each node to retain the node's original state information, enhancing the ability to preserve local information during node state updates.

[0028] Preferably, the graph edges are provided with one or more of the following attributes: (1) Edge type attribute, used to distinguish torque transmission relationship, gear meshing relationship, bearing support relationship, shaft connection relationship and mechanical constraint relationship; (2) Side direction attribute, used to characterize the direction of power flow, the direction of fault impact transmission, or the priority direction of information propagation; (3) Edge weight attribute, used to characterize the coupling strength, contact strength, transmission influence degree or working condition related propagation strength between key transmission components.

[0029] During model training, the edge weights of existing edges in the physical master topology graph can be modified or adaptively updated. However, edge weight updates are limited to a pre-determined set of physical topology edges, and arbitrary connection edges unrelated to real mechanical connections, power transmission, gear meshing, or support constraints are not generated. Therefore, the graph propagation process is always constrained by the real mechanical relationships of the coal mining machine transmission system, avoiding purely data-driven graph structures that produce information propagation paths inconsistent with the actual transmission mechanism.

[0030] The physical master topology constructed in the above manner ensures that the subsequent fault propagation process is strictly limited by the actual power transmission path and mechanical coupling relationship of the coal mining machine transmission system, thereby improving the accuracy of fault propagation modeling, component-level positioning accuracy, and the reliability of diagnostic results.

[0031] It should be noted that the number of nodes, edge connection relationships, edge direction, and edge attributes can be adjusted accordingly for different transmission stages, gear arrangements, bearing support methods, coupling structures, and output shaft structures.

[0032] S2, acquire the multi-modal timing signals within a preset time period during the operation of the coal mining machine transmission system.

[0033] Specifically, for the multi-source heterogeneous monitoring information involved in the operation of the coal mining machine transmission system, this embodiment of the invention collects multi-modal time-series signals related to the operating status of each key transmission component or system.

[0034] The multimodal timing signals include at least: (1) vibration signals, used to characterize the mechanical impact, wear, loosening, cracks, eccentricity and meshing abnormalities of key transmission components; (2) motor current signals, used to characterize the characteristics of drive load changes, power response, motor operating status and electromechanical coupling status; (3) (3) Speed ​​signal, used to characterize the changes in operating speed, periodic disturbances and working condition switching information of the coal mining machine transmission system; (4) Torque signal, used to characterize the changes in transmission load, impact load and power transmission status; (5) Load or working condition information, used to characterize the operating environment and load background of the coal mining machine.

[0035] In one specific embodiment, S2 includes: S21, acquire the original vibration signal, original motor current signal, original speed signal, original torque signal and original load signal within a preset time period during the operation of the coal mining machine transmission system, and form the original multimodal time sequence signal.

[0036] S22, the original multimodal time series signal is denoised, outlier removed, normalized, time aligned and segmented by a sliding window to obtain a multimodal time series signal. The multimodal time series signal includes at least vibration signal, motor current signal, speed signal, torque signal and load signal.

[0037] Among them, time alignment is used to ensure that different modal signals have a consistent time reference within the same time window; sliding window segmentation is used to divide continuous long time series signals into sample segments suitable for model processing; normalization is used to reduce the amplitude differences between different modal signals, which facilitates subsequent unified modeling.

[0038] Optionally, the original multimodal time-series signals can be resampled to address the issue of inconsistent sampling frequencies among different sensors. Specifically, this can be achieved through interpolation, downsampling, upsampling, or uniform window lengths to adjust the original time-series signals of different modes to the same time scale or sample length. For example, if the original vibration signal has a higher sampling frequency, while the original rotational speed signal and original torque signal have lower sampling frequencies, resampling ensures that the original signals of each mode are aligned within the same time window, facilitating subsequent node mapping and multimodal fusion.

[0039] In a further preferred embodiment, to enhance the ability to characterize weak faults under complex operating conditions, frequency domain features or time-frequency joint features can be constructed based on the above-mentioned time domain preprocessing. For example, the frequency domain signal or time-frequency domain signal of the multimodal time series signal can be obtained through fast Fourier transform, short-time Fourier transform, wavelet transform or other equivalent methods to enhance the model's ability to perceive local shocks, periodic evolution and meshing anomaly features.

[0040] S3 maps multimodal time-series signals to the corresponding component nodes in the physical master topology graph, forming a set of multi-source state observations of the nodes.

[0041] In this model, the multimodal time-series signals are not directly input as a whole, but are mapped to corresponding component nodes in the physical master topology diagram according to the sensor installation location, signal physical meaning, component affiliation, and power transmission path. In a specific embodiment, S3 includes: mapping the multimodal time-series signals to corresponding component nodes in the physical master topology diagram according to the sensor installation location, signal physical meaning, component affiliation, and power transmission path; specifically: For vibration sensors installed on critical transmission components, the vibration signals in the multimodal time-series signals are directly mapped to the corresponding component nodes; For sensors installed between adjacent critical transmission components or near support locations, multimodal timing signals are mapped to one or more related component nodes based on the measurement point location, support relationship, or power transmission path. For motor current signals, speed signals, torque signals and load signals, they are broadcast to motor nodes, drive shaft nodes, gear nodes or other relevant transmission component nodes according to their physical objects and power transmission paths; For component nodes without direct sensors, a multi-source state observation set is constructed using the states of adjacent nodes, global operating condition signals, and missing status indicators. For example, when a secondary gear node lacks a direct sensor, the states of the primary and tertiary gear nodes with which it meshes, as well as the bearing nodes with which it supports, can be used as neighborhood compensation information. Simultaneously, global operating condition features such as motor current, speed, torque, and load signals are introduced as auxiliary inputs, and a missing status indicator is set for the secondary gear node to indicate that it lacks direct sensor observations. This constitutes the input representation for the unobserved component node.

[0042] Through the above mapping method, each component node no longer represents just an abstract signal channel, but rather a state unit with a clear mechanical meaning in the coal mining machine transmission system. This state unit simultaneously includes mechanical vibration state, electrical load state, speed and torque state, and operating background information, thus providing a foundation for subsequent fault propagation modeling under physical topology constraints.

[0043] Specifically, for any component node in the physical master topology diagram, based on the installation location, measurement point location, or functional attributes of its corresponding component, vibration signals, motor current signals, speed signals, torque signals, and operating condition signals (load signals) associated with the operating state of that component are selected to form a multi-source state observation set for that component node. Among them, the multi-source state observation set of the target component node includes at least: (1) the local vibration signal of the component corresponding to the target component node; (2) the motor current signal or local current response signal associated with the target component node; (3) the speed signal, torque signal, or load signal associated with the transmission chain segment where the target component node is located; (4) the operating condition background signal associated with the operating state of the target component node; and (5) a missing identifier characterizing whether any mode of the target component node is missing. The missing identifier can be represented by a binary vector to indicate whether each mode is available. For example, for a component node containing five modes—vibration signal, motor current signal, speed signal, torque signal, and operating condition signal—its missing identifier can be represented as [m1,m2,m3,m4,m5], where mi=1 indicates that the corresponding mode is available, and mi=0 indicates that the corresponding mode is missing. For example, [1,1,1,1,1] indicates that all five modes are available; [0,1,1,1,1] indicates that the vibration mode is missing, but the other modes are available; [0,0,0,0,1] indicates that the component node has no direct sensor data and only has global operating condition information.

[0044] It should be noted that if the frequency domain signal or time-frequency domain signal of the multimodal timing signal is also obtained in step S2 above, the embodiment of the present invention will also map the frequency domain signal or time-frequency domain signal of the multimodal timing signal to the corresponding component node of the physical main topology diagram.

[0045] In this embodiment of the invention, the multimodal timing signals are not simply spliced ​​into an input model. Instead, they are mapped to corresponding component nodes according to sensor installation location, signal physical meaning, component affiliation, and power transmission path. For example, vibration signals can be mapped to corresponding gear, bearing, or shaft nodes, and global or semi-global signals such as current, speed, torque, and load can be broadcast to relevant component nodes according to the power transmission path. In this way, each component node represents the comprehensive operating state of a real component, rather than an abstract signal channel. This method organizes multi-source heterogeneous signals in the coal mining machine transmission system into node state inputs with clear component meanings, thereby avoiding the problems of unclear component semantics and difficulty in fault location caused by simply splicing all modal signals.

[0046] S4. Perform time-series feature encoding on the time-series signals of different modes in the multi-source state observation set of each component node to obtain the deep state features of different modes of each component node.

[0047] Specifically, time-series signals of different modes in a multi-source state observation set belonging to the same key transmission component node are encoded with time-series features to obtain deep state features such as vibration deep state features, current deep state features, speed deep state features, torque deep state features, and operating condition deep state features.

[0048] The temporal feature coding is implemented using a temporal feature coding network, which can be a one-dimensional convolutional neural network, a recurrent neural network, a long short-term memory network, a gated recurrent unit, a self-attention-based temporal coding network, a state-space temporal coding network, a convolutional-attention hybrid coding network, or any combination thereof.

[0049] like Figure 3 As shown, it is a schematic diagram of the process of converting multimodal time-series signals into deep state features of different modes in an embodiment of the present invention.

[0050] S5, input the deep state features of different modes of each component node into the node fusion unit to perform intra-node state consistency fusion, and obtain the fused node state representation of each component node.

[0051] Specifically, since different modal signals reflect different aspects of the operating state of the same key transmission component, simple splicing can easily lead to information redundancy, modal conflicts, and noise superposition. Therefore, this embodiment of the invention designs a multi-source state consistency fusion mechanism within a node to form a fused node state representation that can reflect the true operating state of the component, based on the deep state characteristics of different modalities of the component node.

[0052] In practice, the deep state features of the different modalities mentioned above are input into the fusion unit within the node, the importance weights of the different modalities for the current component node state representation are calculated, and the fusion is performed according to the weights to obtain the fused node state representation of the component node.

[0053] The intra-node state consistency fusion addresses the issue of varying fault susceptibility of the same component under different observation modes. For example, gear meshing anomalies may manifest as impact characteristics in vibration modes, load fluctuations in current modes, and periodic disturbances in speed or torque modes. This embodiment of the invention uses an intra-node fusion mechanism to collaboratively express the deep state characteristics of these different modes, making the fused node state closer to the actual operating state of the component.

[0054] Preferably, the fusion mechanism of the fusion unit within the node adopts a cross-modal attention mechanism, which is used to adaptively learn the cooperative, complementary and redundant relationships between different modalities, thereby highlighting the key modalities that are more sensitive to fault identification and suppressing invalid or weakly correlated modalities that are heavily polluted by noise.

[0055] In alternative implementations, the intra-node fusion mechanism may also adopt a gated fusion mechanism, a dynamic weighted fusion mechanism, a splicing fusion mechanism, a tensor fusion mechanism, a low-rank fusion mechanism, or other equivalent fusion structures, but all of them maintain the basic technical idea of ​​"cooperative fusion of multi-source state observations for the same transmission component node".

[0056] More preferably, the fusion mechanism can dynamically adjust the fusion weights of different modal features by combining the current operating condition information, so as to enhance the model's adaptability under variable load, variable speed and complex operating conditions.

[0057] like Figure 4 As shown, it is a schematic diagram of the intra-node state consistency fusion performed by the intra-node fusion unit in an embodiment of the present invention.

[0058] S6. Within the adjacency relationship defined by the physical main topology graph, the fused node state representation of each component node is propagated and aggregated based on the fault impact information between adjacent component nodes according to edge direction, edge type and / or edge weight, so as to obtain the updated node state representation of each component node.

[0059] Specifically, for any target component node, it only receives information from adjacent component nodes with which it has a real power transmission, gear meshing, bearing support, shaft connection, or mechanical constraint relationship, and does not receive information from arbitrary nodes with which it has no real mechanical relationship. Therefore, the information propagation process between nodes conforms to the actual dynamic connection of the coal mining machine transmission system, rather than arbitrary fully connected propagation or purely statistical correlation propagation.

[0060] During information propagation, the degree of influence of adjacent component nodes on the target node is determined based on edge direction, edge type, and edge weight. Among them: edge direction is used to characterize the direction of power flow or the direction in which the impact of faults propagates preferentially; edge type is used to distinguish different mechanical relationships such as torque transmission, gear meshing, support constraints, and shaft connections; edge weight is used to characterize the coupling strength between components, the degree of transmission influence, or the propagation strength related to operating conditions.

[0061] In one specific embodiment, in step S6, the updated node state of a target component node is jointly determined by its own state and the states of its physical adjacent nodes. The contribution of the adjacent node states is jointly determined by the edge direction, edge type, edge weight, and node state similarity. After multi-layer propagation, the node states gradually integrate fault impact information from the local and global scope of the coal mining machine transmission chain, thereby characterizing the process of fault propagation, coupling, and diffusion along the transmission system in the feature space.

[0062] Preferably, the above propagation process can be implemented by a graph attention network or a graph Transformer network that integrates edge attributes and direction constraints. For the target component node, by comprehensively considering the characteristics of adjacent nodes, edge direction, edge type, and edge weight information, the information propagation weight (influence weight) between the target component node and each adjacent node is calculated. The information transmitted by different adjacent nodes is adaptively weighted and aggregated, and then the neighborhood aggregation and node state update are completed to obtain the updated node state representation of the target component node.

[0063] Taking the primary gear node as an example, when there is a meshing abnormality in the primary gear, its fault impact only propagates along the edges in the physical main topology graph that have a real mechanical relationship with it. These include the torque transmission edge between the input shaft and the primary gear, the gear meshing edge between the primary gear and the secondary gear, and the support constraint edge between the primary gear and the bearing. The model calculates the influence weights of adjacent nodes on the target node based on edge direction, edge type, and edge weight, and performs weighted aggregation of the states of adjacent nodes to update the state of the target node. For nodes that do not have a direct mechanical connection or transmission constraint relationship with the primary gear, no direct information propagation occurs, thus ensuring that the fault propagation path conforms to the actual transmission chain structure of the coal mining machine.

[0064] In alternative implementations, the above propagation process can also be implemented using graph convolutional networks, GraphSAGE, GIN, relational graph neural networks, edge-aware graph neural networks, graph Transformers, or other equivalent graph learning models. However, all of the above models require message passing within the scope of nodes and edges defined by the physical master topology graph.

[0065] The method described in this invention ensures that graph propagation can only occur within real physical edges. During model training, the edge weights of existing physical edges can be updated to represent the changes in the strength of fault propagation between components under different operating conditions, but non-physical edges unrelated to the actual mechanical structure cannot be added. In other words, the model can learn "which real edge has a greater impact," but it cannot arbitrarily connect edges. Through this physical topology constraint propagation mechanism, this invention avoids information propagation paths inconsistent with the actual mechanical structure generated by ordinary data-driven graph models, improving the accuracy of fault propagation modeling and the physical rationality of diagnostic results.

[0066] like Figure 5 As shown, it is a schematic diagram of graph learning propagation under physical topological constraints in an embodiment of the present invention.

[0067] S7 performs graph-level readout and node-level discrimination operations based on the updated node status representation of each component node, and outputs the system-level fault category and component-level fault location, fault probability or abnormality degree respectively.

[0068] In one specific embodiment, S7 includes: S71, perform graph-level readout operation on the updated node state representation of all component nodes to obtain a global state representation that characterizes the overall operating state of the coal mining machine transmission system, and perform system-level diagnosis based on the global state representation to output the overall machine health status or fault category judgment result.

[0069] S72 performs node-level diagnosis for each component node in the physical master topology graph based on its updated node state representation, and outputs the health probability, failure probability, failure type or abnormality degree of the corresponding key transmission component.

[0070] Through the above two-level diagnostic architecture, the embodiments of the present invention can not only realize the monitoring and early warning of the overall operating status of the coal mining machine transmission system, but also further identify the specific transmission components where potential faults are located, meeting the actual needs of "first discovering abnormalities, and then locating the source of faults" in engineering sites.

[0071] like Figure 6 As shown, it is a schematic diagram of the two-level diagnostic output of system level and component level in an embodiment of the present invention.

[0072] Preferably, system-level output can be implemented using global average pooling, global max pooling, attention readout, weighted readout, or other graph-level readout methods; component-level output can be implemented using binary classification, multi-class classification, probability scoring, anomaly score, or fault confidence output methods.

[0073] Furthermore, it can also output diagnostic confidence, fault severity score, or fault risk level to enhance usability in engineering applications.

[0074] It should be noted that the model involved in this embodiment of the invention can be divided into two core modules. The first part is the node state representation learning module, which takes multimodal time-series signals and a physical master topology graph as input. After node mapping, time-series feature encoding, intra-node state consistency fusion, and physical topology constraint graph propagation, it outputs the updated node state representation of each component node. The second part is the fault diagnosis output module, which takes the updated node state representation of each component node as input. It includes a system-level fault classifier and a node-level classifier, which are used to output the overall fault category (system-level fault category) and the specific component fault probability / fault type, respectively. The two parts can be trained end-to-end as a whole diagnostic network.

[0075] Considering the potential issues in the underground operation environment of coal mining machines, such as loose or damaged vibration sensors, abnormal current acquisition, interrupted speed or torque signals, loss of communication at local measurement points, and abnormal continuous time slice data, this embodiment of the invention constructs missing samples corresponding to the aforementioned engineering scenarios during the model training phase, rather than simply performing ordinary random data augmentation.

[0076] The missing samples include one or more of the following: (1) Modal-level missing samples, that is, one of the vibration, current, speed, torque or working condition modes in a certain component node is masked to simulate a certain type of sensor failure or a certain type of signal acquisition abnormality; (2) Node-level missing samples, that is, all modal features of a certain component node are masked to simulate the failure of the measuring point near the component, communication interruption or node observation unavailable; (3) Time slice-level missing samples, that is, signal segments within a certain continuous time period are masked to simulate communication packet loss, acquisition abnormality or short-term data interruption.

[0077] In one specific embodiment, the method further includes: S8, constructing modality-level missing, node-level missing, and time-slice-level missing samples during the model training phase, and performing robust training by combining masking reconstruction and consistency constraints.

[0078] During training, the model uses non-missing modes, adjacent component node states, and physical topology relationships to compensate for missing states, and uses masking reconstruction loss and diagnostic consistency loss to ensure that the diagnostic results under complete input and missing input remain consistent.

[0079] In a further preferred embodiment, the model not only needs to complete system-level and component-level fault diagnosis tasks, but also needs to restore or reconstruct the masked feature representation based on the physical topological neighborhood context and other available modal information, thereby forming a joint training mechanism for incomplete observation scenarios.

[0080] Therefore, even if some sensors fail, the quality of certain modal data deteriorates, or local time slice data is missing during actual deployment, the model can still complete relatively reliable system-level diagnosis and component-level positioning based on the remaining available information and physical neighborhood relationships.

[0081] Furthermore, in order to simultaneously consider the accuracy of overall fault identification, the ability to locate component faults, and the robustness under conditions of incomplete data, this embodiment of the invention adopts a joint training strategy to optimize the model.

[0082] Robust training mechanisms can be implemented using random masking, mask reconstruction, consistency constraint training, noise injection, input zeroing, modality completion, neighborhood compensation, or other equivalent incomplete observation enhancement methods.

[0083] Specifically, a joint objective function consisting of system-level diagnostic loss and node-level diagnostic loss is constructed, and graph-level output and node-level output are optimized simultaneously through end-to-end training.

[0084] Preferably, the joint objective function can be expressed as: L = λ1L sys + λ2L node ; Among them, L sys L represents the system-level diagnostic loss. node λ1 and λ2 represent the node-level diagnostic loss, where λ1 and λ2 are the weighting coefficients of the corresponding loss terms.

[0085] Furthermore, when introducing robust training under missing conditions, a robust constraint loss can be added to the above joint objective function to form an extended objective function: L = λ1L sys + λ2L node + λ3L rec + λ4L cons ; Where: L rec L represents the reconstruction loss for the masked features; cons λ3 represents the consistency constraint loss of the diagnostic results under complete and missing input conditions; λ4 and λ5 are the corresponding weight coefficients.

[0086] This joint optimization approach can effectively balance the overall system status identification, key component location, and robust diagnostic capabilities under incomplete observation conditions.

[0087] In alternative implementations, system-level diagnostic loss and node-level diagnostic loss can be cross-entropy loss, weighted cross-entropy loss, FocalLoss or other equivalent classification loss; reconstruction loss can be mean squared error loss, mean absolute error loss or other equivalent reconstruction loss; consistency constraint loss can be KL divergence, mean squared error constraint or other equivalent consistency constraint methods.

[0088] like Figure 7 As shown, it is a schematic diagram of the robust training mechanism for missing conditions in an embodiment of the present invention.

[0089] In a further embodiment, to improve the reliability of diagnostic results in the on-site operation and maintenance of coal mining machines, the method provided by the present invention includes S9, that is, not only outputting the fault category, but also outputting the diagnostic interpretation results corresponding to the physical structure of the coal mining machine transmission system. The diagnostic interpretation results include key abnormal components, key contribution modes, key propagation edges, and fault impact propagation paths.

[0090] Among them, key anomalous components are determined by node-level failure probability, anomalousness score, or sensitivity contribution; key contributing modes are determined by multi-source state fusion weights within a node, modal attention weights during multi-source fusion, or modal perturbation sensitivity; key propagation edges are determined by edge weights, propagation contributions, attention weights, or propagation contributions from adjacent nodes during graph propagation; and the failure impact propagation path is obtained by combining key anomalous nodes and key propagation edges according to the connection relationship of the physical main topology graph. For example, if the gear node has the highest failure probability and its vibration mode fusion weight is the largest, and the propagation edge weight between the gear and bearing is relatively high, then the gear can be identified as a key anomalous component, vibration as a key contributing mode, and the gear-bearing edge as a key propagation edge.

[0091] The fault propagation path corresponds to the actual power transmission, gear meshing, or support constraint relationships in the coal mining machine's transmission system. By analyzing the above diagnostic interpretation results, maintenance personnel can understand why the model determines that a certain component is abnormal, which modal signals contribute significantly to the diagnostic results, and along which transmission chain path the fault's impact may spread.

[0092] In alternative implementations, interpretability analysis can also be achieved by combining attention heatmaps, gradient attribution, perturbation sensitivity analysis, key time slice identification, key component sorting, or fault propagation path visualization.

[0093] like Figure 8 As shown, it is a schematic diagram of interpretability analysis based on propagation weight and modal contribution in an embodiment of the present invention.

[0094] The method provided in this embodiment of the invention has at least the following technical effects: (1) Since the embodiments of the present invention not only utilize vibration signals, but also combine multimodal data such as current, speed, torque and working condition information, and map them to component nodes in the physical main topology diagram according to the component correspondence, compared with the method of simply splicing multimodal signals into the model, the present invention can form a node state representation with clear component semantics, improve the integrity of fault feature expression and diagnostic stability under complex working conditions, and thus improve the integrity of fault characterization and noise resistance under complex working conditions.

[0095] (2) Since the graph structure in the embodiments of the present invention is constructed based on the actual power transmission path, gear meshing relationship and bearing support relationship of the coal mining machine transmission system, and the graph propagation is limited to the physical topology edge set, the fault information propagation process in the embodiments of the present invention is more in line with the actual mechanical transmission mechanism and the component-level fault location result is more physically reasonable than the method of constructing graph structure based on statistical correlation, sensor distance or sample similarity.

[0096] (3) Since the embodiments of the present invention construct a two-level output structure that combines system-level diagnosis and component-level positioning, the model can not only judge the overall health status of the machine, but also identify specific faulty components, which meets the application requirements of "first discovering the abnormality and then locating the fault source" in the on-site operation and maintenance of the coal mining machine, and is more in line with the needs of on-site fault diagnosis and maintenance decision-making of the coal mining machine.

[0097] (4) Since the embodiments of the present invention construct modal-level missing, node-level missing and time-slice-level missing samples for downhole sensor failure scenarios during the training phase, and use physical neighborhood nodes and remaining available modes for compensation inference, the system-level diagnosis and component-level positioning capabilities can still be maintained when vibration sensor fails, current signal is abnormal or local measurement point communication is interrupted.

[0098] (5) Since the embodiments of the present invention output key abnormal components, key contribution modes, key propagation edges and fault impact propagation paths, the above explanation results correspond to the real physical structure of the coal mining machine transmission system, which makes it easier for operation and maintenance personnel to understand the source of faults, propagation paths and diagnostic basis, improves the credibility of model engineering applications, and makes it easier for on-site operation and maintenance personnel to understand and use.

[0099] (6) The embodiments of the present invention take into account diagnostic accuracy, physical mechanism constraints, component positioning capability, robustness of missing data and engineering interpretability, and are more suitable for intelligent fault diagnosis of complex transmission systems such as coal mining machines. They can also be extended to industrial equipment with a clear transmission chain structure such as tunneling machines, hoists, conveyors, reducers and gearboxes.

[0100] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A multimodal fault diagnosis method for a coal mining machine transmission system based on physical topology constraints, characterized in that, include: S1. Based on the mechanical structure, component composition, assembly relationship and power transmission path of the coal mining machine transmission system, construct the physical main topology diagram of the coal mining machine transmission system; S2, acquire the multimodal timing signals within a preset time period during the operation of the coal mining machine transmission system; S3 maps multimodal time-series signals to the corresponding component nodes of the physical master topology graph, forming a set of multi-source state observations of the nodes; S4. Perform time-series feature encoding on the time-series signals of different modes in the multi-source state observation set of each component node to obtain the deep state features of different modes of each component node. S5, input the deep state features of different modes of each component node into the node fusion unit to perform intra-node state consistency fusion, and obtain the fused node state representation of each component node; S6. Within the adjacency relationship defined by the physical main topology graph, the fused node state representation of each component node is propagated and aggregated based on the fault impact information between adjacent component nodes according to edge direction, edge type and / or edge weight, so as to obtain the updated node state representation of each component node. S7 performs graph-level readout and node-level discrimination operations based on the updated node status representation of each component node, and outputs the system-level fault category and component-level fault location, fault probability or abnormality degree respectively.

2. The multimodal fault diagnosis method for a coal mining machine transmission system based on physical topology constraints according to claim 1, characterized in that, S1 includes: abstracting key transmission components in the coal mining machine transmission system as graph nodes, and abstracting the torque transmission relationship, gear meshing relationship, bearing support relationship, shaft connection relationship and mechanical constraint relationship between the graph nodes as graph edges, generating a physical main topology graph of the coal mining machine transmission system that includes component nodes, mechanical connection edges, edge direction, edge type and edge weight; The key transmission components include one or more of the following: motor, drive shaft, primary gear, secondary gear, tertiary gear, bearing, coupling, and output shaft; The graph edges are assigned one or more of the following attributes: Edge type attribute is used to distinguish torque transmission relationship, gear meshing relationship, bearing support relationship, shaft connection relationship and mechanical constraint relationship; The edge direction attribute is used to characterize the direction of power flow, the direction of fault impact transmission, or the priority direction of information propagation; Edge weights are used to characterize the coupling strength, contact strength, degree of transmission influence, or condition-related propagation strength between key transmission components.

3. The multimodal fault diagnosis method for a coal mining machine transmission system based on physical topology constraints according to claim 1, characterized in that, S2 includes: S21, acquire the original vibration signal, original motor current signal, original speed signal, original torque signal and original load signal within a preset time period during the operation of the coal mining machine transmission system, and form the original multimodal time sequence signal; S22, the original multimodal time series signal is denoised, outlier removed, normalized, time aligned and segmented by a sliding window to obtain a multimodal time series signal. The multimodal time series signal includes at least vibration signal, motor current signal, speed signal, torque signal and load signal.

4. The multimodal fault diagnosis method for coal mining machine transmission system based on physical topology constraints according to claim 1, characterized in that, S3 includes: Based on the sensor installation location, signal physical meaning, component affiliation, and power transmission path, multimodal time-series signals are mapped to corresponding component nodes in the physical master topology diagram. Specifically: For vibration sensors installed on critical transmission components, the vibration signals in the multimodal time-series signals are directly mapped to the corresponding component nodes; For sensors installed between adjacent critical transmission components or near support locations, multimodal timing signals are mapped to one or more related component nodes based on the measurement point location, support relationship, or power transmission path. For motor current signals, speed signals, torque signals and load signals, they are broadcast to motor nodes, drive shaft nodes, gear nodes or other relevant transmission component nodes according to their physical objects and power transmission paths; For component nodes without direct sensors, a multi-source state observation set for the node is constructed using the status of adjacent nodes, global operating condition signals, and missing identifiers.

5. The multimodal fault diagnosis method for a coal mining machine transmission system based on physical topology constraints according to claim 4, characterized in that, The multi-source state observation set of the target component node includes: local vibration signal of the component corresponding to the target component node; motor current signal or local current response signal related to the target component node; speed signal, torque signal or load signal related to the transmission chain segment where the target component node is located; operating background signal related to the operating state of the target component node; and a missing identifier characterizing whether any mode of the target component node is missing.

6. The multimodal fault diagnosis method for a coal mining machine transmission system based on physical topology constraints according to claim 1, characterized in that, In S5, the fusion mechanism of the fusion unit within the node adopts a cross-modal attention mechanism, a gated fusion mechanism, a dynamic weighted fusion mechanism, a splicing fusion mechanism, or a tensor fusion mechanism.

7. The multimodal fault diagnosis method for a coal mining machine transmission system based on physical topology constraints according to claim 1, characterized in that, In step S6, for the target component node, its updated node state representation is jointly determined by its own state and the states of its physical adjacent nodes. The contribution of the adjacent node states is jointly determined by the edge direction, edge type, and edge weight. By using a graph attention network or graph Transformer network that integrates edge attributes and direction constraints, the influence weight of adjacent nodes on the target component node is calculated based on the edge direction, edge type, and edge weight. The states of adjacent nodes are then weighted and aggregated to obtain the updated node state representation of the target component node.

8. The multimodal fault diagnosis method for a coal mining machine transmission system based on physical topology constraints according to claim 1, characterized in that, S7 includes: S71, perform graph-level readout operation on the node state representation after all component nodes are updated to obtain a global state representation that characterizes the overall operating state of the coal mining machine transmission system, and perform system-level diagnosis based on the global state representation to output the overall machine health status or fault category judgment result. S72 performs node-level diagnosis for each component node in the physical master topology graph based on its updated node state representation, and outputs the health probability, failure probability, failure type or abnormality degree of the corresponding key transmission component.

9. The multimodal fault diagnosis method for a coal mining machine transmission system based on physical topology constraints according to claim 1, characterized in that, The method further includes: S8 constructs modality-level missing, node-level missing, and time-slice-level missing samples during the model training phase, and combines masking reconstruction and consistency constraints for robust training.

10. The multimodal fault diagnosis method for a coal mining machine transmission system based on physical topology constraints according to claim 1, characterized in that, The method further includes: S9 outputs diagnostic interpretation results corresponding to the physical structure of the coal mining machine transmission system. The diagnostic interpretation results include key abnormal components, key contribution modes, key propagation edges, and fault impact propagation paths.

Images