Environmental perception graph relation reasoning method for mechanical transmission system

Abstract

The invention belongs to the technical field of intelligent operation and maintenance of industrial equipment, and discloses an environment perception graph relation reasoning method for a mechanical transmission system. The initial full-connection graph generation is based on a multi-sensor time sequence generation full-connection graph of synchronous acquisition. According to the stable sub-graph discovery, a stable sub-graph structure is automatically discovered among multiple sensors, and the characteristics which are still stable under the change of working conditions are represented, learned and explored without changing the environment. And repeating stable sub-graph discovery and environment invariant representation learning, forming a'discovery-evaluation-optimization 'closed-loop normal form, obtaining a relation sub-graph after optimal screening, and converting implicit sensor correlation into an explicit graph structure. According to the method, system-level panoramic perception of the mechanical transmission system is realized by utilizing parallel observation of multiple sensors in spatial distribution.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for industrial equipment, and in particular to a method for reasoning about environmental perception graphs for mechanical transmission systems. Background Technology

[0002] Mechanical transmission systems are widely used in critical equipment such as high-speed trains, industrial conveyor systems, and wind turbine generators. Because these complex systems typically operate in harsh environments, they are highly susceptible to failure. To ensure normal equipment operation, industrial production often establishes equipment maintenance departments, and equipment fault diagnosis is gradually shifting from traditional signal analysis methods to intelligent diagnostic methods based on machine learning. Comprehensive monitoring of mechanical transmission systems typically utilizes multi-sensor distributed monitoring. Simultaneously, model training data often comes from other machines under different operating conditions or of the same type. Due to the diverse operating conditions and varying physical characteristics of machines, the difference in data distribution between the training and application phases reduces the effectiveness of the diagnostic model. A fault diagnosis model based on domain generalization is trained offline on multi-source, multi-sensor data. After training, the model is applied to online equipment monitoring. Summary of the Invention

[0003] Based on the above technical background, the technical problem to be solved by the present invention is the system-level fault diagnosis problem of mechanical transmission system in dynamic environment, that is, to simultaneously model the coupling relationship between components under multi-sensor conditions and realize health status monitoring with robustness to environmental changes.

[0004] The technical solution of the present invention is as follows: A method for inferring environmental perception graph relationships in mechanical transmission systems, comprising the following steps:

[0005] Step 1: Initial fully connected graph construction;

[0006] Step 2: Establish a stable subgraph discovery mechanism, including a sensor importance perception mechanism and a sensor relationship selection mechanism; obtain stable relationship subgraphs from the multi-source fully connected graph through the sensor relationship selection mechanism; and obtain filtered relationship subgraphs from the stable relationship subgraphs based on the sensor importance perception mechanism.

[0007] Step 3: Environment-invariant representation learning; simulate operating condition changes through cross-domain perturbation to obtain data collected under different operating conditions; introduce sensitivity regularization to ensure that the learned representation is robust to environmental changes, thereby reducing dependence on specific operating conditions;

[0008] Repeat steps 2 and 3 to form a closed-loop paradigm of "discovery-evaluation-optimization" and obtain the optimal filtered relational subgraph, which is used to transform the implicit sensor correlations into an explicit graph structure.

[0009] Step 1 specifically involves: sensors located on different components of the mechanical transmission system; generating a fully connected graph based on synchronously measured multi-sensor time series data; each node in the fully connected graph corresponds to a sensor, and its node characteristics are constituted by the signals collected by that sensor; assuming... A fully connected graph is defined as follows:

[0010]

[0011] in, This represents the set of nodes corresponding to N sensors. It is an adjacency matrix; node characteristics Given a time-series signal acquired by sensor n; and a given original multi-sensor time-series sample. , is mapped onto the initial fully connected graph;

[0012]

[0013] in, Let represent a fully connected adjacency matrix initialized with all 1s. For the first In the first working condition Sensors for each sample The input signal; Indicates by the first In the first working condition An initial fully connected graph consisting of samples.

[0014] The sensor importance perception mechanism specifically involves using an importance coefficient to quantify the direct influence of neighboring nodes on the state of the target node in the current task, thereby capturing the dynamic correlation between nodes.

[0015] Initial node features The linear mapping to the new feature space is as follows:

[0016]

[0017] This represents a learnable parameter matrix;

[0018]

[0019] in, Indicates the features after splicing. This represents the embedding of the nth node in the jth sample within the kth working condition; This represents the embedding of the m-th node in the j-th sample within the k-th working condition; Indicates feature splicing;

[0020] Through a learnable importance vector And activated by the LeakyReLU function. Generate unnormalized importance scores ;

[0021]

[0022] In obtaining Next, for the target node All neighboring nodes The importance score is normalized using the softmax function to obtain the importance coefficient. ;

[0023]

[0024] in, Represents a node The set of neighboring nodes; in the case of a multi-head mechanism, the above steps are executed in parallel. Each attention head calculates its own set of importance coefficients and transformed neighbor features; the outputs of all attention heads are concatenated to form a node. Importance perception representation :

[0025] .

[0026] The sensor relationship selection mechanism is specifically as follows: for the first... One sample, node With nodes The edge features between them are constructed by concatenating the node embeddings of the two nodes;

[0027]

[0028] in, Indicates the first In the first working condition The first sample Embedded nodes; Indicates the first In the source domain, the first The first sample Embedded nodes; Represents a node With nodes Edge features between them; Indicates feature splicing;

[0029] A shared linear layer is applied to each edge feature to compute the edge score:

[0030]

[0031] in, Indicates the first In the first working condition Connecting nodes in each sample With nodes Attention score on the edge Represents a learnable linear transformation; assembles edge scores into a symmetric edge score matrix:

[0032]

[0033] A score-ranked edge selection method is used to extract the top q edges with the highest information content; a mask ratio is defined. The number of selected edges is determined by the following formula:

[0034]

[0035] set up This indicates the top scorers. Indexes of edges; construct a binary edge selection mask. ;

[0036]

[0037] Indicates "the The node and the first Edge index numbers between nodes;

[0038] The edges retained after masking are set as candidate stable interactions between sensors, and the resulting stable adjacency matrix is ​​defined as follows:

[0039]

[0040] Stable subgraph discovery mechanism The selected subgraph obtained through learning is represented as follows:

[0041]

[0042] Indicates from the The first working condition The first sample The node embeddings are obtained after passing through the importance attention mechanism.

[0043] The mask ratio .

[0044] When constructing stable relational subgraphs, appropriately retaining more edges can promote more robust representation learning; however, generally speaking, when... When the value is greater than 0.6, the model's diagnostic accuracy will be reduced because it contains too many edges that are irrelevant to the diagnostic task.

[0045] The environment-invariant representation learning specifically involves: learning the representation of selected subgraphs. Graph convolutional networks are used to extract high-dimensional representations. This step involves feature extraction on the selected graph structure to generate node embeddings fused under sensor relationships, which are then used for downstream fault state identification.

[0046]

[0047] In the diagnostic representation learning process, each operating condition is treated as an independent environment. For each environment Calculate the classification loss separately:

[0048]

[0049] in, Indicates environment The number of samples in The logits output by the model. For the corresponding real tags, This represents the cross-entropy loss function; a perturbation-based strategy is introduced during training to simulate changes in operating conditions.

[0050] The overall training objective consists of two parts: (1) the average empirical risk for all environments. (2) Sensitivity regularization term This reduces reliance on specific environmental relevance; formally expressed as follows:

[0051]

[0052]

[0053]

[0054] By jointly optimizing the above objectives, the state of the transmission system can be accurately identified under diverse operating conditions.

[0055] The perturbation-based strategy specifically involves introducing a scaling factor into the predicted logits. To assess the model's loss sensitivity under such perturbations; for each environment The loss after the disturbance is calculated as follows:

[0056]

[0057] exist Calculate the disturbance loss with respect to the scaling factor. The gradient is used to obtain environment-specific sensitivity:

[0058]

[0059] The square of the gradient Norms are used to quantify sensitivity scores:

[0060] .

[0061] The beneficial effects of this invention are:

[0062] This invention provides an environmental perception graph relationship reasoning method for mechanical transmission systems, which aims to achieve robust system-level health monitoring under varying operating conditions while maintaining the interpretability and generalizability of the diagnostic model.

[0063] The proposed stable subgraph discovery mechanism includes sensor importance assessment and relationship selection, which automatically constructs stable relational subgraphs from multi-sensor time-series signals, transforming implicit sensor correlations into explicit graph structures.

[0064] We propose an environment-invariant representation learning strategy. By simulating changes in operating conditions through cross-domain perturbations and introducing sensitivity regularization, we ensure that the learned representations are robust to environmental changes, thereby reducing dependence on specific operating conditions.

[0065] This invention establishes a closed-loop paradigm of "discovery-evaluation-optimization", which organically combines subgraph discovery, environmental assessment and structural optimization to form a distributed fault diagnosis model that can simultaneously possess generalization ability and interpretability. Attached Figure Description

[0066] Figure 1 This is a schematic diagram of step one of the methods proposed in this invention;

[0067] Figure 2 This is a schematic diagram of step two of the method proposed in this invention;

[0068] Figure 3 This is a schematic diagram of step three of the method proposed in this invention;

[0069] Figure 4 This is a schematic diagram of the "discovery-evaluation-optimization" closed-loop paradigm of this invention. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0071] This invention proposes a method for inferring relationships from an environmental perception graph in mechanical transmission systems. Taking a high-speed train bogie system as an example, the bogie consists of a frame, traction motor, axle box, gearbox, axle, wheels, and loading device. Vibration sensors are installed at different key locations. Figure 1 As shown, in the initial graph construction stage, a fully connected graph is generated based on synchronously measured multi-sensor time series data, providing a comprehensive dependency space for subsequent learning processes. Each node in the graph corresponds to a sensor, and its node features are composed of the signals collected by that sensor. Let... A fully connected graph is defined as follows:

[0072]

[0073] in, This represents the set of nodes corresponding to N sensors. This is an adjacency matrix. Node characteristics. It is a sensor The acquired time-series signal. Given a raw multi-sensor time-series sample. It is mapped onto the initial fully connected graph;

[0074]

[0075] in Let represent a fully connected adjacency matrix initialized with all 1s. This is the first In the first working condition Sensors for each sample The input signal.

[0076] This invention employs a sensor importance perception mechanism to achieve this dynamic adjustment. Importance coefficients quantify the direct influence of neighboring nodes on the target node's state in the current task, thereby capturing the dynamic correlation between nodes. Specifically, initial node features... It is linearly mapped to a new feature space, as shown in the following formula:

[0077]

[0078] in, This indicates feature concatenation. The concatenated features... Then, a learnable importance vector is used. And activated by the LeakyReLU function. Generate unnormalized importance scores. ,

[0079]

[0080] In obtaining Next, for the target node All neighboring nodes The importance score is normalized using the softmax function to obtain the importance coefficient. ;

[0081]

[0082] in, Represents a node The set of neighboring nodes. In the case of a multi-head mechanism, the above steps are executed in parallel. Each attention head calculates its own set of importance coefficients and transformed neighbor features. The outputs of all heads are then concatenated to form a node. Importance perception representation:

[0083] .

[0084] This invention employs a sensor relationship selection mechanism, which not only enhances the robustness of the graph structure but also improves the interpretability of the relationship subgraph, enabling the identification of which sensor interactions play a key role in fault detection. Specifically, for the first... One sample, node With nodes The edge features between them are constructed by concatenating their node embeddings:

[0085]

[0086] in, This represents the embedding of the nth node in the j-th sample of the k-th source domain (i.e., the working condition). A shared linear layer is applied to each edge feature to compute the edge score:

[0087]

[0088] in, Indicates the first In the domain, the first Connecting nodes in each sample With nodes Attention score on the edge This represents a learnable linear transformation. Subsequently, these attention scores are assembled into a symmetric edge score matrix:

[0089]

[0090] A score-based edge selection method extracts the top q edges with the highest information content. A mask ratio is defined. The number of selected edges is determined by the following formula:

[0091]

[0092] set up This indicates the top scorers. The index of the edge. Then, construct the binary edge selection mask. ,

[0093]

[0094] This mask preserves the most important information. There are 3 edges, which are assumed to be candidate stable interactions between sensors. The resulting stable adjacency matrix is ​​defined as:

[0095]

[0096] Stable subgraph discovery mechanism The selected subgraph obtained through learning is represented as follows:

[0097]

[0098] This invention proposes an environment-invariant representation learning method for learning robust, invariant features that remain robust under changing operating conditions. (Pair graph) Representation of selected subgraphs Graph convolutional networks are used to extract high-dimensional representations. This step involves feature extraction on the selected graph structure to generate node embeddings fused under sensor relationships, which are then used for downstream fault state identification.

[0099]

[0100] In the diagnostic representation learning process, each operating condition is treated as an independent environment. In each environment, the diagnostic model must first achieve accurate identification of the fault state. Therefore, for each environment... Classification loss needs to be calculated separately:

[0101]

[0102] in, Indicates environment The number of samples in The logits output by the model. For the corresponding real tags, This represents the cross-entropy loss function. A perturbation-based strategy is introduced during training to simulate changes in operating conditions. Specifically, a scaling factor is introduced onto the predicted logits. The model's loss sensitivity under such perturbations is evaluated. For each environment... The loss after the disturbance is calculated as follows:

[0103]

[0104] exist Calculate the disturbance loss with respect to the scaling factor. The gradient is used to obtain environment-specific sensitivity:

[0105]

[0106] The square of the gradient Norms are used to quantify sensitivity scores:

[0107]

[0108] The overall training objective consists of two parts: (1) the average empirical risk for all environments. (2) Sensitivity regularization term This reduces reliance on specific environmental relevance. The formal expression is as follows:

[0109]

[0110]

[0111]

[0112] By jointly optimizing the above objectives, the model is guided to minimize overfitting to environment-specific spurious correlation patterns, thereby enabling accurate identification of the transmission system state under diverse operating conditions.

[0113] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for environment perception graph relationship reasoning for mechanical transmission system, characterized in that, The steps include the following: Step 1, initial full connection graph construction; Step 2, establish a stable subgraph discovery mechanism, including a sensor importance perception mechanism and a sensor relationship selection mechanism; obtain a stable relationship subgraph from the multi-source full connection graph through the sensor relationship selection mechanism; Based on the sensor importance perception mechanism, a filtered relationship subgraph is obtained from the stable relationship subgraph; Step 3, environment invariant representation learning; simulate working condition changes through cross-domain disturbance to obtain data collected under different working conditions; introduce sensitivity regularization to ensure that the learned representation is robust to environmental changes, thereby reducing dependence on specific working conditions; Repeat steps 2 and 3 to form a "discovery-evaluation-optimization" closed-loop paradigm to obtain an optimal filtered relationship subgraph for converting implicit sensor correlations into explicit graph structures.

2. The environment perception graph relationship reasoning method for a mechanical transmission system according to claim 1, characterized in that, The step 1 is specifically: sensors are located on different components of a mechanical transmission system, a full connection graph is generated based on time series of multiple sensors in synchronous measurement; each node in the full connection graph corresponds to a sensor, and a node feature of the node is composed of signals collected by the sensor; let be a full connection graph, which is defined as follows: ； wherein, denotes a set of nodes corresponding to N sensors, is an adjacency matrix; node features is a time series signal collected by sensor n; given a raw multi-sensor time series sample mapped onto an initial fully connected graph; ； in, Let represent a fully connected adjacency matrix initialized with all 1s. For the first In the first working condition Sensors for each sample The input signal; Indicates by the first In the first working condition An initial fully connected graph consisting of samples.

3. The method of claim 1, wherein, The sensor importance perception mechanism specifically quantifies the direct influence strength of neighbor nodes on the state of the target node in the current task, thereby capturing the dynamic correlation between nodes; Initial node features Linear mapping to new feature space, formula as follows: ； This represents a learnable parameter matrix; ； wherein, denotes the concatenated features, denotes the first node embedding of the first sample in the first working condition; denotes the first node embedding of the first sample in the first working condition; denotes the concatenation; through a learnable importance vector and passed through a LeakyReLU activation function to generate unnormalized importance scores ; ； In obtaining Next, for the target node All neighboring nodes The importance score is normalized using the softmax function to obtain the importance coefficient. ; ； wherein, represents a set of neighbor nodes of the node ; in the case of employing the multi-head mechanism, the above steps are performed in parallel times; each attention head respectively calculates its own set of importance coefficients and transformed neighbor features; the outputs of all attention heads are spliced to form an importance-aware representation of the node :​ 。 4. The environment perception graph relationship reasoning method for a mechanical transmission system according to claim 3, characterized in that, The sensor relation selection mechanism is specifically that, for the first sample, the node characteristics between the nodes and the edge are constructed by splicing the node embeddings of both. ； wherein, denotes the node embedding of the i-th sample in the j-th working condition; denotes the node embedding of the i-th sample in the j-th working condition; denotes the node embedding of the i-th sample in the j-th working condition; denotes the node embedding of the i-th sample in the j-th working condition; denotes the node embedding of the i-th sample in the j-th working condition; denotes the node embedding of the i-th sample in the j-th working condition; denotes the node embedding of the i-th sample in the j-th working condition; denotes the edge feature between node i and node j; denotes the edge feature between node i and node j; denotes the edge feature between node i and node j; denotes the edge feature between node i and node j; denotes the edge feature between node i and node j; A shared linear layer is applied to each edge feature to calculate the edge score: ； where, denotes the th sample in the th working condition, the attention score of the edge connecting node and node denotes a learnable linear transformation; and assembling the edge scores into a symmetric edge score matrix: ； The edge selection method based on score ranking extracts the top edges with the largest amount of information; the mask ratio is defined The number of selected edges is determined by the following formula: ； set up This indicates the top scorers. Indexes of edges; construct a binary edge selection mask. ; ； represents an edge index number between the "first node and the "second node represents an edge index number between the "first node and the "second node​ The edges retained after masking are set as candidate stable interactions between sensors, and the resulting stable adjacency matrix is defined as: ； Stable subgraph discovery mechanism The learned refined subgraph representation is represented as: ； represents the node embedding of the i-th node of the i-th sample from the j-th working condition after the importance attention mechanism. represents the node embedding of the i-th node of the i-th sample from the j-th working condition after the importance attention mechanism. represents the node embedding of the i-th node of the i-th sample from the j-th working condition after the importance attention mechanism. represents the node embedding of the i-th 5. The mechanical transmission system oriented environment perception graph relationship reasoning method according to claim 4, characterized in that, The mask ratio .

6. The mechanical transmission system oriented environment perception graph relationship reasoning method according to claim 1, characterized in that, The environment-invariant representation learning specifically involves: learning the representation of selected subgraphs. Graph convolutional networks are used to extract high-dimensional representations. ; This step performs feature extraction on the selected graph structure to generate node embeddings fused under sensor relationships for downstream fault state identification; ； In the diagnostic representation learning process, each working condition is treated as an independent environment ; for each environment , the classification loss is calculated separately: ； where, denotes the number of samples in environment k, are logits output by the model, are the corresponding true labels, denotes the cross-entropy loss function; a perturbation-based strategy is introduced during the training process to simulate the working condition changes; The overall training objective consists of two parts: (1) the average empirical risk over all environments to ensure the predictive performance; (2) a sensitivity regularization term to reduce the dependence on environment-specific correlations; formally expressed as follows: ； ； ； Through joint optimization of the above targets, the state of the transmission system is accurately identified under diversified working conditions.

7. The mechanical transmission system oriented environment perception graph relationship reasoning method according to claim 6, characterized in that, The perturbation-based strategy specifically involves introducing a scaling factor on the predicted logits , evaluating the loss sensitivity of the model under such perturbations; for each environment , the perturbed loss is calculated as follows: ； At The gradient of the perturbation loss with respect to the scaling factor is computed at the point, resulting in an environment-specific sensitivity: ； the square of this gradient The norm is used to quantify the sensitivity score: 。

Images