An adaptive graph diagnosis method and system integrating structure perception and dynamic propagation

Abstract

The present invention discloses an adaptive graph diagnosis method and system that integrates structural perception and dynamic propagation, comprising: collecting vibration acceleration signals under typical fault conditions; dividing the vibration acceleration signal data into multiple time windows, using the sampled data of each time window as the node features of the graph, and constructing graph structure data; based on the graph structure data, using the structural perception similarity modeling mechanism to measure the structural similarity between nodes and dynamically adjust the feature aggregation strategy; using the local adaptive residual feature propagation mechanism to perform multi-layer feature propagation and fusion on the feature aggregation strategy; inputting the fused node representation into the classification module, outputting the corresponding fault type, and realizing intelligent fault diagnosis. While maintaining computational efficiency, the ability to identify weak fault features in complex industrial signals is significantly improved. Effective representation of multi-source time series signals, feature enhancement, and high-precision classification of fault types are achieved.

Description

Technical Field

[0001] The present invention relates to the field of industrial intelligent monitoring and graph neural network modeling, and specifically to an adaptive graph diagnosis method and system that integrates structure perception and dynamic propagation. Background Art

[0002] With the widespread deployment of high-end equipment in key sectors such as aerospace, intelligent manufacturing, electric power, and rail transit, its core operating components (such as bearings, gears, and rotors) are constantly exposed to high loads, high speeds, and complex environmental interference, which can easily generate weak or even nonlinear early fault signals. To ensure stable equipment operation and extend its service life, achieving high-precision, robust, and low-latency fault identification and health assessment of key components has become a core technical issue in intelligent manufacturing and equipment operation and maintenance.

[0003] Traditional data-driven diagnostic methods rely primarily on machine learning algorithms such as support vector machines and decision trees. These rely on artificially constructed features and are unable to adapt to the complexity of signal patterns brought about by changes in operating conditions. In recent years, deep learning technology has developed rapidly, and methods such as convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been widely used in the field of fault diagnosis. CNNs excel at extracting local spatial features from raw vibration or acoustic signals, while RNNs are more suitable for capturing trends in temporal changes. However, CNNs are limited by their convolutional receptive field and struggle to capture long-distance dependencies across windows. RNNs, on the other hand, suffer from problems such as vanishing gradients and time-consuming training when modeling long sequences.

[0004] As an emerging non-Euclidean data processing tool, graph neural networks (GNNs) demonstrate powerful representational capabilities in complex system modeling by modeling topological relationships and attribute interactions between nodes. Especially in industrial scenarios, sensor data often exhibits characteristics such as heterogeneity, multi-source, and structural inconsistency. The introduction of GNNs provides a new modeling approach for such problems. However, existing GNN methods generally focus on homogeneous graph structures, assuming that nodes are of the same type and neighbor labels are consistent. This makes it difficult to cope with real-world scenarios with variable equipment operating conditions and heterogeneous signal sources. In addition, as the number of layers in mainstream GNNs increases, node features tend to become consistent over multiple rounds of aggregation, resulting in "oversmoothing" and the loss of key fault differentiation information, which seriously affects model performance.

[0005] Therefore, there is an urgent need to design a new intelligent diagnosis framework for heterogeneous graph structures that takes into account both multi-frequency feature fusion and deep propagation effectiveness, so as to break through the application bottleneck of existing methods. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] To solve the above technical problems, the present invention provides the following technical solutions: an adaptive graph diagnosis method integrating structural perception and dynamic propagation, comprising: collecting vibration acceleration signals under typical fault conditions;

[0008] Dividing the vibration acceleration signal data into multiple time windows, using the sampled data of each time window as the node features of a graph, and constructing graph structure data;

[0009] Based on the graph structure data, a structure-aware similarity modeling mechanism is used to measure the structural similarity between nodes and dynamically adjust the feature aggregation strategy;

[0010] Using the local adaptive residual feature propagation mechanism, the feature aggregation strategy is propagated and fused in multiple layers;

[0011] The fused node representation is input into the classification module, which outputs the corresponding fault type to achieve intelligent fault diagnosis.

[0012] As a preferred solution of the adaptive graph diagnosis method integrating structural perception and dynamic propagation described in the present invention, the typical fault states include but are not limited to healthy state, outer ring fault, inner ring fault, rolling element fault, and composite fault.

[0013] As a preferred solution of the adaptive graph diagnosis method integrating structure perception and dynamic propagation described in the present invention, wherein: the graph structure data is composed of a feature matrix and an adjacency matrix;

[0014] During the feature matrix construction phase, the time series signal is sampled through a specified time window, and the feature vectors of each time window are stacked into columns to obtain a feature matrix;

[0015] The adjacency matrix is ​​composed of time sequence edges and similarity edges. The time sequence edges include: if two nodes are adjacent in the time sequence, they are recorded as 1; otherwise, they are recorded as 0.

[0016] The similarity edge includes calculating the similarity of the feature vectors of any two nodes; if the similarity reaches a preset value, it is recorded as 1; otherwise, it is recorded as 0.

[0017] As a preferred solution of the adaptive graph diagnosis method integrating structure perception and dynamic propagation described in the present invention, wherein: the structure perception similarity modeling mechanism includes: in a sensor network, node ν i The structure-aware similarity between and neighboring nodes is expressed as:

[0018] Among them, d ij Represents node ν i With neighbor node ν jThe feature similarity measure between Represents node ν i The neighbor node is the node ν i directly connected nodes;

[0019] The squared term of the structural perception similarity is used as an additional feature input to model the nonlinear relationship through the multi-layer perceptron:

[0020]

[0021] Among them, ψ' i The nonlinear expression result of the structure-aware similarity weight of the nodes; MLP sasm Represents the multilayer perceptron used in the structure-aware similarity modeling module for nonlinear mapping of input features.

[0022] As a preferred solution of the adaptive graph diagnosis method integrating structure perception and dynamic propagation described in the present invention, the local adaptive residual feature propagation mechanism includes removing the inter-layer nonlinear operation in the standard GNN structure and introducing the residual information propagation strategy:

[0023] The nodes in the first layer are represented as: H (1) =LARFP (1) (F,X)=FX;

[0024] The nodes in layer l are represented as:

[0025] Where X is the original input feature; γ∈[0,1] represents the hyperparameter; Represents a graph filter that represents the topological structure of the graph by normalizing the adjacency matrix; l represents the index of the layer; LARFP (l) represents the local adaptive residual feature propagation operation of the lth layer; Indicates the index of the layer before layer l; represents the normalized degree matrix; represents the normalized adjacency matrix;

[0026] Apply low-pass filtering and high-pass filtering respectively, and obtain the feature representation of the lth layer:

[0027]

[0028] Through the learnable weight matrix, the d-dimensional features are projected into the z-dimensional feature space:

[0029]

[0030] in, is a learnable weight matrix; Represents a real matrix space with d rows and c columns; Represents the low-frequency information of the lth layer; Represents the high-frequency information of the lth layer; F L Represents a low-pass filter; F G represents a high-pass graph filter; ξ∈[0,1] is a hyperparameter; I represents the identity matrix; represents a nonlinear transformation function; Represents the low-frequency features of the lth layer after projection transformation; represents the high-frequency features of the first layer after projection transformation; A represents the adjacency matrix; D represents the degree matrix;

[0031] During the dimension transformation process, an additional transformation of the original feature X is introduced:

[0032]

[0033] in, is a learnable weight matrix; Represents the original features after projection transformation.

[0034] As a preferred solution of the adaptive graph diagnosis method integrating structure perception and dynamic propagation described in the present invention, the multi-layer feature propagation and fusion includes: Perform multi-stage fusion to obtain the final node representation;

[0035] Node v i Fusion weight calculation: [λ I ,λ L ,λ G ]=MLP λ ([ψ',ψ' 2 ]);

[0036] in, Represents the fusion weight of the node in the initial layer feature; Represents the fusion weight of the node in the low-pass filter layer feature; Represents the fusion weight of the node in the high-pass filter layer feature; MLP λ represents the multilayer perceptron for computing λ; ψ' = [ψ'1, ψ'2, ..., ψ' n ]; n represents the number of nodes; Represents n-dimensional real number space;

[0037] For the lth layer, the fusion of all nodes is expressed as:

[0038]

[0039] Among them, ⊙ represents element-by-element multiplication, Represents λ I The first column of Represents λ L The first column of Represents λ G The first column of .

[0040] As a preferred solution of the adaptive graph diagnosis method integrating structure perception and dynamic propagation of the present invention, the fused node representation includes: after fusing multiple layers of features, the features of each layer are concatenated to form the final node representation:

[0041]

[0042] in, is a learnable output weight matrix, C represents the number of fault categories, L represents the number of layers, z represents the dimension after conversion, and ∥ represents the feature concatenation operation; Z (L) It is the last node representation of the Lth layer.

[0043] An adaptive graph diagnosis system that integrates structural perception and dynamic propagation and adopts the method described in the present invention is characterized by: an acquisition module that acquires vibration acceleration signals under typical fault conditions; a processing module that divides the vibration acceleration signal data into multiple time windows, uses the sampled data of each time window as the node features of the graph, and constructs graph structure data; an adjustment module that measures the structural similarity between nodes based on the graph structure data using a structural perception similarity modeling mechanism and dynamically adjusts the feature aggregation strategy; a calculation module that uses a local adaptive residual feature propagation mechanism to perform multi-layer feature propagation and fusion on the feature aggregation strategy; and a classification module that inputs the fused node representation and outputs the corresponding fault type to achieve intelligent fault diagnosis.

[0044] A computer device comprises: a memory and a processor; the memory stores a computer program, wherein: when the processor executes the computer program, the steps of any one of the methods of the present invention are implemented.

[0045] A computer-readable storage medium stores a computer program, wherein: when the computer program is executed by a processor, the steps of any one of the methods of the present invention are implemented.

[0046] Beneficial effects of the present invention: The adaptive graph diagnosis method provided by the present invention, which integrates structure perception and dynamic propagation, is based on a graph neural network and constructs a hybrid graph structure containing time-dependent edges and similarity edges. Combining structure-aware similarity modeling with a multi-frequency feature residual propagation mechanism, it significantly improves the ability to identify weak fault features in complex industrial signals while maintaining computational efficiency. In response to the data heterogeneity and structural complexity of sensor signals during the operation of high-end industrial equipment, as well as the over-smoothing problem in deep model training, an adaptive heterogeneous graph intelligent diagnosis method integrating structure perception and dynamic residual propagation is provided to achieve effective representation of multi-source time series signals, feature enhancement, and high-precision classification of fault types. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0048] Figure 1 This is an overall flow chart of an adaptive graph diagnosis method integrating structure perception and dynamic propagation provided by the first embodiment of the present invention;

[0049] Figure 2 Comparative results of the time consumed for model training in an adaptive graph diagnosis method integrating structure perception and dynamic propagation provided by the second embodiment of the present invention;

[0050] Figure 3 The performance of GCN and the proposed method at different network layers in an adaptive graph diagnosis method that integrates structure perception and dynamic propagation provided by the second embodiment of the present invention;

[0051] Figure 4 Verification results of a model ablation experiment on an adaptive graph diagnosis method that integrates structure perception and dynamic propagation, provided in the second embodiment of the present invention. DETAILED DESCRIPTION

[0052] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0053] Example 1, with reference to Figure 1, as one embodiment of the present invention, provides an adaptive graph diagnosis method integrating structure perception and dynamic propagation, comprising:

[0054] Step 1: Conduct vibration test experiments on typical bearing faults that are most commonly used in high-end rotating machinery, and collect vibration acceleration signals under typical fault conditions.

[0055] The acceleration signals of the bearing include vibration acceleration signals under various typical fault conditions, such as healthy state, outer ring fault, inner ring fault, rolling element fault, and combined fault.

[0056] Step 2: Divide the vibration acceleration signal data into multiple time windows and construct graph structure data. The graph structure data consists of a feature matrix and an adjacency matrix. The adjacency matrix is ​​composed of time sequence edges and similarity-based edges.

[0057] A hybrid strategy graph data construction method based on time series and similarity measurement aims to effectively convert time series signals into graph data to meet the needs of fault diagnosis tasks. The method mainly includes two key steps: feature matrix construction and adjacency matrix construction. First, in the feature matrix construction stage, the time series signal is sampled through a specified time window, and the sampled data of each time window is used as the node feature of the graph. Assume that the length of the time series signal is L, and it is divided into several time windows, where the data in each window is represented as Where d is the feature dimension of each time window, n is the number of nodes, and n = L / d.

[0058] Therefore, the feature matrix The construction of provides an efficient representation of the temporal evolution of the signal by stacking the feature vectors of each time window into columns.

[0059] The construction of the adjacency matrix is ​​a crucial step in graph data processing, which is defined as the connection relationship between nodes in the graph. In this method, the construction of the adjacency matrix is ​​divided into two major steps, namely, based on time sequence edges and similarity edges, to better capture the temporal dependencies and global patterns in the signal. The time sequence edges mainly capture local short-term dependencies by modeling temporal continuity. Specifically, it is stipulated that the n nodes in each type of fault data are connected in sequence, that is, the i-th node and the i+1-th node form an edge connection to construct the local sequential structure of the graph. Mathematically, the adjacency matrix of the time sequence edges can be expressed as A ts , where each node i is connected to node i+1, and the elements in the matrix satisfy:

[0060]

[0061] i represents the current node, and j represents any node.

[0062] This time-sequential construction method enables graph data to effectively capture the local evolution and short-term dependencies of time series signals.

[0063] Compared with local time dependence, the long-range dependence and periodic characteristics in bearing vibration signals often produce similar vibration patterns at different time points in the signal. This information is crucial for fault diagnosis. Therefore, in the process of constructing the adjacency matrix, a similarity-based edge construction method is introduced to reveal the dependency relationship between remote nodes by calculating the similarity between nodes. Given any two nodes ν i With node ν j , calculate the similarity measure S(i,j) of the feature vectors of the two, which can be calculated using cosine similarity or Euclidean distance.

[0064] By setting the similarity threshold θ, when S(i,j)>θ, an edge connection is formed between node i and node j. Mathematically, the adjacency matrix A of the similarity edge is sim (i,j) is expressed as:

[0065]

[0066] In summary, this hybrid strategy first effectively models the local continuity of the signal using time-sequential edges, capturing the gradual evolution of the vibration signal, which is important for identifying sudden faults. Second, by introducing similarity edges, it can effectively identify long-range dependencies and global patterns in the signal. This is particularly true during the long-term operation of bearings, where periodic fault patterns often exhibit similarities across time steps, and simple time-sequential modeling often fails to reveal these long-range dependencies. Combining these two approaches, the hybrid strategy provides richer and more accurate data input for GNN applications in fault diagnosis, thereby comprehensively improving the performance of GNN models in bearing fault diagnosis, especially for processing complex bearing vibration signals with periodic and nonlinear characteristics.

[0067] Step 3: Based on the graph structure data, the structure-aware similarity modeling mechanism (SASM) is used to measure the structural similarity between nodes and dynamically adjust the feature aggregation strategy.

[0068] The Structural-Aware Similarity Modeling (SASM) method is used to measure the similarity of different nodes in topological structure and adaptively adjust the feature fusion method during information propagation. Specifically, SASM is mainly used to measure the consistency of the target node and its neighborhood in the feature space, thereby providing guidance for subsequent GNN propagation. Assume that in a sensor network, the node ν i The structure-aware similarity between and neighboring nodes is expressed as:

[0069]

[0070] Among them, d ij =S(i,j) represents node v i With neighbor node v j The feature similarity measure between Represents node v i The neighbor node is the node ν i directly connected nodes.

[0071] Structural-aware similarity reflects the information consistency of a node in its local topology, that is, whether its neighborhood has similar characteristic patterns. In homogeneous environments, nodes with high structural similarity can often directly rely on neighbor information to enhance their own characteristics. However, in heterogeneous environments, the characteristics of neighbors may differ significantly, so it is necessary to optimize information propagation strategies to reduce the impact of erroneous information. For example, in a fault diagnosis task, if a node in a sensor network records an abnormal vibration signal while most of its neighboring nodes are in a normal state, then traditional GNNs that directly aggregate neighbor information may weaken the abnormal characteristics of the node, thereby affecting the final classification accuracy. Therefore, in heterogeneous environments, SASM can effectively help the model determine when to enhance neighborhood characteristics and when to rely more on its own information.

[0072] In addition, to further enhance the expressive power of structure-aware similarity, we introduce nonlinear mapping to capture high-order information of feature distribution. Specifically, we use the squared term of structure-aware similarity as an additional feature input and model the nonlinear relationship through a multi-layer perceptron:

[0073]

[0074] Among them, ψ' i The nonlinear expression result of the structure-aware similarity weight of the nodes; MLP sasm Represents the multilayer perceptron used in the structure-aware similarity modeling module for nonlinear mapping of input features.

[0075] This improvement enables structure-aware similarity to not only measure the first-order similarity between a node and its neighbors, but also to encode the variance information of the similarity distribution, thereby improving its ability to distinguish different types of nodes in complex environments. For example, in the traditional similarity calculation method, if a node v1 has two neighbors v2 and v3, and d 12 =0,d 13 =1, then its structural similarity calculation value ψ1 may be the same as the value calculated by another node v4 (whose neighbors all have similar features), resulting in the inability to effectively distinguish nodes of different categories.i ,The model can identify the distribution characteristics of different neighborhoods, and thus more accurately model the local structural differences between nodes.,Through SASM, we can classify the structure of nodes before information propagation, so that homogeneous nodes,prioritize neighborhood information, while heterogeneous nodes rely more on their own,features, thereby optimizing the feature propagation strategy in fault diagnosis tasks.

[0076] Step 4: Perform multi-layer feature propagation and fusion based on the Locally Adaptive Residual Feature Propagation (LARFP) mechanism to balance the propagation of high-frequency and low-frequency information and alleviate the over-smoothing phenomenon in deep graph neural networks.

[0077] The structure-aware similarity modeling mechanism is the first step. Its primary task is to dynamically adjust the information aggregation strategy by measuring the structural similarity between nodes. The core goal of this step is to optimize feature aggregation based on the feature similarity between nodes and their neighbors, thereby improving the model's ability to distinguish between different node features. This similarity modeling is essential before feature aggregation, as it provides a reasonable, similarity-based feature representation for subsequent feature propagation.

[0078] Next, after similarity modeling is complete, the local adaptive residual feature propagation mechanism is applied to the multi-layer feature propagation and fusion stage. The LARFP mechanism guides the flow of information through residual paths during the information propagation process, avoiding the oversmoothing problem found in traditional GNN models and further enhancing the effectiveness of multi-layer information fusion. Only after similarity modeling can the LARFP mechanism effectively propagate and fuse features based on the correct features.

[0079] Therefore, the order of these two steps is determined by their respective functions: similarity modeling provides important structural information for feature propagation, which in turn uses this information to perform multi-level feature fusion. Reversing the order may lead to a lack of basis for the feature aggregation strategy, affecting the performance and stability of the entire model.

[0080] In fault diagnosis tasks, multi-layer feature propagation of sensor data is key to improving model performance. However, traditional GNNs rely primarily on simple aggregation of inter-layer information, which can lead to the loss of important fault features and further cause oversmoothing in deep models. To ensure sufficient information interaction between different layers while preventing the model from losing representational power during deep propagation, we propose a new feature propagation strategy—locally adaptively guided residual feature propagation. This mechanism ensures the validity of information at each layer and enhances learning capabilities for complex industrial signals.

[0081] In the standard GNN structure (such as GCN, GAT), the basic form of information propagation is:

[0082]

[0083] Among them, H (l) is the node representation of the lth layer, W (l) is the trainable parameter matrix, is a nonlinear activation function, F is a graph filter, and the adjacency matrix is ​​normalized To represent the topology of the graph.

[0084] Inspired by SGC, we remove inter-layer nonlinear operations and design a new residual information propagation strategy: In the standard GNN structure, we remove inter-layer nonlinear operations and introduce a residual information propagation strategy:

[0085] The nodes in the first layer are represented as: H (1) =LARFP (1) (F,X)=FX.

[0086] The nodes in layer l are represented as:

[0087] Where X is the original input feature; γ∈[0,1] represents the hyperparameter; Represents a graph filter that represents the topological structure of the graph by normalizing the adjacency matrix; l represents the index of the layer; LARFP (l) represents the local adaptive residual feature propagation operation of the lth layer; Indicates the index of the layer before layer l; represents the normalized degree matrix; Represents the normalized adjacency matrix; the core idea of ​​this method is that each layer contains a certain proportion of the original input information (1-γ)X to ensure that the model does not lose the original features during deep propagation; the accumulated information of the first l-1 layers is introduced in a residual way This allows information from different layers to be comprehensively utilized, thus maintaining sufficient differentiation between messages at different layers without causing excessive smoothing of the information. Through the above node representation update strategy, the LARFP mechanism can dynamically maintain input information and avoid information attenuation during deep propagation.

[0088] In traditional GCN architectures, GNNs primarily use filters normalized by the adjacency matrix to propagate information. However, this approach typically focuses solely on low-pass filtering, ignoring high-frequency information. In industrial fault diagnosis scenarios, low-pass information helps capture global patterns, while high-pass information enhances the ability to capture abnormal signals. Therefore, using only low-pass information can miss important fault features. Therefore, enhanced filters are introduced in the LARFP mechanism to achieve the joint propagation of low-pass and high-pass information.

[0089] Apply low-pass filtering and high-pass filtering respectively, and obtain the feature representation of the lth layer:

[0090]

[0091] After the low-pass and high-pass information extraction is completed, the model needs to further transform these features to reduce the computational complexity, improve the information expression ability, and ensure that different types of features can be reasonably optimized during the fusion process. and high-pass filtering The generated features are still in a high-dimensional space (d-dimensional). Directly using high-dimensional features for node classification is computationally intensive and may result in redundant information propagation. Therefore, we first project the d-dimensional features into a more compact z-dimensional feature space using a learnable weight matrix, enabling more efficient information fusion. The specific calculation is as follows:

[0092]

[0093] in, It is a learnable weight matrix that reduces the feature dimension while retaining the most discriminative low-pass and high-pass information; Represents a real matrix space with d rows and c columns; Represents the low-frequency information of the lth layer; shows the high-frequency information of the first layer; F L Represents a low-pass filter; F G represents a high-pass graph filter; ξ∈[0,1] is a hyperparameter; I represents the identity matrix; represents a nonlinear transformation function; Represents the low-frequency features of the lth layer after projection transformation; represents the high-frequency features of the lth layer after projection transformation; A represents the adjacency matrix; D represents the degree matrix.

[0094] During the dimension transformation process, an additional transformation of the original feature X is introduced to ensure that the model does not completely lose the initial state information during information propagation:

[0095]

[0096] in, is a learnable weight matrix; Represents the original features after projection transformation. Perform multi-stage fusion to obtain the final node representation.

[0097] In GNNs, feature fusion is a crucial step in improving model accuracy and robustness. In particular, in fault diagnosis tasks, features between nodes can differ significantly, making feature fusion at different layers crucial. Traditional feature fusion methods typically use fixed weighting strategies. However, in complex fault diagnosis tasks, the similarity between nodes varies from task to task. Therefore, designing an adaptive feature fusion mechanism based on structure-aware similarity is an effective solution. This method dynamically adjusts the fusion weights of features at different levels by incorporating the structure-aware similarity of nodes, achieving excellent classification results in both homogeneous and heterogeneous graphs.

[0098] In order to effectively fuse features at different levels, we first need to calculate an adaptive fusion weight for each node, that is, the contribution of each node's features between different layers. This fusion weight is calculated based on structure-aware similarity and can dynamically adjust the influence of different layers based on the local structural features of the node. Specifically, for node v i Fusion weight calculation:

[0099] [λ I λ L ,λ G ]=MLP λ ([ψ',ψ' 2 ])

[0100] in, Represents the fusion weight of the node in the initial layer feature; Represents the fusion weight of the node in the low-pass filter layer feature; Represents the fusion weight of the node in the high-pass filter layer feature; MLP λ represents the multilayer perceptron for computing λ; ψ' = [ψ'1, ψ'2, ..., ψ' n ]; n represents the number of nodes; Represents n-dimensional real number space.

[0101] In this way, the final representation of the node is able to combine low-pass and high-pass information, and the weights are adaptively adjusted according to the structure-aware similarity to effectively fuse features from different layers.

[0102] After calculating the fusion weight of each node, we further fuse multiple layers of features. In the feature fusion process of layer l, we first extract the node v through low-pass filtering and high-pass filtering. i The characteristics of low frequency and high frequency Then, node v i The final representation Z (l) Calculated by weighted low-pass and high-pass features:

[0103]

[0104] Among them, ⊙ represents element-by-element multiplication, Represents λ I The first column of Represents λ L The first column of Represents λ G This dynamic fusion weight mechanism based on the similarity of node structure perception ensures that the performance of nodes at different feature levels can be better integrated, thereby improving the classification ability of the model under multiple fault modes.

[0105] Step 5: Input the final fused node representation into the classification module, output the corresponding fault type, and realize intelligent fault diagnosis.

[0106] After multiple layers of feature fusion, we concatenate the features of each layer to form the final node representation. This representation is further passed to the output layer for classification prediction. The final concatenation of the node representation is:

[0107]

[0108] in, is a learnable output weight matrix, C represents the number of fault categories, L represents the number of layers, z represents the dimension after conversion, and ∥ represents the feature concatenation operation; Z (L) It is the last node representation of the Lth layer.

[0109] During this process, the final representation of the node reflects multi-level and multi-scale feature information, greatly enhancing the ability to identify fault patterns. Therefore, the proposed multi-layer feature fusion mechanism based on structure-aware similarity dynamically adjusts the weights of feature fusion by incorporating local structure-aware similarity, ensuring effective fusion of node information during multi-layer propagation. This method can flexibly adapt to structural changes in both homogeneous and heterogeneous graphs, significantly improving classification accuracy in fault diagnosis tasks.

[0110] On the other hand, this embodiment also provides an adaptive graph diagnosis system that integrates structure perception and dynamic propagation, which includes:

[0111] The acquisition module collects vibration acceleration signals under typical fault conditions; the processing module divides the vibration acceleration signal data into multiple time windows, uses the sampling data of each time window as the node features of the graph, and constructs graph structure data; the adjustment module uses the structure-aware similarity modeling mechanism based on the graph structure data to measure the structural similarity between nodes and dynamically adjust the feature aggregation strategy; the calculation module uses the local adaptive residual feature propagation mechanism to perform multi-layer feature propagation and fusion on the feature aggregation strategy; the classification module inputs the fused node representation and outputs the corresponding fault type to realize intelligent fault diagnosis.

[0112] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program code.

[0113] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0114] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0115] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0116] Example 2, reference Figure 2-Figure 4 , is an embodiment of the present invention, which provides an adaptive graph diagnosis method that integrates structure perception and dynamic propagation. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0117] To ensure the effectiveness of model training and comparability of experimental results, this paper performed unified data preprocessing and parameter configuration. First, the original vibration signal was segmented using a sliding time window of fixed length 1024. The data in each time window was treated as the input features of a single node, and each fault type was ensured to contain 100 node samples to ensure a balanced category distribution. All node data was standardized before input into the model to improve the convergence speed and stability of model training. The dataset was divided into training, validation, and test sets in a 6:2:2 ratio, which were used for model parameter training, hyperparameter optimization, and final performance evaluation, respectively. The model was trained using the Adam optimizer, with an initial learning rate of 0.001, a maximum iteration number of 200, a batch size of 64, and an early stopping strategy to prevent overfitting.

[0118] In order to fully verify the effectiveness and advancement of the proposed model in fault diagnosis tasks, this paper designed a systematic comparative experiment covering multiple mainstream model algorithms to ensure the scientificity and rigor of the comparative results. Specifically, the experiment selected three representative comparative algorithms: (1) the classic model based on non-graph structure - Multi-layer Perceptron (MLP). This method only uses the feature information of the node itself and cannot capture the topological dependency, which is used as the benchmark for non-graph structure modeling; (2) the classic GNN model based on homogeneous graphs, including graph convolutional network (GCN) and graph attention network (GAT). This type of method can realize local feature aggregation based on graph structure information, but generally has problems such as over-reliance on low-pass filtering and insufficient adaptability to heterogeneous graph structures; (3) graph learning models for heterogeneous graphs and complex topological structure optimization, such as MixHop, GPRGNN and H2GCN. These methods have made improvements in multi-order feature aggregation and information propagation mechanisms, aiming to improve the representation ability of complex fault modes. The comparative models cover multiple levels from traditional single-node feature learning to complex heterogeneous graph modeling, verifying that the baseline setting is sufficient and has broad representativeness.

[0119] Table 1 Test results of the baseline algorithm and the proposed method on different datasets

[0120]

[0121]

[0122] Experimental results, shown in Table 1, show that the proposed model significantly outperforms existing methods in classification accuracy on three representative datasets, demonstrating excellent generalization and robustness. Specifically, on the DIRG_Bearing dataset, the proposed model achieved a top classification accuracy of 97.62%, significantly exceeding the state-of-the-art models GPRGNN and H2GCN. The proposed method also performed well on the other two datasets, further demonstrating its strong adaptability and robustness in diverse data environments. These experimental results fully demonstrate the effectiveness and broad applicability of the proposed method under diverse complex operating conditions and multimodal sensor data. The overall performance of the compared models shows that the traditional MLP model, due to its lack of consideration of the topological structure and dependencies between nodes, exhibits relatively low classification accuracy on all three datasets. While classic GNNs such as GCN and GAT can effectively aggregate information between nodes by leveraging graph structures, their information propagation relies primarily on low-pass filtering, which limits their performance in heterogeneous graphs and complex fault modes. In particular, while GAT incorporates an attention mechanism to enhance feature representation, it still suffers from insufficient generalization when handling heterogeneous data and complex operating conditions. Improved models such as MixHop and GPRGNN have enhanced their ability to model complex graph structures by introducing multi-scale aggregation and high-order neighbor information. However, they still have limitations in heterogeneous graph information fusion and mitigating oversmoothing issues. H2GCN has improved its heterogeneous graph processing capabilities, but due to its imperfect deep feature fusion mechanism, the model's improvement in feature representation and classification accuracy is limited. In contrast, the model proposed in this paper effectively alleviates the oversmoothing and information loss issues of traditional GNN models by introducing LARFP and a multi-layer feature fusion mechanism based on structure-aware similarity. Furthermore, SASM further enhances the model's adaptability to heterogeneous graph data, enabling it to maintain high classification accuracy and robustness under various complex operating conditions and multimodal signal environments. Comprehensive experimental results demonstrate that our method not only outperforms existing mainstream models in classification accuracy but also possesses stronger generalization and adaptability to complex environments, fully demonstrating its potential for application in high-end equipment health monitoring and intelligent fault diagnosis.

[0123] Comparative analysis of training time:

[0124] In actual industrial applications such as fault diagnosis and health monitoring, the computational efficiency of the model is one of the important indicators to measure its practicality. Especially in the real-time monitoring and dynamic control tasks of high-end equipment, the model needs to have not only high classification accuracy, but also fast training and inference speed to meet the system's requirements for high responsiveness and low latency. Therefore, this paper further compares and analyzes the performance of the proposed method and the typical baseline model in terms of training time, focusing on its computational efficiency and engineering application potential. In the experimental setup, all models are trained under the same data set, hyperparameter configuration and hardware environment to ensure the fairness and comparability of the results. The specific results of training time are as follows: Figure 2 As shown in the figure, the classic GCN and GAT take 3.28 seconds and 4.16 seconds, respectively, to train. Both models employ layer-by-layer convolution or adjacency information aggregation, and perform nonlinear activation or attention weight calculations at each layer, resulting in high computational complexity. This further reduces training efficiency, especially when processing large-scale graph data.

[0125] In contrast, SGC significantly simplifies the computational process by removing multiple layers of nonlinear activation operations, retaining only graph filters and linear transformations, achieving an optimal training time of 0.93 seconds. However, while the SGC model offers advantages in computational efficiency, its overly simplified feature propagation mechanism limits its ability to model complex patterns and heterogeneous data, making it difficult to meet the requirements of fault identification in complex multimodal scenarios. Building on the aforementioned methods, the model proposed in this paper introduces LARFP and a multi-layer feature fusion mechanism based on structure-aware similarity, enhancing the model's expressiveness and discriminative capabilities while controlling the network's computational complexity. Experimental data shows that the proposed model trains in 1.42 seconds, significantly outperforming mainstream GNN models such as GCN and GAT, and approaching the lightweight SGC model. This result demonstrates that the proposed method, while improving model robustness and generalization, effectively alleviates the computational bottlenecks common in deep graph-structured models, enabling efficient training under complex model architectures. This comparative analysis also demonstrates that the proposed model balances feature modeling capabilities with computational efficiency, demonstrating excellent real-time performance and engineering application potential in tasks such as processing heterogeneous graph structures of multimodal sensor data and dynamic fault classification under complex operating conditions. It is particularly well-suited for large-scale deployment in high-end equipment health monitoring systems, intelligent manufacturing equipment, and industrial IoT environments, meeting the multiple requirements of model accuracy, computational efficiency, and resource consumption in practical applications.

[0126] Analysis of over-smoothing problem:

[0127] Over-smoothing is a common problem in deep GNN models when stacking multiple layers. As the number of network layers increases, node features gradually become homogenized, significantly reducing the ability to distinguish between nodes and severely impacting the model's classification performance. This is particularly true in complex scenarios such as high-end equipment fault diagnosis, where fault signatures often exhibit subtle but critical differences. The over-smoothing effect of traditional GNN models further exacerbates confusion about fault types, reducing the practicality and reliability of diagnostic systems. Therefore, alleviating over-smoothing is crucial for improving model depth and generalization performance.

[0128] In order to verify the effectiveness of the proposed method in alleviating the over-smoothing phenomenon, this paper selects the classic GCN as the comparison baseline and evaluates the classification accuracy trend of GCN and the proposed method under different layer settings. The experimental results are shown in Figure 2. Figure 3 As shown in the figure, the GCN model can still maintain a certain degree of classification accuracy when the model is shallow (2-3 layers). However, as the number of model layers increases, the classification performance of GCN decreases rapidly. When the number of layers increases to 10, the accuracy rate drops to 57.36%. This shows that in the deep structure of the GCN model, the node representation tends to be homogenized, making it difficult to retain effective feature information. The over-smoothing problem is significant, and the model's expressive power is greatly weakened.

[0129] In contrast, the proposed model maintains high classification performance even in multi-layer structures. In shallow structures, model accuracy steadily improves with the number of layers, reaching an optimal accuracy of 97.62% for a 5-layer model. Even with further increases in the number of layers, the model maintains high performance. Even in extremely deep structures with 9 and 10 layers, accuracy remains high, with no significant performance degradation. This demonstrates that the proposed method, through LARFP and a multi-layer feature fusion mechanism based on structure-aware similarity, effectively alleviates the information loss and feature oversmoothing problems encountered in traditional GNNs during deep stacking. The local similarity-guided residual path dynamically adjusts feature transfer and the ratio of residual information between different layers, ensuring that the initial features of nodes are preserved during deep propagation. The structure-aware multi-layer fusion achieves a balance between high- and low-frequency information, further enhancing the discriminative power of node representations. Therefore, the proposed model not only exhibits excellent classification capabilities in shallow network structures, but also effectively avoids oversmoothing in deep network architectures, demonstrating stronger feature representation and model robustness.

[0130] SASM plug-and-play module capability analysis:

[0131] To further validate the versatility and effectiveness of the proposed SASM module, this section designed a series of plug-and-play experiments, integrating the SASM module into current mainstream GNN models (GPRGNN and H2GCN) as well as the main model architecture proposed in this paper. A comprehensive comparative analysis was conducted on three representative datasets. The overall experimental results show that the SASM module, as a lightweight plug-and-play unit, significantly improves the classification performance of different models, maintaining consistent performance gains across different datasets and operating conditions. Specifically, after introducing the SASM module, the classification accuracy of the GPRGNN model on the three different datasets increased from 94.12%, 90.11%, and 87.27% to 95.57%, 91.43%, and 88.06%, respectively. Similarly, the H2GCN model showed a slight performance improvement after integrating the SASM module. These results demonstrate the wide adaptability and effectiveness of the SASM module across different GNN architectures. This module introduces a structure-aware similarity evaluation mechanism that dynamically measures and adjusts the weight of information propagation between nodes. This effectively alleviates information confusion in heterogeneous graph data and improves the discriminative ability of node representations. Furthermore, the SASM module boasts a lightweight structure, low computational overhead, and no significant training burden. It offers excellent pluggability and scalability, making it suitable for rapid integration with various GNN architectures.

[0132] Ablation experiment:

[0133] In order to fully verify the effectiveness and necessity of the key modules in the model proposed in this paper, this section designed and implemented ablation experiments to systematically analyze the different components of the model. The focus is on the impact of the SASM module and the local adaptive guided residual feature propagation mechanism on the overall model performance. By removing individual modules for comparative experiments while keeping other parameters and experimental conditions consistent, we deeply analyze their effects on model classification accuracy, feature representation ability, and adaptability to complex working conditions. The specific experimental results are as follows: Figure 4 As shown. First, after removing the SASM module, the model no longer has a dynamic information propagation mechanism based on local structural perception similarity, and cannot adaptively adjust the information fusion strategy according to the structural heterogeneity between the node and its neighbors. Experimental results show that in the absence of the SASM module, the classification performance of the model drops to 96.14%, 93.36% and 91.59%, respectively. Although the overall performance is still higher than some existing comparison models, it is still lower than the complete model. This result verifies the important role of the SASM module in enhancing the distinguishability of node features and improving the classification accuracy of heterogeneous graph data. Through dynamic weight adjustment, SASM makes the model more flexible and robust when processing multimodal sensor data under complex working conditions, significantly alleviating the category confusion problem in the process of node information propagation.

[0134] Secondly, when the LARFP module is removed, the model lacks an effective residual information regulation mechanism, resulting in a weakened information transmission capability in the deep network structure and a decrease in feature expression capability. Especially when the number of GNN layers increases, traditional information aggregation methods often cause severe over-smoothing, resulting in a decrease in node representation capability. Experimental data show that after removing the LARFP module, the accuracy of the model on three typical data sets decreased by 1.04%, 0.96%, and 1.06% compared to the complete model. This phenomenon shows that the LARFP module, by introducing a residual path guided by local similarity, not only effectively preserves the original feature information of the node, but also improves the model's deep feature fusion capability, thereby maintaining strong node discrimination and feature expression capabilities in high-level graph structures, alleviating the performance degradation problem caused by over-smoothing.

[0135] Comprehensive ablation experiment results show that the complete model integrates two key modules, SASM and LARFP, and achieves the best classification accuracy. Compared with the model without a single module, the complete model shows better performance in multiple data sets and multiple working conditions. This fully verifies the scientificity and rationality of the design of the two modules, and also shows that the synergistic effect of the two in the model is significant. SASM is responsible for accurately modeling the structural similarity between nodes and improving the quality of information fusion under heterogeneous graph data, while LARFP enhances the deep representation ability and robustness of the model through residual information guidance and multi-layer information regulation. The synergistic effect of the two makes the proposed model not only have excellent classification performance, but also have good generalization ability and engineering application potential.

[0136] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. An adaptive graph diagnosis method integrating structure perception and dynamic propagation, characterized in that: include: Collect vibration acceleration signals under typical fault conditions; Dividing the vibration acceleration signal data into multiple time windows, using the sampled data of each time window as the node features of a graph, and constructing graph structure data; Based on the graph structure data, a structure-aware similarity modeling mechanism is used to measure the structural similarity between nodes and dynamically adjust the feature aggregation strategy; Using the local adaptive residual feature propagation mechanism, the feature aggregation strategy is propagated and fused in multiple layers; The fused node representation is input into the classification module, which outputs the corresponding fault type to achieve intelligent fault diagnosis; The structure-aware similarity modeling mechanism includes: in a sensor network, nodes The structure-aware similarity between and neighboring nodes is expressed as: ; in, Representation node With neighboring nodes The feature similarity measure between Representation node The neighbor node is the neighbor node of directly connected nodes; The squared term of the structural perception similarity is used as an additional feature input to model the nonlinear relationship through the multi-layer perceptron: ； in, The nonlinear expression result of the structure-aware similarity weight of the nodes; Represents the multi-layer perceptron used in the structure-aware similarity modeling module for nonlinear mapping of input features; The local adaptive residual feature propagation mechanism includes removing the inter-layer nonlinear operation in the standard GNN structure and introducing the residual information propagation strategy: The nodes in layer 1 are represented as: ; The nodes in layer l are represented as: ; in, is the original input feature; represents a hyperparameter; Represents a graph filter, which represents the topological structure of the graph by normalizing the adjacency matrix; l represents the index of the layer; represents the local adaptive residual feature propagation operation of the lth layer; Indicates the index of the layer before layer l; represents the normalized degree matrix; represents the normalized adjacency matrix; Apply low-pass filtering and high-pass filtering respectively, and obtain the feature representation of the lth layer: ； ； ； ； Through the learnable weight matrix, the d-dimensional features are projected into the z-dimensional feature space: ； ； in, is a learnable weight matrix; Represents a real matrix space with d rows and c columns; Represents the low-frequency information of the lth layer; Represents the high-frequency information of the lth layer; represents a low-pass filter; represents a high-pass graph filter; is a hyperparameter; I represents the identity matrix; represents a nonlinear transformation function; Represents the low-frequency features of the lth layer after projection transformation; represents the high-frequency features of the first layer after projection transformation; A represents the adjacency matrix; D represents the degree matrix; During the dimension transformation process, an additional transformation of the original feature X is introduced: ； in, is a learnable weight matrix; Represents the original features after projection transformation.

2. The adaptive graph diagnosis method integrating structure perception and dynamic propagation according to claim 1, characterized in that: The typical fault states include, but are not limited to, healthy state, outer ring fault, inner ring fault, rolling element fault, and combined fault.

3. The adaptive graph diagnosis method integrating structure perception and dynamic propagation according to claim 2, characterized in that: The graph structure data consists of a feature matrix and an adjacency matrix; During the feature matrix construction phase, the time series signal is sampled through a specified time window, and the feature vectors of each time window are stacked into columns to obtain a feature matrix; The adjacency matrix is ​​composed of time sequence edges and similarity edges; the time sequence edges include: if two nodes are adjacent in the time sequence, they are recorded as 1; Otherwise, it is recorded as 0; The similarity edge includes calculating the similarity of the feature vectors of any two nodes; if the similarity reaches a preset value, it is recorded as 1; Otherwise, it is recorded as 0.

4. The adaptive graph diagnosis method integrating structure perception and dynamic propagation according to claim 3, characterized in that: The multi-layer feature propagation and fusion includes: 、 、 Perform multi-stage fusion to obtain the final node representation; node Fusion weight calculation: ; in, Represents the fusion weight of the node in the initial layer feature; Represents the fusion weight of the node in the low-pass filter layer feature; Represents the fusion weight of the node in the high-pass filter layer feature; represents a multilayer perceptron for computing λ; ; n represents the number of nodes; Represents n-dimensional real number space; For the lth layer, the fusion of all nodes is expressed as: ； in, represents element-wise multiplication, express The first column of express The first column of express The first column of .

5. The adaptive graph diagnosis method integrating structure perception and dynamic propagation according to claim 4, characterized in that: The fused node representation includes fusing multiple layers of features and concatenating the features of each layer to form the final node representation: ； in, is the learnable output weight matrix, C represents the number of fault categories, L represents the number of layers, and z represents the dimension after conversion. Represents feature concatenation operation; It is the last node representation of the Lth layer.

6. An adaptive graph diagnosis system that integrates structure perception and dynamic propagation using the method of any one of claims 1 to 5, characterized in that: Acquisition module, collecting vibration acceleration signals under typical fault conditions; A processing module divides the vibration acceleration signal data into multiple time windows, uses the sampled data of each time window as a node feature of a graph, and constructs graph structure data; An adjustment module, based on the graph structure data, uses a structure-aware similarity modeling mechanism to measure the structural similarity between nodes and dynamically adjusts the feature aggregation strategy; A computing module, utilizing a local adaptive residual feature propagation mechanism to perform multi-layer feature propagation and fusion on the feature aggregation strategy; The classification module inputs the fused node representation and outputs the corresponding fault type to achieve intelligent fault diagnosis.

7. A computer device comprising: memory and processor; The memory stores a computer program, wherein the processor implements the steps of any one of the methods according to claims 1 to 5 when executing the computer program.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images