A mechanical equipment residual life prediction method, device, medium and equipment

Abstract

This invention discloses a method, apparatus, medium, and equipment for predicting the remaining useful life of mechanical equipment, relating to the field of equipment remaining useful life prediction. The method includes: acquiring multimodal data of the mechanical equipment under test; extracting intra-modal and inter-modal features from the multimodal data to form a feature set; adjusting the weight distribution of each modality in the feature set to determine the weighted fusion of temporal evolution features; determining a structurally constrained local dynamic graph based on a physical prior graph and an attribute association graph; inputting the physical prior graph, attribute association graph, and structurally constrained local dynamic graph into a multi-channel graph convolutional neural network to determine the fused spatial features; globally aggregating the spatial features to determine topological features; concatenating the temporal evolution features with the topological features to construct a spatiotemporal joint feature vector; and mapping the spatiotemporal joint feature vector layer by layer to obtain the predicted remaining useful life value of the mechanical equipment at future times.

Description

Technical Field

[0001] This invention relates to the field of equipment remaining life prediction, and particularly to a method, apparatus, medium, and equipment for predicting the remaining life of mechanical equipment. Background Technology

[0002] As industrial systems continue to evolve towards large-scale and intelligent production, mechanical equipment has gradually evolved from single devices into complex systems composed of multiple subsystems and components, exhibiting typical characteristics such as strong coupling, multi-scale operation, and dynamic evolution.

[0003] Against this backdrop, the widespread deployment of multi-source sensing technology has resulted in equipment operation data exhibiting high-dimensionality, high-frequency, and highly heterogeneous characteristics, creating a prominent contradiction between "data explosion" and "difficulty in extracting effective knowledge." Remaining useful life (RUL), as a key technology for achieving predictive maintenance and ensuring the safe operation of systems, is crucial for improving the reliability of complex industrial systems.

[0004] Currently, mainstream multimodal RUL fusion methods can be categorized into four types: feature-level fusion, model-level fusion, decision-level fusion, and hybrid-level fusion. Feature-level fusion primarily involves uniformly encoding the original multimodal signals or their feature representations in the input layer or shallow network. Examples include parallel multiscale attention convolutional neural networks (PMSACNN), weighted fusion and kernel principal component analysis (KPCA), and high-dimensional kernel density estimation (KDE) to construct joint representations. However, these methods often neglect deep semantic relationships between modalities, making it difficult to fully capture cross-modal coupling relationships in the degradation process of complex equipment. This limits the collaborative expressive ability of multimodal information during fusion, thus affecting the accuracy of remaining life prediction for mechanical equipment. Summary of the Invention

[0005] This invention provides a method, apparatus, medium, and device for predicting the remaining life of mechanical equipment, to solve the aforementioned problems existing in the prior art, namely, how to achieve accurate prediction of the remaining life of mechanical equipment in the prior art. This invention provides a method for predicting the remaining life of mechanical equipment, the method comprising: Acquire multimodal data corresponding to the mechanical equipment under test; wherein, the multimodal data includes vibration data, sound data, current data and infrared thermal images; By constructing receptive fields of different dimensions, intramodal features are extracted from multimodal data, and a cross-modal attention mechanism is introduced to dynamically capture complementary information between different modalities in multimodal data and determine intermodal features. Based on intramodal and intermodal features, a feature set is constructed. By introducing a dynamic fusion mechanism, the weight allocation of each modal data in the feature set is adaptively adjusted to determine the temporal evolution features after weighted fusion. Based on the physical connections between various sensors in the mechanical equipment, a physical prior graph is constructed. Based on the similarity of the temporal evolution characteristics of each sensor in the mechanical equipment, an attribute association graph is constructed. Based on the physical prior graph and the attribute association graph, a structural constraint matrix is ​​determined. Based on the structural constraint matrix and the dynamic evolution weights determined based on the sensor node characteristics of the mechanical equipment, a structurally constrained local dynamic graph is constructed. The evolutionary features corresponding to the physical prior graph, the attribute association graph, and the structurally constrained local dynamic graph are extracted and concatenated to determine the fused spatial features. The fused spatial features are then globally aggregated to determine the topological structure features. The weighted and fused temporal evolution features and topological features are nonlinearly concatenated to construct a spatiotemporal joint feature vector. By mapping the spatiotemporal joint feature vector layer by layer, the predicted remaining life of the mechanical equipment at future moments is obtained.

[0006] Optionally, the topological features are determined by constructing a multi-channel graph convolutional neural network (GCN). The GCN includes a feature fusion module and a global aggregation module. The feature fusion module extracts and concatenates the evolutionary features corresponding to the physical prior graph, attribute association graph, and structurally constrained local dynamic graph to determine the fused spatial features. The global aggregation module introduces Chebyshev polynomials to globally aggregate the fused spatial features to determine the topological features.

[0007] Optionally, topological features can be determined by globally aggregating the fused spatial features using Chebyshev polynomials, specifically including: ; in, Topological features, For the normalized Laplace matrix, It is a Chebyshev polynomial of order j. Here, J represents the learnable coefficients, and J is the order of the polynomial. The spatial characteristics after fusion.

[0008] Optionally, obtaining the weighted fusion of temporal evolution features specifically includes: Capture feature sets using global average pooling The global context is used, and adaptive importance scores for each feature branch are calculated using nonlinear transformations. s i : ; in, W h1 and W h2 These are the weight matrices of the two fully connected layers within the feature fusion module. ; The adaptive importance scores are normalized using the Softmax function to obtain the adaptive fusion weights. a i Specifically, it includes: ; Based on adaptive fusion weights, the fused temporal evolution feature representation is obtained in the following manner: ; in, Here, N is the feature mapping function, and N is the total number of feature components to be fused. H i For the first i There are 10 initial feature components to be fused.

[0009] Optionally, the step of constructing a structurally constrained local dynamic graph based on the structural constraint matrix and the dynamic evolution weights determined by the sensor node characteristics of the mechanical equipment specifically includes: The sensor node features are mapped to a high-dimensional representation space using a shared weight matrix W, and then the correlation energy between the central node and its neighboring nodes is determined by the following formula through a nonlinear transformation. e ij : ; in, h i , h j Represents a node i and nodes j Input features, a || represents a learnable attention vector, and || represents vector concatenation. Let be the transpose of the learnable attention vector a; Based on correlation energy, for neighborhoods satisfying the constraint matrix M... conduct Softmax Normalization is used to determine the dynamic evolution weights: ; The structural constraint matrix M is multiplied element-wise with the dynamic evolution weights, and the structural constraint local dynamic graph is determined using the following formula. : ; in, The dynamic adjacency matrix represents the structure-constrained local dynamic graph at time t. This represents the influence weight of node j on node i at time t. This represents element-wise product.

[0010] Optionally, the acquisition of the physical prior graph and attribute association graph specifically includes: Based on the structural connectivity of mechanical equipment, the topological adjacency matrix of the physical prior graph is constructed in the following manner. A p,ij : ; The multi-channel temporal evolution features are set to x respectively. i and x j The attribute similarity of multi-channel temporal evolution features is determined by using cosine similarity metric. S ij : Based on attribute similarity S ij By setting a threshold The adjacency matrix of the attribute association graph is determined as follows: A s,ij : .

[0011] Optionally, the mean square error (MSE) can be minimized as the objective function to minimize the error between the predicted value and the actual lifetime value; min-max normalization can be used to standardize the vibration data, sound data, and current data.

[0012] This invention provides a device for predicting the remaining life of mechanical equipment, comprising: The acquisition module is used to acquire multimodal data corresponding to the mechanical equipment under test; wherein, the multimodal data includes vibration data, sound data, current data and infrared thermal images; The temporal evolution module is used to extract intra-modal features from multimodal data by constructing receptive fields of different dimensions, and to introduce a cross-modal attention mechanism to dynamically capture complementary information between different modalities in multimodal data and determine inter-modal features. Based on intra-modal and inter-modal features, a feature set is constructed. By introducing a dynamic fusion mechanism, the weight allocation of each modality in the feature set is adaptively adjusted to determine the weighted fused temporal evolution features. The hybrid topology module is used to construct a physical prior map based on the physical connections between various sensors in a mechanical device; construct an attribute association map based on the similarity of the temporal evolution features of various sensors in the mechanical device; determine the structural constraint matrix based on the physical prior map and the attribute association map; construct a structurally constrained local dynamic map based on the structural constraint matrix and the dynamic evolution weights determined based on the sensor node features of the mechanical device; extract and concatenate the evolution features corresponding to the physical prior map, the attribute association map, and the structurally constrained local dynamic map to determine the fused spatial features; and perform global aggregation on the fused spatial features to determine the topological structure features. The prediction module is used to nonlinearly concatenate the weighted fusion of temporal evolution features and topological features to construct a spatiotemporal joint feature vector; by mapping the spatiotemporal joint feature vector layer by layer, the predicted value of the remaining life of the mechanical equipment at future moments is obtained.

[0013] The present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting the remaining life of mechanical equipment.

[0014] The present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described method for predicting the remaining life of mechanical equipment.

[0015] Compared to existing technologies, the beneficial effects of this invention are as follows: This invention provides a method for predicting the remaining life of mechanical equipment. This method extracts multi-domain hybrid features from vibration, sound, and current time-series signals, and constructs a multimodal data sample set with infrared thermal images to achieve a unified expression of multi-source information. Based on this, a temporal convolutional network is used to construct intramodal feature representations, and a parallel dilated convolutional structure is used to form a multi-scale receptive field, effectively coupling high-frequency impact information with low-frequency degradation trends to mine fine-grained intramodal features in multimodal signals. Simultaneously, a cross-modal attention mechanism is introduced to extract intermodal features to achieve dynamic interaction and semantic alignment, and a cross-modal adaptive fusion strategy is used for weighted integration to alleviate heterogeneous representation conflicts and enhance feature expression capabilities. Secondly, a hybrid topological representation mechanism is constructed, including a physical prior graph, a global static attribute association graph, and a structurally constrained local dynamic graph. A multi-channel graph convolutional neural network is used to learn the hybrid topological features, achieving synergistic optimization of "prior knowledge guidance" and "data-driven modeling." Finally, the temporal evolution features and topological features are deeply integrated to establish a unified spatiotemporal collaborative modeling framework, so as to capture high-order spatiotemporal dependencies in complex systems and achieve accurate prediction of the remaining life of mechanical equipment. Attached Figure Description

[0016] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0017] Figure 1 A flowchart illustrating a method for predicting the remaining life of mechanical equipment provided in an embodiment of the present invention; Figure 2 This is a flowchart of the temporal evolution module for cross-modal attention provided in an embodiment of the present invention; Figure 3 This is a flowchart of dilated convolution processing provided in an embodiment of the present invention; Figure 4 This is a flowchart of the cross-attention projection operator processing provided in an embodiment of the present invention; Figure 5 This is a flowchart illustrating the construction process of a structurally constrained local dynamic graph provided in an embodiment of the present invention. Figure 6 This invention provides a test platform for the full life cycle degradation of a reducer, as described in the embodiments of the present invention. Figure 7 A flowchart illustrating the construction process of degraded data samples provided in this embodiment of the invention; Figure 8 The results of different comparison methods for dataset 1 provided in the embodiments of the present invention; Figure 9 The results of different comparison methods for dataset 2 provided in the embodiments of the present invention; Figure 10 A schematic diagram of the adjacency matrix representation of the physical prior graph provided in an embodiment of the present invention; Figure 11 This is a schematic diagram of the attribute association graph interaction relationship and adjacency matrix representation provided in an embodiment of the present invention; Figure 12 This is a schematic diagram of the structural constraint-type interaction relationship and adjacency matrix representation provided in an embodiment of the present invention; Figure 13 This is a graph showing the comparison results of ablation performance in an embodiment of the present invention. Figure 14 A schematic diagram of a computer device for predicting the remaining life of mechanical equipment provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0019] The technical solution of the present invention and how the technical solution of the present invention solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present invention will now be described with reference to the accompanying drawings.

[0020] Figure 1 This is a flowchart of a method for predicting the remaining life of mechanical equipment provided in an embodiment of the present invention, such as... Figure 1 As shown in this embodiment, a method for predicting the remaining life of mechanical equipment includes: S1: Acquire multimodal data corresponding to the mechanical equipment under test; wherein, the multimodal data includes vibration data, sound data, current data and infrared thermal image.

[0021] For example, the present invention first performs characterization embedding processing on multimodal data of vibration sound current and infrared thermal image; specifically, the multimodal sequence input sample is defined as The finite non-empty set of all modes of the sample at time t is M represents the total number of modalities. Four modalities—vibration, sound, current, and infrared thermal imaging—are used to perform multimodal monitoring and RUL prediction for complex systems of pediatric deformity devices. Let represent the sample value of the i-th mode at time t.

[0022] For time-series data of vibration, sound, and current, min-max normalization is first used for scale standardization. Then, the original features of each modal sample are mapped to an embedding vector. (1) For infrared thermal image modalities, a pre-trained convolutional neural network ResNet18 is used to extract features in order to capture the semantic structural information of the image: (2) in, This represents the feature extraction model after removing the final classification layer, and its output dimension. Through the above mapping and representation embedding mechanism, inputs from different modalities are mapped to a latent feature space of a unified dimension by a modal feature encoder, and their corresponding embedding vectors are used to realize subsequent temporal information flow, spatial structure flow modeling and RUL prediction tasks.

[0023] S2: By constructing receptive fields of different dimensions, intramodal features are extracted from multimodal data, and a cross-modal attention mechanism is introduced to dynamically capture complementary information between different modalities in multimodal data and determine intermodal features; based on intramodal and intermodal features, a feature set is constructed; by introducing a dynamic fusion mechanism, the weight allocation of each modal data in the feature set is adaptively adjusted to determine the temporal evolution features after weighted fusion.

[0024] For example, parallel multi-scale dilated convolutional layers can be used to construct receptive fields of different dimensions to couple high-frequency impact features and low-frequency evolution trends in mechanical systems, thereby mining multi-scale degradation information in multi-modal monitoring signals and extracting intra-modal features. This invention extracts inter-modal features through a cross-attention mechanism and then characterizes high-order temporal evolution features through a cross-modal adaptive fusion processor.

[0025] For example, in response to the significant heterogeneity of multimodal data such as vibration, acoustics, current and infrared thermal images in terms of time scale, physical properties and semantic space, there are challenges such as inconsistent sampling frequencies, feature distribution shifts and semantic expression differences among different modalities, which makes it difficult to effectively align and fuse multimodal information.

[0026] To address this, this invention proposes a Cross-modal Attention-based Temporal Evolution Module. This module aims to capture fine-grained nonlinear degradation features in multimodal signals from mechanical equipment through multi-scale receptive fields. Within a unified temporal modeling framework, it achieves deep interaction and collaborative representation of multimodal features, thereby alleviating heterogeneous representation conflicts and fully leveraging the representational capabilities of single modalities and the complementary information of multimodal data. This results in unified modeling and collaborative enhancement of multimodal temporal evolution features. The module includes intramodal feature extraction, intermodal interaction modeling, and Cross-modal Adaptive Fuser (CAF), with the following structure: Figure 2 As shown.

[0027] First, the intra-modal feature extractor employs TCN for intra-modal feature extraction. Dilated convolution effectively captures multi-scale features, and the complementary multi-scale information and temporal modeling enhance the characterization of long-term temporal dependencies in multi-modal data. This fully mines multi-scale degradation features in multi-modal monitoring signals, capturing fine-grained local temporal information and long-term dependent global semantic features, and achieving precise purification of key information. The inter-modal feature extractor introduces a cross-modal attention mechanism to extract inter-modal features, promoting information interaction and complementary feature extraction between different modalities, thereby enhancing modal expressive power. Dynamically modeling the dependencies between different modalities and achieving semantic alignment of heterogeneous features through the cross-modal attention mechanism alleviates representational conflicts, thus improving the expressive power and robustness of key degradation information.

[0028] Specifically, in intramodal feature extraction, considering the coexistence of high-frequency transient impacts and low-frequency evolution trends in mechanical system signals, a parallel multi-scale dilated convolution structure is introduced. This multi-scale dilated convolution layer addresses the limited receptive field of traditional convolutions, employing dilated convolutions with different dilation factors (d) to capture multi-scale degradation information. For sensor signals with different sampling frequencies, dilated convolution expands the receptive field without increasing the number of parameters. The dilated convolution structure is as follows: Figure 3 As shown.

[0029] For the input sequence x, the dilation convolution operation of the i-th branch is defined as: (3) in, Indicates the expansion rate The convolution operator, f i For the corresponding convolution kernel, i For feature branch indexes of parallel multi-scale convolutional structures, l This is the index for the number of convolutional layers. Different sets of dilation rates can be set. Each branch can obtain a differentiated temporal receptive field, thereby enabling multi-scale modeling capabilities for short-term local dynamics and long-term degradation trends.

[0030] Subsequently, the output features of each branch are fused to obtain a multi-scale temporal representation. H multi : (4) By using a multi-scale dilated convolutional structure, different branches correspond to different temporal receptive fields, thereby achieving collaborative modeling of short-term impact characteristics and long-term degradation trends.

[0031] Obtaining multi-scale time-series representations of various modalities H multi Subsequently, in order to further explore the correlation between signals from different sensors and achieve semantic alignment, a cross-modal attention mechanism (CMA) was introduced to dynamically capture complementary information between heterogeneous data (inter-modal features) and alleviate representation conflicts between different modalities.

[0032] like Figure 4 As shown, for modes i and j, a linear transformation is used to map them to a high-dimensional feature space to obtain the query matrix. Q i Key matrix K j Sum matrix V j : (5) in, This is a learnable weight matrix. Subsequently, the attention weight distribution of mode i to mode j is calculated, and... V j We perform weighted aggregation to obtain the cross-modal representation of mode i after enhancement. :

[0033] (6) in, d k This is the feature dimension scaling factor.

[0034] To capture multi-layered cross-modal interaction information, a multi-head attention structure is adopted: (7) in, This represents the output of the k-th attention. d head For each head's feature dimension, It is a mapping matrix after multi-head splicing, responsible for fusing multi-spatial information back to the original feature dimension.

[0035] Through this mechanism, the model can adaptively align the feature distributions of different modalities, extract key degradation and evolution features, and enhance the robustness of the model under complex conditions.

[0036] For example, high-quality multimodal representations can be systematically constructed from the perspective of multi-scale and multi-level feature expression, obtaining rich and effective modal representation information, and comprehensively improving the model's ability to capture and represent information about complex cross-modal relationships. Cross-modal adaptive fusion, targeting multimodal mixed features, designs a dynamic fusion mechanism to adaptively adjust the weight allocation, alignment methods, and interaction strategies between different modal data, achieving effective fusion, alignment, and collaborative learning of multimodal data in the same latent space. This provides a better feature foundation and decision-making basis for multimodal learning models in complex task scenarios.

[0037] For example, to achieve efficient integration of multimodal hybrid features, a cross-modal adaptive fusion module (CAF) is designed. This module constructs a dynamic weight allocation mechanism for locally degenerate features extracted within a modality and interactive features extracted between modalities, thereby achieving deep alignment and collaborative modeling in the latent space.

[0038] Let the feature set to be fused be... This includes feature vectors that contain both intra-modal and inter-modal elements. First, global average pooling is used to capture the global context, and then an adaptive importance score for each feature branch is calculated using a non-linear transformation.

[0039] (8) in, W h1 and W h2 These are the weight matrices of the two fully connected layers within the feature fusion module. .

[0040] To ensure the competition and complementarity of weights among different features, the importance scores are normalized using the Softmax function to obtain adaptive fusion weights. (9) Finally, the fused temporal evolution feature representation is obtained through weighted fusion: (10) in, This is a feature mapping function used to project heterogeneous features onto a unified metric space. i Here, N represents the feature branch index of the parallel multi-scale convolutional structure, N is the total number of feature branches, and α is the feature branch index. i For adaptive fusion weights, H i The initial feature components extracted from the i-th branch.

[0041] Through the CAF module, the model can dynamically adjust the contribution of each mode based on the real-time quality of the input data, providing a higher quality decision-making basis for subsequent remaining lifetime prediction or fault diagnosis tasks.

[0042] S3: Based on the physical connection relationships between various sensors in the mechanical equipment, construct a physical prior graph; based on the similarity of the temporal evolution characteristics of various sensors in the mechanical equipment, construct an attribute association graph; based on the physical prior graph and the attribute association graph, determine the structural constraint matrix; based on the structural constraint matrix and combined with the dynamic evolution weights determined based on the sensor node characteristics of the mechanical equipment, construct a structurally constrained local dynamic graph; extract and concatenate the evolution features corresponding to the physical prior graph, the attribute association graph, and the structurally constrained local dynamic graph to determine the fused spatial features; perform global aggregation on the fused spatial features to determine the topological structure features.

[0043] For example, this invention constructs three types of graph structure hybrid topological features. The first is a physical prior graph that constructs a static physical prior knowledge topology to set the transmission boundary of information, ensuring that the prediction process conforms to the association logic of the mechanical system. The second is a global static attribute correlation graph that uses cosine similarity to calculate the adjacency matrix. The third is a graph structure that fuses the physical prior graph and the attribute correlation graph to determine a structurally constrained local dynamic graph representation.

[0044] For example, to fully characterize the spatial dependencies among multiple sensors and adapt to the time-varying characteristics of mechanical system degradation, a hybrid topology representation mechanism is constructed. This mechanism consists of a physical prior graph, a global static attribute association graph, and a structurally constrained local dynamic graph. Figure 3 Part of the composition involves establishing a collaborative modeling framework between "prior knowledge guidance" and "data-driven learning," enabling the model to achieve a balance between structural invariance and dynamic adaptability, thereby improving prediction accuracy and engineering interpretability. The specific acquisition process of the physical prior graph, the global static attribute association graph, and the structurally constrained local dynamic graph may include:

[0045] (1) Physical prior diagram representation The spatial layout and structural connections of sensors in mechanical systems contain important prior knowledge. However, traditional data-driven methods often overlook this physical structural information, potentially leading to spurious modeling that does not conform to engineering logic. To address this, the physical composition, spatial geometric distribution, and functional coupling logic of mechanical equipment are explicitly encoded into a fixed graph structure, constructing a static physical prior graph. This effectively ensures that the model prediction process conforms to the objective correlation logic of the mechanical system, achieving physical consistency constraints for unconstrained data-driven modeling. This graph remains unchanged under different samples and operating conditions, and is embedded in the neural network structure as a domain knowledge constraint to limit the propagation path of feature information. Based on the structural connectivity of the mechanical system, a topological adjacency matrix of the physical prior graph is constructed:

[0046] (11) This matrix defines the "physical boundary" of feature flow. During graph convolution, this mask forces constraint information to propagate only along physically reasonable paths, thereby suppressing spurious correlations that violate engineering logic and achieving physical consistency constraints for purely data-driven modeling.

[0047] (2) Attribute association graph representation In addition to explicit physical connections, there may be implicit correlations between different sensors caused by operating environments or functional synchronization. To uncover this long-term consistency at the statistical attribute level, a data-driven global static attribute correlation graph is constructed.

[0048] Let the multi-channel time-series evolution characteristics be: x i and x j The attribute similarity is measured using cosine similarity: (12) This leads to the attribute association matrix: (13) By setting a threshold Obtain the adjacency matrix of the attribute association graph A s Only significantly relevant connections are retained. This figure reflects the global synchronous evolution relationship among sensors, can identify sensor groups with similar functional degradation trends, and enhances the model's ability to perceive changes in the global state of the system.

[0049] (3) Structural constraint type local dynamic diagram representation To characterize the complex nonlinear coupling relationships during the degradation process of a mechanical system, this invention proposes a structurally constrained local dynamic diagram, such as... Figure 5 As shown, this method follows the design principle of "fixed skeleton, evolving weights." The topological connections of the graph are pre-locked by the prior structure, while the interaction strength between nodes is dynamically updated over time. Specifically, physical prior knowledge and attribute associations are used to limit the boundaries of information propagation, avoiding spurious correlations that violate engineering logic from purely data-driven modeling. At the same time, a dynamic attention evolution operator is introduced to adaptively learn the association strength between nodes, thereby characterizing the constantly changing nonlinear interaction relationships during the degradation process.

[0050] Within this framework, the structural relationships and functional couplings of the mechanical system are explicitly integrated into the graph topology construction process, achieving a unity of structural stability and relational dynamism. The graph's connection skeleton is jointly determined by the physical prior graph and the attribute association graph, and is pre-locked through a masking mechanism, thereby constructing a structural constraint matrix to limit the evolution range of subsequent dynamic weights.

[0051] First, the structural constraint matrix M is constructed by combining the physical prior graph and the global static attribute association graph: (14) in, This indicates a logical OR operation, meaning that dynamic feature exchange is allowed only if there is a physical connection or strong attribute correlation between two nodes.

[0052] At each time step t, the interaction strength between node i and node j is no longer determined by the static distance, but is adaptively learned through an evolutionary attention mechanism. Specifically, firstly, the node features are mapped to a high-dimensional representation space using a shared weight matrix W, and then the association energy between the center node and its neighboring nodes is calculated through a nonlinear transformation.

[0053] (15) in, h i , h j Let represent the input features of nodes i and j, a represent the learnable attention vector, and || represent vector concatenation.

[0054] To ensure the sparsity of the graph structure and prevent interference from invalid information, only neighborhoods satisfying the constraint matrix M are considered. Perform Softmax normalization to obtain the final dynamic evolution weights: (16) The final dynamic adjacency matrix is ​​a deep fusion of physical structure constraints and data-driven attention weights, and its matrix element representation at time step t is as follows: (17) in, , This represents element-wise product.

[0055] The structurally constrained dynamic adjacency matrix exhibits the characteristic of real-time evolution with the degradation state. Its element changes characterize the nonlinear migration process of sensor signal correlation strength as a mechanical component transitions from health to failure. Compared to traditional global static graph methods, this strategy effectively portrays the time-varying and non-stationary characteristics of relation strength during degradation, significantly improving the model's robustness to sudden disturbances and complex noise environments. Simultaneously, structural constraints ensure that dynamic learning always adheres to the physical framework of the mechanical system, while the weight evolution mechanism accurately captures the evolutionary trend of multi-channel coupling strength, enabling the model to perceive subtle changes in the internal interconnected structures of the system. This provides a physically meaningful characterization of degradation evolution for remaining lifetime prediction.

[0056] For example, after constructing a hybrid topology of physical prior graph, attribute association graph, and structural constraint dynamic graph, this invention designs a multi-view graph convolutional network (MG-TCN) to effectively integrate spatial dependency information from different topological channels. This network achieves deep modeling of the spatial coupling features of complex mechanical systems through multi-graph parallel convolution and feature fusion mechanisms.

[0057] Suppose the hybrid topology contains K=3 channels, corresponding to the physical prior graph, the global static attribute association graph, and the structural constraint type local dynamic graph, respectively. The adjacency matrix of the k-th channel is denoted as... A k In the l-th layer graph convolution, spatial feature aggregation is performed independently for each channel:

[0058] (18) in, The LeakyReLU activation function is used. This indicates the introduction of a self-connected adjacency matrix. Its degree matrix, For the first l Layer node embedding, whose initial input is temporal multi-domain features; This is the learnable weight matrix for this channel.

[0059] To fully preserve the evolutionary features of the physical prior graph, the global static attribute association graph, and the structurally constrained local dynamic graph, a feature concatenation method is used to fuse the outputs of each channel. The fused spatial features are represented as follows: (19) Z1 represents the spatial feature representation obtained by channel aggregation based on the physical prior graph, reflecting the spatial dependency information that conforms to the physical topological constraints of the mechanical structure. Z2 represents the spatial feature representation obtained by channel aggregation based on the attribute association graph, reflecting the spatial dependency information between nodes based on the similarity of long-period statistical attributes. Z3 represents the spatial feature representation obtained by channel aggregation based on the structural constraint-type local dynamic graph, reflecting the spatial dependency information dynamically generated under physical constraints as the working conditions evolve.

[0060] The concatenated feature vector integrates multi-source spatial representations from different topological perspectives, providing multi-source spatial topological structure information support for subsequent predictions.

[0061] To further capture long-range spatial dependencies across components, Chebyshev polynomials are introduced to globally aggregate the concatenated feature Hspat, resulting in enhanced topological features: (20) in, For the normalized Laplace matrix, It is a Chebyshev polynomial of order j. Here, represents the learnable coefficients, and J is the polynomial order. This higher-order convolution can characterize system-level macroscopic degradation patterns, enhancing the model's ability to express complex spatial dependencies.

[0062] S4: Nonlinearly concatenate the weighted fusion of temporal evolution features and topological features to construct a spatiotemporal joint feature vector; by mapping the spatiotemporal joint feature vector layer by layer, obtain the predicted remaining life of the mechanical equipment at future moments.

[0063] For example, in obtaining the temporal evolution representation respectively Enhanced characterization of spatial flow topology This section then designs a feature-level fusion strategy to achieve full integration of spatiotemporal information.

[0064] To simultaneously preserve the subtle fluctuations in temporal evolution and the spatial coupling at the system level, temporal evolution features and topological structure features are nonlinearly concatenated to construct a spatiotemporal feature fusion vector: (twenty one) in, This is a feature splicing operator. This operation aligns the deep abstraction of the time dimension with the structural topology of the spatial dimension, forming a multidimensional coverage of the health state of the mechanical system.

[0065] Subsequently, the fusion features will be The input is fed into a regression prediction head constructed from a multilayer perceptron (MLP). The transformation from the feature space to the RUL value space is achieved through layer-by-layer mapping:

[0066] (twenty two) in, Let be the predicted RUL value at time t; W r1 , b r1 Weights and biases for predicting the head hidden layer; ReLU nonlinear activation function; W r2 ,b r2 These are the parameters for the output layer.

[0067] To train the aforementioned end-to-end model, mean squared error (MSE) is used as the objective function for parameter optimization. The weights of the spatiotemporal two-stream network are updated by minimizing the deviation between the predicted and actual lifetime values. (twenty three) in,N This represents the total number of training samples. and Let represent the true label and the model prediction value of the i-th sample, respectively. The regularization coefficient is . The set of learnable parameters for the model. This is the L2 regularization term, used to adjust the balance between empirical risk and structural risk, preventing the model from overfitting during training.

[0068] For example, given the limited availability of publicly available multimodal system-level datasets, this invention validates the proposed method on a typical single-device multimodal dataset. Although the experimental object is a single device, its operation still involves multi-sensor coupling, multimodal information interaction, and complex degradation and evolution characteristics, which can reflect key modeling problems in complex mechanical equipment systems to a certain extent. Therefore, this experimental setup not only verifies the effectiveness of the proposed method in multimodal collaborative representation and spatiotemporal modeling, but also provides a feasible basis for its extension to larger-scale complex system scenarios.

[0069] The full life cycle degradation test platform for speed reducers mainly includes key components such as motors, speed reducers, and loads. Its structural composition and sensor arrangement are as follows: Figure 6 As shown in the figure. This invention focuses on the performance degradation characteristics of a reducer during operation, comprehensively collecting multi-source sensing information such as vibration, sound, current, and infrared images to construct a realistic and complete degradation dataset. For vibration signal monitoring, four horizontal and vertical accelerometers are arranged at the input and output ends of the reducer to synchronously collect the multi-directional vibration responses of the bearings at both ends. Sound signals are acquired by arranging sound sensors at the input and output ends of the reducer to achieve real-time monitoring of bearing operating noise. Current signals are directly collected from the input and output currents of the drive motor to characterize changes in electromagnetic load and transmission resistance. Simultaneously, a miniature infrared thermal imager is installed on the side connected to the magnetic powder brake to record the temperature distribution of the reducer housing and its changes over time, enabling visualized thermal feature tracking of the mechanical degradation process. Data acquisition channels and signal types are shown in Table 1.

[0070] Table 1 Data acquisition channels and signal types of the reducer For example, the experiment collected full-life degradation data for three reducers under three typical operating conditions: Condition 1 at 30Hz / 195N, Condition 2 at 35Hz / 180N, and Condition 3 at 40Hz / 165N. The sampling frequency was 20480Hz, each sampling lasted 1 second, and the sampling interval was 30 seconds. As the service life progressed, various degradation characteristics gradually emerged: vibration amplitude reached 6–10 times that of the initial normal state, the equipment was accompanied by significant abnormal noise, the drive current fluctuated significantly, and the local temperature in the infrared image showed a significant increase. These multi-source signal characteristics collectively characterized the evolution process of the reducer from a healthy state to a failure state. The experimental data preparation method and data scale are shown in Table 2.

[0071] Table 2. Experimental data preparation methods and data scale After data acquisition, the raw data is preprocessed, including outlier removal and signal cleaning. Multiple noise reduction strategies are employed for vibration, sound, and current signals to improve signal effectiveness and stability. Subsequently, 45 types of features, including time-domain, frequency-domain, time-frequency-domain, complexity-domain, and relative similarity features, are extracted from the denoised signals. Feature selection is based on comprehensive evaluation criteria. To better evaluate the comprehensive score, the impact of correlation on equipment performance is given priority in the comprehensive evaluation indicators. The final feature score is defined as follows:

[0072] (twenty four) in , and Set them to 0.4, 0.3, and 0.3 respectively. Corr As a correlation indicator, Mon As a monotonicity indicator, Rob As a robustness index, the three eigenvalue ranges are: .

[0073] For each sensing channel, the optimal 10-dimensional features are selected to form a feature subset. For infrared thermal images, representative thermal features are extracted through operations such as cropping, noise suppression, and contrast enhancement to highlight the temporal evolution information of the local temperature rise in the reducer. Finally, a degradation data sample set is constructed based on the characteristics of the multimodal signals.

[0074] For example, representative thermal features can be extracted from infrared thermal images through operations such as cropping, noise suppression, and contrast enhancement to highlight the temporal evolution information of the local temperature rise in the reducer. Finally, a degradation data sample set is constructed based on the characteristics of the multimodal signals, and its sample structure diagram is shown below. Figure 7 As shown.

[0075] For example, the proposed method can be validated on Anaconda3 2020.10 and Matlab2022a. The experiment uses a ResNet18-based network structure as the backbone model. For data preprocessing, the visual modality samples 3 frames from each video segment; if the number of frames is less than 3, it is repeatedly sampled and padded; if the number of frames is greater than 3, they are randomly discarded, retaining 3 frames. The image size is normalized to 112×112 and standardized. The audio modality sampling frequency is set to 16kHz, processed into a spectral vector of length 45000, and truncated or zero-padding is performed during loading. The training set is divided into 70% and the test set into 30%. The learning rate in the CME module... The number of iterations for each discriminator update is set to 5. The activation functions used are ReLU and Tanh. The cross-modal fusion module has 128-dimensional intermediate hidden layers, and the activation function is ReLU. The optimizer used is Adam, with an initial learning rate of 0.001. Dropout is set to 0.3, and the maximum number of iterations is set to 60. A threshold can be set. It is 0.5.

[0076] Generally, Gated Graph Convolutional Networks (GGCNs) construct a spatial-temporal map from multi-sensor signals as input to a prediction model. Gated graph convolutional layers are built to accurately extract wear features by simultaneously modeling the temporal and spatial dependencies in the multi-sensor signals. The extracted features are then fed into a quantile regression layer to estimate remaining service life.

[0077] Hierarchical Attention Graph Convolutional Network (HAGCN): This network uses a bidirectional long short-term memory network to model the temporal dependencies of sensor measurements. Based on the spatiotemporal graph generated by the sensor network, a regularized self-attention graph pooling method is designed to achieve effective information fusion from the sensors.

[0078] The Deep Adaptive Transformer and Graph Attention Network Enhancement Model (GAT-DAT) is used to construct graph data through the correlation between sensors. The GAT submodule fuses node features to extract spatial correlations, and the DAT submodule effectively extracts the temporal features of the data through a self-attention mechanism and adaptively predicts the remaining service life of the mechanical equipment.

[0079] A Spatiotemporal Fusion Attention Approach Based on Graph Neural Networks (STFA) is proposed: This approach employs a spatial GNN to fuse the spatial features and structural information of a device, and introduces an improved attention mechanism to fuse temporal features. The fused features are then input into a fully connected layer for remaining lifetime prediction.

[0080] The Dual-Stream Spatiotemporal Fusion Network (DS-STFN) transforms sensor monitoring data into grid and graph structure formats. An enhanced temporal convolutional network with an improved self-attention mechanism extracts temporal features from the grid-structured data. A graph converter module is designed to capture spatial features from the graph-structured data. Finally, feature fusion is performed on the spatiotemporal features to predict the RUL (Relative Limiting Rank).

[0081] To comprehensively evaluate the consistency and deviation between model predictions and actual values, four commonly used indicators are introduced, including: the coefficient of determination (R²). 2 The root mean square error (RMSE), mean absolute percentage error (MAPE), and mean absolute error (MAE) are also mentioned. Among them, R... 2 The metrics are used to measure a model's ability to explain data variance; RMSE and MAE assess prediction bias from the perspectives of squared error and absolute error, respectively; MAPE measures the proportional deviation of the model's prediction error from the true value from the perspective of relative error. These metrics together constitute a reliable evaluation system for model prediction performance, defined as follows:

[0082] (25) (26) (27) (28) To better illustrate the uncertainty of the prediction results, a 95% confidence interval is constructed. Uncertainty analysis mainly includes three key metrics: Prediction Interval Coverage Probability (PICP), Prediction Interval Normalized Average Width (PINAW), and Continuous Ranked Probability Score (CRPS). Specifically, PICP statistically assesses the reliability of the prediction interval by measuring the frequency with which actual values ​​fall within the estimated range. PINAW quantifies the average dispersion of the prediction interval, while CRPS provides a comprehensive assessment of the overall accuracy and probabilistic calibration characteristics of the prediction distribution. The mathematical formulas for these metrics are as follows:

[0083] (29) in, n For the sample size, y i These are actual measured values. Let I be the i-th measurement interval. I is an index function, if... yi If the value falls within the prediction range, it equals 1; otherwise, it equals 0.

[0084] (30) in, n For the sample size, y max and y min These represent the maximum and minimum values ​​of the actual measured data, respectively. and These are the upper and lower bounds of the i-th prediction interval, respectively.

[0085] (31) In the formula, F(x) is the cumulative distribution function (CDF) of the predicted distribution, and y is the actual observed value. For the Heaviside step function ( (1 if true, 0 otherwise). CRPS measures the difference between the predicted CDF and the actual observed values, providing a comprehensive score for the accuracy and clarity of the predicted distribution.

[0086] Table 3 shows the performance comparison results of different remaining lifetime prediction methods on datasets 1R_1 and 2R_2. It can be observed that, compared with existing mainstream remaining lifetime prediction methods, the proposed method achieves optimal performance under both operating conditions. On dataset R_1, the proposed method R... 2 The proposed method achieves an improvement of 0.89%–7.17%, a reduction in RMSE of 26.61%–63.41%, a reduction in MAPE of 40.86%–72.00%, and a reduction in MAE of 31.56%–65.32%. On the R_2 dataset, the proposed method achieves a significant improvement in performance. 2 The improvement ranged from 1.11% to 5.49%, while the RMSE decreased by 33.72% to 62.34%, the MAPE decreased by 35.25% to 66.90%, and the MAE decreased by 34.88% to 61.97%. Figure 8 and Figure 9 The results of different comparison methods for dataset 1 are shown.

[0087] Table 3. Performance comparison of different remaining lifetime prediction methods on two datasets. For example, since the infrared thermal imager's monitoring field of view covers the entire reducer experimental platform, the infrared image temperature signal (N9) is defined as a global supervision node in the physical prior graph. A hierarchical physical prior network is constructed by establishing bidirectional connections between N9 and all local sensor nodes. This topology allows the model to capture local mechanical damage (such as vibration anomalies caused by pitting or spalling) while simultaneously performing cross-dimensional corrections to the overall degradation of the system through global thermal evolution. The graph consists of nine nodes, with the connections between nodes strictly adhering to the mechanical transmission logic and physical coupling characteristics of the experimental platform. Specifically, it is divided into four levels, as shown in Table 4.

[0088] Table 4. Node hierarchy and logical edge definition of the physical prior graph like Figure 10 The figure shows the adjacency matrix representation of the physics prior graph. Figure 11 This represents the interaction relationships and adjacency matrix of the attribute association graph. The graph structure learned through static graph learning is as follows: Figure 12 As shown, global static graphics are used to describe the long-term stable structural relationships within the device, which helps in understanding the potential spatial correlations between sensors.

[0089] This invention designed nine ablation experiments (F1-F9) to systematically verify the contributions of temporal evolution features, physical prior graphs, attribute association graphs, and structural constraints to the prediction performance of the remaining useful life (RUL) of complex devices. The experiments first constructed a baseline group containing only a single dimension (F1 temporal only, F2 hybrid topology only). Then, through a step-by-step stacking approach, the gain effects of fusing single-topology features (F3-F5), dual-topology features (F6-F8), and temporal evolution features were examined. Finally, F9, as the full model, integrates temporal evolution features and hybrid topology features to verify the best performance of the multimodal data-driven model under physical prior guidance, attribute association enhancement, and structural constraint guarantees. The ablation experiment setup and comparison results are shown in Tables 5 and 6, respectively. The visualization of the ablation study performance comparison results is shown in Table 6. Figure 13 As shown.

[0090] Table 5 Ablation Experiment Setup Table 6 Comparison of ablation experiment results (%) For example, this invention utilizes a dynamic weight evolution mechanism to capture the nonlinear degradation of coupling strength of each channel during the process of equipment transitioning from health to failure. This solves the problem of physical and data mismatch caused by the inability of a single static graph to simultaneously consider physical priors and data characteristics, achieving deep synergy between "prior knowledge guidance" and "data-driven learning." A multi-channel graph convolutional neural network is used to train topological features, and finally, temporal evolution features and topological features are fused for complex system modeling and remaining lifetime prediction.

[0091] This invention proposes a spatiotemporal graph neural network-based modeling method for complex systems, incorporating attribute features and hybrid topological representations. It extracts multi-domain hybrid features from vibration, sound, and current time-series signals, constructs a multimodal data sample set with infrared thermal images, and achieves a unified representation of multi-source information. Based on this, a cross-modal attention temporal evolution module is designed. Intramodal feature representations are constructed through a temporal convolutional network, and a parallel dilated convolutional structure is used to form a multi-scale receptive field, effectively coupling high-frequency impact information with low-frequency degradation trends to mine fine-grained intramodal features in multimodal signals. Simultaneously, a cross-modal attention mechanism is introduced to extract intermodal features, achieving dynamic interaction and semantic alignment. Weighted integration is performed through a cross-modal adaptive fusion strategy to alleviate heterogeneous representation conflicts and enhance feature representation capabilities. Furthermore, a hybrid topological representation mechanism is constructed, including a physical prior graph, a global static attribute association graph, and a structurally constrained local dynamic graph. A multi-channel graph convolutional neural network is used to learn the hybrid topological features, achieving synergistic optimization of "prior knowledge guidance" and "data-driven modeling." Finally, temporal evolution features and topological structure features are deeply integrated to establish a unified spatiotemporal collaborative modeling framework. This framework captures high-order spatiotemporal dependencies in complex systems, enabling accurate characterization of equipment degradation processes and prediction of remaining lifetime. Compared to existing multimodal learning methods, the proposed method improves classification accuracy by 0.76-8.98% and F1 score by 1.47-8.87%. Furthermore, the model exhibits superior performance and good generalization ability.

[0092] The above are one or more embodiments of the mechanical equipment remaining life prediction method provided in this specification. Based on the same idea, this specification also provides a corresponding mechanical equipment remaining life prediction device, including: The acquisition module is used to acquire multimodal data corresponding to the mechanical equipment under test; wherein, the multimodal data includes vibration data, sound data, current data and infrared thermal images; The temporal evolution module is used to extract intra-modal features from multimodal data by constructing receptive fields of different dimensions, and to introduce a cross-modal attention mechanism to dynamically capture complementary information between different modalities in multimodal data and determine inter-modal features. Based on intra-modal and inter-modal features, a feature set is constructed. By introducing a dynamic fusion mechanism, the weight allocation of each modality in the feature set is adaptively adjusted to determine the weighted fused temporal evolution features. The hybrid topology module is used to construct a physical prior map based on the physical connections between various sensors in a mechanical device; construct an attribute association map based on the similarity of the temporal evolution features of various sensors in the mechanical device; determine the structural constraint matrix based on the physical prior map and the attribute association map; construct a structurally constrained local dynamic map based on the structural constraint matrix and the dynamic evolution weights determined based on the sensor node features of the mechanical device; extract and concatenate the evolution features corresponding to the physical prior map, the attribute association map, and the structurally constrained local dynamic map to determine the fused spatial features; and perform global aggregation on the fused spatial features to determine the topological structure features. The prediction module is used to nonlinearly concatenate the weighted fusion of temporal evolution features and topological features to construct a spatiotemporal joint feature vector; by mapping the spatiotemporal joint feature vector layer by layer, the predicted value of the remaining life of the mechanical equipment at future moments is obtained.

[0093] Specific limitations regarding the remaining useful life prediction device for mechanical equipment can be found in the limitations of the remaining useful life prediction method for mechanical equipment mentioned above, and will not be repeated here. Each module in the aforementioned remaining useful life prediction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0094] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described method for predicting the remaining life of mechanical equipment.

[0095] The present invention also provides Figure 14 The schematic diagram of the computer device shown is as follows: Figure 14 As shown, at the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the mechanical equipment remaining life prediction method provided in the above embodiments.

[0096] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this invention.

Claims

1. A method for predicting the remaining life of a mechanical device, characterized by, include: Acquire multimodal data corresponding to the mechanical equipment under test; wherein, the multimodal data includes vibration data, sound data, current data and infrared thermal images; By constructing receptive fields of different dimensions, intramodal features are extracted from multimodal data, and a cross-modal attention mechanism is introduced to dynamically capture complementary information between different modalities in multimodal data and determine intermodal features. Based on intramodal and intermodal features, a feature set is constructed. By introducing a dynamic fusion mechanism, the weight allocation of each modal data in the feature set is adaptively adjusted to determine the temporal evolution features after weighted fusion. Based on the physical connections between various sensors in the mechanical equipment, a physical prior graph is constructed. Based on the similarity of the temporal evolution characteristics of each sensor in the mechanical equipment, an attribute association graph is constructed. Based on the physical prior graph and the attribute association graph, a structural constraint matrix is ​​determined. Based on the structural constraint matrix and the dynamic evolution weights determined based on the sensor node characteristics of the mechanical equipment, a structurally constrained local dynamic graph is constructed. The evolutionary features corresponding to the physical prior graph, the attribute association graph, and the structurally constrained local dynamic graph are extracted and concatenated to determine the fused spatial features. The fused spatial features are then globally aggregated to determine the topological structure features. The weighted and fused temporal evolution features and topological features are nonlinearly concatenated to construct a spatiotemporal joint feature vector. By mapping the spatiotemporal joint feature vector layer by layer, the predicted remaining life of the mechanical equipment at future moments is obtained.

2. The method of claim 1, wherein the mechanical device is a bearing. The topological features are determined by constructing a multi-channel graph convolutional neural network (GCN). The GCN includes a feature fusion module and a global aggregation module. The feature fusion module extracts and concatenates the evolutionary features corresponding to the physical prior graph, attribute association graph, and structurally constrained local dynamic graph to determine the fused spatial features. The global aggregation module introduces Chebyshev polynomials to globally aggregate the fused spatial features to determine the topological features.

3. The method of claim 2, wherein the step of determining the remaining life of the mechanical device is performed by a computer. By introducing Chebyshev polynomials to globally aggregate the fused spatial features, topological structure features are determined, specifically including: ； in, Topological features, For the normalized Laplace matrix, It is a Chebyshev polynomial of order j. Here, J represents the learnable coefficients, and J is the order of the polynomial. The spatial characteristics after fusion.

4. The method for predicting the remaining life of mechanical equipment as described in claim 1, characterized in that, The acquisition of the weighted fusion-derived temporal evolution features specifically includes: Capture feature sets using global average pooling The global context is used, and adaptive importance scores for each feature branch are calculated using nonlinear transformations. s i : ； in, W h1 and W h2 These are the weight matrices of the two fully connected layers within the feature fusion module. ; The adaptive importance scores are normalized using the Softmax function to obtain the adaptive fusion weights. a i Specifically, it includes: ； Based on adaptive fusion weights, the fused temporal evolution feature representation is obtained in the following manner: ； in, Here, N is the feature mapping function, and N is the total number of feature components to be fused. H i For the first i There are 1 initial feature components to be fused.

5. The method for predicting the remaining life of mechanical equipment as described in claim 1, characterized in that, The construction of a structurally constrained local dynamic graph based on the structural constraint matrix and the dynamic evolution weights determined by the sensor node characteristics of the mechanical equipment specifically includes: The sensor node features are mapped to a high-dimensional representation space using a shared weight matrix W, and then the correlation energy between the central node and its neighboring nodes is determined by the following formula through a nonlinear transformation. e ij : ； in, h i , h j Represents a node i and nodes j Input features, a || represents a learnable attention vector, and || represents vector concatenation. Let be the transpose of the learnable attention vector a; Based on correlation energy, for neighborhoods satisfying the constraint matrix M... conduct Softmax Normalization is used to determine the dynamic evolution weights: ； The structural constraint matrix M is multiplied element-wise with the dynamic evolution weights, and the structural constraint local dynamic graph is determined using the following formula. : ； in, The dynamic adjacency matrix represents the structure-constrained local dynamic graph at time t. This represents the influence weight of node j on node i at time t. This represents element-wise product.

6. The method for predicting the remaining life of mechanical equipment as described in claim 1, characterized in that, The acquisition of the physical prior graph and attribute association graph specifically includes: Based on the structural connectivity of mechanical equipment, the topological adjacency matrix of the physical prior graph is constructed in the following manner. A p,ij : ； The multi-channel temporal evolution features are set to x respectively. i and x j The attribute similarity of multi-channel temporal evolution features is determined by using cosine similarity metric. S ij : Based on attribute similarity S ij By setting a threshold The adjacency matrix of the attribute association graph is determined as follows: A s,ij : 。 7. The method for predicting the remaining life of mechanical equipment as described in claim 1, characterized in that, The mean square error (MSE) is used as the objective function to minimize the error between the predicted value and the actual lifespan value; min-max normalization is used to standardize the vibration data, sound data, and current data.

8. A device for predicting the remaining life of mechanical equipment, characterized in that, include: The acquisition module is used to acquire multimodal data corresponding to the mechanical equipment under test; wherein, the multimodal data includes vibration data, sound data, current data and infrared thermal images; The temporal evolution module is used to extract intra-modal features from multimodal data by constructing receptive fields of different dimensions, and to introduce a cross-modal attention mechanism to dynamically capture complementary information between different modalities in multimodal data and determine inter-modal features. Based on intra-modal and inter-modal features, a feature set is constructed. By introducing a dynamic fusion mechanism, the weight allocation of each modal data in the feature set is adaptively adjusted to determine the weighted fusion temporal evolution features. The hybrid topology module is used to construct a physical prior map based on the physical connections between various sensors in a mechanical device; construct an attribute association map based on the similarity of the temporal evolution features of various sensors in the mechanical device; determine the structural constraint matrix based on the physical prior map and the attribute association map; construct a structurally constrained local dynamic map based on the structural constraint matrix and the dynamic evolution weights determined based on the sensor node features of the mechanical device; extract and concatenate the evolution features corresponding to the physical prior map, the attribute association map, and the structurally constrained local dynamic map to determine the fused spatial features; and perform global aggregation on the fused spatial features to determine the topological structure features. The prediction module is used to nonlinearly concatenate the weighted fusion of temporal evolution features and topological features to construct a spatiotemporal joint feature vector; by mapping the spatiotemporal joint feature vector layer by layer, the predicted value of the remaining life of the mechanical equipment at future time is obtained.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method for predicting the remaining life of mechanical equipment as described in any one of claims 1-7.

10. A computer device, characterized in that, The device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the remaining life prediction method for mechanical equipment according to any one of claims 1-7.

Images