A method, device, medium and equipment for predicting the remaining life of a rotating machine

Abstract

The application discloses a residual life prediction method and device of a rotating machine, a medium and equipment, relates to the field of residual life prediction, and comprises the following steps: acquiring multi-modal data corresponding to a rotating machine to be measured; wherein the multi-modal data comprises vibration data, sound data, current data and an infrared thermal image; constructing a denoising neural network CAWC-UNet based on a UNet network; performing a denoising enhancement operation on the multi-modal data through the pre-trained denoising neural network CAWC-UNet to obtain enhanced multi-modal data; performing cross-modal fusion on the enhanced multi-modal data to determine fused features, determining corresponding bidirectional state features for representing a dynamic process of equipment degradation according to the fused features; performing pooling and regression prediction on the bidirectional state features to predict the residual life of the rotating machine to be measured and obtain a residual useful life RUL prediction value.

Description

Technical Field

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

[0002] As a core power component of industrial systems, the health status of rotating machinery directly impacts the operational safety and maintenance costs of complex engineering equipment. Remaining Useful Life (RUL) prediction, a crucial component of Prognostics and Health Management (PHM), aims to characterize degradation mechanisms and estimate failure processes by analyzing operational data, thereby achieving predictive maintenance and ensuring engineering reliability. Currently, with the deployment of multi-source sensors, the digitalization of industrial monitoring, and the development of artificial intelligence, multimodal monitoring is gradually replacing single-modal system sensing methods, reflecting equipment operating status from multiple physical perspectives such as vibration, sound, current, and infrared thermometry. The engineering complementarity of different modal data helps to describe the complete evolution process of equipment from minor wear and localized damage to macroscopic degradation from a more comprehensive perspective, significantly improving the reliability and generalization ability of rotating machinery RUL estimation.

[0003] To address the cross-modal fusion problem in multimodal lifetime prediction tasks, existing research typically extracts unique representations of each modality's data to form single-modal features. These representations are then mapped to a unified, shared semantic subspace. Within this subspace, feature-level, model-level, decision-level, or hybrid methods are employed to achieve joint fusion and interactive modeling of multimodal features, thereby uncovering the potential information in multimodal monitoring data. Furthermore, cross-modal fusion methods emphasize collaborative learning and knowledge transfer between modalities, promoting information sharing and improving overall lifetime prediction performance by constructing modal mapping relationships.

[0004] However, existing methods suffer from feature redundancy and semantic noise when dealing with heterogeneous modalities, resulting in high computational costs and a lack of efficient global semantic modeling capabilities during the fusion process. This makes it difficult for the model to capture deep semantic relationships and long-term temporal dependencies between modalities in complex degradation scenarios, thus limiting the stability of the prediction model and making it difficult to guarantee the accuracy of lifetime prediction. Summary of the Invention

[0005] This invention provides a method, apparatus, medium, and device for predicting the remaining life of rotating machinery, to solve the aforementioned problems in the prior art, namely, how to improve the accuracy of remaining life prediction for rotating machinery in the prior art. This invention provides a method for predicting the remaining life of rotating machinery, the method comprising: Acquire multimodal data corresponding to the rotating machinery under test; wherein, the multimodal data includes vibration data, sound data, current data and infrared thermal images; Infrared thermal images are used as target modes, Gaussian noise is added to the target modes, and vibration data, sound data and current data are used as modal features. Based on the acquired RUL task labels and modal features, cross-modal conditional embedding is determined. A denoising neural network, CAWC-UNet, is constructed based on the UNet network. The pre-trained CAWC-UNet is used to denoise and enhance multimodal data, yielding enhanced multimodal data. The CAWC-UNet includes an encoder, a bottleneck layer, a decoder, and a skip connection layer. The encoder, based on multi-layer residual blocks and downsampling operations, extracts multi-scale target modal features from the input target modality after Gaussian noise is added, the time step, and the cross-modal conditional embedding. The bottleneck layer introduces multi-head self-attention (MHA) and cross-modal cross-attention mechanisms to construct the global dependency between the time step, the cross-modal conditional embedding, and the multi-scale target modal features, determining the enhanced features. The decoder upsamples the enhanced features to determine the upsampled features, and then fuses the corresponding layer features of the encoder with the upsampled features of the decoder through the skip connection layer to determine the enhanced multimodal data. The enhanced multimodal data is fused across modes to determine the fused features. Based on the fused features, corresponding bidirectional state features for characterizing the dynamic process of equipment degradation are determined. The remaining service life (RUL) of the rotating machinery under test is predicted by pooling and regression prediction of the bidirectional state features.

[0006] Optionally, adding Gaussian noise to the target mode specifically includes: By gradually moving the infrared thermal image over T time steps Adding a small amount of Gaussian noise, we obtain the sequence. The transition distribution at each step is defined as: ; in, These are pre-defined noise scheduling parameters, representing hyperparameters of the noise variance added at different step sizes t. I It is the identity matrix. It follows a normal distribution; Noise samples at any step size t Depend on Export: ; The noise sample at any step size t is obtained using the following formula. : ; in, Indicates the initial state of the original data. This represents the data state after t steps of noise addition, where T represents the total number of steps in the diffusion process, and t represents the current step number. This represents the total proportion of original information retained from step 1 to step t. This represents standard Gaussian noise.

[0007] Optionally, determining the cross-modal conditional embedding based on the acquired RUL task labels and modal features specifically includes: Based on the acquired RUL task labels and modal features, the condition set is determined using the following formula: ; in, Represents modal features, where y is the label for the RUL prediction task. C For the condition set; Based on the set of conditions, the cross-modal conditional embedding is determined using the following formula: ; ; in, Representing modal features and RUL task labels, For learnable parameter matrix, For bias terms, For cross-modal conditional embedding, M For the total number of modes, For the first j The weight of each condition on the currently generated task. This indicates element-wise multiplication.

[0008] Optionally, an attention fusion bottleneck MBT module and a BiLSTM network are used to perform cross-modal fusion on the enhanced multimodal data; wherein, the attention fusion bottleneck MBT module performs cross-modal fusion on the enhanced multimodal data to determine the fused features; and the BiLSTM network determines the corresponding bidirectional state features for characterizing the dynamic process of device degradation based on the fused features.

[0009] Optionally, based on the acquired noise prediction loss and prediction error, a loss function is determined, and minimizing the loss function is used as the training optimization objective to train the denoising neural network CAWC-UNet; wherein, the acquisition of the loss function specifically includes: The noise prediction loss is obtained using the following formula: ; in, For noise prediction loss, For expectation operator, This is actual Gaussian noise. For the noise predicted by the model, For noisy prediction targets, As the original input for reference, The noisy feature state after processing; The prediction error is obtained using the following formula: ; in, For prediction error, n The total number of samples, i For sample index, Let be the model's predicted output value for the i-th sample. The true label value of the i-th sample; Based on the noise prediction loss and prediction error, the loss function is obtained using the following formula: ; in, L Total It is a loss function. It's a hyperparameter.

[0010] Optionally, the vibration data, sound data, and current data are normalized; the external thermal image features are cropped, resized, and denoised.

[0011] This invention provides a device for predicting the remaining life of rotating machinery, comprising: The acquisition module is used to acquire multimodal data corresponding to the rotating machinery under test; wherein, the multimodal data includes vibration data, sound data, current data and infrared thermal images; The preprocessing module is used to take the infrared thermal image as the target mode, add Gaussian noise to the target mode, and take the vibration data, sound data and current data as modal features. Based on the acquired RUL task label and modal features, the cross-modal conditional embedding is determined. A cross-modal enhancement module is used to construct a denoising neural network CAWC-UNet based on the UNet network. The pre-trained CAWC-UNet is used to denoise and enhance multimodal data, obtaining enhanced multimodal data. The CAWC-UNet includes an encoder, a bottleneck layer, a decoder, and a skip connection layer. The encoder, based on multi-layer residual blocks and downsampling operations, extracts multi-scale target modal features from the input target modality after adding Gaussian noise, the time step, and the cross-modal conditional embedding. The bottleneck layer introduces multi-head self-attention (MHA) and cross-modal cross-attention mechanisms to construct a global dependency between the time step, the cross-modal conditional embedding, and the multi-scale target modal features, determining the enhanced features. The decoder upsamples the enhanced features to determine the upsampled features, and then fuses the corresponding layer features of the encoder with the upsampled features of the decoder through the skip connection layer to determine the enhanced multimodal data. The cross-modal fusion module and prediction module are used to perform cross-modal fusion on the enhanced multimodal data, determine the fused features, and determine the corresponding bidirectional state features to characterize the dynamic process of equipment degradation based on the fused features. By pooling and regression prediction on the bidirectional state features, the remaining service life (RUL) of the rotating machinery under test is predicted to obtain the RUL prediction value.

[0012] 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 rotating machinery.

[0013] 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 rotating machinery.

[0014] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a method for predicting the remaining life of rotating machinery. This method adds Gaussian noise to infrared thermal images when performing cross-modal enhancement on multimodal data, and determines cross-modal conditional embedding based on the acquired RUL task labels and modal features. This enables structural reconstruction and cross-modal enhancement of low-quality modes, thereby achieving low-quality mode enhancement or missing mode completion. This means that even if the target mode has missing or low-quality issues, the model can utilize information from other high-quality modes to complete and enhance the feature representation of the target mode, thereby improving the accuracy of RUL prediction. Furthermore, by fusing the enhanced cross-modal data, efficient cross-modal information flow and feature integration, as well as cross-modal representation learning and temporal modeling capabilities, are achieved, thus enabling accurate prediction of the remaining life of rotating machinery. Attached Figure Description

[0015] 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.

[0016] Figure 1 A flowchart illustrating a method for predicting the remaining life of rotating machinery, provided as an embodiment of the present invention; Figure 2 This is a schematic diagram of low-quality, unbalanced multimodal data provided in an embodiment of the present invention. Figure 3 This is a diagram illustrating the cross-modal enhancement module architecture of the multimodal conditional denoising diffusion probability model provided in an embodiment of the present invention. Figure 4 This is a diagram illustrating the cross-modal adaptive weighted condition constraint architecture provided in an embodiment of the present invention. Figure 5 This is a diagram of the noise prediction network architecture of CAWC-UNet provided in an embodiment of the present invention; Figure 6 The MBT architecture diagram provided in this embodiment of the invention; Figure 7 An architecture diagram of BiLSTM provided for embodiments of the present invention; Figure 8 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 9 This is a flowchart illustrating the construction of degraded data samples provided in an embodiment of the present invention. Figure 10 A visualization of the prediction results of different RUL prediction methods on Dataset-1 provided in this embodiment of the invention; Figure 11 A visualization of the prediction results of different RUL prediction methods on Dataset-2 provided in this embodiment of the invention; Figure 12 This is a schematic diagram of the evaluation results before and after cross-modal enhancement provided in an embodiment of the present invention; Figure 13 Performance visualization of different modal combination experiments provided in Dataset 1 of this invention; Figure 14 Performance visualization of different modal combination experiments provided in this embodiment of the invention; Figure 15 The model performance under different modality missing rates provided in the embodiments of the present invention; Figure 16 A schematic diagram of a computer device for predicting the remaining life of rotating machinery provided in an embodiment of the present invention. Detailed Implementation

[0017] 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.

[0018] 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.

[0019] Example 1 Figure 1 This is a flowchart of a method for predicting the remaining life of rotating machinery provided in an embodiment of the present invention, as shown below. Figure 1 As shown in this embodiment, a method for predicting the remaining life of rotating machinery includes: S1: Acquire multimodal data corresponding to the rotating machinery under test; wherein, the multimodal data includes vibration data, sound data, current data and infrared thermal image.

[0020] For example, a multimodal sequence input sample is defined as The finite non-empty set of all modes of the sample at time t is M To determine the total number of modes, four modes—vibration, sound, current, and infrared thermal imaging—were used for multimodal monitoring and RUL prediction of complex rotating machinery systems. Let represent the sample value of the i-th mode at time t.

[0021] For multimodal 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 ResNet50 is used to extract high-dimensional visual features to capture the semantic structural information of the image. (2); in, The infrared thermal image input represents time t. This represents the feature extraction model after removing the final classification layer, outputting the infrared image feature dimensions. , It is a high-dimensional feature representation of infrared images. It is the feature space. Through the above mapping and representation embedding mechanism, the inputs of different modalities are mapped to a latent feature space of a unified dimension through the modal feature encoder, and the corresponding embedding vectors are used to realize subsequent cross-modal association modeling and fusion and RUL prediction tasks.

[0022] S2: Using infrared thermal images as the target mode, Gaussian noise is added to the target mode, and vibration data, sound data, and current data are used as modal features. Based on the acquired RUL task labels and modal features, cross-modal conditional embedding is determined.

[0023] For example, the forward diffusion process q is a Markov chain in which the diffusion is progressively moved to a lower quality mode over T time steps. (In this embodiment, an infrared thermal image) A small amount of Gaussian noise is added to obtain the sequence. The transition distribution at each step is defined as follows:

[0024] (3); in, These are pre-defined noise scheduling parameters, representing hyperparameters of the noise variance added at different step sizes t. I It is the identity matrix. It follows a normal distribution. Due to the properties of the recursive formula, the noise sample at any step size t... Can be directly from Export:

[0025] (4); Right now (5); in, Indicates the initial state of the original data. This represents the data state after t steps of noise addition, where T represents the total number of steps in the diffusion process, and t represents the current step number. This represents the total proportion of original information retained from step 1 to step t. This represents standard Gaussian noise.

[0026] For example, to alleviate the problem of low-quality imbalance in multimodal data, this invention proposes a cross-modal enhancement module for the multi-modal conditional denoising diffusion probability model (MC-DDPM), the architecture of which is as follows: Figure 3As shown in the diagram, in this module, low-quality modalities are set as the target modalities and used as the initial input sequence for the diffusion process. Gaussian noise is gradually injected through the forward diffusion process, causing the structural information of the target modality to gradually degenerate into an isotropic Gaussian distribution, thus obtaining a potential representation of the unstructured state. During the backward diffusion process, a task-guided and cross-modal adaptive weighting strategy is introduced. A cross-modal conditional embedding is constructed by fusing RUL task labels and other high-quality modal features, and this embedding is used as a conditional constraint input to the denoising neural network CAWC-UNet. This allows the model to perform structured recovery based on conditional information, rather than unconditionally generating data from random noise.

[0027] This process effectively reduces interference from irrelevant noise, enabling knowledge transfer and semantic consistency completion between modalities. During training, the network minimizes the mean square error between predicted and actual noise, gradually approximating the true modal representation in the reconstruction process. It reconstructs and enhances the structural information of the target modality in the latent space, thereby providing a more complete, reliable, and semantically consistent modal representation for subsequent feature fusion networks.

[0028] During the backdiffusion process, the model introduces task guidance and cross-modal adaptive weighting strategies from Gaussian noise. It combines the RUL task label and other high-quality auxiliary modalities besides the target modality to construct cross-modal adaptive weighting conditional constraints, the architecture of which is as follows: Figure 4 As shown. The condition set C is defined as:

[0029] (6); in, Represents a set of conditions. Let y represent other high-quality auxiliary modalities besides the target modality, and y be the label for the RUL prediction task. The conditional embedding of the cross-modal adaptive weighting mechanism is represented as:

[0030] (7); (8); in, Representing modal features and RUL task labels, For learnable parameter matrix, For bias terms, For cross-modal conditional embedding, M is the total number of modes. Let be the weight of the j-th condition for the current generated task. This indicates element-wise multiplication.

[0031] Its conditional transition distribution is modeled as follows: (9); The Gaussian form for each step is: (10); in, The mean is a Gaussian distribution. is the standard deviation of the Gaussian distribution.

[0032] S3: Based on the UNet network, a denoising neural network CAWC-UNet is constructed. The pre-trained CAWC-UNet is used to denoise and enhance multimodal data, obtaining the enhanced multimodal data. The CAWC-UNet includes an encoder, a bottleneck layer, a decoder, and a skip connection layer. The encoder, based on multi-layer residual blocks and downsampling operations, extracts multi-scale target modal features from the input target modality after adding Gaussian noise, the time step, and the cross-modal conditional embedding. The bottleneck layer introduces multi-head self-attention (MHA) and cross-modal cross-attention mechanisms to construct the global dependency between the time step, the cross-modal conditional embedding, and the multi-scale target modal features, determining the enhanced features. The decoder upsamples the enhanced features to determine the upsampled features, and then fuses the corresponding layer features of the encoder with the upsampled features of the decoder through the skip connection layer to determine the enhanced multimodal data.

[0033] Exemplary noise prediction network It is used to learn to enhance and complete target modal information from Gaussian noise and CAWC inputs. CAWC-UNet is the core module of the denoising network, mainly composed of an encoder, a bottleneck layer, and a decoder, and its architecture is as follows: Figure 5 As shown, the encoder progressively embeds the target mode, time step, and cross-modal conditions of the input after adding Gaussian noise, extracting multi-scale target modal features through multiple residual blocks and downsampling. A sine-cosine positional encoding method is used to construct the time step embedding. :

[0034] (11); in, This is the original time step encoding vector. , For embedded dimensions, and It is a periodic function; this periodic embedding allows the model to capture the relative position information of different numbers during the diffusion process.

[0035] To enhance expressive ability, Feed it into a two-layer fully connected network with an activation function: (12); in, This indicates the final diffusion time step embedding. and This represents the weight matrix of the fully connected layer. and This represents the bias term of the fully connected layer. The GELU activation function is used to enhance nonlinear expression, resulting in a smoother final diffusion time-step embedding (SE) that is more sensitive to changes in noise level. The network can dynamically adjust the denoising intensity according to the current noise level at different diffusion stages.

[0036] The bottleneck layer introduces multi-head attention (MHA) and cross-modal attention. The MHA mechanism considers different attention distributions and enables the model to capture information from multiple viewpoints. It establishes global dependencies between time steps, cross-modal conditional embeddings, and the target modality.

[0037] (13); in Let be the dimension of the key vector. , , In this way, the bottleneck layer can selectively inject conditional constraint information into the features of the target modality, that is, let the network "learn to judge" at what time step, at what spatial location, and at what semantic scale, which modality's conditional information needs to be used to influence the representation of the target modality, while other information is weakened or ignored.

[0038] The decoder progressively upsamples and generates a denoised signal, then fuses encoder features via skip connections to achieve fine-grained reconstruction. Furthermore, mean As a noise function, it can be utilized by a trained neural network. Redefining:

[0039] (14); Variance is redefined as: (15); By indirectly predicting the conditional posterior distribution in formula (10) ,get: (16); in, Indicates noise scheduling parameters, This represents the noise variance hyperparameter. From step 1 to step t The cumulative product of values.

[0040] Through the above iterations, the complete target mode is finally reconstructed, thereby achieving effective enhancement of the low-quality mode.

[0041] S4: Perform cross-modal fusion on the enhanced multimodal data to determine the fused features. Based on the fused features, determine the corresponding bidirectional state features used to characterize the dynamic process of equipment degradation. By pooling and regression prediction on the bidirectional state features, predict the remaining service life (RUL) of the rotating machinery under test and obtain the RUL prediction value.

[0042] Optionally, the cross-modal fusion module includes an attention fusion bottleneck MBT module and a BiLSTM network. The attention fusion bottleneck MBT module performs cross-modal fusion on the enhanced multimodal data to determine the fused features. The BiLSTM network then determines the corresponding bidirectional state features used to characterize the dynamic process of device degradation based on the fused features.

[0043] For example, to model the deep correlation and semantic interaction between multimodal information such as vibration, sound, current, and infrared thermal images, this invention proposes a novel network architecture that integrates MBT and BiLSTM as a cross-modal fusion module. The MBT architecture uses the "attention fusion bottleneck" to achieve cross-modal fusion and interaction at multiple levels, enabling temporal features and image features to be stitched together through cross-modal fusion.

[0044] Compared to traditional self-attention fusion mechanisms, this model forces information between different modalities to be transmitted through a small number of bottleneck token latent variables. It requires the model to organize and aggregate relevant information in each modality and share necessary correlation information between modalities. This strategy improves fusion performance while effectively reducing computational complexity and storage overhead, achieving efficient cross-modal information flow and feature integration. In the later stages of the fusion phase, the resulting unified modal representation is input into a BiLSTM network. BiLSTM, through a bidirectional structure, utilizes both past and future information to model the global dependencies of the degradation process, capturing the dynamic dependencies of the equipment's degradation state over time. Finally, pooling layers and multilayer perceptron layers are used to regress the fused temporal features, achieving high-precision prediction of the remaining life of rotating machinery.

[0045] Overall, the proposed cross-modal fusion module fully combines the cross-modal global modeling capability of MBT with the temporal memory advantage of BiLSTM, enabling collaborative information aggregation in both modal and temporal dimensions. This enhances the model's expressive power, generalization ability, and robustness towards multimodal data, thereby optimizing the RUL prediction performance of rotating machinery.

[0046] An exemplary cross-modal bottleneck mechanism architecture diagram is shown below. Figure 6 As shown. Specifically, assume the feature sequences of the four modalities are... , , , , in , , , These represent the number of tokens for vibration, sound, current, and infrared image modes, respectively. For feature dimensions.

[0047] To capture pairwise interactions between modalities and avoid the quadratic complexity of pairwise attention, six bottleneck tokens are introduced for the four modalities: This introduces B fusion bottlenecks. The subscripts va, vc, vr, ac, ar, and cr correspond to the modal interactions of (vibration, sound), (vibration, current), (vibration, infrared), (sound, current), (sound, infrared), and (current, infrared), respectively. Each bottleneck token sequence has a length of B and is used to compress and transmit potential correlation information between two modes. The bottleneck token sequence is defined as follows:

[0048] (17); Among them, Z va Represents the bottleneck token sequence. Z va Each specific bottleneck token in the process; The other five bottleneck tokens are defined similarly. Therefore, the overall input sequence can be represented as:

[0049] (18); Where Z represents the total input sequence of the Transformer layer, and || represents sequence concatenation.

[0050] Within the Transformer layer, modal features interact across modalities via bottleneck tokens, restricting all cross-modal attention paths in the model. For example, for the l-th layer, the token computation process for the interaction between vibration and sound modes can be formalized as follows:

[0051] (19); (20); in, , , These represent the vibration, sound, and bottleneck token of the l-th layer, respectively. , This is the set of parameters for this modal interaction. and Only through bottlenecks Exchange information and This represents the updated vibration characteristics and bottleneck characteristics of intermediate states. Bottleneck token. The update was performed twice: first, the vibration mode information was updated (Formula 19), and then the sound mode information was updated (Formula 20). Simultaneously, this invention controls the total number of bottleneck tokens to be much smaller than the total number of potential units for each mode. and This mechanism ensures that all cross-modal attention flows pass through these units, forcing the model to compress redundant features within each modality and share only the information necessary for the task. This reduces computational complexity while maintaining or even enhancing cross-modal fusion performance. Similarly, the interactions between the remaining five modal pairs can be modeled and propagated using corresponding bottleneck tokens.

[0052] For example, BiLSTM includes two independent computation branches: forward and backward. The forward layer performs forward computation at each step and stores the output of the forward hidden layer. The backward layer performs computation and saves the output of the backward hidden layer at each step. The architecture of BiLSTM is as follows: Figure 7 As shown, by layering and stacking bottleneck interactions, high-order cross-modal correlation features between vibration, sound, current, and infrared images are gradually captured. In the final fusion stage, this invention only retains the updated modal token sequence:

[0053] (twenty one); in, This represents the final fused feature sequence. l The number of layers in the Transformer is stacked, and the bottleneck token acts as an interaction intermediary, not directly entering the final prediction layer, thus ensuring that the fused representation is compact and task-relevant.

[0054] Subsequently, the fused bottleneck token time series is input into BiLSTM to capture the time dependency structure during device degradation: (twenty two); (twenty three); in, This represents the output of the forward hidden layer at time t. This represents the output of the backward hidden layer at time t. and represents the activation functions for the forward and backward directions, respectively, and w1, w2, w3, and w4 represent the weight parameters inside the BiLSTM.

[0055] By concatenating the outputs of the forward and backward layers, bidirectional state features are obtained: (twenty four); in, This represents the final output feature of the BiLSTM hidden layer. The corresponding activation function is represented by w5 and w6, which represent the weights used to splice / fuse the bidirectional states. and These are the outputs of the forward and backward LSTMs, respectively. It is the final output of the hidden layer; That is the corresponding activation function.

[0056] Finally, pooling and regression prediction are performed on the feature sequences: (25); in, For average pooling polymerization operation, This is the regression prediction head for the multilayer perceptron, used to output the final RUL estimate.

[0057] S4: Pooling and regression prediction of bidirectional state features are performed through pooling layers and multi-layer perceptron layers to predict the remaining life of the rotating machinery under test and obtain the predicted value of the remaining life (RUL).

[0058] For example, this invention employs a composite loss function comprising two core components for end-to-end training. Each component is designed to facilitate efficient representation learning under conditions of low-quality imbalanced multimodality and enhance the robustness and accuracy of the RUL prediction task. Specifically, it includes: (1) Modal reconstruction loss. In order to incorporate the conditions into the denoising model, the conditional information is included in the optimization objective. In the multimodal conditional backdiffusion process, the training objective is to minimize the noise estimation loss. L MCDDPM The acquisition of, specifically includes:

[0059] (26); in, For noise prediction loss, For expectation operator, This is actual Gaussian noise. For the noise predicted by the model, For noisy prediction targets, As the original input for reference, The noisy feature state after processing; By minimizing noise prediction loss L MCDDPMIt can be trained to obtain a diffusion model that can reconstruct missing or low-quality modalities under multimodal conditions, ensuring that the model gradually recovers the target modality under given conditions, and providing high-quality modal representations for subsequent cross-modal fusion and RUL prediction.

[0060] (2) RUL Prediction Loss. The mean squared error loss is used to estimate the error between the predicted RUL and the true RUL, based on the MSE loss function. L RUL Defined as:

[0061] (27); in, Let n be the prediction error, n be the total number of samples, and i be the sample index. Let be the model's predicted output value for the i-th sample. The true label value of the i-th sample; In summary, the training optimization objective of the network proposed in this invention is to minimize the combination of reconstruction loss and prediction error during training, which facilitates RUL prediction on low-quality imbalanced multimodal data. The loss function can be expressed as: (28); in, L Total It is the total training loss (loss function) of the model. It is a hyperparameter used to balance the weights of mode reconstruction and prediction tasks.

[0062] Example 2 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 8 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.

[0063] Table 1 Data acquisition channels and signal types of the reducer The experiment collected full-life degradation data for three gear reducers under three typical operating conditions: Condition 1 at 30 Hz / 195 N, Condition 2 at 35 Hz / 180 N, and Condition 3 at 40 Hz / 165 N. The sampling frequency was 20480 Hz, 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 gear reducer from a healthy state to a failure state. The data sets of the gear reducers under different operating conditions are described in Table 2.

[0064] Table 2. Dataset Description of Gear Reducers under Different Operating Conditions 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, 51 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:

[0065] (29); in, , and The values ​​are set to 0.4, 0.3, and 0.3 respectively. Corr is the correlation metric, Mon is the monotonicity metric, and Rob is the robustness metric. The ranges of the three eigenvalues ​​are... .

[0066] 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, and its sample structure diagram is shown below. Figure 9 As shown.

[0067] To comprehensively verify the RUL prediction performance of the proposed method under low-quality, imbalanced multimodal data conditions, this section conducts test experiments on rotating machinery degradation datasets under two typical operating conditions. First, a systematic comparative experiment is performed with existing mainstream RUL prediction methods to evaluate the prediction performance of the proposed method under different operating conditions. Second, for the cross-modal enhancement module, the feature distribution and model performance changes before and after enhancement are analyzed, and various combination experiments (single-modal, bimodal, trimodal, and quadruple-modal) are designed to verify the robustness and generalization ability of the model under different modal combinations. Simultaneously, the impact of the core module on the overall performance is quantitatively evaluated through ablation experiments, and scenarios with different modal missing rates are constructed to examine the performance retention ability of the proposed model under low-quality, imbalanced modalities and modal information missing conditions. This section uses infrared images as the low-quality modality to conduct cross-modal enhancement and RUL prediction verification, and comprehensively evaluates the performance of the proposed method and the comparative methods from the perspectives of prediction accuracy, consistency, and error control. The experimental comparison results of different RUL prediction methods are shown in Table 3.

[0068] Table 3. Experimental comparison results of different RUL prediction methods In Table 3, BA (BiGRU-Attention): BiGRU captures the dependencies between time series, and the attention mechanism focuses on key moment features, thereby improving the effectiveness of feature extraction and prediction accuracy; TCB (Transformer-CNN-BiGRU): A hybrid framework integrating Transformer, CNN, and BiGRU. Transformer extracts global dependencies, CNN captures local patterns, BiGRU models temporal information, and uses the Grey Wolf Optimization (GWO) algorithm to automatically adjust hyperparameters to improve performance; BT (BiLSTM-Transformer): A hybrid model combining bidirectional LSTM and Transformer. BiLSTM excels at modeling bidirectional dependencies in time series, and Transformer introduces a self-attention mechanism to capture long-range dependencies and feature interactions; CTB: A multi-layer structure integrating CNN, Transformer, and BiLSTM. CNN extracts local spatial features, Transformer models global dependencies, and BiLSTM further learns temporal dynamics, forming an efficient integrated feature representation framework; MHT (MultiscaleHourglass-Transformer): Proposes a multi-scale hourglass Transformer network (MHT) for RUL prediction. Multi-timescale feature fusion is achieved through a one-dimensional convolutional hourglass structure, followed by the capture of cross-scale global features by a Transformer and pyramid self-attention mechanism; TCNT: The TCN transformer combines a Temporal Convolutional Network (TCN) with a Transformer. TCN is used to extract stable temporal features, while the Transformer enhances global context awareness through an attention mechanism, thereby improving the modeling ability for complex sequence tasks; ATCN (Attention-based Temporal Convolutional Network): An improved self-attention mechanism is combined with TCN. The attention module strengthens time-step features highly correlated with RUL, TCN captures long-term dependencies, and optimizes feature weights through channel attention (SE mechanism).

[0069] like Figure 10 and Figure 11 As shown, Figure 10 (a) to (h) and Figure 11Figures (a) to (h) show the prediction results of eight different RUL prediction methods on Dataset 1 and Dataset 2, respectively, further validating their performance from the perspectives of predicted trajectories and confidence intervals. The yellow solid line represents the actual RUL trajectory, the blue solid line represents the model's predicted value, the blue shaded area corresponds to the 95% confidence interval of the predicted value, and the red bar residuals at the bottom reflect the error between the predicted and actual values. Overall, BA, BT, TCNT, and ATCN all exhibit significant prediction bias at different stages of operation, with large deviations between their prediction curves and the actual values, and wide error fluctuation ranges, indicating insufficient ability to handle low-quality and modally imbalanced data input. In contrast, TCB, CTB, and MHT models improve fitting ability to some extent, but error accumulation still occurs in the later decay stage of the RUL, and the confidence intervals expand irregularly in local stages, indicating insufficient learning ability for nonlinear degradation trends.

[0070] Compared with the methods mentioned above, the method proposed in this invention exhibits the best prediction stability and accuracy. The predicted curve closely matches the actual RUL trajectory, the residual amplitude is significantly reduced, and the 95% confidence interval remains smoothly converged throughout the entire operating cycle, demonstrating the model's enhanced robustness and stronger generalization ability to low-quality, imbalanced multimodal input data. Furthermore, the model can more sensitively capture changes in degradation trends during equipment operation, avoiding prediction drift and response hysteresis, thus demonstrating excellent dynamic response capability and prediction performance in full lifecycle RUL estimation. In summary, the experimental results fully verify the effectiveness and engineering potential of the proposed method in modeling complex degradation mechanisms and fusing low-quality, imbalanced multimodal information in RUL prediction tasks.

[0071] like Figure 12 As shown, this illustrates the changes in feature distribution and performance metrics of the model on different work condition datasets before and after cross-modal feature enhancement. For dataset 1 and dataset 2, respectively... Figure 12 (a) to Figure 12 (c) and Figure 12 (d) to Figure 12 (f). For the feature representation before enhancement ( Figure 12 (a) and Figure 12 As can be seen from (d), samples of different health states form multiple isolated clusters in the feature space. Although there is some discriminatory power among the samples, there is a lack of continuous evolutionary trajectories and global consistency; samples of different remaining lifespan stages are discretely scattered in space. This indicates that under unenhanced conditions, multimodal features still suffer from insufficient alignment and semantic gaps, making it difficult for the model to characterize the gradual evolution of the device during the real degradation process at the feature level, and limiting the RUL representation capability. In contrast, the enhanced feature space distribution ( Figure 12 (b) and Figure 12 (e) shows a more significant improvement. The sample distribution gradually converges into a continuous degradation trajectory with a manifold structure, exhibiting a smooth and gradual evolutionary trend with changes in remaining lifetime. This indicates that the enhancement module effectively improves the alignment and representation capabilities between different modal features, enabling the model to capture the implicit common structure of the degradation process, making the feature space more realistically reflect continuous degradation behavior and lifetime evolution laws, and forming a more interpretive representation of the real degradation mechanism. Performance index comparison bar chart ( Figure 12 (c) and Figure 12 (f) further validated the effectiveness of the enhancement module. To facilitate the visualization of the four types of indicators, 1−R² was used instead of R² to draw a three-dimensional bar chart. The results show that the enhanced model's R² is significantly improved, while RMSE, MAPE, and MAE all decrease significantly. The prediction error distribution is more concentrated and stable, reflecting enhanced model fitting ability and reduced error fluctuation. This indicates that cross-modal enhancement not only improves the quality of feature representation but also significantly improves prediction accuracy and robustness. In summary, the cross-modal feature enhancement module achieves significant improvements in feature structure continuity, modal information consistency, and prediction reliability, effectively promoting the model's ability to represent the real degradation mechanism, thereby improving the overall performance of rotating machinery remaining life prediction.

[0072] To evaluate the impact of different modalities on RUL prediction performance, this invention designed multiple sets of modal combination experiments involving four modalities: vibration, sound, current, and infrared thermal imaging. Specifically, this invention constructed 14 experimental combinations, including single-modal, dual-modal, tri-modal, and quadruple-modal experiments, to systematically analyze the complementarity between different modalities and whether increasing the number of modalities can continuously improve model performance. Among them, M1–M4 are single-modal experiments used to measure the independent contribution of each modality to the prediction task; M5–M10 are dual-modal combinations used to observe the synergistic effect between different modalities; M11–M13 are tri-modal fusion experiments used to explore the high-dimensional correlation expression capability of multimodal models; and M14 is a quadruple-modal complete input used to verify whether full-modal fusion can achieve optimal prediction performance. The experimental setup is shown in Table 4.

[0073] Table 4 Modal Combination Settings and Experiment ID Figure 13 (a) and Figure 14 (a) shows the comparative visualization results of the eight models in datasets 1 and 2 on four metrics: 1-R², RMSE, MAPE, and MAE. Figure 13 (b) and Figure 14(b) shows the performance metrics of the models for Dataset 1 and Dataset 2 proposed in this invention under different modality combinations. The overall trend indicates that, compared to prediction schemes based on a single modality, the model exhibits stronger robustness and prediction consistency after introducing multimodal data fusion. In the experimental comparison between Dataset-1 and Dataset-2, as the number of modalities gradually increases, 1−R², RMSE, MAPE, and MAE all show a significant decreasing trend, indicating that multimodal fusion can effectively improve prediction accuracy and reduce errors. Furthermore, compared to existing modality fusion methods, the cross-modal enhancement module proposed in this invention significantly improves feature alignment and completion capabilities under conditions of modality loss and noise interference, enabling the model to output more stable and reliable prediction results even under different operating conditions. It should be noted that in some specific modality combinations, the improvement in prediction performance is limited, and even slight fluctuations may occur. Further analysis revealed that this is mainly due to noise accumulation and redundancy in the data acquisition process of rotating machinery sensors. For example, acoustic modes may have a low signal-to-noise ratio at specific speeds, and some modes may have overlapping features, leading to insufficient fusion gain. Even so, under most experimental configurations, multimodal fusion still significantly outperforms single-modal prediction, and with the assistance of the cross-modal enhancement module, the prediction performance maintains a steady upward trend. Overall, the experimental results verify the effectiveness and reliability of the proposed method under real-world operating conditions. The cross-modal enhancement module, combined with a conditional adaptive weighting strategy, not only alleviates the performance degradation caused by low-quality modes or missing modes but also improves the model's ability to uniformly express information from different modes, thus providing a more robust and reliable technical path for industrial equipment life prediction tasks.

[0074] This invention addresses the problem of low-quality, imbalanced multimodal data encountered by rotating machinery in complex operating environments. Based on vibration, sound, current, and infrared thermal modes, it constructs an interpretable RUL prediction method. The cross-modal enhancement module introduces MC-DDPM, combining task guidance and adaptive weighting strategies to construct conditional constraints during back-diffusion, achieving structural reconstruction and cross-modal enhancement of low-quality modes, thus alleviating the problem of low-quality, imbalanced modes. The cross-modal fusion module combines MBT and BiLSTM, taking into account dependencies and collaborative learning in both modal and temporal dimensions, achieving efficient cross-modal information flow and feature integration, as well as cross-modal representation learning and temporal modeling capabilities. Experimental results on a real-world rotating machinery lifecycle degradation dataset demonstrate that, compared to existing mainstream RUL prediction methods, the proposed method achieves a 0.43–3.30% improvement in R², a 18.01–63.89% reduction in RMSE, a 19.38–78.27% reduction in MAPE, and a 25.49–65.85% reduction in MAE. This method not only significantly improves the accuracy of RUL prediction but also exhibits stronger robustness and generalization ability. Furthermore, it demonstrates higher reliability and practicality in cross-modal enhancement evaluation, experiments with different modal combinations, interpretability analysis of modal global contributions, and experiments with missing modalities, effectively adapting to the prediction needs of complex industrial environments.

[0075] The above describes one or more embodiments of a method for predicting the remaining life of rotating machinery. Based on the same idea, this specification also provides a corresponding device for predicting the remaining life of rotating machinery, including: The acquisition module is used to acquire multimodal data corresponding to the rotating machinery under test; wherein, the multimodal data includes vibration data, sound data, current data and infrared thermal images; The preprocessing module is used to take the infrared thermal image as the target mode, add Gaussian noise to the target mode, and take the vibration data, sound data and current data as modal features. Based on the acquired RUL task label and modal features, the cross-modal conditional embedding is determined. A cross-modal enhancement module is used to construct a denoising neural network CAWC-UNet based on the UNet network. The pre-trained CAWC-UNet is used to denoise and enhance multimodal data, obtaining enhanced multimodal data. The CAWC-UNet includes an encoder, a bottleneck layer, a decoder, and a skip connection layer. The encoder, based on multi-layer residual blocks and downsampling operations, extracts multi-scale target modal features from the input target modality after adding Gaussian noise, the time step, and the cross-modal conditional embedding. The bottleneck layer introduces multi-head self-attention (MHA) and cross-modal cross-attention mechanisms to construct a global dependency between the time step, the cross-modal conditional embedding, and the multi-scale target modal features, determining the enhanced features. The decoder upsamples the enhanced features to determine the upsampled features, and then fuses the corresponding layer features of the encoder with the upsampled features of the decoder through the skip connection layer to determine the enhanced multimodal data. The cross-modal fusion module and prediction module are used to perform cross-modal fusion on the enhanced multimodal data, determine the fused features, and determine the corresponding bidirectional state features to characterize the dynamic process of equipment degradation based on the fused features. By pooling and regression prediction on the bidirectional state features, the remaining service life (RUL) of the rotating machinery under test is predicted to obtain the RUL prediction value.

[0076] Specific limitations regarding the remaining life prediction device for rotating machinery can be found in the limitations of the remaining life prediction method for rotating machinery described above, and will not be repeated here. Each module in the aforementioned remaining life prediction device for rotating machinery 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 corresponding operations of each module.

[0077] 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 rotating machinery.

[0078] The present invention also provides Figure 16 The schematic diagram of the computer device shown is as follows: Figure 16 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 remaining life prediction method for rotating machinery provided in the above embodiments.

[0079] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0080] 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 rotating machinery, characterized in that, include: Acquire multimodal data corresponding to the rotating machinery under test; wherein, the multimodal data includes vibration data, sound data, current data and infrared thermal images; Infrared thermal images are used as target modes, Gaussian noise is added to the target modes, and vibration data, sound data and current data are used as modal features. Based on the acquired RUL task labels and modal features, cross-modal conditional embedding is determined. A denoising neural network, CAWC-UNet, is constructed based on the UNet network. The pre-trained CAWC-UNet is used to denoise and enhance multimodal data, yielding enhanced multimodal data. The CAWC-UNet includes an encoder, a bottleneck layer, a decoder, and a skip connection layer. The encoder, based on multi-layer residual blocks and downsampling operations, extracts multi-scale target modal features from the input target modality after Gaussian noise is added, the time step, and the cross-modal conditional embedding. The bottleneck layer introduces multi-head self-attention (MHA) and cross-modal cross-attention mechanisms to construct the global dependency between the time step, the cross-modal conditional embedding, and the multi-scale target modal features, determining the enhanced features. The decoder upsamples the enhanced features to determine the upsampled features, and then fuses the corresponding layer features of the encoder with the upsampled features of the decoder through the skip connection layer to determine the enhanced multimodal data. The enhanced multimodal data is fused across modes to determine the fused features. Based on the fused features, corresponding bidirectional state features for characterizing the dynamic process of equipment degradation are determined. The remaining service life (RUL) of the rotating machinery under test is predicted by pooling and regression prediction of the bidirectional state features.

2. The method for predicting the remaining life of rotating machinery as described in claim 1, characterized in that, The addition of Gaussian noise to the target mode specifically includes: By gradually moving the infrared thermal image over T time steps Adding a small amount of Gaussian noise, we obtain the sequence. The transition distribution at each step is defined as: ； in, These are pre-defined noise scheduling parameters, representing hyperparameters of the noise variance added at different step sizes t. I It is the identity matrix. It follows a normal distribution; Noise samples at any step size t Depend on Export: ； The noise sample at any step size t is obtained using the following formula. : ； in, Indicates the initial state of the original data. This represents the data state after t steps of noise addition, where T represents the total number of steps in the diffusion process, and t represents the current step number. This represents the total proportion of original information retained from step 1 to step t. This represents standard Gaussian noise.

3. The method for predicting the remaining life of rotating machinery as described in claim 1, characterized in that, The determination of cross-modal conditional embedding based on the acquired RUL task labels and modal features specifically includes: Based on the acquired RUL task labels and modal features, the condition set is determined using the following formula: ； in, Represents modal features, where y is the label for the RUL prediction task. C For the condition set; Based on the set of conditions, the cross-modal conditional embedding is determined using the following formula: ； ； in, Representing modal features and RUL task labels, For learnable parameter matrix, For bias terms, For cross-modal conditional embedding, M The total number of modes, For the first j The weight of each condition on the currently generated task. This indicates element-wise multiplication.

4. The method for predicting the remaining life of rotating machinery as described in claim 1, characterized in that, The enhanced multimodal data is fused across modally using an attention fusion bottleneck MBT module and a BiLSTM network. Specifically, the attention fusion bottleneck MBT module performs cross-modal fusion on the enhanced multimodal data to determine the fused features. The BiLSTM network determines corresponding bidirectional state features to characterize the dynamic process of device degradation based on the fused features.

5. The method for predicting the remaining life of rotating machinery as described in claim 1, characterized in that, Based on the acquired noise prediction loss and prediction error, a loss function is determined, and minimizing the loss function is used as the training optimization objective to train the denoising neural network CAWC-UNet; wherein, the acquisition of the loss function specifically includes: The noise prediction loss is obtained using the following formula: ； in, For noise prediction loss, For expectation operator, This is actual Gaussian noise. For the noise predicted by the model, For noisy prediction targets, As the original input for reference, The noisy feature state after processing; The prediction error is obtained using the following formula: ; in, For prediction error, n The total number of samples, i For sample index, Let be the model's predicted output value for the i-th sample. The true label value of the i-th sample; Based on the noise prediction loss and prediction error, the loss function is obtained using the following formula: ； in, L Total It is a loss function. It's a hyperparameter.

6. The method for predicting the remaining life of rotating machinery as described in claim 1, characterized in that, The vibration data, sound data, and current data are normalized; the external thermal image features are cropped, resized, and denoised.

7. A device for predicting the remaining life of rotating machinery, characterized in that, include: The acquisition module is used to acquire multimodal data corresponding to the rotating machinery under test; wherein, the multimodal data includes vibration data, sound data, current data and infrared thermal images; The preprocessing module is used to take the infrared thermal image as the target mode, add Gaussian noise to the target mode, and take the vibration data, sound data and current data as modal features. Based on the acquired RUL task label and modal features, the cross-modal conditional embedding is determined. A cross-modal enhancement module is used to construct a denoising neural network CAWC-UNet based on the UNet network. The pre-trained CAWC-UNet is used to denoise and enhance multimodal data, obtaining enhanced multimodal data. The CAWC-UNet includes an encoder, a bottleneck layer, a decoder, and a skip connection layer. The encoder, based on multi-layer residual blocks and downsampling operations, extracts multi-scale target modal features from the input target modality after adding Gaussian noise, the time step, and the cross-modal conditional embedding. The bottleneck layer introduces multi-head self-attention (MHA) and cross-modal cross-attention mechanisms to construct a global dependency between the time step, the cross-modal conditional embedding, and the multi-scale target modal features, determining the enhanced features. The decoder upsamples the enhanced features to determine the upsampled features, and then fuses the corresponding layer features of the encoder with the upsampled features of the decoder through the skip connection layer to determine the enhanced multimodal data. The cross-modal fusion module and prediction module are used to perform cross-modal fusion on the enhanced multimodal data, determine the fused features, and determine the corresponding bidirectional state features to characterize the dynamic process of equipment degradation based on the fused features. By pooling and regression prediction on the bidirectional state features, the remaining service life (RUL) of the rotating machinery under test is predicted to obtain the RUL prediction value.

8. 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 rotating machinery as described in any one of claims 1-6.

9. A computer device, characterized in that, It 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 rotating machinery according to any one of claims 1-6.

Images