Mechanical health state degradation trend prediction method and system of time sequence implicit diffusion model
By designing a temporal implicit diffusion model and utilizing a bidirectional gated cyclic unit and a vector quantization variational autoencoder model, the problem of predicting the health status degradation trend under limited data conditions was solved, achieving efficient prediction of the health status of mechanical equipment and improving the model's temporal dependence and prediction accuracy.
Patent Information
- Application Number
- CN202511445207.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-11
- Publication Date
- 2026-02-13
AI Technical Summary
Existing deep learning-based methods for predicting health status degradation trends suffer from overfitting and insufficient generalization when faced with limited available data in real-world industrial scenarios. Furthermore, diffusion models are computationally burdensome, fail to capture sufficient temporal dependency information, and have poor hard constraints.
A temporal implicit diffusion model is designed, which maps backtracking window data to a low-dimensional temporal hidden vector through a bidirectional gated cyclic unit. Combining a vector quantization variational autoencoder model and a diffusion model, probabilistic conditional constraints are introduced to predict future degradation trends.
It improves the model's ability to capture temporal dependencies and predict accuracy, and can predict future health status degradation trends using only historical monitoring data of mechanical equipment, reducing computational burden and enhancing the model's generalization performance.
Smart Images
Figure CN121524618A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of neural networks, data processing, and mechanical equipment health state prediction, and particularly relates to a mechanical health state degradation trend prediction method and system based on a time series implicit diffusion model. BACKGROUND
[0002] With the promotion of the construction of unmanned intelligent factories, the rapid development of artificial intelligence and industrial big data is deeply changing the operation and maintenance mode of manufacturing industry. Traditional periodic maintenance strategies have been difficult to meet the real-time, high-precision and personalized needs. And the deep learning driven prognostics and health management (PHM) technology provides an effective solution. As one of the most critical tasks in PHM, implementing a reasonable predictive maintenance (PM) strategy not only can prolong the service life of equipment, but also can guarantee the production continuity and economic benefits.
[0003] At present, the remaining useful life prediction based on life label is the most widely used method in industrial PM, because it can directly predict the remaining life of equipment through a deep model. However, in many actual industrial scenarios, it is often unrealistic to obtain complete life cycle data and perform RUL labeling. For example, some equipment may exist in a single form, and without full life cycle data, RUL prediction cannot be implemented. At present, the degradation trend forecasting (DTF) technology is a feasible solution, which only needs to be based on historical monitoring data, and with the help of mainstream time series prediction models, the degradation trend in the future period can be predicted. Although existing DTF-oriented research has achieved remarkable results, when facing limited available degradation data in actual industrial scenarios, traditional deep time series models may still have problems of overfitting and insufficient generalization The generative modeling technique provides a direct and effective solution to the data scarcity problem due to its strong feature distribution reconstruction capability. Among them, the generative adversarial network has been widely used to alleviate the plight of data shortage, but unfortunately, the performance of the generative adversarial network is often restricted by the mode collapse problem caused by the adversarial mechanism and the selection of the regularizer. To break through these limitations, a diffusion model with stronger training stability and generalization is introduced, however, the existing generation scheme based on the diffusion model is still limited to improving the generalization performance of the model through data augmentation, and these methods fail to effectively capture the inherent time series dynamic characteristics of the data. Therefore, the present application redesigns the diffusion model learning paradigm for DTF. Specifically, in the training stage, the historical data is used as a conditional constraint (backtracking window) to guide the reconstruction of the predicted data (prediction window), and then in the inference stage, the last playback window data is used to predict the future prediction window data to realize time series prediction. However, the diffusion model has problems such as large computational burden, insufficient capture of time series dependent information, and poor hard conditional constraints. SUMMARY
[0004] The embodiment of the present application provides a mechanical health state degradation trend prediction method and system of a time series implicit diffusion model, a time series pre-enhancement module based on a bidirectional gate recurrent unit is designed, and backtracking window data is mapped to a low-dimensional time series hidden vector. In the inference stage, the prediction of the future degradation trend is realized by the constraint of the hidden vector.
[0005] The embodiment of the present application provides a mechanical health state degradation trend prediction method of a time series implicit diffusion model, comprising: Acceleration signals generated by stiffness fluctuation caused by motion contact of mechanical equipment parts in various states are monitored and acquired, sensitive feature indexes are constructed based on the acceleration signals, the feature indexes are segmented into multiple segments according to a certain step length, each segment of data is segmented into left and right two segments according to a certain proportion, the first segment on the left is taken as backtracking window data, and the second segment on the right is taken as prediction window data, so as to obtain a plurality of groups of feature index data sets corresponding one by one between the backtracking window and the prediction window; A bidirectional gate recurrent unit model is constructed, based on the feature index data set, the backtracking window data is taken as input, the prediction window data is taken as output, and the bidirectional gate recurrent unit is trained; All backtracking window data in the feature index data set is input into the trained bidirectional gate recurrent unit, and a low-dimensional time series hidden vector of an intermediate layer of the bidirectional gate recurrent unit is obtained: A vector quantization variational auto-encoding model is constructed, the reconstruction process of the prediction window data is learned through compression and recovery, and then the prediction window data is compressed to a low-dimensional discrete hidden space; A diffusion model is constructed in the low-dimensional discrete hidden space, a time series implicit diffusion model network architecture is formed, and the low-dimensional time series hidden vector As a constraint on the training and inference sampling process of the implicit diffusion model, the prediction window data serves as the input and reconstruction output of the temporal implicit diffusion model. According to the set probability, the coefficients of the constraints of the temporal implicit diffusion model are sampled and determined to be 0 or 1, so as to guide the composite training of the temporal implicit diffusion model under both constrained and unconstrained conditions. The temporal implicit diffusion model is trained based on the aforementioned feature index dataset. After the temporal implicit diffusion model is trained, the first step in the reverse diffusion process of the temporal diffusion model is estimated in the low-dimensional discrete hidden space. n The noise intensity of the step; By combining the estimated noise intensity, a low-dimensional vector is reconstructed through sampling; Given backtracking window data, the corresponding low-dimensional temporal hidden vector is obtained, the low-dimensional vector is reconstructed, and the decoder of the vector quantization variational autoencoder model is used to restore the reconstructed low-dimensional vector to the original data dimension, thereby obtaining the estimated prediction window data and realizing the prediction of degradation trend.
[0006] This application provides a mechanical health state degradation trend prediction system based on a time-series implicit diffusion model, including a processor and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it implements the steps of the mechanical health state degradation trend prediction method based on the time-series implicit diffusion model as described above.
[0007] This application implements both conditional and unconditional sampling in a single model to capture temporal dependency information of data from both global and local dimensions. A temporal pre-augmentation module based on a bidirectional gated recurrent unit is designed to map backtracking window data to a low-dimensional temporal hidden vector. During the inference phase, the future degradation trend is predicted through constraints imposed by this hidden vector. This method only requires historical monitoring acceleration signals from mechanical equipment to predict the degradation trend of health status over a future period. Furthermore, the bidirectional gated recurrent unit enhances the model's ability to capture temporal dependencies, while the introduction of probabilistic conditional constraints improves the accuracy of model predictions.
[0008] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description
[0009] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of the basic process of the mechanical health status degradation trend prediction method of the temporal implicit diffusion model in the embodiments of this application; Figure 2 This is a schematic diagram of the main structure of the temporal implicit diffusion model for predicting the mechanical health status degradation trend using the temporal implicit diffusion model, as described in the embodiments of this application. Figure 3 This diagram illustrates the detailed structural parameters of the prediction network of the temporal implicit diffusion model, which is used to predict the mechanical health status degradation trend of the temporal implicit diffusion model in the embodiments of this application. Figure 4 This is a schematic diagram of a high-power gear transmission experimental device, which serves as an application example of the temporal implicit diffusion model for predicting the mechanical health status degradation trend in this application. Detailed Implementation
[0010] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
[0011] This application provides a method for predicting the degradation trend of mechanical health status using a time-series implicit diffusion model, such as... Figure 1 As shown, it includes the following steps: In step S1, the acceleration signal generated by the stiffness fluctuation caused by the motion contact of mechanical equipment parts under various states is monitored and acquired. Based on the acceleration signal, a sensitive feature index is constructed. The feature index is divided into multiple segments according to a certain step size. Each segment of data is divided into left and right segments according to a certain ratio. The first segment on the left is used as the backtracking window data and the second segment on the right is used as the prediction window data, so as to obtain a feature index dataset with multiple backtracking windows and prediction windows corresponding one-to-one. In step S2, a bidirectional gated recurrent unit model is constructed. Based on the feature index dataset, the backtracking window data is used as input and the prediction window data is used as output to train the bidirectional gated recurrent unit. In step S3, all backtracking window data from the feature index dataset are input into the trained bidirectional gated recurrent unit to obtain the low-dimensional temporal hidden vector of the intermediate layer of the bidirectional gated recurrent unit: In the formula: This refers to the bidirectional gated loop unit. This represents the network weight parameters of the bidirectional gated recurrent unit model. t This represents the current time step of the feature index dataset. Indicates the first t The low-dimensional temporal hidden vector of the intermediate layer of the bidirectional gated cyclic unit model described in the time step.
[0012] In step S4, a vector quantization variational autoencoder model is constructed. By compressing the prediction window data and reconstructing the prediction window data through recovery learning, the prediction window data is compressed into a low-dimensional discrete hidden space.
[0013] In step S5, a diffusion model is constructed in the low-dimensional discrete hidden space to form a temporal implicit diffusion model network architecture, and the low-dimensional temporal hidden vectors are... As a constraint on the training and inference sampling process of the implicit diffusion model, the prediction window data serves as the input and reconstruction output of the temporal implicit diffusion model.
[0014] In step S6, the coefficients of the temporal implicit diffusion model constraint conditions are sampled to determine whether they are 0 or 1 according to a set probability, so as to guide the composite training of the temporal implicit diffusion model under constrained or unconstrained conditions. For example, the coefficients of the temporal implicit diffusion model constraint conditions are determined to be 0 or 1 according to a certain probability through Bernoulli sampling.
[0015] In step S7, the temporal implicit diffusion model is trained based on the feature index dataset. For example, in some embodiments, the temporal implicit diffusion model is optimized using noise matching loss, where the noise matching loss is: in, This represents the loss of the temporal implicit diffusion model; Represents the mathematical expectation; L This indicates the length of the backtracking window data; H This indicates the length of the prediction window data. Indicates a uniform distribution; This represents the maximum time step of the feature index dataset; n This represents the number of diffusion steps in the diffusion model. N This represents the maximum number of diffusion steps in the diffusion model. This represents a low-dimensional vector compressed into the low-dimensional discrete hidden space; Indicates in tThe low-dimensional vector compressed at time step 1 is diffused to the 1st dimensional vector through a forward diffusion process of a diffusion model. n The perturbation vector of the step; This indicates that the mean is 0 and the variance is 0. I , where, I Represents a unit vector; This represents a trainable U-Net denoising network; This represents the parameters of the U-Net network. , Represents a Gaussian noise vector; These represent predefined parameters representing the forward diffusion process in the diffusion model; where, .
[0016] In step S8, after the temporal implicit diffusion model is trained, the first step of the backward diffusion process of the temporal diffusion model is estimated in the low-dimensional discrete hidden space. n The noise intensity of the step; In step S9, the low-dimensional vector is reconstructed by combining the estimated noise intensity. For example, the low-dimensional vector can be reconstructed by sampling through stepwise denoising sampling.
[0017] In step S10, given the backtracking window data, the corresponding low-dimensional temporal hidden vector is obtained, the low-dimensional vector is reconstructed, and the decoder of the vector quantization variational autoencoder model is used to restore the reconstructed low-dimensional vector to the original data dimension to obtain the estimated prediction window data, thereby realizing the prediction of degradation trend.
[0018] The method in this application is designed with a temporal pre-augmentation module based on a bidirectional gated recurrent unit. It maps backtracking window data to a low-dimensional temporal hidden vector, and during the inference phase, the future degradation trend is predicted through constraints imposed by this hidden vector. This method only requires historical monitoring acceleration signals from mechanical equipment to predict the degradation trend of health status over a future period. Furthermore, the bidirectional gated recurrent unit enhances the model's ability to capture temporal dependencies, while the introduction of probabilistic conditional constraints improves the accuracy of the model's predictions.
[0019] In some embodiments, the process of obtaining a feature index dataset that corresponds one-to-one with multiple backtracking windows and prediction windows specifically includes: Acceleration signals from mechanical equipment under long-term operation are collected using an accelerometer, and a sequence of sensitive feature indicators is constructed based on the acceleration signals. ; The feature index sequence is smoothed by applying a moving average, resulting in a smoothed feature index sequence. The moving average expression is: In the formula,k This indicates the number of time periods in the moving average.
[0020] The feature index sequence is segmented into a backtracking window. and prediction window data segment , X and Y One-to-one correspondence; In the final prediction stage, the last segment of the feature index sequence is of length 1. L data segment As input to the model using backtracking window data, the model estimates and predicts the window data. This enables the prediction of degradation trends.
[0021] In some embodiments, training the bidirectional gated recurrent unit based on the feature index dataset, using backtracking window data as input and prediction window data as output, includes: Based on the aforementioned feature index dataset, the backtracking window data is used as input, the prediction window data is used as output, and the bidirectional gated recurrent unit is trained using the least squares error loss to obtain the trained bidirectional gated recurrent unit model, wherein the least squares error loss is: In the formula: This represents the least squares error loss; This represents the prediction window data. express t The estimated value of the prediction window data at the given time.
[0022] In some embodiments, the training loss of the vector quantization variational autoencoder model is: In the formula, This represents the training loss of the vector quantization variational autoencoder model; E This represents the encoder of the vector quantization variational autoencoder model; D This represents the decoder of the vector quantization variational autoencoder model; This indicates a gradient-stopping operation; express t The prediction window data at the specified time; This represents the balance coefficient.
[0023] In some embodiments, the temporal implicit diffusion model diffuses to the perturbation vector at step n during the forward diffusion process. Its forward diffusion process is as follows: In the formula: Represents a Gaussian noise vector; This represents the predefined parameters of the forward diffusion process in the diffusion model; where, .
[0024] In some embodiments, the time-series diffusion model is estimated during the reverse diffusion process. n The noise intensity of the step is: In the formula, This represents the noise intensity estimate for the reverse diffusion process; This represents the balance coefficient.
[0025] In some embodiments, in S900, the noise intensity estimation combined with the reverse diffusion process, and the sampling reconstruction of the low-dimensional vector based on the estimated noise intensity, are achieved through progressive denoising sampling, wherein progressive denoising sampling is as follows: In the formula, This indicates that the diffusion model in the forward diffusion process is in the first... n The predefined variance of the step; ; e This represents the noise vector obtained by sampling from a Gaussian distribution.
[0026] In some embodiments, given backtracking window data, predicting degradation trends includes: Input all backtracking window data from the feature index dataset of the given backtracking window data into the bidirectional gated recurrent unit to obtain the low-dimensional temporal hidden vector. The hidden vector is compressed into a low-dimensional discrete hidden space by the encoder of the vector quantization variational autoencoder model. Gaussian noise vector is obtained by sampling from a Gaussian distribution. Acquired through a progressive denoising sampling process This process is repeated until the fully denoised reconstructed low-dimensional vector is obtained. ; The fully denoised reconstructed low-dimensional vector is then decoded using a vector quantization variational autoencoder model. Restore to the original data dimension to obtain the predicted prediction window data. .
[0027] This application also proposes an implementation example of a method for predicting the degradation trend of mechanical health status using a time-series implicit diffusion model, including the following steps: S10 monitors and acquires acceleration signals generated by stiffness fluctuations caused by the motion contact of mechanical equipment parts under various conditions over a long period of time, and constructs sensitive feature indicators based on the acceleration signals. The feature index sequence is smoothed by using a moving average to reduce noise interference, resulting in a smoothed feature index sequence. The feature index sequence is divided into multiple segments according to a certain step size. Each segment is further divided into left and right sub-segments according to a certain ratio. The first left sub-segment is used as the backtracking window data, and the second sub-segment is used as the prediction window data. This results in multiple feature index datasets with one-to-one correspondence between the backtracking window and the prediction window. The gear under test in the gear transmission system is any one of spur gears, bevel gears, helical gears, and herringbone gears. The gear under test is installed in the parallel shaft gearbox of the gear transmission system, and power transmission is achieved through two-stage reduction.
[0028] S20, Construct a bidirectional gated recurrent unit (BRU) model. Based on the feature index dataset, use backtracking window data as input and prediction window data as output. Train the BRU using the least squares error loss to obtain the trained BRU model. The least squares error loss expression is: In the formula: This represents the least squares error loss; This represents the prediction window data. express t Estimate the value of the forecast window data at any given time.
[0029] S30: Input all backtracking window data from the feature index dataset into the bidirectional gated recurrent unit to obtain the low-dimensional temporal hidden vector of the intermediate layer of the bidirectional gated recurrent unit model: In the formula: This indicates a bidirectional gated loop unit. This represents the network weight parameters of the bidirectional gated recurrent unit model. t This represents the current time step of the feature index dataset. Indicates the first t Low-dimensional temporal hidden vectors in the intermediate layer of the time-step bidirectional gated cyclic unit model.
[0030] S40, construct a vector quantization variational autoencoder model. This model compresses and reconstructs the prediction window data by learning the process of compressing and recovering the prediction window data, thereby compressing the prediction window data into a low-dimensional discrete hidden space. The training loss expression for the vector quantization variational autoencoder model is: In the formula: This represents the training loss of a vector quantized variational autoencoder model. E The encoder represents the vector quantization variational autoencoder model; D The decoder represents the vector quantization variational autoencoder model; This indicates a gradient-stopping operation; express t Predict window data in real time; Indicates the balance coefficient; S50 constructs a diffusion model in a low-dimensional discrete hidden space, forming a temporal implicit diffusion model network architecture, which integrates low-dimensional temporal hidden vectors. As a constraint on the training and inference sampling process of the temporal implicit diffusion model, the prediction window data serves as both the input and the reconstruction output of the temporal implicit diffusion model.
[0031] S60, according to a certain probability, the coefficients of the constraints of the temporal implicit diffusion model are determined to be 0 or 1 through Bernoulli sampling, thereby guiding the composite training of the temporal implicit diffusion model under both constrained and unconstrained conditions. S70, The temporal implicit diffusion model is trained based on the feature index dataset, and then optimized using noise matching loss. The noise matching loss expression is as follows: In the formula: This represents the loss of the temporal implicit diffusion model; Represents the mathematical expectation; L Indicates the length of the backtracking window data; H Indicates the length of the prediction window data; Indicates a uniform distribution; This represents the maximum time step in the feature index dataset. n This indicates the number of diffusion steps in the diffusion model; N This represents the maximum number of diffusion steps in the diffusion model; This represents a low-dimensional vector after being compressed into the discrete hidden space of dimension 1; Indicates in t The low-dimensional vector compressed at time step 1 is diffused to the 1st dimensional vector through the forward diffusion process of the diffusion model. n The perturbation vector of the step; This indicates that the mean is 0 and the variance is . I , where, I Represents a unit vector; This represents a trainable U-Net denoising network; Indicates the U-Net network parameters; Represents a Gaussian noise vector; These represent predefined parameters representing the forward diffusion process in the diffusion model; where, .
[0032] S80, after training the temporal implicit diffusion model, in the low-dimensional discrete hidden space, the reverse diffusion process of the diffusion model is performed.n The noise intensity of the step is estimated by the following expression: In the formula: This represents the noise intensity estimate for the reverse diffusion process; This represents the balance coefficient.
[0033] S90, combining the noise intensity estimation from the reverse diffusion process, reconstructs the low-dimensional vector through progressive denoising sampling. The expression for the progressive denoising sampling process is: In the formula, This indicates that the diffusion model represents the first step in the forward diffusion process. n The predefined variance of the step; ; e This represents the noise vector obtained by sampling from a Gaussian distribution; S100, obtain the low-dimensional temporal hidden vector through step S30. The encoder of the vector quantization variational autoencoder model compresses the hidden vectors into a low-dimensional discrete hidden space.
[0034] Furthermore, Gaussian noise vectors are sampled from the Gaussian distribution. The noise reduction sampling process in step S90 is used to obtain the data. This continues until a fully denoised reconstructed low-dimensional vector is obtained. .
[0035] Finally, a decoder using a vector quantization variational autoencoder model reconstructs the low-dimensional vector back to the original data dimension, yielding the prediction window data. Complete the prediction of the degradation trend of mechanical equipment.
[0036] Specifically, taking a certain gear transmission system as an example for further explanation, the sampling frequency of the acceleration signal of the gear transmission system under test is 25.6kHz.
[0037] In the verification process of the embodiments, this example selects traditional time series prediction methods and time series prediction methods based on generative models as references. The specific methods for comparison are: 1) a health status degradation trend prediction method based on gated recurrent units (GRUs); 2) a health status degradation trend prediction method based on bidirectional gated recurrent units (BiGRUs); 3) a health status degradation trend prediction method based on Transformer models; and 4) a health status degradation trend prediction method based on a generative model-based autoregressive denoising diffusion model (TimeGrad). Among them, the method proposed in this invention is a temporal latent diffusion model (TLDM) based on generative models.
[0038] Specifically applied to predicting the health degradation trend of gear transmission systems, this invention is validated using data from a high-power gear transmission test bench. A schematic diagram of the test bench is shown below. Figure 4 As shown. The structure of the temporal implicit diffusion model used is as follows. Figure 2 As shown, the specific process and parameters are as follows: Figure 3 As shown.
[0039] The experimental platform includes a drive motor 1, a torque sensor 2, a coupling 3, an acceleration sensor 4, a parallel shaft gearbox under test 5, a speed-increasing planetary gearbox 6, a speed-reducing planetary gearbox 7, a test gearbox 8, and a load motor 9. The power for the experimental platform is provided by the drive motor 1. This invention uses vibration data with a sampling frequency of 25.6 kHz for analysis. During the test in the embodiment, according to 3... The criterion is that the variance fluctuation of the characteristic index series should be less than 3. The characteristic index values were used as historical data, i.e., when the equipment was operating normally, to predict the degradation trend at different lengths in the future using TLDM. Specifically, two different backtracking window lengths were selected: 512 (L512) and 256 (L256). Four different prediction window lengths were also selected: 384 (H384), 256 (H256), 192 (H192), and 128 (H128). Six prediction tasks were formed by combining the backtracking and prediction window data. In the prediction bias calculation stage, the root mean square error between the actual data and the final prediction window data was used as the prediction bias.
[0040] Based on the constructed feature index sequence, and after end-to-end training and testing using various methods, the final deviation values (root mean square error of the prediction results ± standard deviation of the deviation values) of the predicted health status degradation trend are shown in Table 1. As can be seen from the table, the diffusion model-based time series prediction method outperforms traditional GRU, BiGRU, and Transformer time series prediction methods, especially in prediction tasks with long prediction window data segments. Among them, the TLDM proposed in this invention exhibits the best prediction performance, with an average prediction deviation of 0.0067, which is more than 30% higher than TimeGrad (prediction deviation reduced by 0.003). In prediction tasks with larger step sizes (e.g., a playback window data length of 512 and a prediction window data length of 384), TLDM improves prediction performance by more than 20% compared to TimeGrad (prediction deviation reduced by 0.0035), and by more than 40% compared to the traditional time series prediction method GRU (prediction deviation reduced by 0.0063). This fully demonstrates the superior performance of the TLDM proposed in this application in predicting the trend of health status degradation of mechanical equipment under limited data and large step length.
[0041] Table 1. Prediction results of health status degradation trend for the gear transmission system embodiment. This application presents a method for predicting the degradation trend of mechanical health status based on a time-series implicit diffusion model. The constructed degradation trend prediction model only requires historical monitoring data to predict the future health status degradation trend, enabling early maintenance and support decisions. On the one hand, it solves the dilemma of difficulty in effectively obtaining full life cycle data for some high-value mechanical equipment. On the other hand, the introduction of a generative model reduces the prediction bias of the health status degradation trend prediction model with limited available data, thus improving the applicability of the model.
[0042] This application presents a method for predicting the degradation trend of mechanical health status based on a temporal implicit diffusion model. It captures temporal dependency information within historical data through a bidirectional gated recurrent unit module and appends this dependency information to the training and sampling generation processes of the temporal implicit diffusion model. Combined with probabilistic conditional constraints, this improves the generalization and effectiveness of the temporal implicit diffusion model. Furthermore, this method establishes a degradation state prediction method based on a generative model, significantly enhancing the model's predictive performance with limited data.
[0043] This application also proposes a mechanical health state degradation trend prediction system based on a temporal implicit diffusion model, including a processor and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it implements the steps of the mechanical health state degradation trend prediction method based on the temporal implicit diffusion model as described above.
[0044] It should be noted that, in the embodiments of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0045] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0046] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.
[0047] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims. All of these forms are within the protection scope of this application.
Claims
1. A method for predicting the degradation trend of mechanical health status using a temporal implicit diffusion model, characterized in that, include: Monitor and acquire acceleration signals generated by stiffness fluctuations caused by the movement and contact of mechanical equipment parts under various conditions. Construct sensitive feature indicators based on the acceleration signals. Divide the feature indicators into multiple segments according to a certain step size. Each segment is divided into left and right segments according to a certain ratio. Use the first segment on the left as the backtracking window data and the second segment on the right as the prediction window data to obtain a feature indicator dataset with one-to-one correspondence between multiple backtracking windows and prediction windows. A bidirectional gated recurrent unit model is constructed. Based on the feature index dataset, the backtracking window data is used as input and the prediction window data is used as output to train the bidirectional gated recurrent unit. Input all backtracking window data from the feature index dataset into the trained bidirectional gated recurrent unit to obtain the low-dimensional temporal hidden vector of the intermediate layer of the bidirectional gated recurrent unit: A vector quantization variational autoencoder model is constructed, and the reconstruction process of the prediction window data is learned through compression and recovery, thereby compressing the prediction window data into a low-dimensional discrete hidden space. A diffusion model is constructed in the low-dimensional discrete hidden space to form a temporal implicit diffusion model network architecture, which incorporates the low-dimensional temporal hidden vectors. As a constraint on the training and inference sampling process of the implicit diffusion model, the prediction window data serves as the input and reconstruction output of the temporal implicit diffusion model. According to the set probability, the coefficients of the constraints of the temporal implicit diffusion model are sampled to be 0 or 1, so as to guide the composite training of the temporal implicit diffusion model under both constrained and unconstrained conditions. The temporal implicit diffusion model is trained based on the aforementioned feature index dataset. After the temporal implicit diffusion model is trained, the _th_ step in the reverse diffusion process of the temporal diffusion model is estimated in the low-dimensional discrete hidden space. n The noise intensity of the step; By combining the estimated noise intensity, a low-dimensional vector is reconstructed through sampling; Given backtracking window data, the corresponding low-dimensional temporal hidden vector is obtained, the low-dimensional vector is reconstructed, and the decoder of the vector quantization variational autoencoder model is used to restore the reconstructed low-dimensional vector to the original data dimension, thereby obtaining the estimated prediction window data and realizing the prediction of degradation trend.
2. The method for predicting the degradation trend of mechanical health status using a time-series implicit diffusion model as described in claim 1, characterized in that, The process of obtaining a feature index dataset that corresponds one-to-one with multiple backtracking windows and prediction windows specifically includes: Acceleration signals from mechanical equipment under long-term operation are collected using an accelerometer, and a sequence of sensitive feature indicators is constructed based on the acceleration signals. ; The feature index sequence is smoothed by applying a moving average, resulting in a smoothed feature index sequence. ; The feature index sequence is segmented into a backtracking window. and prediction window data segment , X and Y One-to-one correspondence; In the final prediction stage, the last segment of the feature index sequence is of length 1. L data segment As input to the model using backtracking window data, the model estimates and predicts the window data. This enables the prediction of degradation trends.
3. The method for predicting the degradation trend of mechanical health status using a time-series implicit diffusion model as described in claim 1, characterized in that, Training the bidirectional gated recurrent unit based on the aforementioned feature index dataset, using backtracking window data as input and prediction window data as output, includes: Based on the aforementioned feature index dataset, the backtracking window data is used as input, the prediction window data is used as output, and the bidirectional gated recurrent unit is trained using the least squares error loss to obtain the trained bidirectional gated recurrent unit model, wherein the least squares error loss is: In the formula: This represents the least squares error loss; This represents the prediction window data. express t The estimated value of the prediction window data at the given time.
4. The method for predicting the degradation trend of mechanical health status using a time-series implicit diffusion model as described in claim 3, characterized in that, The training loss of the vector quantization variational autoencoder model is: In the formula, This represents the training loss of the vector quantization variational autoencoder model; E This represents the encoder of the vector quantization variational autoencoder model; D This represents the decoder of the vector quantization variational autoencoder model; This indicates a gradient-based stopping operation; express t The prediction window data at the specified time; This represents the balance coefficient.
5. The method for predicting the degradation trend of mechanical health status using a time-series implicit diffusion model as described in claim 4, characterized in that, The temporal implicit diffusion model diffuses to the perturbation vector at step n during the forward diffusion process. Its forward diffusion process is as follows: In the formula: Represents a Gaussian noise vector; This represents the predefined parameters of the forward diffusion process in the diffusion model; where, .
6. The method for predicting the mechanical health status degradation trend using a time-series implicit diffusion model as described in claim 5, characterized in that, Training the time-series implicit diffusion model based on the aforementioned feature index dataset includes: The temporal implicit diffusion model is optimized using a noise matching loss, which is: in, This represents the loss of the temporal implicit diffusion model; Represents the mathematical expectation; L This indicates the length of the backtracking window data; H This indicates the length of the prediction window data. Indicates a uniform distribution; This represents the maximum time step of the feature index dataset; n This represents the number of diffusion steps in the diffusion model. N This represents the maximum number of diffusion steps in the diffusion model. This represents a low-dimensional vector compressed into the low-dimensional discrete hidden space; Indicates in t The low-dimensional vector compressed at time step 1 is diffused to the 1st dimensional vector through a forward diffusion process of a diffusion model. n The perturbation vector of the step; This indicates that the mean is 0 and the variance is 0. I , where, I Represents a unit vector; This represents a trainable U-Net denoising network; This represents the parameters of the U-Net network.
7. The method for predicting the mechanical health status degradation trend using a time-series implicit diffusion model as described in claim 5, characterized in that, Estimate the first time series diffusion model during the reverse diffusion process. n The noise intensity of the step is: In the formula, This represents the noise intensity estimate for the reverse diffusion process; This represents the balance coefficient.
8. The method for predicting the degradation trend of mechanical health status using a time-series implicit diffusion model as described in claim 5, characterized in that, Based on the estimated noise intensity, the low-dimensional vector is reconstructed through progressive denoising sampling, where progressive denoising sampling is as follows: In the formula, This indicates that the diffusion model in the forward diffusion process is in the first... n The predefined variance of the step; ; e This represents the noise vector obtained by sampling from a Gaussian distribution.
9. The method for predicting the degradation trend of mechanical health status using a time-series implicit diffusion model as described in claim 1, characterized in that, Given backtracking window data, predicting degradation trends includes: Input all backtracking window data from the feature index dataset of the given backtracking window data into the bidirectional gated recurrent unit to obtain the low-dimensional temporal hidden vector. The hidden vector is compressed into a low-dimensional discrete hidden space by the encoder of the vector quantization variational autoencoder model. Gaussian noise vector is obtained by sampling from a Gaussian distribution. Acquired through a progressive denoising sampling process This process is repeated until the fully denoised reconstructed low-dimensional vector is obtained. ; The fully denoised reconstructed low-dimensional vector is then decoded using a vector quantization variational autoencoder model. Restore to the original data dimension to obtain the predicted prediction window data. .
10. A system for predicting the degradation trend of mechanical health status using a temporal implicit diffusion model, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, which, when executed by the processor, implements the steps of the method for predicting the mechanical health state degradation trend using a temporal implicit diffusion model as described in any one of claims 1 to 9.