High-dimensional hierarchical time sequence modeling method for predicting residual life of industrial rotating equipment
By employing information bottleneck theory and manifold learning with off-diagonal covariance constraints, the problem of capturing degradation trends in the prediction of the remaining useful life of rotating equipment under high-dimensional data was solved, achieving efficient and accurate prediction of the remaining useful life.
Patent Information
- Application Number
- CN202511898160.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-16
- Publication Date
- 2026-04-28
AI Technical Summary
Existing technologies struggle to effectively capture degradation trends in predicting the remaining service life of rotating equipment under high-dimensional data, and end-to-end deep learning methods lack the integration of physical mechanisms, leading to inconsistent prediction results.
We construct a global representation using the information bottleneck theory, combine it with a block interaction mechanism and manifold learning off-diagonal covariance prior constraints, and optimize feature extraction and fusion through a weighted loss function to achieve efficient remaining lifetime prediction.
It significantly reduces the prediction error of remaining lifetime under high-dimensional multi-source data, provides accurate and lightweight technical support, and improves prediction accuracy and efficiency.
Smart Images
Figure CN121935823A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a high-dimensional hierarchical time series modeling method for predicting the remaining life of industrial rotating equipment, belonging to the field of fault prediction and health management technology. Background Technology
[0002] In the context of intelligent manufacturing, unplanned downtime is a core bottleneck restricting the efficient operation of modern industrial systems. The resulting productivity losses, soaring maintenance costs, and safety hazards seriously threaten enterprises' core competitiveness. To ensure production continuity, improve asset utilization, and achieve predictive maintenance, fault prediction and health management (PHM) technology has become a key pillar of intelligent industrial operations and maintenance. Among these technologies, predicting the remaining useful life (RUL) of critical equipment such as rotating machinery is the core technological foundation for PHM systems to achieve accurate decision-making and closed-loop health management. Accurate RUL prediction can guide users to implement proactive maintenance, thereby avoiding catastrophic failures and optimizing maintenance resource allocation.
[0003] Current methods for predicting Relative Limiting Dimensions (RUL) data to address the challenges of high-dimensional data mainly fall into two categories: traditional methods based on preprocessing and end-to-end deep learning methods. Preprocessing-based methods rely on expert knowledge, using signal processing, statistical analysis, and feature engineering to reduce dimensionality and extract features from the original data. While interpretable, these methods suffer from inherent flaws such as a disconnect between the optimization objective and the prediction task, and poor generalization ability. Their knowledge-intensive nature makes them heavily reliant on domain experience, making them difficult to adapt to industrial scenarios with different physical mechanisms, requiring extensive manual redesign for new applications. End-to-end deep learning methods can automatically learn features from the original data, avoiding tedious feature engineering and possessing powerful adaptive representation learning capabilities. However, these methods are entirely data-driven, lacking the integration and guidance of domain physical mechanisms (such as thermodynamic degradation laws and friction and wear mechanisms). When training data is limited or operating conditions deviate, they are prone to producing physically inconsistent prediction results, limiting their reliable application in strong-mechanism industrial PHM scenarios.
[0004] Existing methods that attempt to embed domain knowledge into deep learning models, such as adding physical consistency constraints through loss functions or regularizing the feature space using simple Gaussian distributions, often rely on simplification assumptions and generally fail to effectively model the inherent temporal dependencies in degraded data, resulting in limited performance improvements for RUL prediction. Therefore, constructing a prediction framework that can adaptively learn degradation features and deeply integrate prior physical knowledge from high-dimensional, multi-source time-series monitoring data has become an urgent technical requirement for improving the accuracy of industrial equipment life prediction and achieving reliable intelligent operation and maintenance. Developing new methods that integrate knowledge and adaptive learning advantages is of significant engineering importance for overcoming current technological limitations and promoting the practical application of PHM technology in industrial scenarios. Summary of the Invention
[0005] This invention proposes a high-dimensional hierarchical time series modeling method for predicting the remaining useful life of industrial rotating equipment. This method utilizes information bottleneck theory to construct a global representation based on information bottleneck constraints, fusing and mapping multi-source time series data to a low-dimensional feature space, thereby capturing a low-dimensional yet information-rich global representation. A block-based interaction mechanism is constructed to extract local features and perform hierarchical feature fusion, achieving global-local collaborative enhancement of representation fusion. A time series representation based on manifold learning and off-diagonal covariance prior constraints is constructed, embedding time series degradation features into the global representation. This method solves the key problem of existing methods struggling to capture degradation trends in remaining useful life prediction tasks under high-dimensional data challenges, significantly reducing the error and parameter quantity in remaining useful life prediction under high-dimensional multi-source data, and providing accurate and lightweight technical support for the application of equipment remaining useful life prediction.
[0006] The technical solution adopted in this invention is a high-dimensional hierarchical time series modeling method for predicting the remaining service life of industrial rotating equipment. It constructs a global representation based on information bottleneck constraints, a collaborative enhancement representation based on block interaction mechanism, and a time series representation based on manifold learning off-diagonal covariance prior constraints. It also establishes a weighted loss function to achieve high-performance and low-cost prediction of the remaining service life of industrial rotating equipment.
[0007] Includes the following steps:
[0008] Step 1: Construct a global representation based on information bottleneck constraints;
[0009] Step 11: Install and deploy 28 sensors to collect multi-source time-series monitoring data related to the degradation trend of industrial equipment throughout its entire lifecycle; process the data using a sliding window to obtain training and testing samples with dimensions of [missing information]. 60 is the window length for each sample. Given a set of real numbers; features are extracted from each sample using a feature extraction module that includes an embedding layer, a convolutional layer, and a non-linear activation function, resulting in a feature dimension of . The deep features are extracted from each sensor, with 128 representing the feature dimensions of each sensor after feature extraction. Attention pooling is used to fuse the deep features from these 28 sensors at the feature level. For a dimensionality-reduced global representation 8 is the predefined low-dimensional global representation dimension;
[0010] Step 12, global representation after dimensionality reduction using multivariate Gaussian prior constraints. The learning process will transform the low-dimensional global representation The mapping is to a parameterized standard multivariate Gaussian distribution, where the mean vector is... Used to construct the covariance matrix The vector is ;
[0011] Step 13: Calculate the KL divergence loss constraint mean vector using a predefined multivariate Gaussian prior. and diagonal matrix Learning; given sample dimensions and predicted output To maximize mutual information At the same time, minimize mutual information The objective function for optimizing the information bottleneck is constructed as follows:
[0012] (1)
[0013] in To weigh and Weight of importance For For input, These are the weights of the output deep network model;
[0014] Mutual information is derived based on the nonnegativity of KL divergence. The lower bound and mutual information The upper bound is shown below:
[0015] (2);
[0016] (3);
[0017] in, This represents the input data for the model; The label to be predicted. This represents the compressed representation of the data after passing through the model encoder. express and The true joint distribution, Represents conditional distribution The variational approximation, express and The true joint distribution, Indicates a given hour The true posterior distribution, Indicates marginal distribution Variational approximation;
[0018] Substituting the two variational boundaries mentioned above into the original objective function yields a parameterized multivariate Gaussian distribution. Random sampling is then performed from this distribution to generate a low-dimensional and information-rich global representation for subsequent feature learning and prediction. ;
[0019] Step 14: Using the global representation model constructed in steps 11 and 12 and the optimization objective defined in step 13, model training can automatically execute feature extraction, fusion, parameterization, and resampling processes to obtain a low-dimensional and information-rich global representation. ;
[0020] Step 2: Construct a collaborative enhancement representation based on a block-based interaction mechanism;
[0021] Step 21: Process the shallow features of the multi-source sensor after embedding layer processing. Divide into blocks; along the second dimension The blocks are uniformly divided into 32 sub-feature blocks, represented as follows:
[0022] (4);
[0023] Wherein, the dimension of each sub-feature block is , Along the second dimension;
[0024] Step 22, global representation Matrix obtained after broadcasting Each sub-feature block obtained in step 21 With matrix By combining the elements, an enhanced interactive input can be formed:
[0025] (5);
[0026] Step 23, combine each spliced feature Input is fed into a shared linear interaction module, and the output is the interacted sub-features. ;
[0027] Step 24: Output the 32 sub-features after interactive processing in Step 23. Reassemble along the feature dimensions to restore the complete feature tensor:
[0028] (6);
[0029] Step 3: Construct a temporal representation based on manifold learning of off-diagonal covariance prior constraints;
[0030] Step 31: Convert the low-dimensional global representation output by the attention pooling module The parameterization is performed as a Gaussian mixture model with three components, resulting in a low-dimensional global representation. The probability distribution is as follows:
[0031] (7);
[0032] in, For the first The mixing weights of the Gaussian components satisfy the following conditions: ; For the first The mean vector of Gaussian components; For the first The covariance matrix of Gaussian components; For the first One multivariate Gaussian component;
[0033] Step 32: The low-dimensional global representation is transformed through a parameterization layer. Mapped to an output vector with a total dimension of 135. ;
[0034] Step 33, the output vector obtained in step 32 The complete posterior Gaussian mixture model distribution is defined according to the preset weight dimension 3, mean dimension 24, and off-diagonal covariance matrix dimension 108; the KL divergence between the posterior Gaussian mixture model and the prior Gaussian mixture model estimated by manifold learning is calculated according to the method described in step 31; and a low-dimensional global representation containing rich temporal degradation information is obtained by sampling from the posterior Gaussian mixture model. ;
[0035] Step 4: Establish the weighted loss function;
[0036] Step 41, calculate Huber's losses. and loss Weighted portfolio for robust regression loss The expression is:
[0037] (8);
[0038] Step 42, establish the KL divergence regularization loss based on the Gaussian mixture model, the expression of which is:
[0039] (9);
[0040] in, To represent the first output from the model parameterization layer Posterior mixture weights of Gaussian components; , These are the 1st, 2nd, and 3rd digits of the posterior distribution. The mean vector and covariance matrix of each Gaussian component; , The first and second prior distributions estimated based on manifold learning are respectively the first and second... The mean vector and covariance matrix of each Gaussian component;
[0041] Step 43: Add the robust regression loss from Step 41 to the KL divergence regularization loss term based on the Gaussian mixture model from Step 42 to obtain the total loss function, expressed as:
[0042] (10);
[0043] Step 5, prediction of remaining useful life;
[0044] The features obtained after encoding in steps 1, 2, and 3 Input the decoder network and output the remaining lifetime prediction value; after continuous iteration and optimization of the total loss function in step 4, output the accurate remaining lifetime prediction value.
[0045] This method first constructs a global representation based on information bottleneck constraints, then constructs a collaborative enhancement representation based on a block interaction mechanism, then constructs a time series representation based on manifold learning off-diagonal covariance prior constraints, and finally predicts the remaining useful life by optimizing the total loss function. By iteratively executing steps (1) to (5), the representation modeling of multi-source high-dimensional time series data can be realized, and finally, the remaining useful life prediction results with lower error and higher efficiency can be obtained. Attached Figure Description
[0046] Figure 1 This is a general overview of a time-aware high-dimensional hierarchical feature fusion method for predicting remaining useful life.
[0047] Figure 2 The graph shows the predicted remaining service life of engine 11.
[0048] Figure 3 The graph shows the predicted remaining service life of engine 14.
[0049] Figure 4 The graph shows the predicted remaining service life of engine 15. Detailed Implementation
[0050] A high-dimensional hierarchical time series modeling method for predicting the remaining useful life of industrial rotating equipment is proposed. This method utilizes information bottleneck theory to construct a global representation based on information bottleneck constraints, fusing and mapping multi-source time series data to a low-dimensional feature space, thereby capturing a low-dimensional yet information-rich global representation. A block-based interaction mechanism is constructed to extract local features and perform hierarchical feature fusion, achieving global-local collaborative enhancement of representation fusion. A time series representation based on manifold learning and off-diagonal covariance prior constraints is constructed, embedding time series degradation features into the global representation. This method solves the key problem of existing methods struggling to capture degradation trends in remaining useful life prediction tasks under high-dimensional data challenges, significantly reducing the error and parameter quantity in remaining useful life prediction under high-dimensional multi-source data, and providing accurate and lightweight technical support for equipment remaining useful life prediction applications.
[0051] The process involves constructing a global representation based on information bottleneck constraints, a collaborative enhancement representation based on a block interaction mechanism, and a time-series representation based on manifold learning-based off-diagonal covariance prior constraints. A weighted loss function is then established to achieve high-performance, low-cost prediction of the remaining service life of industrial rotating equipment. The steps include:
[0052] Step 1: Construct a global representation based on information bottleneck constraints;
[0053] Step 11: Install and deploy 28 sensors to collect multi-source time-series monitoring data related to the degradation trend of industrial equipment throughout its entire lifecycle; process the data using a sliding window to obtain training and testing samples with dimensions of [missing information]. 60 is the window length for each sample. Given a set of real numbers; features are extracted from each sample using a feature extraction module that includes an embedding layer, a convolutional layer, and a non-linear activation function, resulting in a feature dimension of . The deep features are extracted from each sensor, with 128 representing the feature dimensions of each sensor after feature extraction. Attention pooling is used to fuse the deep features from these 28 sensors at the feature level. For a dimensionality-reduced global representation 8 is the predefined low-dimensional global representation dimension;
[0054] Step 12, global representation after dimensionality reduction using multivariate Gaussian prior constraints. The learning process will transform the low-dimensional global representation The mapping is to a parameterized standard multivariate Gaussian distribution, where the mean vector is... Used to construct the covariance matrix The vector is ;
[0055] Step 13: Calculate the KL divergence loss constraint mean vector using a predefined multivariate Gaussian prior. and diagonal matrix Learning; given sample dimensions and predicted output To maximize mutual information At the same time, minimize mutual information The objective function for optimizing the information bottleneck is constructed as follows:
[0056] (1);
[0057] in To weigh and Weight of importance For For input, These are the weights of the output deep network model;
[0058] Mutual information is derived based on the nonnegativity of KL divergence. The lower bound and mutual information The upper bound is shown below:
[0059] (2);
[0060] (3);
[0061] in, This represents the input data for the model; The label to be predicted. This represents the compressed representation of the data after passing through the model encoder. express and The true joint distribution, Represents conditional distribution The variational approximation, express and The true joint distribution, Indicates a given hour The true posterior distribution, Indicates marginal distribution Variational approximation;
[0062] Substituting the two variational boundaries mentioned above into the original objective function yields a parameterized multivariate Gaussian distribution. Random sampling is then performed from this distribution to generate a low-dimensional and information-rich global representation for subsequent feature learning and prediction. ;
[0063] Step 14: By training the model using the global representation model constructed in steps 11 and 12 and the optimization objective defined in step ③, the feature extraction, fusion, parameterization, and resampling processes can be automatically executed to obtain a low-dimensional and information-rich global representation. ;
[0064] Step 2: Construct a collaborative enhancement representation based on a block-based interaction mechanism;
[0065] Step 21: Process the shallow features of the multi-source sensor after embedding layer processing. Divide into blocks; along the second dimension The blocks are uniformly divided into 32 sub-feature blocks, represented as follows:
[0066] (4);
[0067] Wherein, the dimension of each sub-feature block is , Along the second dimension;
[0068] Step 22, global representation Matrix obtained after broadcasting Each sub-feature block obtained in step 21 With matrix By combining the elements, an enhanced interactive input can be formed:
[0069] (5);
[0070] Step 23, combine each spliced feature Input is fed into a shared linear interaction module, and the output is the interacted sub-features. ;
[0071] Step 24: Output the 32 sub-features after interactive processing in step 23. Reassemble along the feature dimensions to restore the complete feature tensor:
[0072] (6);
[0073] Step 3: Construct a temporal representation based on manifold learning of off-diagonal covariance prior constraints;
[0074] Step 31: Convert the low-dimensional global representation output by the attention pooling module The parameterization is performed as a Gaussian mixture model with three components, resulting in a low-dimensional global representation. The probability distribution is as follows:
[0075] (7);
[0076] in, For the first The mixing weights of the Gaussian components satisfy the following conditions: ; For the first The mean vector of Gaussian components; For the first The covariance matrix of Gaussian components; For the first One multivariate Gaussian component;
[0077] Step 32: The low-dimensional global representation is transformed through a parameterization layer. Mapped to an output vector with a total dimension of 135. ;
[0078] Step 33, the output vector obtained in step 32 The complete posterior Gaussian mixture model distribution is defined according to the preset weight dimension 3, mean dimension 24, and off-diagonal covariance matrix dimension 108; the KL divergence between the posterior Gaussian mixture model and the prior Gaussian mixture model estimated by manifold learning is calculated according to the method described in step 31; and a low-dimensional global representation containing rich temporal degradation information is obtained by sampling from the posterior Gaussian mixture model. ;
[0079] Step 4: Establish the weighted loss function;
[0080] Step 41, calculate Huber's losses. and loss Weighted portfolio for robust regression loss The expression is:
[0081] (8);
[0082] Step 42, establish the KL divergence regularization loss based on the Gaussian mixture model, the expression of which is:
[0083] (9);
[0084] in, To represent the first output from the model parameterization layer Posterior mixture weights of Gaussian components; , These are the 1st, 2nd, and 3rd digits of the posterior distribution. The mean vector and covariance matrix of each Gaussian component; , The first and second prior distributions estimated based on manifold learning are respectively the first and second... The mean vector and covariance matrix of each Gaussian component;
[0085] Step 43: Add the robust regression loss from Step 41 to the KL divergence regularization loss term based on the Gaussian mixture model from Step 42 to obtain the total loss function, expressed as:
[0086] (10);
[0087] Step 5, prediction of remaining useful life;
[0088] The features obtained after encoding in steps 1, 2, and 3 Input the decoder network and output the remaining lifetime prediction value; after continuous iteration and optimization of the total loss function in step 4, output the accurate remaining lifetime prediction value.
[0089] This method first constructs a global representation based on information bottleneck constraints, then designs a global-local collaborative enhancement representation fusion method based on a block interaction mechanism, then develops a time-aware off-diagonal covariance prior estimation method based on manifold learning, and finally predicts remaining useful life by optimizing the joint loss function. The iterative steps 1-4 can realize the representation modeling of multi-source high-dimensional time series data, and finally obtain a remaining useful life prediction result with lower error and higher efficiency.
Claims
1. A high-dimensional hierarchical time series modeling method for predicting the remaining life of industrial rotating equipment, characterized in that, Includes the following steps: Step 1: Construct a global representation based on information bottleneck constraints; Step 11: Install and deploy 28 sensors to collect multi-source time-series monitoring data related to the degradation trend of industrial equipment throughout its entire lifecycle; process the data using a sliding window to obtain training and testing samples with dimensions of [missing information]. 60 is the window length for each sample. It is the set of real numbers; Step 12, global representation after dimensionality reduction using multivariate Gaussian prior constraints. The learning process will transform the low-dimensional global representation The mapping is to a parameterized standard multivariate Gaussian distribution, where the mean vector is... Used to construct the covariance matrix The vector is ; Step 13: Calculate the KL divergence loss constraint mean vector using a predefined multivariate Gaussian prior. and diagonal matrix Learning; given sample dimensions and predicted output To maximize mutual information At the same time, minimize mutual information Construct an objective function to optimize information bottlenecks; Step 14: Using the global representation model constructed in steps 11 and 12 and the optimization objective defined in step 13, model training can automatically execute feature extraction, fusion, parameterization, and resampling processes to obtain a low-dimensional and information-rich global representation. ; Step 2: Construct a collaborative enhancement representation based on a block-based interaction mechanism; Step 21: Process the shallow features of the multi-source sensor after embedding layer processing. Divide into blocks; along the second dimension The blocks are uniformly divided into 32 sub-feature blocks, represented as follows: (4)； Wherein, the dimension of each sub-feature block is , Along the second dimension; Step 22, global representation Matrix obtained after broadcasting Each sub-feature block obtained in step 21 With matrix By combining the elements, an enhanced interactive input can be formed: (5)； Step 23, combine each spliced feature Input is fed into a shared linear interaction module, and the output is the interacted sub-features. ; Step 24: Output the 32 sub-features after interactive processing in Step 23. Reassemble along the feature dimensions to restore the complete feature tensor: (6)； Step 3: Construct a temporal representation based on manifold learning of off-diagonal covariance prior constraints; Step 31: Convert the low-dimensional global representation output by the attention pooling module The parameterization is performed as a Gaussian mixture model with three components, resulting in a low-dimensional global representation. The probability distribution; Step 32: The low-dimensional global representation is transformed through a parameterization layer. Mapped to an output vector with a total dimension of 135. ; Step 33, the output vector obtained in step 32 The complete posterior Gaussian mixture model distribution is defined according to the preset weight dimension 3, mean dimension 24, and off-diagonal covariance matrix dimension 108; the KL divergence between the posterior Gaussian mixture model and the prior Gaussian mixture model estimated by manifold learning is calculated according to the method described in step 31; and a low-dimensional global representation containing rich temporal degradation information is obtained by sampling from the posterior Gaussian mixture model. ; Step 4: Establish the weighted loss function; Step 41, calculate Huber's losses. and loss Weighted portfolio for robust regression loss The expression is: (8)； Step 42, establish the KL divergence regularization loss based on the Gaussian mixture model, the expression of which is: (9)； in, To represent the first output from the model parameterization layer Posterior mixture weights of Gaussian components; , These are the 1st, 2nd, and 3rd digits of the posterior distribution. The mean vector and covariance matrix of each Gaussian component; , The first and second prior distributions estimated based on manifold learning are respectively the first and second... The mean vector and covariance matrix of each Gaussian component; Step 43: Add the robust regression loss from Step 41 to the KL divergence regularization loss term based on the Gaussian mixture model from Step 42 to obtain the total loss function, expressed as: (10)； Step 5, prediction of remaining useful life; The features obtained after encoding in steps 1, 2, and 3 Input the decoder network and output the remaining lifetime prediction value; after continuous iteration and optimization of the total loss function in step 4, output the accurate remaining lifetime prediction value.
2. The high-dimensional hierarchical time series modeling method for predicting the remaining life of industrial rotating equipment according to claim 1, characterized in that, In step 11, a feature extraction module containing an embedding layer, a convolutional layer, and a non-linear activation function is used to extract features from each sample, resulting in a feature dimension of [missing value]. The deep features are extracted from each sensor, with 128 representing the feature dimensions of each sensor after feature extraction. Attention pooling is used to fuse the deep features from these 28 sensors at the feature level. For a dimensionality-reduced global representation , where 8 is the predefined low-dimensional global representation dimension.
3. The high-dimensional hierarchical time series modeling method for predicting the remaining life of industrial rotating equipment according to claim 1, characterized in that, In step 13, the objective function for optimizing the information bottleneck is constructed as follows: (1)； in To weigh and Weight of importance For For input, These are the weights of the output deep network model; Mutual information is derived based on the nonnegativity of KL divergence. The lower bound and mutual information The upper bound is shown below: (2)； (3)； in, This represents the input data for the model; The label to be predicted. This represents the compressed representation of the data after passing through the model encoder. express and The true joint distribution, Represents conditional distribution The variational approximation, express and The true joint distribution, Indicates a given hour The true posterior distribution, Indicates marginal distribution Variational approximation; Substituting the two variational boundaries mentioned above into the original objective function yields a parameterized multivariate Gaussian distribution. Random sampling is then performed from this distribution to generate a low-dimensional and information-rich global representation for subsequent feature learning and prediction. .
4. The high-dimensional hierarchical time series modeling method for predicting the remaining life of industrial rotating equipment according to claim 1, characterized in that, In step 31, low-dimensional global representation The probability distribution is as follows: (7)； in, For the first The mixing weights of the Gaussian components satisfy the following conditions: ; For the first The mean vector of Gaussian components; For the first The covariance matrix of Gaussian components; For the first Each multivariate Gaussian component.
5. The high-dimensional hierarchical time series modeling method for predicting the remaining life of industrial rotating equipment according to claim 1, characterized in that, First, a global representation based on information bottleneck constraints is constructed. Second, a collaborative enhancement representation based on a block interaction mechanism is constructed. Third, a time series representation based on manifold learning off-diagonal covariance prior constraints is constructed. Finally, the remaining useful life is predicted by optimizing the total loss function. Steps 1 to 5 are executed iteratively to realize the representation modeling of multi-source high-dimensional time series data, and finally obtain the remaining useful life prediction results with lower error and higher efficiency.