A life prediction method for multi-stage degradation process cross-domain alignment

Abstract

The application discloses a kind of life prediction methods for multi-stage degradation process cross-domain alignment, comprising the following steps: constructing time sample window, extracting the window level latent degradation feature of the time sample window using time convolution network;Based on the extracted latent degradation feature trajectory, turning point detection is carried out, and the stage boundary of the degradation process is identified;Using the Transformer encoder, the high-dimensional time sequence feature modeling of the original monitoring data sequence is carried out;The Transformer encoding features of different device individuals are grouped according to the degradation stage, and the maximum mean difference constraint is introduced within the same degradation stage to realize the alignment of the cross-domain feature distribution in each stage;The Transformer encoding features after alignment are input into the prediction module, and the remaining useful life prediction result of the object to be tested is output. It can avoid the misalignment problem between samples in different degradation stages, improve the life prediction accuracy and generalization ability of the model under complex working conditions, individual differences and unseen objects, and has good engineering application value.

Description

Technical Field

[0001] This invention belongs to the field of mechanical remaining life prediction technology, specifically relating to a life prediction method for cross-domain alignment of multi-stage degradation processes. Background Technology

[0002] Under complex operating conditions and long-term service, mechanical equipment and its critical components will inevitably experience performance degradation, eventually leading to functional failure. Accurate prediction of their remaining service life is crucial for ensuring safe equipment operation, optimizing maintenance strategies, and reducing operating costs.

[0003] Existing data-driven lifetime prediction methods typically rely on the assumption that training and test data satisfy similar distributions. However, in practical applications, due to the combined effects of manufacturing errors, assembly differences, operating condition fluctuations, load variations, and random degradation behavior, significant distributional shifts often exist between different individual devices, resulting in insufficient generalization prediction capabilities of existing models in cases where no individual devices have been observed or in cross-operating condition scenarios. Simultaneously, many degrading objects typically exhibit distinct stage-specific characteristics during their lifetime evolution, with significant differences in statistical properties, feature distributions, and degradation rates corresponding to different degradation stages.

[0004] If this stage correlation is ignored during the modeling process and the overall feature distribution is directly aligned uniformly, it is easy to cause misalignment mapping of samples at different degradation stages in the feature space, thereby weakening the model's ability to represent consistent degradation patterns and affecting the accuracy of lifetime prediction.

[0005] Therefore, how to construct a lifespan prediction method that can simultaneously perceive the characteristics of the degradation stage and achieve cross-individual feature distribution alignment within the same degradation stage has become a key problem that urgently needs to be solved in the current related fields. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a lifetime prediction method for cross-domain alignment of multi-stage degradation processes, which addresses the shortcomings of the prior art. This method solves the problem that the prior art is prone to causing misalignment mapping of samples at different degradation stages in the feature space, thereby weakening the model's ability to represent consistent degradation patterns and affecting the accuracy of lifetime prediction.

[0007] This invention adopts the following technical solution: a lifetime prediction method for cross-domain alignment of multi-stage degradation processes, comprising the following steps: The original monitoring data sequence of the object under test is segmented by a sliding time window to construct a time-series sample window, and a temporal convolutional network is used to extract the window-level potential degradation features of the time-series sample window; Turning point detection is performed based on the extracted potential degradation feature trajectories, the stage boundaries of the degradation process are identified, multiple continuous degradation stages are divided, and each time series sample window is assigned a corresponding degradation stage label. A Transformer encoder is used to model the high-dimensional temporal features of the original monitoring data sequence to obtain Transformer encoded features. Based on the degradation stage label, the Transformer encoded features of different devices are grouped according to the degradation stage. Within the same degradation stage, the maximum mean difference constraint is introduced to achieve cross-domain feature distribution alignment step by step. The aligned Transformer encoded features are input into the prediction module, which outputs the prediction result of the remaining useful life of the object under test.

[0008] Preferably, the temporal convolutional network adopts a causal convolutional structure to satisfy causality constraints, extracts potential temporal features in the degradation process through dilated convolution and residual connection structures, and outputs window-level potential degradation features. When extracting potential degradation features, it only relies on current and historical monitoring data.

[0009] Preferably, the inflection point detection is a kernel function-based inflection point detection, which determines the boundary of the degradation stage by identifying structural changes in potential degradation feature trajectories.

[0010] Preferably, the Transformer encoder is composed of multiple stacked Transformer blocks. Each Transformer block includes a multi-head self-attention layer, a position feedforward network, and residual connections and layer normalization modules in each sub-layer.

[0011] Preferably, when the Transformer encoder performs high-dimensional temporal feature modeling on the original monitoring data sequence, it first performs layer normalization on the original monitoring data sequence, and then processes it through a multi-head self-attention layer and a position feedforward network to obtain the Transformer encoded features.

[0012] Preferably, the step-by-step cross-domain feature distribution alignment is as follows: calculate the maximum mean difference of individual coding features of different devices in each degradation stage, and sum the maximum mean differences of all stages to obtain the total loss of step-by-step cross-domain alignment.

[0013] Preferably, the prediction module uses a two-layer fully connected network as the prediction head, with the first layer used to compress and encode the feature dimension and the second layer used to output the remaining lifetime scalar.

[0014] Preferably, during model training, the weighted sum of the remaining lifetime prediction loss and the total loss of phased cross-domain alignment is used as the total loss function, and the joint optimization of the model is completed by optimizing the total loss.

[0015] Preferably, the weighting relationship between lifetime prediction loss and total cross-domain alignment loss in stages is adjusted by a preset balance coefficient.

[0016] Preferably, the degradation stage is divided into three stages: early slow degradation, mid-term accelerated degradation, and late rapid degradation.

[0017] Compared with the prior art, the present invention has at least the following beneficial effects: This application first extracts potential degradation features through sliding time window segmentation and temporal convolutional networks, providing a reliable basis for degradation stage identification and avoiding noise interference from direct modeling of raw data. Based on feature trajectory detection of inflection points, degradation stages are divided and labeled to actively perceive the stage-specific nature of degradation, addressing the problems of ignoring stage correlations and sample misalignment. A Transformer encoder is used to model high-dimensional temporal features, enhancing the expressive power of degradation temporal dependencies. Stage-by-stage cross-domain alignment is performed within the same degradation stage according to stage labels, addressing the deficiency of individual device distribution offsets and eliminating distribution differences across individuals and operating conditions. Based on the aligned features, lifetime prediction results are output, simultaneously improving the model's generalization ability and prediction accuracy, solving the dual key problems of insufficient generalization and inaccurate prediction in existing models.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the following description of the relative embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the method in an embodiment of the present invention; Figure 2 This is a prediction of the remaining service life of an example harmonic reducer at each time step throughout its entire life cycle, as shown in this embodiment of the invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present 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 the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "one side," "one end," and "one side," etc., indicating the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0023] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0024] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0025] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0026] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0027] The accompanying drawings illustrate various structural schematic diagrams according to embodiments disclosed in this invention. These drawings are not to scale, and some details have been enlarged for clarity, and some details may have been omitted. The shapes of the various regions and layers shown in the drawings, as well as their relative sizes and positional relationships, are merely exemplary and may deviate from reality due to manufacturing tolerances or technical limitations. Furthermore, those skilled in the art can design regions / layers with different shapes, sizes, and relative positions as needed.

[0028] This invention provides a lifetime prediction method for cross-domain alignment in multi-stage degradation processes. After extracting window-level potential degradation features using a temporal convolutional network, transition point detection is performed based on the potential degradation feature trajectories to identify degradation stage boundaries and complete stage division. Then, a Transformer encoder is used to model high-dimensional features of the original monitoring data sequence, and a maximum mean difference constraint is introduced within the same degradation stage for different individuals to achieve stage-by-stage cross-domain feature alignment. Finally, the remaining lifetime prediction result is output based on the aligned features, significantly improving the model's domain generalization ability.

[0029] Please see Figure 1 As shown, the present invention provides a lifetime prediction method for cross-domain alignment of multi-stage degradation processes, comprising the following steps: S1. Perform stage-based perception degradation modeling. The original monitoring data sequence of the object under test is segmented by a sliding time window to construct a time-series sample window. A temporal convolutional network is used to extract the window-level potential degradation features of the time-series sample window. The original monitoring data sequence is the full lifecycle sensing monitoring sequence of key components of mechanical equipment. The sliding time window segmentation involves continuously segmenting the monitoring sequence according to a fixed window length.

[0030] Specifically, it includes the following: S11. Obtain the original monitoring data sequence of the object to be tested, and segment the original monitoring data using a sliding time window method to construct a series of sample windows arranged in chronological order. Let the first... Each sample window is represented as:

[0031] in, Indicates the length of the time window; This indicates that the original monitoring data sequence is in the 1st... The input sample corresponding to each time point.

[0032] S12. A Temporal Convolutional Network (TCN) is used to learn the potential degradation features of the monitoring data. The TCN uses a causal convolutional structure to satisfy causal constraints. It extracts the potential temporal features in the degradation process through dilated convolution and residual connection structures, and outputs window-level potential degradation features. When extracting potential degradation features, it only relies on current and historical monitoring data.

[0033] Specifically, let the mapping relationship between the input sequence and the output sequence be:

[0034] Where, f( ) represents the overall mapping function of TCN. Represents the input sequence. This indicates the output sequence.

[0035] To satisfy the causality constraint, TCN at time... The output of depends only on the current and historical inputs, and its single-layer causal convolution can be represented as:

[0036] in, Indicates the kernel size. Indicates the expansion factor. This represents the convolution kernel parameters.

[0037] By using dilated convolution and residual connection structures, latent temporal features in the degradation process are extracted, and corresponding latent degradation feature representations are formed.

[0038] S2. Based on the extracted potential degradation feature trajectories, inflection point detection is performed to identify the stage boundaries of the degradation process, complete the degradation stage division, and assign corresponding degradation stage labels to each time series sample window.

[0039] Among them, the turning point detection is a kernel function-based turning point detection method. It determines the boundary of the degradation stage by identifying structural changes in the potential degradation feature trajectory, and divides the degradation process into different degradation stages.

[0040] In this embodiment, a three-stage degradation process is used as an example for modeling: an early slow degradation stage, a mid-term accelerated degradation stage, and a late rapid degradation stage. Each sample window is assigned a corresponding stage label.

[0041] S3. Use a Transformer encoder to model the high-dimensional time-series features of the original monitoring data sequence to obtain Transformer encoded features.

[0042] The Transformer encoder consists of multiple stacked Transformer blocks. Each Transformer block contains a multi-head self-attention layer, a position feedforward network, and residual connections and layer normalization modules in each sub-layer.

[0043] When the Transformer encoder performs high-dimensional temporal feature modeling on the original monitoring data sequence, it first performs layer normalization on the original monitoring data sequence, and then processes it through a multi-head self-attention layer and a position feedforward network to obtain the encoded features.

[0044] The specific implementation steps are as follows: S31. After completing the stage-aware degradation modeling, construct a Transformer-based encoder to learn high-dimensional feature representations from the original monitoring data sequence.

[0045] In the specific implementation, the original monitoring data sequence is first subjected to layer normalization; then the processed features are input into a multi-head self-attention layer, and the outputs of each attention head are concatenated and linearly transformed to obtain the final feature representation. For the ... A Transformer block, whose attention sublayer can be represented as:

[0046] in, , , These represent the query matrix, key matrix, and value matrix, respectively. This represents the dimension of the key vector. After processing by a multi-head self-attention and feedforward network, the corresponding Transformer encoded features are obtained, which serve as the input to the subsequent prediction module.

[0047] S4. Based on the degradation stage label, the Transformer coding features of different devices are grouped according to the degradation stage. The maximum mean difference constraint is introduced within the same degradation stage to achieve cross-domain feature distribution alignment step by step.

[0048] The step-by-step cross-domain feature distribution alignment involves calculating the maximum mean difference of the coding features of different devices in each degradation stage, and summing the maximum mean differences of all stages to obtain the total loss of the step-by-step cross-domain alignment.

[0049] Specifically, after completing the Transformer-based feature encoding, each feature is associated with the stage label corresponding to its respective degradation stage.

[0050] Suppose that the feature sets of two different individuals in the source domain are respectively and Based on stage labels, it can be and Each stage is divided into multiple subsets, with each subset corresponding to a degeneracy stage, as detailed below:

[0051] in, and These respectively represent different individuals in the source domain belonging to the first... A set of features for each stage of degradation.

[0052] In real-world degradation scenarios, different individuals often exhibit significant distributional differences at different degradation stages. To address this, a cross-domain, stage-by-stage feature alignment mechanism is adopted, which aligns the feature distributions of different individuals within each degradation stage.

[0053] Specifically, the maximum mean discrepancy (MMD) is used as the alignment criterion. For the first... Each degradation stage has a staged alignment loss defined as follows:

[0054] in, Represents the kernel mapping function. This represents the regenerated nucleus Hilbert space.

[0055] Furthermore, the MMD losses from all degradation stages are summed to obtain the total staged cross-domain alignment loss:

[0056] By minimizing the maximum mean difference within the same degradation stage of different device individuals, feature distribution alignment across stages can be achieved, thereby improving the model's generalization ability for lifetime prediction under complex operating conditions and unseen individual conditions.

[0057] S5. Lifetime Prediction Output: Input the aligned Transformer encoded features into the prediction module and output the prediction result of the remaining useful life (RUL) of the object under test.

[0058] In this embodiment, the prediction module uses a two-layer fully connected network as the prediction head. The first layer is used to compress and encode the feature dimension, and the second layer is used to output the RUL scalar.

[0059] Specifically, let the predicted result output by the model be... The corresponding real lifespan label is Then the lifetime prediction loss can be expressed as:

[0060] in, This represents the total number of training samples.

[0061] Furthermore, during model training, the total loss function is a weighted sum of the remaining lifetime prediction loss and the total loss of staged cross-domain alignment. The model is then jointly optimized by optimizing the total loss. Specifically:

[0062] in, This represents the balance coefficient, used to adjust the weighting relationship between the lifetime prediction task and the phased cross-domain alignment task.

[0063] This embodiment uses accelerated life test data of a harmonic reducer to verify the proposed method. The experiment utilizes an OMRON NJ301-1100 programmable logic controller (PLC) to acquire the internal current signal of the harmonic reducer in real time throughout its entire life cycle. The sampling frequency is set to 2 kHz, and the single sampling duration is 1 s. The specific operation is as follows: For the collected raw monitoring data sequence, a sample window arranged in chronological order is first constructed using a sliding time window method, and then input into a temporal convolutional network (TCN) to extract window-level potential degradation features. Subsequently, in the potential degradation feature space, kernel-based inflection point detection is combined to identify the degradation stage boundaries, thus completing the stage division of the degradation process.

[0064] This embodiment uses a three-stage degradation process as an example for modeling: early slow degradation, mid-stage accelerated degradation, and late rapid degradation. Based on this, the original monitoring data sequence is input into a Transformer encoder to learn a high-dimensional temporal feature representation. The encoder consists of two stacked Transformer blocks. Each Transformer block contains a multi-head self-attention layer and a position-feedforward network with six attention heads. Residual connections and layer normalization modules are incorporated into each sub-layer to enhance the model's ability to express temporal dependencies.

[0065] The encoder hidden layer dimension is set to 64, and the activation function is ReLU. Based on the aforementioned stage segmentation results, each sample window is assigned a corresponding stage label, and the maximum mean difference (MMD) constraint is introduced within the same degradation stage according to the stage label to achieve cross-individual feature distribution alignment.

[0066] Finally, the aligned features are input into a prediction head consisting of two fully connected layers. The first layer is used to compress the features, with a hidden feature dimension of 32, and the second layer is used to output the remaining lifespan prediction result, with a hidden feature dimension of 1.

[0067] During model training, the batch size was set to 128, the learning rate to 1e-3, the number of training epochs to 100, and the loss function weight factor... The optimizer is set to 1e-2, and Adan is used as the optimizer. The total loss function is jointly optimized by weighting the lifetime prediction loss and the staged cross-domain alignment loss.

[0068] The above implementation process can effectively verify the lifetime prediction capability of the present invention in multi-stage degradation processes and cross-individual distribution migration scenarios.

[0069] Figure 2 The Remaining Life (RUL) prediction results for each time step throughout the entire lifespan of a harmonic reducer are described. The results show that after extracting window-level potential degradation features using a temporal convolutional network, transition point detection is performed based on the potential degradation feature trajectories to identify degradation stage boundaries and complete stage division. Then, a Transformer encoder is used to model high-dimensional features of the original monitoring data sequence, and a maximum mean difference constraint is introduced within the same degradation stage for different individuals to achieve stage-by-stage cross-domain feature alignment. Finally, based on the aligned features, the remaining lifespan prediction results are output, yielding satisfactory lifespan prediction results. This significantly improves the model's domain generalization ability and outperforms a series of comparative methods in terms of prediction results.

[0070] In summary, this application extracts window-level potential degradation features through a temporal convolutional network, then detects transition points based on the potential degradation feature trajectories to identify degradation stage boundaries and complete stage division. Next, a Transformer encoder is used to model high-dimensional features of the original monitoring data sequence, and a maximum mean difference constraint is introduced within the same degradation stage for different individuals to achieve cross-domain feature alignment across stages. Finally, based on the aligned features, the remaining lifetime prediction results are output, significantly improving the model's domain generalization ability.

[0071] Furthermore, this invention is not limited to the prediction of harmonic reducers (RU) in industrial robots, but can also be extended to other automated equipment and systems.

[0072] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A lifetime prediction method for cross-domain alignment in multi-stage degradation processes, characterized in that, Includes the following steps: The original monitoring data sequence of the object under test is segmented by a sliding time window to construct a time-series sample window, and a temporal convolutional network is used to extract the window-level potential degradation features of the time-series sample window; Turning point detection is performed based on the extracted potential degradation feature trajectories to identify the stage boundaries of the degradation process, divide multiple continuous degradation stages, and assign corresponding degradation stage labels to each time series sample window. A Transformer encoder is used to model the high-dimensional temporal features of the original monitoring data sequence to obtain Transformer encoded features. Based on the degradation stage label, the Transformer encoded features of different devices are grouped according to the degradation stage. The maximum mean difference constraint is introduced within the same degradation stage to achieve cross-domain feature distribution alignment step by step. The aligned Transformer encoded features are input into the prediction module, which outputs the prediction result of the remaining useful life of the object under test.

2. The lifetime prediction method for cross-domain alignment of multi-stage degradation processes according to claim 1, characterized in that, The temporal convolutional network adopts a causal convolutional structure to satisfy causality constraints. It extracts potential temporal features in the degradation process through dilated convolution and residual connection structures, and outputs window-level potential degradation features. When extracting potential degradation features, it only relies on current and historical monitoring data.

3. The lifetime prediction method for cross-domain alignment of multi-stage degradation processes according to claim 1, characterized in that, The inflection point detection is a kernel function-based inflection point detection that determines the boundary of the degradation stage by identifying structural changes in potential degradation feature trajectories.

4. The lifetime prediction method for cross-domain alignment of multi-stage degradation processes according to claim 1, characterized in that, The Transformer encoder is composed of multiple stacked Transformer blocks. Each Transformer block contains a multi-head self-attention layer, a position feedforward network, and residual connections and layer normalization modules in each sub-layer.

5. The lifetime prediction method for cross-domain alignment of multi-stage degradation processes according to claim 4, characterized in that, When the Transformer encoder performs high-dimensional temporal feature modeling on the original monitoring data sequence, it first performs layer normalization on the original monitoring data sequence, and then processes it through a multi-head self-attention layer and a position feedforward network to obtain the Transformer encoded features.

6. The lifetime prediction method for cross-domain alignment of multi-stage degradation processes according to claim 1, characterized in that, The step-by-step cross-domain feature distribution alignment is as follows: calculate the maximum mean difference of individual coding features of different devices in each degradation stage, and sum the maximum mean differences of all stages to obtain the total loss of step-by-step cross-domain alignment.

7. The lifetime prediction method for cross-domain alignment of multi-stage degradation processes according to claim 6, characterized in that, The prediction module uses a two-layer fully connected network as the prediction head. The first layer is used to compress and encode the feature dimension, and the second layer is used to output the remaining lifetime scalar.

8. The lifetime prediction method for cross-domain alignment of multi-stage degradation processes according to claim 7, characterized in that, During model training, the total loss function is a weighted sum of the remaining lifetime prediction loss and the total loss of phased cross-domain alignment. The joint optimization of the model is completed by optimizing the total loss.

9. The lifetime prediction method for cross-domain alignment of multi-stage degradation processes according to claim 8, characterized in that, The weighting relationship between lifetime prediction loss and total loss of phased cross-domain alignment is adjusted by pre-setting a balance coefficient.

10. The lifetime prediction method for cross-domain alignment of multi-stage degradation processes according to claim 1, characterized in that, The degradation process is divided into three stages: early slow degradation, mid-term accelerated degradation, and late rapid degradation.

Images