Equipment residual life prediction data preprocessing method based on multi-scale dynamic label re-calibration

Abstract

The invention provides an equipment residual life prediction data preprocessing method based on multi-scale dynamic label re-calibration. The method comprises the following steps: S1, multi-scale time delay embedding; s2, kernel principal component analysis health index construction; s3, zero-phase finite impulse response filtering is carried out; s4, performing multi-scale HI fusion and dynamic slope calibration; and S5, dynamic segmentation RUL label reconstruction is carried out. According to the method, through multi-scale time delay embedding, the performance of degradation dynamics on different time granularities is captured, and the information content of health indexes is enriched. KPCA and zero-phase FIR filtering are combined, a smooth and continuous health index curve capable of reflecting nonlinear degradation essence is effectively extracted, and a foundation is laid for accurate FPT recognition. According to the method, the FPT is calibrated and positioned through the dynamic slope, and the segmented RUL tag is reconstructed according to the FPT, so that the limitation of a fixed threshold method is overcome, and the tag can more truly reflect the degradation behavior of individual equipment.

Description

Technical Field

[0001] This invention belongs to the field of industrial data preprocessing and machine learning label optimization technology, specifically involving a data preprocessing method for predicting the remaining life of equipment based on multi-scale dynamic label recalibration. Background Technology

[0002] In data-driven equipment remaining lifetime (RUL) prediction, the accuracy of the labels directly determines the model's performance. Traditional methods often use fixed thresholds (such as uniformly defining the 125 cycles before failure as the degradation starting point) to define RUL labels, ignoring the significant differences in degradation rates and patterns among individual devices. This "one-size-fits-all" labeling strategy leads to label distribution bias, meaning the labels cannot accurately reflect the true degradation state of the equipment, thus misleading model learning and limiting its prediction accuracy.

[0003] Existing technologies attempt to identify the starting point of degradation by constructing health indicators and detecting change points, but they mostly rely on single-scale health indicators and global thresholds, which do not adequately consider individual differences in equipment and multi-scale dynamic characteristics in the degradation process, making it difficult to accurately capture the turning point of the degradation stage. Summary of the Invention

[0004] The purpose of this invention is to provide a method for preprocessing equipment remaining life prediction data based on multi-scale dynamic tag recalibration.

[0005] A preprocessing method for equipment remaining lifetime prediction data based on multi-scale dynamic label recalibration includes the following steps: S1, multi-scale time delay embedding; S2, Construction of health indicators using principal component analysis; S3, zero-phase finite impulse response filter; S4, Multi-scale HI fusion and dynamic slope calibration; S5, dynamic segmented RUL tag reconstruction.

[0006] Furthermore, S1 specifically refers to: Set a set of delay parameters And the embedding dimension d, the observation vector at each time point t in the original sequence. Perform parallel multi-scale embedding to construct a high-dimensional phase space vector: Where k represents the different latency scales of the index.

[0007] Furthermore, in S1 when When N is the total sequence length, to ensure consistent sequence length after reconstruction, an end-padding strategy is adopted, using the last observation value. Filling missing windows, this operation is defined as: Further, S2 specifically involves: independently performing kernel principal component analysis on the embedded data at each scale obtained in S1. The kernel principal component analysis is performed through a nonlinear mapping function. The data is projected into a high-dimensional feature space, and then linear PCA is performed in that space using a kernel function. Implicitly compute the inner product in the high-dimensional feature space, avoiding the complexities of explicit computation. .

[0008] Furthermore, the sample The projection onto the k-th principal component is given by the following equation: in, To determine the normalized eigenvector corresponding to the k-th largest eigenvalue of the kernel matrix, the first principal component with the highest variance explanation rate is selected. A health indicator curve that characterizes the health status of equipment.

[0009] Furthermore, the filter output in S3 From the input signal With filter impulse response The convolution sum is given as: Furthermore, in S3, a Type I FIR filter that maintains strictly linear phase in both the full passband and stopband is preferably used.

[0010] Further, S4 includes: calculating the variance of each scale HI curve over its entire life cycle, using this variance as a weight to perform a linear weighted average of multiple HI curves to obtain a fused HI curve; calculating the first-order numerical derivative of the fused HI curve using the central difference method, searching for the point with the largest slope change during the equipment life cycle, and identifying this point as FPT.

[0011] Furthermore, in S5, prior to FPT, the device is in a healthy phase where performance degradation is negligible; therefore, the RUL tag maintains a constant maximum value during this phase (e.g., Starting from the FPT point, the device enters the accelerated degradation phase, and the RUL tag decreases linearly to 0 over time. This process is represented as: The resulting label sequence is the dynamically reconstructed RUL label, which is used for training the subsequent prediction model.

[0012] The beneficial effects of this invention are as follows: 1. This invention captures the dynamics of degradation at different time granularities through multi-scale time delay embedding, enriching the information content of health indicators. Combined with KPCA and zero-phase FIR filtering, it effectively extracts smooth, continuous health indicator curves that reflect the nonlinear nature of degradation, laying the foundation for accurate FPT identification.

[0013] 2. This invention uses dynamic slope calibration to locate the FPT and reconstructs the segmented RUL tag accordingly, overcoming the limitations of the fixed threshold method and enabling the tag to more accurately reflect the degradation behavior of individual devices. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the multi-scale time delay embedding of the present invention; Figure 3 This is a schematic diagram of the HI curve extracted by KPCA in this invention; Figure 4 This is a schematic diagram comparing the HI curves before and after FIR filtering in this invention; Figure 5 This is a schematic diagram of the multi-scale HI fusion curve of the present invention; Figure 6 This is a schematic diagram of the HI slope curve and FPT recognition of the present invention; Figure 7 This is a schematic diagram comparing the dynamic label reconstruction and fixed threshold label of the present invention; Figure 8(a) shows the effect of the present invention in dividing the life cycle by the identified FPT; Figure 8(b) is a comparison chart A between the dynamically reconstructed label of the present invention and the traditional fixed threshold label; Figure 8(c) is a comparison of the dynamically reconstructed label of the present invention and the traditional fixed threshold label. Detailed Implementation

[0015] The present invention will now be further described with reference to the accompanying drawings.

[0016] Figure 1 This is a flowchart illustrating the method of the present invention, showing the complete steps from raw data to reconstructed labels. The following embodiments use the prediction of remaining service life of aero-engines as an application scenario, and take the FD001 subset of the C-MAPSS dataset as an example for explanation.

[0017] A preprocessing method for equipment remaining lifetime prediction data based on multi-scale dynamic label recalibration includes the following steps: S1, Data preparation and sensor selection, multi-scale time delay embedding; like Figure 1As shown, the input to this method is multi-sensor time-series data recorded during equipment operation; taking the NASA C-MAPSS dataset as an example, this dataset contains simulation data of turbofan engines under various failure modes; firstly, Pearson correlation coefficients are calculated for all 21 sensor signals and remaining lifetime (RUL) labels in the dataset, and the analysis results are as follows. Figure 2 As shown. Based on this analysis, sensor channels significantly related to the degradation process are selected as inputs for subsequent processing; in this embodiment, the ultimately selected sensors include sensors numbered 2, 3, 4, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20, and 21, totaling 14 channels. This step aims to eliminate redundant or irrelevant sensor signals and focus on key degradation information.

[0018] Reference Figure 1 and Figure 3 This step aims to capture the dynamic characteristics of device degradation at different time scales, based on the Takens embedding theorem. The specific implementation is as follows: Parameter settings: Set the embedding dimension To construct a multi-scale representation, a set of delay parameters is defined. This means that four embedding sequences from different time perspectives will be constructed in parallel. Embedding operation: for each time point 14-dimensional sensor reading vector (in Embedding vectors of four scales are constructed in parallel according to the following formula: in, Indexes have different latency scales. For example... Figure 3 As shown, the embedding sequences at different scales reflect the data trajectories at different time granularities; Sequence boundary handling: when When N is the total length of the sequence, an end-padding strategy is used, employing the last observation. Pad off any excess portion to ensure that all embedded sequence lengths match the original sequence. This operation is defined as: S2, Construction of health indicators based on kernel principal component analysis: such as Figure 1 and Figure 4 As shown, this step aims to extract a one-dimensional index from the aforementioned multi-scale, high-dimensional nonlinear embedded data that can smoothly and continuously represent the overall health status of the equipment. The specific steps are as follows: Method Principle: Given that equipment degradation processes typically exhibit strong nonlinearity, linear principal component analysis (PCA) may struggle to effectively extract the dominant degradation trend. Therefore, this invention employs kernel principal component analysis (KPCA). KPCA utilizes a nonlinear mapping function... The data is projected into a high-dimensional feature space (Reproducing Kernel Hilbert Space, RKHS), and then linear PCA is performed in that space, which enables the capture of nonlinear features; Specific calculations: For the embedded data obtained in S1 at each scale (e.g., corresponding to...) sequence Perform KPCA independently; first, calculate the kernel matrix. Its elements The kernel function is given by the kernel function; in this example, the radial basis function (RBF) is chosen. Then, the eigenvalues ​​of the kernel matrix are solved to obtain the eigenvectors. ;sample In the The projections onto the principal components are given by the following equation: Health indicator extraction: Select the first principal component with the largest variance explained. As a health indicator (HI) curve at this scale; such as Figure 4 As shown, the HI curves extracted by KPCA can roughly reflect the degradation trend of the equipment, but may still contain noise and fluctuations; this step generates an HI curve for each scale.

[0019] S3, Zero-phase Finite Impulse Response Filter: (e.g., ...) Figure 1 and Figure 5 As shown, this step aims to smooth the HI curve obtained in step 3, suppress high-frequency noise, and strictly maintain the phase of the signal (i.e., the timing relationship), laying the foundation for subsequent accurate slope calculation. The specific steps are as follows: Filter Selection and Design: This invention selects a Finite Impulse Response (FIR) filter, specifically a Type I FIR low-pass filter. An FIR filter is a non-recursive linear time-invariant system. Its greatest advantage is that it can be designed to have strictly linear phase characteristics. This means that all frequency components of the filtered signal have the same delay, preventing phase distortion and thus strictly preserving the timing shape of the degradation process. Based on engineering experience, a filter of order I is designed... (i.e., filter length) The low-pass filter has a cutoff frequency set to a value that can effectively filter out high-frequency noise but still retain the main degradation trend.

[0020] Filtering operation: Use each HI curve obtained from S2 as the input signal. The impulse response of the designed FIR filter Perform discrete convolution operation. The filter output. Calculated by the following formula: Figure 5 The comparison of the HI curves before and after filtering is visually demonstrated. It can be seen that the filtered curve becomes smoother, effectively removing high-frequency noise.

[0021] S4, multi-scale health indicator fusion, dynamic slope calibration and failure inflection point identification; like Figure 1 and Figure 6 As shown, this step aims to comprehensively utilize the complementary information contained in HI curves at different scales to generate a more robust and representative comprehensive HI curve. Fusion Strategy: Calculate the variance of each filtered HI curve over its entire lifetime. The larger the variance, the more significant the dynamic changes of the curve, and the richer the degradation information it may contain. Use this variance as a weight to perform a linear weighted average of the four filtered HI curves (corresponding to four delay scales). Effect: such as Figure 6 As shown, the HI curve obtained by weighted fusion (Fused HI) is smoother than any single-scale HI curve and can better reflect the overall degradation trend, reducing the impact of local fluctuations. like Figure 1 and Figure 7 As shown, this step (dynamic slope calibration and failure inflection point identification) aims to accurately locate the inflection point (FPT) from a healthy state to an accelerated degradation state of the device based on the fused HI curve. The main contents are as follows: Slope Calculation: For the fused HI curve, the first-order numerical derivative (slope) is calculated using the central difference method. The central difference method has higher computational accuracy than forward or backward differencing. The calculation formula is as follows: A smooth degradation rate curve (HI Slope) is calculated, as shown below. Figure 7 As shown.

[0022] FPT Identification: The transition of equipment from a healthy state to a degraded state is usually accompanied by a significant increase in the degradation rate, i.e., a peak in the slope of the HI curve. Empirically, the search range for FPT is limited to 35% to 55% of the total lifespan; within this range, the global maximum point on the slope curve is identified as the Failure Tolerance Point (FPT).

[0023] S5, Dynamic Segmented Remaining Lifetime Label Reconstruction: (e.g., ...) Figure 1As shown in Figure 8, this step, based on the identified FPT, reconstructs a physically meaningful and individualized RUL tag for each device unit to replace the traditional, inaccurate fixed threshold tag; the specific steps are as follows: Refactoring Strategy: Healthy Phase (from start of operation to FPT point): During this phase, device performance degradation is considered negligible, and the RUL tag maintains a constant maximum value, denoted as... In the general settings of the C-MAPSS dataset, take... Degradation phase (from FPT point to failure point): RUL from It begins to decrease linearly to 0.

[0024] Mathematical representation: The reconstructed RUL tag function can be represented as: Performance Comparison: Figure 8(a) shows the effect of lifecycle segmentation based on the identified FPT. Figures 8(b) and 8(c) compare dynamic reconstruction labels with traditional fixed threshold labels (e.g., linearly decreasing from cycle 125). The inflection point of the dynamic labels coincides with the physical degradation starting point revealed by the HI curve, while the fixed labels may deviate significantly from reality and fail to reflect individual differences.

[0025] Method Validation: The effectiveness of the method in this invention was systematically validated on the C-MAPSS and N-CMAPSS datasets. The dataset with reconstructed labels using the method of this invention (denoted as G+M in the ablation study) was compared with the dataset using the original fixed labels (denoted as G), and both were used to train the same prediction model. Key experimental results are shown in the table below: Table 1. Performance improvements (RMSE and Score) of the label recalibration method on various subsets of C-MAPSS

[0026] Data shows that after applying the preprocessing method of this invention, the RMSE of the prediction model decreased by an average of 11.72%, and the score decreased by an average of 20.74%, significantly improving prediction accuracy and engineering applicability. This demonstrates that the method of this invention can effectively improve the performance of subsequent prediction models by providing supervision signals that are more consistent with the physical degradation process.

[0027] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for preprocessing equipment remaining life prediction data based on multi-scale dynamic label recalibration, characterized in that, Includes the following steps: S1, multi-scale time delay embedding; S2, Construction of health indicators using principal component analysis; S3, zero-phase finite impulse response filter; S4, Multi-scale HI fusion and dynamic slope calibration; S5, dynamic segmented RUL tag reconstruction.

2. The method for preprocessing equipment remaining life prediction data based on multi-scale dynamic label recalibration according to claim 1, characterized in that, Specifically, S1 is: Set a set of delay parameters And the embedding dimension d, the observation vector at each time point t in the original sequence. Perform parallel multi-scale embedding to construct a high-dimensional phase space vector: Where k represents the different latency scales of the index.

3. The method for preprocessing equipment remaining life prediction data based on multi-scale dynamic label recalibration according to claim 2, characterized in that, When in S1 When N is the total sequence length, to ensure consistent sequence length after reconstruction, an end-padding strategy is adopted, using the last observation value. Filling missing windows, this operation is defined as:

4. The method for preprocessing equipment remaining life prediction data based on multi-scale dynamic label recalibration according to claim 3, characterized in that, Specifically, S2 involves independently performing kernel principal component analysis on the embedded data at each scale obtained in S1. Kernel principal component analysis is performed through a nonlinear mapping function. The data is projected into a high-dimensional feature space, and then linear PCA is performed in that space using a kernel function. Implicitly compute the inner product in the high-dimensional feature space, avoiding the complexities of explicit computation. .

5. The method for preprocessing equipment remaining life prediction data based on multi-scale dynamic label recalibration according to claim 4, characterized in that, The sample The projection onto the k-th principal component is given by the following equation: in, To determine the normalized eigenvector corresponding to the k-th largest eigenvalue of the kernel matrix, the first principal component with the highest variance explanation rate is selected. A health indicator curve that characterizes the health status of equipment.

6. The method for preprocessing equipment remaining life prediction data based on multi-scale dynamic label recalibration according to claim 1, characterized in that, The filter output in S3 From the input signal With filter impulse response The convolution sum is given as:

7. The method for preprocessing equipment remaining life prediction data based on multi-scale dynamic label recalibration according to claim 6, characterized in that, The S3 preferably uses a Type I FIR filter that maintains a strictly linear phase in both the full passband and stopband.

8. A method for preprocessing equipment remaining life prediction data based on multi-scale dynamic label recalibration as described in claim 7, characterized in that, S4 includes: calculating the variance of each scale HI curve over its entire life cycle, using this variance as a weight to perform a linear weighted average of multiple HI curves to obtain a fused HI curve; calculating the first-order numerical derivative of the fused HI curve using the central difference method, searching for the point with the largest slope change during the equipment life cycle, and identifying this point as the FPT.

9. A method for preprocessing equipment remaining life prediction data based on multi-scale dynamic label recalibration as described in claim 8, characterized in that, In S5, prior to FPT, the device is in a healthy phase where performance degradation is negligible; therefore, the RUL tag maintains a constant maximum value during this phase (e.g., Starting from the FPT point, the device enters the accelerated degradation phase, and the RUL tag decreases linearly to 0 over time. This process is represented as: The resulting label sequence is the dynamically reconstructed RUL label, which is used for training the subsequent prediction model.

Images