Degradation trend migration prediction-based cross-equipment variable-working-condition residual life prediction method and system

Abstract

The invention belongs to the field of intelligent operation and maintenance and fault prediction of electromechanical equipment, and discloses a cross-equipment variable-working-condition residual life prediction method and system based on degradation trend migration prediction, and the method comprises the steps: obtaining equipment operation data, and constructing a health index curve; dividing an equipment performance degradation process into a healthy operation stage and a degradation stage; and building a residual life prediction model, inputting a historical health index curve and covariable characteristics of the degradation stage, and outputting a future trend value of the health index. Carrying out model training by utilizing source domain full life cycle data and target domain partial data, extracting features by using a Transform encoder, and realizing feature distribution alignment through an adaptive domain loss mechanism; the BiLSTM network is pre-trained by using the source domain features, and then fine tuning is performed by using the target domain data; and obtaining the residual service life value by predicting the time when the health index curve reaches the failure threshold value. According to the method, the problem of insufficient model generalization ability caused by data difference is effectively solved, and the residual service life prediction accuracy is remarkably improved.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of intelligent operation and maintenance and fault prediction technology for electromechanical equipment, and particularly relates to a method and system for predicting the remaining life of equipment under varying operating conditions based on degradation trend migration prediction. Background Technology

[0002] With the development of industrial digitalization and intelligentization, intelligent maintenance technology based on remaining useful life prediction has become crucial for ensuring reliable equipment operation. In real-world industrial scenarios, due to production safety constraints, it is difficult to obtain degradation data for the entire lifecycle of target equipment. Typically, knowledge transfer can only be achieved using full lifecycle data from existing similar equipment. This raises the problem of remaining useful life (RUL) prediction in cross-equipment, variable operating condition scenarios. In this scenario, the distribution of equipment monitoring data varies significantly, posing a significant challenge to the generalization performance of the model.

[0003] Traditional RUL prediction methods and existing transfer learning methods often suffer from the following problems when facing complex cross-device variable operating condition scenarios: (1) Direct feature alignment of the original signal can easily lead to distortion of the representation of key degradation stages, resulting in negative transfer; (2) Traditional domain adaptation methods are difficult to effectively balance the contribution of loss measurement based on different distribution differences, affecting the feature alignment effect; (3) When the target domain data is scarce, the model is difficult to learn effective degradation rules, resulting in large prediction errors. Therefore, there is an urgent need for a new method that can effectively cope with the differences in data distribution across equipment variable operating conditions and achieve accurate RUL prediction.

[0004] Based on the above analysis, the urgent technical problem to be solved by the existing technology is: how to design a RUL prediction method that can finely align the degradation stage features of the source domain and the target domain, adaptively balance multi-objective optimization, and effectively adapt to the individual degradation patterns of the target equipment under the constraint of scarce target domain data, so as to improve the accuracy, generalization and practicality of the prediction model in cross-device variable operating condition scenarios.

[0005] The patent application "A Bearing Remaining Service Prediction Method Based on Multi-Scale Bayesian Convolutional Transformer Model" (CN202311568493.4) ​​filed by Nanjing University of Aeronautics and Astronautics proposes a method for predicting the remaining service life of bearings based on a multi-scale Bayesian convolutional Transformer model. The innovation of this method lies mainly in the combination of three key technologies: "multi-scale," "Bayesian," and "convolutional Transformer." Specifically, it uses convolutional kernels of different sizes to process the original vibration signal in parallel, simultaneously capturing local features at different time scales during bearing degradation and enhancing feature representation capabilities. The weight parameters in the convolutional neural network are treated as probability distributions and trained using variational inference and other methods. This allows the model to quantify the uncertainty of the prediction, outputting a prediction distribution (such as mean and variance), rather than just a point estimate. Finally, the powerful self-attention mechanism of the Transformer encoder is utilized to capture long-distance dependencies and global temporal patterns between multi-scale features. The main challenges and shortcomings of this method lie in its engineering implementation: high model complexity, strong data dependence, unverifiable reliability of uncertainty estimation, and poor interpretability. Summary of the Invention

[0006] The technical problem to be solved by this invention and the remaining life prediction methods in the prior art mainly rely on a large amount of complete lifecycle tag data of specific equipment. However, in actual industrial scenarios, equipment often operates under varying conditions, and there are individual differences between different equipment of the same type, resulting in inconsistent data feature distributions between the target domain equipment (the device under test) and the source domain equipment (historical data equipment) (i.e., there is domain offset).

[0007] Existing cross-device prediction methods struggle to effectively overcome distributional differences when target domain label data is scarce. They also often neglect the accurate identification of device degradation initiation points, leading to premature intervention in predictions during the device health phase or failure to accurately adapt to specific degradation patterns in the target domain during migration. This severely impacts the accuracy and generalization ability of remaining lifetime predictions.

[0008] The present invention aims to solve the technical problems of low remaining lifetime prediction accuracy and poor generalization ability in the above-mentioned prior art under cross-device and variable operating conditions, due to large differences in feature distribution and scarcity of target domain data.

[0009] To address the aforementioned technical problems, the present invention provides a method for predicting the remaining life of equipment under varying operating conditions based on degradation trend migration prediction. This method includes:

[0010] Acquire lifecycle operation data of source domain devices and current stage operation data of target domain devices, and construct a sequence of health indicators characterizing the performance degradation state of the devices;

[0011] The starting point of degradation of the health indicator sequence is determined based on statistical rules, and data of the degradation stage are extracted.

[0012] A degradation trend prediction model is constructed, which includes a feature extraction unit, a domain adaptation unit, and a trend prediction unit.

[0013] The degradation trend prediction model is trained in two stages using the degradation stage data: the first stage aligns the feature distributions of the source domain and the target domain based on the domain adaptation mechanism, and the second stage fine-tunes the parameters of the trend prediction unit using the target domain data.

[0014] The real-time degradation data of the target domain device is input into the trained model to generate a prediction sequence of future health indicators, and the time span during which the prediction sequence reaches the preset failure threshold is calculated as the remaining service life.

[0015] Furthermore, the construction of the health indicator sequence characterizing the performance degradation state of the equipment includes: multi-domain feature extraction and screening of the original operating data, dimensionality reduction and fusion of the screened features, and generating a normalized one-dimensional health indicator sequence by calculating the distance or reconstruction error between the current state and the initial health state.

[0016] The determination of the degradation start point based on statistical rules specifically includes: calculating the statistical characteristics of health indicators during the healthy operation phase of the equipment and setting a control upper limit; when the monitored health indicator value exceeds the control upper limit for a preset number of consecutive times, the moment of the first exceedance is determined as the degradation start point.

[0017] Furthermore, the feature extraction unit is constructed using an attention mechanism network with long-range dependency capture capability; the trend prediction unit is constructed using a bidirectional recurrent neural network structure to predict future temporal change trends based on the extracted features.

[0018] The first stage of training specifically involves: using data from the source and target domains as input, extracting deep features through the feature extraction unit, and using the loss function of the domain adaptation unit to constrain the feature distribution and obtain domain-invariant features.

[0019] Furthermore, the domain adaptation unit employs an adaptive domain loss mechanism, which is configured to: dynamically calculate the weights of the maximum mean difference loss and the adversarial loss during training; establish an inverse correlation adjustment rule based on the changing trend of the loss value, automatically decaying the corresponding weight coefficient when one of the loss values ​​is detected to increase, and vice versa, so as to balance the contribution of the maximum mean difference loss and the adversarial loss to the model gradient.

[0020] Furthermore, the adaptive domain loss mechanism calculates the changing trend of each loss value in the current training round using the exponential moving average method, and the total loss function in the domain adaptation stage is composed of the superposition of the predicted loss and the weighted domain adaptation loss.

[0021] Furthermore, the second stage uses target domain data to fine-tune the parameters of the trend prediction unit, specifically including: freezing the network layer parameters of the feature extraction unit and the domain adaptation unit; and using the domain-invariant features of the target domain to perform backpropagation training on the unfrozen trend prediction unit to update its network weights.

[0022] Furthermore, the time span during which the calculated prediction sequence reaches the preset failure threshold specifically includes: using an iterative prediction method, taking the prediction output of the previous moment as the input of the next moment, generating a continuous future health indicator trajectory until the trajectory value reaches the failure threshold.

[0023] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any of the above embodiments.

[0024] The present invention also provides a computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the steps of the method described in any of the above embodiments.

[0025] Compared with the prior art, the present invention has the following advantages:

[0026] In the context of existing technologies, this invention effectively solves the technical problems of insufficient accuracy and limited generalization ability in remaining lifetime prediction under cross-device and variable operating conditions by constructing an overall technical architecture of degradation stage interception and two-stage transfer training (domain alignment + fine-tuning).

[0027] First, at the data modeling level, this invention analyzes the entire lifecycle data of equipment based on statistical rules to accurately pinpoint the degradation initiation point, eliminating interference from healthy phase data on model training at the source. This allows the prediction model to focus on learning the true degradation evolution pattern, avoiding noise and bias introduced by early stable operating conditions. Through multi-domain feature fusion and normalization, a single health indicator (HI) is constructed, effectively eliminating differences in the dimensions of different sensors and providing a direct representation of the degree of degradation. Combined with the control upper limit determination method, this significantly reduces the risk of misjudgment caused by noise fluctuations, ensuring the reliability and consistency of the prediction starting point.

[0028] Secondly, at the model training level, this invention proposes a two-stage transfer learning framework. The first stage introduces a domain adaptation mechanism, which systematically eliminates distribution shifts caused by individual differences in different devices and variations in operating conditions by mapping features from the source and target domains to a unified feature subspace. In terms of model architecture, an attention mechanism is employed to capture key dependencies in long-term sequences, and a bidirectional long short-term memory network (BiLSTM) is combined to extract temporal features from different time periods, enabling the model to more fully characterize the evolutionary trends of complex degenerate signals. Based on this, an adaptive domain loss dynamic weight adjustment mechanism is proposed. By constraining different domain alignment loss terms through reverse correlation, a single loss function is avoided from dominating the training process. This ensures distribution alignment while preserving key discriminative information, thereby improving training stability, accelerating model convergence, and significantly enhancing cross-domain generalization ability.

[0029] Furthermore, in the second stage of training, this invention employs a fine-tuning strategy that freezes the feature extraction layer, performing adaptive optimization of the target domain only for the prediction unit. This effectively suppresses overfitting risks under small sample conditions while preserving general degradation features of the source domain, achieving a balance between "generalization ability" and "device-specific characterization." In addition, by generating continuous future degradation trajectories through iterative prediction, compared to traditional single-point prediction methods, it can provide a complete lifetime evolution path, facilitating users' intuitive assessment of device state change trends.

[0030] Second, as supporting evidence of the inventiveness of the claims of this invention, it is also reflected in the following aspects:

[0031] (1) In terms of expected benefits and commercial value, the present invention can significantly improve predictive maintenance capabilities and reduce sudden failures, and is applicable to intelligent operation and maintenance scenarios of high-value equipment such as wind power, rail transit, and aero engines.

[0032] (2) At the technical level, it reduces the cost waste caused by over-overhaul and unnecessary maintenance, and fills the application gap in the field of cross-equipment and variable operating condition RUL prediction;

[0033] (3) In terms of overcoming technical difficulties, it provides a systematic solution to the long-standing problem of cross-domain life prediction, and promotes the transformation of equipment management mode from "post-maintenance" to "pre-prediction";

[0034] (4) In terms of overcoming technical bias, this invention breaks through the inherent assumption of traditional models relying on a large amount of failure data under the same working conditions and equipment, and gets rid of path dependence on a single equipment or fixed working conditions, enabling the model to maintain stable performance in real industrial scenarios where failure samples are scarce and working conditions are complex and changeable. This method can be widely applied to rotating machinery such as bearings, gears, and motors, as well as various electromechanical systems, and has good potential for industry promotion and commercialization. Attached Figure Description

[0035] Figure 1 This is a flowchart of the cross-equipment variable operating condition remaining life prediction method based on degradation trend migration prediction provided in an embodiment of the present invention;

[0036] Figure 2 This is a flowchart of the training and testing process of the Transformer-BiLSTM-DAFT model provided in this embodiment of the invention;

[0037] Figure 3 This is a block diagram of a cross-equipment variable operating condition remaining life prediction system based on degradation trend migration prediction provided in an embodiment of the present invention;

[0038] Figure 4 (a) and (b) are examples of the 3-type methods provided in the embodiments of the present invention. A schematic diagram illustrating the degradation stages of the criteria;

[0039] Figure 5 (a), (b), and (c) are schematic diagrams of the RMSE score, MAE score, and A22-31 experimental results of the present invention, respectively.

[0040] Figure 6 (a), (b), and (c) are schematic diagrams of the RMSE score, MAE score, and B31-21 experimental results of the present invention, respectively.

[0041] Figure 7 (a), (b), and (c) are graphs showing the predicted remaining useful life of the test objects provided in the embodiments of the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0043] like Figure 1 As shown, this embodiment of the invention provides a method for predicting the remaining life of equipment under varying operating conditions based on degradation trend migration prediction. The method specifically includes:

[0044] Step 1: Construction and standardization of health indicators.

[0045] Instead of directly migrating and aligning high-dimensional, heterogeneous raw signals, a health index that can uniformly characterize the performance degradation state of devices is constructed. The specific steps are as follows:

[0046] (1) Data preprocessing and feature engineering: Noise reduction and preliminary feature extraction are performed on the raw operating data collected from the source and target domain devices. A series of features are extracted from multiple perspectives such as time domain and frequency domain, and these features are comprehensively sorted and screened based on monotonicity and trend indicators to select the top few key features that best reflect the degradation pattern of the equipment.

[0047] (2) Health Indicator Construction: Dimensionality reduction and fusion operations are performed on the selected key features. Distance measurement, reconstruction and other methods are used to construct a quantitative indicator (Health Indicator, HI) that can reflect the health status of the equipment by comparing the distance or reconstruction error between the test data and the data of the equipment in the initial health state. Thus, the health status of the equipment is transformed into a one-dimensional time series.

[0048] (3) Health index normalization: The obtained HI curve is subjected to maximum-min normalization to map the health status of the equipment throughout its entire life cycle to the interval [0, 1]. Here, 0 represents that the equipment is in an initial fully healthy state, and 1 represents that the equipment has failed. This step eliminates the dimensional differences of HI under different equipment and different operating conditions, laying a unified foundation for subsequent cross-domain migration.

[0049] Step 2, Degradation stage division.

[0050] To address the interference of data from the healthy operation phase (before significant performance degradation occurs) on degradation trend prediction models, this invention introduces 3... The criteria accurately define the degradation stage, and only use data from the degradation stage for model training and testing. The specific method is as follows:

[0051] The mean and standard deviation of HI data during the healthy operation phase of the calculation device are used to set the upper control limit as μ+3. When the HI value continuously exceeds the control limit for multiple consecutive time points, the first time point exceeding the limit is determined as the degradation start point. This continuous judgment mechanism can effectively avoid misjudgment caused by random noise or instantaneous impact. Finally, only the data between the degradation start point and the equipment failure point is extracted for subsequent model training, eliminating the long-term stable health phase, so that the model can focus more on learning degradation dynamics.

[0052] Step 3, Model building and training.

[0053] Step 3, Model building and training.

[0054] After completing the construction of health indicators and the extraction of degradation stages, this invention constructs a remaining life prediction model for cross-device and variable operating condition scenarios based on the need for degradation trend migration prediction, and completes model learning by adopting a "two-stage transfer training" method. Its specific technical solutions, uses and technical effects are as follows.

[0055] In terms of model structure design, the overall model consists of a feature extraction module, a temporal modeling module, a domain alignment module, and a lifetime prediction module. The feature extraction module takes normalized health indicator time series as input and converts one-dimensional health indicators into high-dimensional representations through embedded mapping. The temporal modeling module employs a structure combining an attention mechanism and a bidirectional long short-term memory network to capture long-term dependencies and temporal evolution characteristics during degradation, thereby improving the ability to model nonlinear and non-stationary degradation trends. The lifetime prediction module outputs the estimated remaining lifetime at the corresponding time point, enabling quantitative prediction of future degradation trajectories.

[0056] In terms of training strategy, this invention employs a two-stage transfer training mechanism. The first stage is the cross-domain alignment training stage, which uses degraded data from the source domain and limited degraded data from the target domain as joint input. While minimizing the loss of the lifetime prediction task, a domain alignment loss constraint is introduced. Through an adaptive domain loss weight adjustment mechanism, the relationship between prediction loss and distribution alignment loss is dynamically balanced, enabling the degradation features of the source and target domains to align within a common subspace. This mitigates the distribution shift caused by individual differences in different devices and variations in operating conditions. The training objective of this stage is to learn degradation representation features with cross-device generalization capabilities.

[0057] The second stage is the target domain fine-tuning stage. After completing cross-domain feature alignment, the feature extraction and temporal modeling modules are frozen, and the lifetime prediction module is fine-tuned only using target domain degradation data. This allows the model to maintain its ability to express general degradation features while further adapting to the specific degradation rate and evolution pattern of the target device. This training method effectively avoids the overfitting problem caused by updating the entire model under conditions of scarce target domain samples.

[0058] In terms of model application and technical effectiveness, this invention, through the aforementioned model construction and training scheme, achieves accurate prediction of the remaining lifespan of target equipment even when failure data in the target domain is extremely limited or incomplete. Simultaneously, relying on degradation stage data training and a cross-domain alignment mechanism, it significantly improves the model's predictive stability and robustness under varying operating conditions and cross-equipment application scenarios. This method not only reduces reliance on a large number of failure samples from similar equipment but also provides a feasible and reliable technical means for predictive maintenance of complex equipment in real-world industrial scenarios.

[0059] like Figure 2The diagram illustrates the training and testing flowchart of the Transformer-Bidirectional Long Short-Term Memory-Domain Adaptation and Fine-Tuning (Transformer-BiLSTM-DAFT) model in this invention. Model training utilizes full-lifecycle degradation data from the source domain device and partial degradation data from the target domain device, specifically historical HI sequences in the degradation stage and selected covariate sequences. The model training process includes two stages: feature extraction based on domain adaptation and trend prediction fine-tuning.

[0060] (1) Domain Adaptation Stage: Using historical HI sequences and covariate sequences as model inputs, the encoder part of the Transformer model is used as a feature extractor shared by the source and target domains. This encoder consists of a multi-layer stacked self-attention mechanism and a feedforward neural network, which can effectively capture long-range dependencies and global contextual information in the input sequence, thereby extracting richer deep features. An adaptive domain loss mechanism is introduced, which dynamically calculates the weights of the maximum mean difference loss and the adversarial loss through exponential moving average (EMA), thereby automatically balancing the proportion of the two losses, achieving efficient alignment of the feature distributions of the source and target domains, and obtaining domain-invariant features.

[0061] The feature extraction calculation formula based on the Transformer encoder is as follows:

[0062] ,

[0063] ,

[0064] in, The output obtained by each self-attention head , , These are the query parameter matrix, key parameter matrix, and value parameter matrix, respectively. This is the learnable output weight matrix.

[0065] The adaptive domain loss mechanism balances the impact of different losses through exponential moving average (EMA) and dynamic weighting based on the loss value. The specific steps are as follows:

[0066] First, in each iteration, the instantaneous values ​​of the adversarial loss and the MMD loss are processed by exponential moving average to generate a smoothed loss estimate that reflects their long-term trend. The mathematical formula is shown below:

[0067] ,

[0068] in, This represents the EMA loss value in the current training epoch. This represents the EMA loss value in the previous training round. This represents the loss value for the current training round. It is the historical weighting coefficient. When it is close to 1, it means that the influence of historical losses accounts for a large proportion. When it is close to 0, it means that the influence of current losses accounts for a large proportion.

[0069] Because the EMA recursively references the EMA value of the previous step in each calculation, and the EMA of the previous step already includes the loss from earlier rounds, it can accumulate all past loss information, rather than relying solely on the current loss value. Based on the smoothed loss value, the weight generation module establishes a reverse association rule: when the EMA value of a certain loss relatively increases, its weight coefficient decreases proportionally; conversely, when the EMA value decreases, its weight coefficient gradually increases. This mechanism achieves a balance between domain confusion capability and distribution alignment performance by dynamically adjusting the contribution ratio of MMD loss and adversarial loss. Its mathematical formula is shown below:

[0070] ,

[0071] ,

[0072] in, The distribution represents the weights of the MMD loss and the adversarial loss. , These represent the EMA loss value of MMD and the EMA loss value of adversarial training in the current training round, respectively, with E representing a minimum value to prevent the denominator from being 0.

[0073] The overall loss function during the domain adaptation phase is:

[0074] ,

[0075] Among them, for The optimal parameters are obtained by performing a grid search.

[0076] (2) Fine-tuning stage: The model first uses the source domain features obtained in the domain adaptation stage to pre-train the BiLSTM prediction module, then freezes some BiLSTM network parameters, and uses the target domain features to fine-tune the model in order to further optimize the model’s generalization in the target domain, thereby achieving better prediction accuracy.

[0077] The prediction model is built upon a BiLSTM network, employing a multi-layered stacked BiLSTM architecture to achieve context-aware modeling of time-series data through bidirectional temporal feature extraction. First, the input data is processed by dual-channel Long Short-Term Memory (LSTM) units for forward and inverse temporal modeling, capturing temporal dependencies within the data. Subsequently, the bidirectional hidden states are concatenated through a feature fusion layer, integrating the forward and inverse information. Finally, the prediction result is output through a fully connected output layer. This design allows the model to fully utilize the contextual information in the time-series data, thereby improving prediction accuracy and robustness.

[0078] The formula for calculating the hidden layer state of the BiLSTM module is as follows:

[0079] ,

[0080] ,

[0081] ,

[0082] in, Let this be the forward hidden state at time t. Let be the backward hidden state at time t. For the splicing operation, the forward and backward hidden states are combined.

[0083] During the fine-tuning phase, only the parameters of the BiLSTM and fully connected layers are updated, and the loss function is the mean squared error.

[0084] Step 4: Calculate the remaining useful life of the target domain devices.

[0085] The historical HI sequence and covariate sequence of a target domain device in the degradation stage are input into a trained Transformer-BiLSTM-DAFT model to obtain a predicted sequence of its future health indicators. By predicting the future changes in the HI curve and comparing them with the device's failure threshold, the device's RUL (Recovery Limiting Rank) can be calculated. When the predicted value reaches the failure threshold, the time difference between the prediction start point and the failure point is the device's RUL value.

[0086] To better measure the relationship between the actual RUL and the predicted RUL, this invention selects relative error (RE) as the main evaluation index. Relative error effectively reflects the difference between the true and predicted values, and through standardization, it increases the comparability between different orders of magnitude. The formula is shown below:

[0087] ,

[0088] like Figure 3 As shown, the present invention provides a cross-equipment variable operating condition remaining life prediction system based on degradation trend migration prediction, which specifically includes:

[0089] The health indicator construction module is used to construct and standardize health indicators;

[0090] The degradation stage segmentation module is used for 3-level degradation. The criteria divide the degradation stages;

[0091] The model training module is used to build and train the Transformer-BiLSTM-DAFT model;

[0092] The lifetime prediction module is used to calculate the remaining lifetime of the target domain device based on the trained model.

[0093] I. Specific application areas or related products of this invention.

[0094] The present invention provides a method and system for predicting the remaining service life of equipment under varying operating conditions across different devices, based on degradation trend migration prediction. This method is primarily applied to the field of intelligent operation and maintenance and predictive maintenance of electromechanical equipment. To verify the effectiveness of the method, a specific embodiment is provided using publicly available datasets. This embodiment uses the XJTU-SY rolling bearing accelerated life dataset as the source domain data and the PHM2012 challenge rolling bearing dataset as the target domain data to simulate the practical engineering problem of predicting the remaining service life of new equipment using existing equipment lifecycle data in a cross-device, varying operating condition scenario.

[0095] Step 1: Construction and standardization of health indicators.

[0096] The bearing vibration signal was denoised, and typical time-domain and frequency-domain features were extracted. Monotonicity and trend indicators were used for comprehensive ranking and screening. Principal Component Analysis (PCA) was then used for dimensionality reduction. The dimensionality-reduced health stage data was used to train a Self-Organizing Maps (SOM) network, with the minimum quantization error used as the bearing's health indicator curve. The obtained HI curve was then subjected to max-min normalization to map the bearing's full life-cycle health indicator values ​​to the [0, 1] interval.

[0097] Step 2, dividing the degradation stage into units.

[0098] Exploit 3 The criteria are used to divide the HI curve into degradation stages and select initial prediction points. Considering that in most bearing datasets, the data at the beginning of the operation phase is usually in a healthy state, the data from the first 20 minutes is selected to calculate 3. The upper and lower thresholds of the criteria. To avoid false alarms, a continuous 5-point anti-false alarm mechanism is introduced. This mechanism effectively eliminates outliers, thereby ensuring the accurate division of the degradation stage.

[0099] Figure 4 The HI curves and degradation stage classification results for two bearings, Data1_1 from the XJTU-SY dataset and PHM1_2 from the PHM 2012 dataset, are presented. It can be observed that in the initial stage, the HI remains relatively stable at 3. Within this range, the bearing is in good health. Over time, HI gradually increases and eventually exceeds 3. The upper limit indicates that the bearing has entered the degradation stage. The HI volatility in the chart breaks through 3. The significant increase after the upper limit further indicates that the bearing has shown obvious signs of degradation. This process also verifies 3. The criteria can identify early signs of degradation in a timely manner. In general, 3 The data segment within the range is considered the healthy phase, during which device performance is stable and HI values ​​fluctuate little; exceeding 3... The data segment with the upper limit and subsequent continuous increase is the degradation stage. During this stage, the HI value increases monotonically with performance degradation and is the core data for predicting the degradation trend.

[0100] Following the same procedure, other bearing data in the XJTU-SY bearing dataset and the PHM2012 bearing dataset were analyzed to determine the degradation initiation point, and based on this, the bearings were divided into normal and degraded states. Table 1 shows the proportion of each bearing degradation stage in the entire life cycle data and the degradation initiation point.

[0101] Table 1 Degradation state parameters of experimental dataset

[0102]

[0103] Step 3, Model building and training unit.

[0104] A Transformer-BiLSTM-DAFT model was constructed. Using the current HI curve as input, the model predicted the future trend of the HI curve. Based on the feature ranking results, the top 5 features were selected as input covariates to assist the model in prediction. The parameters of the Transformer feature extractor network are shown in Table 2.

[0105] Table 2 Transformer Feature Extractor Structure Parameters

[0106]

[0107] The network parameters of the BiLSTM prediction model are shown in Table 3.

[0108] Table 3. BiLSTM Prediction Module Structural Parameters

[0109]

[0110] The learning rate of the neural network was set to 0.001, the batch size to 20, and the Adam algorithm was used as the optimizer with 100 iterations. These hyperparameter settings were intended to optimize the training efficiency and convergence stability of the model in order to fully capture the key features in the data.

[0111] To address the issue of freezing the number of model layers in the BiLSTM prediction model, we conducted traversal experiments to determine the network parameters for the first three layers before freezing, thereby optimizing model performance and improving prediction accuracy.

[0112] To verify whether the degradation trend prediction model based on deep transfer learning constructed in this invention meets the needs of practical applications, 36 sets of transfer experiments were set up to evaluate the effectiveness of the proposed method, as shown in Table 4. Since some datasets have limited available data after the degradation stage division, effective training would be impossible if the data volume is too small. Therefore, the first 50% of the data in the degradation stage was selected as the training data for the target domain to ensure sufficient data to support model training and meet accuracy requirements.

[0113] Table 4 Transfer Learning Experiment Setup

[0114]

[0115] Step 4, Target Domain Device Remaining Service Life Calculation Unit.

[0116] To further illustrate the relationship between trend prediction and device RUL, experiment A11_11 is used as an example to analyze the relationship between health indicators and prediction starting point. Based on the three criteria, the degradation start point of the target domain bearing dataset PHM1_1 is detected at 315 minutes, and the failure location at 466 minutes. Therefore, the bearing progresses from the degradation point to failure in 151 minutes. Then, 30%, 50%, and 70% of the degradation stage data are selected as known data for the target domain for training. Finally, the trained model is used to predict the RUL values ​​at different time points on the test data.

[0117] As can be seen from the results in Table 8, the gap between the actual RUL and the predicted RUL gradually decreases as the prediction starting point advances, indicating that fault progression and data accumulation play a crucial role in improving prediction accuracy. Figure 4The predicted remaining service life (RUL) of the test object is shown in the figure. The obtained curve meets the requirements for RUL prediction in actual research, especially when the rolling bearing is close to failure, the model can provide relatively accurate prediction results. This phenomenon can be attributed to insufficient data in the initial failure stage, making it difficult for the model to accurately predict RUL. However, as the amount of data increases, the prediction accuracy of the model gradually improves. The initial error mainly stems from the uncertainty caused by data scarcity and the interference of random processes. With the continuous accumulation of sensor data, the model can better capture the trend of bearing degradation, and the predicted RUL value gradually approaches the true value.

[0118] Table 8. Results of Actual and Predicted Remaining Useful Life

[0119]

[0120] To better verify the relationship between the HI curve prediction results and RUL, a dataset with a long degradation phase and sufficient data volume was selected for RUL prediction analysis. A dataset was chosen as the source domain for each operating condition category, and the target domain consisted of data samples with a degradation phase duration exceeding 10 minutes. The target domain input data was still divided into training and test sets at a 50 / 50 ratio. As shown in Table 9, analysis was conducted on 12 transfer experiments, with rounding applied to cases where the prediction time was less than 1 minute. In Group A experiments, Data 1_1, Data 2_1, and Data 3_1 datasets were selected as the source domain, and PHM 1_1 dataset was used as the target domain. The results showed that all prediction errors were controlled within 5%. In Group B experiments, PHM 1_1, PHM 2_1, and PHM 3_1 datasets were selected as the source domain, and Data 1_1, Data 2_1, and Data 3_1 datasets were used as the target domain, also yielding good relative error results.

[0121] Table 9. Partial experimental RUL prediction results

[0122]

[0123] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0124] II. Evidence related to the technical effects obtained by the embodiments of the present invention.

[0125] To verify the advantages of the proposed method in predicting the degradation trend of electromechanical equipment, four different degradation trend prediction models were constructed for comparison. Simultaneously, to evaluate the effectiveness of the transfer framework, a baseline model was established: a non-transfer degradation trend prediction model constructed from a Transformer-BiLSTM model, serving as a reference for evaluating the transfer performance of other models. The four selected comparison models are as follows:

[0126] (1) Transformer-BiLSTM-T: This method is a health prediction method based on Transformer-BiLSTM. Except for domain adaptation and fine-tuning, this method adopts the same model structure and hyperparameters as the Transformer-BiLSTM-DAFT model. The model is trained using only a portion of the target domain data and evaluated on the target domain test data.

[0127] (2) Transformer-BiLSTM-S: This method uses the same model structure and hyperparameters as the Transformer-BiLSTM-DAFT model. However, the model is not fine-tuned; it is trained using source domain data and evaluated on target domain test data.

[0128] (3) Transformer-BiLSTM-ST: This method uses the same model structure and hyperparameters as the Transformer-BiLSTM-DAFT model. However, the model is not fine-tuned, but only trained on the prediction model using source domain data and part of the target domain data, and evaluated on the target domain test data.

[0129] (4) FT-MMD-LSTM: This method uses a fully connected network as a feature extractor and aligns the feature distribution differences between the source and target domains through maximum mean difference loss. The LSTM-based prediction model is first pre-trained using source domain features, then fine-tuned using target domain data, and evaluated on target domain test data. The feature extractor parameters of the fully connected layer are (1, 32), (32, 64), and (64, 6), and the number of layers in the LSTM prediction model is the same as that in the model of this invention.

[0130] To ensure the validity of the comparative experiments and to eliminate the influence of feature differences on the experimental results, the same input data was used uniformly. The data used for training included source domain degradation stage data (labeled) and the first 50% of target domain degradation stage data (labeled), while the target domain test data consisted of the remaining data that was not used in training.

[0131] As shown in Table 5, in the transfer experiments of Group A, the Transformer-BiLSTM-DAFT model demonstrated significant advantages in the degradation trend prediction task. The model exhibited the smallest mean and narrowest error range in both Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE), fully validating its excellent prediction accuracy and stability under complex conditions. The Transformer-BiLSTM-T model, trained only on a portion of the target domain data, resulted in a significant decrease in prediction performance, indirectly confirming the limitations of limited training data on model prediction performance and generalization ability. The Transformer-BiLSTM-S model, optimized through domain adaptation, was trained again. However, due to the nonlinear mapping of multi-dimensional temporal features involved in degradation trend prediction, it has higher complexity than the traditional RUL prediction task, leading to a larger error range and insufficient stability compared to other models. After domain adaptation, the Transformer-BiLSTM-ST model trains on both source and target domain data simultaneously. While this alleviates the problem of insufficient target domain samples, the model parameter optimization process is dominated by the source domain due to the scale advantage of the source domain data, resulting in poor prediction accuracy. The FT-MMD-LSTM model employs a similar transfer learning strategy to the model proposed in this invention, but due to limitations in its feature extractor and prediction model, it has limitations in representing features under complex working conditions, resulting in poor adaptability. Therefore, its prediction accuracy is lower than that of the model proposed in this invention.

[0132] Table 5. Evaluation Indicators of Group A Experimental Model

[0133]

[0134] like Figure 5 As shown in (a), under the RMSE evaluation index, the model proposed in this invention exhibits high prediction accuracy and good robustness to noise interference, which is far superior to other comparative models. Figure 5 (b) Further, it is shown that under the MAE evaluation index, the model of this invention exhibits more stable predictive performance, and its predicted results have higher consistency with the true values. Even in experimental scenarios with significant feature differences, the model proposed in this invention maintains good performance. Taking the A22-31 experiment as an example, as... Figure 5As shown in (c), the predicted trends of the proposed model and the FT-MMD-LSTM model maintain a high degree of consistency with the true values, verifying the effectiveness of the proposed model's method of combining domain adaptation and fine-tuning in the degradation trend prediction task. However, the proposed model is closer to the true values ​​in terms of trend. In contrast, although the FT-MMD-LSTM model can also follow the predicted trend, its performance is not as good as the proposed model, exhibiting certain limitations in terms of noise interference and data complexity. In addition, the Transformer-BiLSTM-T model, due to limited data, has a relatively flat predicted trend, failing to fully utilize the model's performance. While the Transformer-BiLSTM-S and Transformer-BiLSTM-ST models show some trend prediction ability, their fitting degree to the HI curve needs improvement due to the significant influence of the source domain data. Although other models also show good prediction results in some experiments, the proposed model has significant advantages over other models in terms of prediction accuracy, stability, and robustness, fully verifying the strong adaptability of the proposed model in complex working conditions and experimental scenarios with large feature differences.

[0135] As shown in Table 6, in the transfer experiments of Group B, the Transformer-BiLSTM-DAFT model proposed in this invention also exhibits significant advantages, especially in RMSE and MAE metrics, where the difference compared to the comparative models is more pronounced. This further verifies the universality and generalization ability of this model in different target domains. In the Group B experiments, the performance gap between the Transformer-BiLSTM-T and Transformer-BiLSTM-S models has narrowed, mainly due to the increased size of the target domain dataset, which alleviates the performance difference caused by the different data volume between the two models. However, the experimental results show that both models are still slightly insufficient in predicting degradation trends. Furthermore, the Transformer-BiLSTM-ST model and the FT-MMD-LSTM model scored similarly in the RMSE and MAE evaluation metrics in the B group experiments, indicating that when the target domain data meets specific conditions, the fine-tuning strategy can enable the model to better adapt to the distribution of the target domain data, thereby improving the model's prediction accuracy. However, compared with the Transformer-BiLSTM-DAFT model proposed in this invention, there is still a certain gap, indicating that this model has better adaptability and prediction accuracy when dealing with complex cross-domain transfer tasks.

[0136] Table 6. Evaluation Indicators of Group B Experimental Model

[0137]

[0138] like Figure 6As shown in (a), under the RMSE evaluation index, the model of this invention exhibits superior accuracy and can still maintain good robustness when faced with significant differences in feature distribution, demonstrating its adaptability in complex environments. Figure 6 (b) Further, it is shown that under the MAE evaluation index, the model proposed in this invention exhibits more stable predictive performance, with a smaller gap between the predicted results and the true values, reflecting its generalization and accuracy on the target domain data. Especially when both the source and target domain data are insufficient and the domain heterogeneity is high, the model proposed in this invention can still maintain good predictive performance. Taking the B31_21 experiment as an example, as... Figure 6 As shown, due to the limited number of source domain samples and the significant feature differences between the source and target domains, only limited trend information can be learned during training. This invention's model rationally fuses and aligns this limited information through an adaptive loss mechanism, thereby better adapting to the target domain data during pre-training and fine-tuning, effectively improving the model's predictive ability in the target domain. In the B31_21 experiment, the Transformer-BiLSTM-T model outperforms the domain-adapted Transformer-BiLSTM-S model in the target domain. However, due to the lack of data from the later stages of target domain degradation, the Transformer-BiLSTM-T model's predictive performance tends to plateau at the end of degradation. In contrast, the Transformer-BiLSTM-ST model relies on the degradation pattern of the source domain data and early degradation information in the target domain; therefore, it fails to provide accurate prediction results when predicting degradation in the later stages of the target domain. The FT-MMD-LSTM model is similar to the label in terms of prediction curve shape, but its accuracy is significantly lower than that of the model proposed in this invention because the neighborhood information between the source and target domains is not fully aligned. Although it can reflect the degradation trend to some extent, it still does not reach the accuracy level of the model proposed in this invention.

[0139] To better evaluate and analyze the impact of each module of the proposed Transformer-BiLSTM-DAFT-based degradation trend prediction model on the overall performance, this invention employs ablation experiments. By progressively removing or modifying different components of the model, the impact on model performance is observed, thereby determining the contribution of each module to the model. As shown in Table 7, six sets of ablation experiments were conducted: A11-11, A21-21, A31-31, B11-11, B21-21, and B31-31. The performance of five models—the complete model, the model with the adaptive domain loss mechanism removed, the model with only fine-tuning retained, the model without degradation stage division, and the model with the Transformer feature extractor replaced—was compared. The meanings of each ablation model are as follows:

[0140] (1) Remove the adaptive domain loss mechanism: Use the HI curves of the first 50% of the degradation stage as the model input, remove the adaptive domain loss mechanism during training, use fixed weights, and keep the rest of the model structure unchanged.

[0141] (2) Only retain fine-tuning strategy: use the HI curves of the first 50% of the degradation stage as model input, cancel the domain adaptation process, use source domain data for pre-training and target domain partial data for fine-tuning during training, and keep the rest of the model structure unchanged.

[0142] (3) No degradation stage division: The HI curve is used as the model input, and the degradation stage division is cancelled. To ensure the effectiveness of the experiment, the training data includes the first 50% of the degradation stage data in addition to the health status data, while the rest of the model structure remains unchanged.

[0143] (4) Replace the Transformer feature extractor: Using the HI curves of the first 50% of the degradation stage as the model input, replace the Transformer feature extractor with a fully connected layer of the same number of layers, while keeping the rest of the model structure unchanged.

[0144] Table 7 Results of ablation experiment evaluation indicators

[0145]

[0146] As shown in Table 7, after removing the adaptive domain loss mechanism, the robustness of the model significantly decreased in experiments with high domain heterogeneity, indicating that this module played a crucial role in aligning the feature differences between the source and target domains. When only the fine-tuning strategy was retained, the model relied on source domain data for pre-training and target domain data for fine-tuning, without employing a domain adaptation mechanism. This resulted in the model failing to effectively align the features of the source and target domains in tasks with significant domain differences, leading to lower prediction performance than the complete model, especially when target domain data was scarce, resulting in lower prediction accuracy. Without degradation stage segmentation, the model input included healthy state data and some degradation stage data. Since healthy state data could mislead the model's prediction of degradation trends, this reduced the model's prediction accuracy. Replacing the Transformer feature extractor with a fully connected layer maintained the model structure, but because the Transformer is better at handling long-term dependencies and complex features in sequence data, the replacement reduced the model's feature extraction capability, leading to significantly inferior performance in tasks handling complex feature interactions. The results of stepwise ablation experiments, with each module added sequentially, further demonstrated the positive contribution of the proposed modules to the model's performance. The addition of each module effectively improved the model's performance and further enhanced the overall model's predictive ability.

[0147] To further illustrate the relationship between trend prediction and device RUL, experiment A11_11 is used as an example to analyze the relationship between health indicators and prediction starting points. Based on 3 According to the criteria, the degradation start point of the target domain bearing dataset PHM1_1 was detected to be 315 min, and the failure location was 466 min. Therefore, the bearing went from the degradation point to the failure state in 151 min. Then, 30%, 50%, and 70% of the degradation stage data were selected as known data in the target domain for training. Finally, the trained model was used to predict the RUL prediction values ​​at different time points on the test data.

[0148] As can be seen from the results in Table 8, the gap between the actual RUL and the predicted RUL gradually decreases as the prediction starting point advances, indicating that fault progression and data accumulation play a crucial role in improving prediction accuracy. Figure 7 The predicted remaining service life (RUL) of the test object is shown in the figure. The obtained curve meets the requirements for RUL prediction in actual research, especially when the rolling bearing is close to failure, the model can provide relatively accurate prediction results. This phenomenon can be attributed to insufficient data in the initial failure stage, making it difficult for the model to accurately predict RUL. However, as the amount of data increases, the prediction accuracy of the model gradually improves. The initial error mainly stems from the uncertainty caused by data scarcity and the interference of random processes. With the continuous accumulation of sensor data, the model can better capture the trend of bearing degradation, and the predicted RUL value gradually approaches the true value.

[0149] Table 8. Results of Actual and Predicted Remaining Useful Life

[0150]

[0151] To better verify the relationship between the HI curve prediction results and RUL, a dataset with a long degradation phase and sufficient data volume was selected for RUL prediction analysis. A dataset was chosen as the source domain for each operating condition category, and the target domain consisted of data samples with a degradation phase duration exceeding 10 minutes. The target domain input data was still divided into training and test sets at a 50 / 50 ratio. As shown in Table 9, analysis was conducted on 12 transfer experiments, with rounding applied to cases where the prediction time was less than 1 minute. In Group A experiments, Data 1_1, Data 2_1, and Data 3_1 datasets were selected as the source domain, and PHM 1_1 dataset was used as the target domain. The results showed that all prediction errors were controlled within 5%. In Group B experiments, PHM 1_1, PHM 2_1, and PHM 3_1 datasets were selected as the source domain, and Data 1_1, Data 2_1, and Data 3_1 datasets were used as the target domain, also yielding good relative error results.

[0152] Table 9. Partial experimental RUL prediction results

[0153]

[0154] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting remaining life across equipment under varying operating conditions based on degradation trend migration prediction, characterized in that, include: Acquire lifecycle operation data of source domain devices and current stage operation data of target domain devices, and construct a sequence of health indicators characterizing the performance degradation state of the devices; The starting point of degradation of the health indicator sequence is determined based on statistical rules, and data of the degradation stage are extracted. A degradation trend prediction model is constructed, which includes a feature extraction unit, a domain adaptation unit, and a trend prediction unit. The degradation trend prediction model is trained in two stages using the degradation stage data: the first stage aligns the feature distributions of the source domain and the target domain based on the domain adaptation mechanism, and the second stage fine-tunes the parameters of the trend prediction unit using the target domain data. The real-time degradation data of the target domain device is input into the trained model to generate a prediction sequence of future health indicators, and the time span during which the prediction sequence reaches the preset failure threshold is calculated as the remaining service life.

2. The method as described in claim 1, characterized in that, The constructed health indicator sequence characterizing the performance degradation state of the equipment includes: Multi-domain feature extraction and filtering are performed on the original operating data. The filtered features are then dimensionality-reduced and fused. By calculating the distance or reconstruction error between the current state and the initial health state, a normalized one-dimensional health indicator sequence is generated. The determination of the degradation start point based on statistical rules specifically includes: calculating the statistical characteristics of health indicators during the healthy operation phase of the equipment and setting a control upper limit; when the monitored health indicator value exceeds the control upper limit for a preset number of consecutive times, the moment of the first exceedance is determined as the degradation start point.

3. The method as described in claim 1, characterized in that, The feature extraction unit is constructed using an attention mechanism network with long-range dependency capture capability; the trend prediction unit is constructed using a bidirectional recurrent neural network structure to predict future temporal change trends based on the extracted features. The first stage of training specifically involves: using data from the source and target domains as input, extracting deep features through the feature extraction unit, and using the loss function of the domain adaptation unit to constrain the feature distribution and obtain domain-invariant features.

4. The method as described in claim 1, characterized in that, The domain adaptation unit employs an adaptive domain loss mechanism, which is configured as follows: The weights of the maximum mean difference loss and the adversarial loss are dynamically calculated during training. An inverse correlation adjustment rule is established based on the changing trend of the loss value. When one of the loss values ​​is detected to increase, its corresponding weight coefficient is automatically reduced, and vice versa, so as to balance the contribution of the maximum mean difference loss and the adversarial loss to the model gradient.

5. The method as described in claim 4, characterized in that, The adaptive domain loss mechanism calculates the trend of each loss value in the current training round using the exponential moving average method. The total loss function in the domain adaptation stage is composed of the predicted loss and the weighted domain adaptation loss.

6. The method as described in claim 1, characterized in that, The second stage involves fine-tuning the parameters of the trend prediction unit using target domain data, specifically including: Freeze the network layer parameters of the feature extraction unit and the domain adaptation unit; The unfrozen trend prediction units are trained by backpropagation using the domain-invariant features of the target domain, and their network weights are updated.

7. The method as described in claim 1, characterized in that, The time span during which the calculated predicted sequence reaches the preset failure threshold specifically includes: An iterative prediction method is adopted, which uses the prediction output of the previous moment as the input of the next moment to generate a continuous trajectory of future health indicators until the trajectory value reaches the failure threshold.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1 to 7.

9. A cross-equipment variable operating condition remaining life prediction system based on degradation trend migration prediction, characterized in that, include: The data processing module is used to construct standardized health indicators and extract data on the degradation stage based on statistical rules; The model building module is used to build deep learning networks that include feature extraction, domain adaptation, and trend prediction functions. The transfer training module is configured to execute a two-stage training strategy. First, it aligns the features of the source domain and the target domain through a dynamically weighted domain adaptation loss. Then, it freezes some network layers and fine-tunes the prediction network using data from the target domain. The lifetime calculation module is used to output an iterative prediction sequence based on the trained model and calculate the remaining lifetime.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 7.

Images