Method and device for predicting remaining service life of equipment

Abstract

This invention discloses a method and apparatus for predicting the remaining useful life of equipment, relating to the field of aerospace technology. The method involves mining common engine degradation information from massive amounts of data on similar engines, mining individual degradation information similar to a target engine from a highly transferable source domain, and establishing a personalized prediction model for the target engine. The proposed method was validated using internationally available turbofan engine data. Comparative results demonstrate that the proposed migration prediction framework achieves higher prediction accuracy than existing deep learning RUL prediction methods.

Description

Technical Field

[0001] The present invention relates to the field of aerospace technology, and in particular to a method and device for predicting the remaining useful life of equipment. Background Art

[0002] Accurate prediction of the remaining life of equipment is crucial for its safe use and reasonable and accurate maintenance. The key to improving the accuracy of equipment remaining life prediction lies in building a prediction model that can accurately recognize the performance degradation patterns of the target equipment. In practice, for the remaining life prediction model, it is usually required to estimate the remaining life as early as possible before a failure. However, the target equipment has a short historical operating data, and its performance degradation patterns have not yet been clearly revealed, making it difficult to build an effective prediction model. To address the above difficulties, this patent introduces transfer learning, a cutting-edge technology in the field of machine learning, into equipment remaining life prediction to achieve the mutual borrowing of information between different devices of the same model.

[0003] Transfer learning is currently being widely applied in various fields. A multi-source transfer learning network (MSTLN) framework addresses the issues of different fault distributions between the target and source domains, as well as the imbalance in the number of healthy and faulty samples. However, most current transfer learning research focuses on categorical data and classification tasks, with relatively little research applied to condition assessment and life prediction. In recent years, a transfer learning algorithm based on a bidirectional long short-term memory (BLSTM) recurrent neural network has been proposed for remaining useful life (RUL) estimation. However, this transfer prediction method primarily transfers a single type of information, with relatively few hybrid transfer methods. Furthermore, transfer learning research on samples from multiple source domains is limited. Existing transfer prediction methods primarily rely on direct application or minor improvements of classification transfer learning methods, with few specifically designed approaches tailored to the characteristics of time series data and prediction tasks. Existing methods do not address the fusion of samples from source domains with varying degrees of transferability. Summary of the Invention

[0004] The technical problem solved by the solution provided by the embodiment of the present invention is how to improve the accuracy of prediction of the remaining useful life of equipment.

[0005] According to an embodiment of the present invention, a method for predicting the remaining useful life of equipment is provided, comprising:

[0006] Acquire operation monitoring data of the target device, and calculate the healthy bulldozer distance and the faulty bulldozer distance between the operation monitoring data of the target device and the screened healthy baseline set and the screened faulty baseline set;

[0007] Converting the healthy bulldozer distance and the faulty bulldozer distance into the health of the target device using a linear weight fusion algorithm, and determining the migration time of the target device using the health of the target device;

[0008] Obtaining the length of the target device degradation period according to the migration time of the target device, and calculating the cosine distance weight and the bulldozer distance weight respectively according to the length of the target device degradation period;

[0009] Calculating the transferable distance between the target device and each device of the same model using the cosine distance weight and the bulldozer distance weight, and screening out multiple devices of the same model that are similar to the target device based on the transferable distance;

[0010] Obtaining device information of each similar device of the same model, and fine-tuning the pre-trained DLSTM-FC prediction network model using the device information of each similar device of the same model to obtain a fine-tuned DLSTM-FC prediction network model;

[0011] Inputting the device degradation sensitive parameter information of the target device into the fine-tuned and trained DLSTM-FC prediction network model for prediction processing to obtain the remaining service life information of the target device;

[0012] Among them, DLSTM-FC refers to a deeper long short-term memory-fully connected layer.

[0013] According to an embodiment of the present invention, a device for predicting the remaining useful life of equipment is provided, comprising:

[0014] a migration time determination module configured to obtain operation monitoring data of a target device and calculate a healthy bulldozer distance and a faulty bulldozer distance between the operation monitoring data of the target device and a screened healthy baseline set and a screened faulty baseline set; convert the healthy bulldozer distance and the faulty bulldozer distance into the health of the target device using a linear weight fusion algorithm, and determine the migration time of the target device using the health of the target device;

[0015] a similar device screening module, configured to obtain the target device degradation period length based on the target device migration time, and calculate a cosine distance weight and a bulldozer distance weight based on the target device degradation period length; calculate the migration distance between the target device and each device of the same model using the cosine distance weight and the bulldozer distance weight, and screen out multiple similar devices of the same model that are similar to the target device based on the migration distance;

[0016] The remaining service life prediction module is used to obtain the device information of each similar device of the same model, and use the device information of each similar device of the same model to fine-tune the pre-trained DLSTM-FC prediction network model to obtain the fine-tuned DLSTM-FC prediction network model; input the device degradation sensitive parameter information of the target device into the fine-tuned DLSTM-FC prediction network model for prediction processing to obtain the remaining service life information of the target device.

[0017] The solution provided by the embodiment of the present invention has the following effects:

[0018] (1) Different from the improvement of a single transfer learning algorithm, a personalized and systematic transfer learning framework for engine RUL prediction is proposed. The migration prediction method is optimized in terms of when to migrate, what to migrate, and how to migrate. A personalized RUL prediction model for the target engine is established, which improves the ability to deconstruct and comprehensively utilize multi-source domain data and the prediction accuracy.

[0019] (2) To address the issue of when to transfer, unlike existing methods that directly predict migration based on the target domain, a migration timing identification algorithm based on WD dual baseline assessment (DBA-WD) is designed to avoid interference with healthy stage operation data that is useless for predictive modeling.

[0020] (3) Regarding what to migrate, the proposed multi-source domain deconstruction method based on TL-EDM measures transferability from three perspectives: degradation degree, degradation rate, and initial health. It can effectively screen out highly transferable source domains and sequences from low transferable density datasets.

[0021] (4) In terms of migration method, a personalized migration learning scheme for engine RUL prediction based on two-stage transfer learning is designed to extract and integrate the general degradation information of the same type of engines and the individual degradation information of the target engine, thereby solving the influence of individual differences and improving the migration effect and prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to understand the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0023] Figure 1 This is a general flow chart of equipment remaining useful life prediction provided by an embodiment of the present invention;

[0024] Figure 2 This is a schematic diagram of performance evaluation and migration time identification based on DBA-WD provided by an embodiment of the present invention;

[0025] Figure 3 Schematic diagram of transferability measurement and multi-source domain deconstruction based on TL-EDM provided by an embodiment of the present invention;

[0026] Figure 4 Schematic diagram of the network structure of the DLSTM migration prediction model provided by an embodiment of the present invention;

[0027] Figure 5 Schematic diagram of RUL migration prediction based on the DLSTM model provided by an embodiment of the present invention;

[0028] Figure 6 This is a schematic diagram of multi-index fusion smoothing scale optimization provided by an embodiment of the present invention;

[0029] Figure 7 1 is a schematic diagram of the optimal scale smoothing result of target engine No. 1 provided in an embodiment of the present invention;

[0030] Figure 8 Schematic diagram of parameter normalization and healthy fault baseline screening provided by an embodiment of the present invention;

[0031] Figure 9 1 is a schematic diagram of an engine performance status evaluation result provided by an embodiment of the present invention;

[0032] Figure 10 3 is a schematic diagram of the migration time identification result of target engine No. 3 provided in an embodiment of the present invention;

[0033] Figure 11 Schematic diagram of PCA results of monitoring parameters provided by an embodiment of the present invention;

[0034] Figure 12 3 is a schematic diagram of the measurement results of the transferable distance of target engine No. 3 provided in an embodiment of the present invention;

[0035] Figure 13 This is a schematic diagram of the multi-source domain deconstruction results of target engine No. 3 provided by an embodiment of the present invention;

[0036] Figure 14 Schematic diagram of training data and weights of target engine No. 3 provided in an embodiment of the present invention;

[0037] Figure 15 3 is a schematic diagram of the fine-tuning training results of target engine No. 3 provided in an embodiment of the present invention;

[0038] Figure 16 3 is a schematic diagram of the RUL prediction result of target engine No. 3 provided in an embodiment of the present invention;

[0039] Figure 17 This is a flow chart of a method for predicting the remaining useful life of equipment provided by an embodiment of the present invention;

[0040] Figure 18 This is a schematic diagram of a device for predicting the remaining useful life of equipment provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] The preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. It should be understood that the preferred embodiments described below are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0042] Figure 17 This is a flow chart of a method for predicting the remaining useful life of equipment provided by an embodiment of the present invention. Figure 17 As shown, including:

[0043] Step S101: Acquire operation monitoring data of a target device, and calculate a healthy bulldozer distance and a faulty bulldozer distance between the operation monitoring data of the target device and a screened healthy baseline set and a screened faulty baseline set;

[0044] Step S102: using a linear weight fusion algorithm to convert the healthy bulldozer distance and the faulty bulldozer distance into the health of the target device, and using the health of the target device to determine the migration time of the target device;

[0045] Step S103: obtaining the length of the target device degradation period according to the migration time of the target device, and calculating the cosine distance weight and the bulldozer distance weight respectively according to the length of the target device degradation period;

[0046] Step S104: Calculating the migration distance between the target device and each device of the same model using the cosine distance weight and the bulldozer distance weight, and screening out multiple devices of the same model that are similar to the target device based on the migration distance;

[0047] Step S105: obtaining device information of each similar device of the same model, and using the device information of each similar device of the same model to fine-tune the pre-trained DLSTM-FC prediction network model to obtain a fine-tuned DLSTM-FC prediction network model;

[0048] Step S106: Inputting the device degradation sensitive parameter information of the target device into the fine-tuned and trained DLSTM-FC prediction network model for prediction processing to obtain the remaining service life information of the target device.

[0049] Among them, the calculation of the healthy bulldozer distance and the faulty bulldozer distance between the operation monitoring data of the target device and the filtered healthy baseline set and faulty baseline set includes: obtaining the historical operation monitoring data of the training device, and filtering the healthy baseline set and the faulty baseline set from the historical operation monitoring data of the training device; calculating the healthy bulldozer distance between the operation monitoring data of the target device and the healthy baseline set, and calculating the faulty bulldozer distance between the operation monitoring data of the target device and the faulty baseline set through the bulldozer distance measurement method.

[0050] Furthermore, the converting the healthy bulldozer distance and the faulty bulldozer distance into the health of the target device by using a linear weight fusion algorithm includes:

[0051]

[0052] Among them, HI T is the health of the target device; Wf is the fault distance weight; Wh is the health distance weight; df is the fault bulldozer distance of the fault baseline set; dh is the health bulldozer distance.

[0053] Furthermore, the calculating of the cosine distance weight and the bulldozer distance weight according to the degradation period length of the target equipment includes:

[0054]

[0055]

[0056] Among them, W cos is the weight of the cosine distance; t l is the degradation period length of the target device; T up ,T lower are the upper and lower limits of the distance weight respectively, and T up +T lower =1;L max ,L min are the maximum and minimum values ​​of the weight change length range respectively; W EMD is the weight of the bulldozer distance, and W cos +W EMD =1.

[0057] Among them, the use of the cosine distance weight and the bulldozer distance weight to calculate the transferable distance between the target device and each device of the same model includes: extracting the first degradation main information from the multidimensional monitoring parameters of the target device through the principal component analysis method, and at the same time extracting the second degradation main information from the multidimensional monitoring parameters of the device of the same model; obtaining the target degradation feature sequence from the first degradation main information, and obtaining the same model degradation feature sequence from the second degradation main information; using the target degradation feature sequence and the same model degradation feature sequence to respectively calculate the cosine distance and bulldozer distance between the target degradation feature sequence and the same model degradation feature sequence; according to the cosine distance, the bulldozer distance, the cosine distance weight and the bulldozer distance weight, calculating the transferable distance between the target device and each device of the same model. Furthermore, the calculation of the transferable distance between the target device and each device of the same model according to the cosine distance, the bulldozer distance, the cosine distance weight and the bulldozer distance weight includes:

[0058] d t =W cos *d cos (Z s ,Z t ,t)+W EMD *d EMD (Z s ,Z t ,t)

[0059] Among them, d t is the migration distance corresponding to the time delay t; d cos is the cosine distance; d EMD is the bulldozer distance; Z s is the degenerate characteristic sequence of the same model; Z t is the target degradation feature sequence; t is the starting point of the delay sliding window.

[0060] Among them, the pre-trained DLSTM-FC prediction network model includes: constructing a DLSTM-FC prediction network model that includes 5 layers of LSTM networks with different model structure parameters and 2 layers of FC networks with different model structure parameters from top to bottom; obtaining device information of multiple devices of the same model, and using the device information of the multiple devices of the same model to pre-train the DLSTM-FC prediction network model to obtain a pre-trained DLSTM-FC prediction network model.

[0061] Specifically, the pre-trained DLSTM-FC prediction network model is fine-tuned using the device information of each device of the same model to obtain the fine-tuned DLSTM-FC prediction network model, including: calculating the similarity value between each similar device of the same model and the target device, and configuring a sample weight value for each similar device of the same model based on each similarity value; using the device degradation sensitive parameter information and device remaining service life information of each similar device of the same model with configured sample weight values ​​as the input and output of the pre-trained DLSTM-FC prediction network model, respectively, and fine-tuning the weight values ​​of the last three network layers in the pre-trained DLSTM-FC prediction network model to obtain the fine-tuned DLSTM-FC prediction network model.

[0062] Furthermore, the DLSTM-FC prediction network model is pre-trained using the device information of the multiple devices of the same model to obtain the pre-trained DLSTM-FC prediction network model, including: pre-training the weight value of each network layer in the DLSTM-FC prediction network model using the device information of the multiple devices of the same model to obtain the weight value of each network layer in the DLSTM-FC prediction network model; randomly initializing the weight values ​​of the last three network layers in the DLSTM-FC prediction network model, and setting the training status of the weight values ​​of the first four network layers in the DLSTM-FC prediction network model to a non-trainable state, and setting the training status of the weight values ​​of the last three network layers to a trainable state, to obtain the pre-trained DLSTM-FC prediction network model; wherein the device information includes: device degradation sensitive parameter information and device remaining service life information.

[0063] Figure 18 Schematic diagram of a device for predicting the remaining useful life of equipment provided by an embodiment of the present invention. Figure 18As shown, it includes: a migration time determination module 201, which is used to obtain the operation monitoring data of the target device, and calculate the healthy bulldozer distance and the faulty bulldozer distance between the operation monitoring data of the target device and the screened healthy benchmark set and faulty benchmark set; using a linear weight fusion algorithm to convert the healthy bulldozer distance and the faulty bulldozer distance into the health of the target device, and using the health of the target device to determine the migration time of the target device; a similar device screening module 202, which is used to obtain the degradation period length of the target device according to the migration time of the target device, and calculate the cosine distance weight and the bulldozer distance weight respectively according to the degradation period length of the target device; using the cosine distance weight and The bulldozer distance weight calculates the migratable distance between the target device and each device of the same model, and based on the migratable distance, screens out multiple similar devices of the same model that are similar to the target device; the remaining service life prediction module 203 is used to obtain the device information of each similar device of the same model, and use the device information of each similar device of the same model to fine-tune the pre-trained DLSTM-FC prediction network model to obtain the fine-tuned DLSTM-FC prediction network model; the device degradation sensitive parameter information of the target device is input into the fine-tuned DLSTM-FC prediction network model for prediction processing to obtain the remaining service life information of the target device.

[0064] like Figure 1 As shown, the present invention comprises migration time identification, highly transferable source domain and sequence screening, and personalized migration prediction. This transferability can be leveraged to deconstruct multiple source domains and predict the RUL of the target engine. By mining both general and individual degradation information, the prediction model for the target engine becomes more personalized and accurate.

[0065] Part 1: Performance Evaluation and Migration Time Identification Based on DBA-WD

[0066] This paper proposes a DBA-WD algorithm that integrates healthy distance and fault distance, such as Figure 2 As shown in Figure 2, a degradation-sensitive feature based on HI is designed. When the degradation-sensitive feature exceeds the threshold, the target engine is considered to be in the degradation stage, which is the migration pre-time.

[0067] Scale-optimized smoothing and denoising: Due to data acquisition errors and the engine operating environment, monitoring data contains a certain amount of noise. Using a scale-optimized data smoothing method, we can remove noise from the raw data while retaining information about engine degradation trends.

[0068] Maximum and minimum value normalization: Engine monitoring data has different physical dimensions and significant differences in data values. Therefore, maximum and minimum normalization is applied to the raw data to unify the values ​​of the monitoring parameters into the same distribution.

[0069] Trend regularization: Different monitoring parameters have different responses to engine performance degradation. Some parameters gradually increase as the engine runs, while others do the opposite. In order to construct an HI that quantitatively characterizes engine performance, the changing trends of the monitoring parameters are regularized as follows:

[0070]

[0071] '

[0072] Where x i,j is the data value of the jth parameter after preprocessing in the i-th cycle, x i,j is the data value after trend normalization.

[0073] In order to realize the analysis of engine operating status or fault deviation and performance evaluation, the engine health and fault baseline sets Gh and Gf are designed:

[0074]

[0075]

[0076] Where P is the number of parameter values, x′ 1,j and x′ N,j The distribution represents the parameter values ​​corresponding to the first and last training. N is the total number of engine cycles. gh and gf are sorted in ascending and descending order. The top M h With M f The engines are defined as healthy and faulty baseline engines. Their first and last cycles form the healthy and faulty baseline sets (Gh and Gf) as follows:

[0077]

[0078] In the formula and Represent the health and failure baselines corresponding to the jth parameter of the kth engine, corresponding to the start and end of life win d The parameter value of the cycle. h With M f Represent the number of healthy and faulty baseline engines respectively.

[0079] During engine performance degradation, monitored parameters fluctuate relatively randomly within a single cycle. Therefore, monitoring parameter values ​​over a specific period can better characterize the engine's short-term performance. Given its ability to measure the distance between two distributions, the WD algorithm was chosen to measure the distance between the target engine and the baseline. By measuring the deviation of the engine's monitored parameter distribution from Gh and Gf, engine performance assessment can be achieved.

[0080] The target engine dataset is represented as The distance between GT and Gh / Gf is calculated as follows:

[0081]

[0082]

[0083] In the formula Respectively represent the i-th cycle and the j-th parameter in the time window win d The monitoring parameter value in the range of i is i=1,2,...,N-win d , where N represents the total number of target engine cycles. and Represent the healthy distance and fault distance respectively. WD (·) represents the bulldozer distance calculation.

[0084] In order to comprehensively evaluate the working status of the engine, multiple monitoring parameters of the target engine are used and expressed as the arithmetic mean of the bulldozer distance:

[0085]

[0086]

[0087] The present invention designs a linear weight fusion algorithm to reduce the evaluation error caused by the reverse distribution. The algorithm converts dh and df into the engine's HI. The fusion equation is as follows:

[0088]

[0089]

[0090]

[0091] Where Wf and Wh represent the health and fault distance weights respectively, HI T is the HI of the target engine.

[0092] In the early stage of engine operation, the HI value of adjacent cycles tends to decrease slightly or remain stable. As the number of engine cycles increases, the degradation rate gradually accelerates, and the engine changes from a healthy state to a degraded state. HI The HI sliding window is cut by the window. The HI within the window is fitted, and the fitting equation is as follows:

[0093] HI′=a i T+b(0-11)

[0094] Where HI' is the fitting HI in the window, T is the number of cycles, a i is the slope and b is the intercept. i is defined as a degradation-sensitive feature. i When it is close to 0, the engine performance is relatively healthy. i When it is less than 0, the engine performance begins to decline.

[0095] When the degradation sensitive feature is less than Thre a When a, the target engine is considered to have entered the degradation stage. In general, the threshold is slightly less than 0 to avoid misjudgment caused by small fluctuations in HI. i Less than Thre a The engine is considered to have entered the degradation stage.

[0096] Part II: Transferability Measurement and Multi-Source Domain Deconstruction Based on TL-EDM

[0097] Figure 3 This is the deconstruction process of multiple source domains based on TL-EDM. Highly transferable source domains have three characteristics: degradation level, degradation rate, and initial health similar to the target engine. To integrate these characteristics, WD and cosine distance are used to measure the transferability of the source domains.

[0098] The present invention uses principal component analysis (PCA) to extract key degradation information from monitoring parameters. Through orthogonal transformation, potentially correlated original variables are converted into mutually independent variables. These converted variables are called principal components. The PCA model identifies the correlation between monitoring parameters and the output-independent principal components.

[0099] This paper selects WD and cosine distance to represent two transferability evaluation indicators: degradation degree and rate, respectively. The weights of these two distances are determined by the target DPC length. To prevent excessive weights from causing an imbalance in the transferability evaluation indicators, upper and lower thresholds are set for the weight of each distance. When the engine's DPC length is Li, the weights of the two distances are calculated as follows:

[0100]

[0101]

[0102] Among them, d-up and Ther d-low are the upper and lower limits of the two distance weights, d-up +Ther d-low =1,L max and L min is the maximum and minimum value of the weight change range, L i is the DPC length of the target engine. For different target engines, W can be adaptively determined. cos and W WD The weight of is used to measure the transferability of the degradation trajectory by weighted integrated distance.

[0103] Due to the difference in production and processing errors, the initial fault degree of the target engine is different. In order to accurately measure the transferability of the source domain, a time delay process is added to . A time delay sliding window is performed on the full life degradation trajectory of the source domain, with the target engine DPC length t l Window length.

[0104] When the starting point of the sliding window is t, the end point is t+t l In addition, t E is the total life cycle length of the source domain engine. s is the source domain degradation trajectory, Z t is the target DPC. The formula for calculating the migration distance is as follows:

[0105] d t =W cos *f cos (Z s ,Z t ,t)+W WD *f WD (Z s ,Z t ,t) (0-14)

[0106] Where W cos With W WD Represent the cosine distance and bulldozer distance weights respectively. cos (·) and f WD (·) represents the cosine distance and bulldozer distance calculation respectively. t is the transferable distance. As time t changes, the transferable distance constitutes a transferable distance set, which is expressed as:

[0107] Deconstructing multiple source domains based on the transferable distance between the source domain and the sequence within the domain can improve the transferable density of the training data, that is, selecting source domains with high transferability and marking the highly transferable sequences in them. Pre-screening highly transferable source domains based on transferability, and then performing secondary screening based on RUL labels. The inverse of the transferable distance represents the transferability of the sequence in the source domain, and its maximum value is defined as the transferability of the source domain.

[0108]

[0109]

[0110] Among them D k represents the set of transferable distances of the k-th source domain engine, γ k Indicates portability.

[0111] The multi-source domain deconstruction process is as follows:

[0112] 1) Source domain pre-screening based on transferability: In the pre-screening, the source domain engines are sorted according to their transferability γ k Sort in descending order and select the first S engines as the pre-screening set.

[0113] 2) Secondary screening of source domains based on RUL labels: The pre-screened source domains have a certain similarity with the target engine. However, the target engine has a relatively short running time, and the source domains have a certain amount of accidental similarity. This paper proposes a secondary screening algorithm to exclude source domain engines with accidental similarity. Based on the pre-screened source domains, according to the optimal migration time t k * Calculate URL label URL k * :

[0114] t k * =argmin t (D k ) (0-17)

[0115]

[0116] Where t k * Represents the delay start time corresponding to the highest transferability sequence, Indicates the full life span of the k-th source domain engine, RUL k * Indicates the RUL label value corresponding to the k-th engine.

[0117] In general, sequences with similar degradation characteristics should have a uniform RUL label distribution. The mean μ and variance σ of the RUL* of the pre-screened source domains are calculated. The secondary selection range of RUL* is [μ-r·σ, μ+r·σ], where r is the ratio coefficient. Engines with RUL* not falling within this range are removed, and the remaining source domains are selected as highly transferable source domains.

[0118] Taking advantage of the LSTM neural network's advantages in time series information mining, we chose LSTM as the primary architecture for building a migration prediction model. Multi-layer architectures are more effective and robust than shallow models in extracting information. Therefore, we stacked multi-layer LSTM networks into a DLSTM model to extract features with varying levels of degradation information from time series.

[0119] The present invention is based on the engine RUL migration prediction model of the DLSTM network. Considering that the model input is a one-dimensional DPC sequence, the computational complexity is reduced and the number of neural network layers is reduced accordingly. At the same time, considering that the target engine has the same model as the source domain, the number of frozen layers is increased accordingly. Finally, the prediction model is determined to be a 7-layer stacked neural network with 4 frozen layers. Combined with the powerful fitting ability of the fully connected layer (FC), the 7 hidden layers are decomposed into 5 LSTM layers and 2 FC layers to establish a complex mapping relationship of the engine "DPC to RUL", such as Figure 4 shown.

[0120] This paper proposes a two-stage transfer learning prediction scheme for multi-source domains, the process of which is as follows: Figure 5 shown.

[0121] 1) First migration phase

[0122] Constructing a DLSTM model and training a universal model: Based on the designed model structure of a 5-layer LSTM network and a 2-layer FC network, a DLSTM-FC engine prediction model is constructed with random initial weights. The model is then trained using the entire multi-source domain data. Each source domain DPC is sequentially concatenated as training input, and the training output is the RUL value corresponding to the DPC sequence. The model's training data is weighted equally, and the trained universal model can identify universal engine degradation information. General layers are frozen and individual layers are initialized: The weights of the first four layers (w1-w4) are set to "untrainable" and the weights of the remaining layers of the model are initialized.

[0123] 2) Secondary migration phase

[0124] Setting the training weights for fine-tuning data: The transferability of DPC sequences corresponding to different initial times t in the source domain varies greatly. To input the transferability time information into the prediction model, the training data weights are designed as follows:

[0125]

[0126] In the formula is the training data weight of the sequence corresponding to the initial delay time t of the k-th engine, is the migration distance.

[0127] Fine-tuning the personalized model layers: Each layer of the personalized model is retrained using the DPC of the highly transferable source domain. During the iterative process of retraining the model, corresponding training data weights are assigned to different position sequences in the source domain. Fine-tuning is performed to determine the weights of the individual degraded information layers (w5-w7), completing the model training process.

[0128] 3) Target engine RUL personalized transfer learning prediction

[0129] The target engine DPC is used as input and the target engine RUL is predicted based on the trained target engine model.

[0130] Case Study

[0131] This paper intends to use a dataset from the 2008 PHM International Conference for method research and validation. This data was obtained using C-MAPSS simulation. The FD001 dataset was selected as validation data for the method. FD001 was collected under the same operating conditions and fault injection pattern, and includes:

[0132] (1) Full life data of multiple performance parameter degradation of 100 sample engines

[0133] (2) Randomly sampled life data of multiple performance parameter degradation of 100 target engines and their remaining service life.

[0134] The FD001 dataset includes 21 available engine gas path performance parameters. During engine performance degradation, engine performance data gradually changes over time, indirectly reflecting the downward trend in engine performance. However, the sensitivity of different parameters to degradation varies over time. Data analysis revealed that seven parameters (T2, P2, P15, epr, farB, Nf_dmd, and PCNFR_dmd) showed no trend. Through investigation and analysis of engine RUL-related information, eight parameters with significant trend variation—T24, T30, Ps30, phi, P30, T50, BPR, and Nf—were selected to predict aircraft engine RUL.

[0135] Data denoising is an important part of data preprocessing. The reference scale self-optimization smoothing method is used to preprocess the selected 8 monitoring parameters. The denoising based on local weighted regression uses different smoothing scales. When the smoothing scale is greater than 75, the denoising effect is not significantly improved. Considering the symmetry and computational complexity, the smoothing scale range is set to [5,77] and the interval step size is set to 4. Figure 6 As shown in Figure 1, the eight parameters of all engines in the dataset are denoised and two smoothing indicators, smoothness and root mean square error (RMSE), are calculated. The results show that the optimal smoothing scale is 45. All parameters are preprocessed with the optimal smoothing scale, and the smoothing results are shown in Figure 1. Figure 7 The results show that the optimal scale smoothing process can filter out the noise in the original parameter data and retain the degradation information.

[0136] like Figure 8 As shown in Figure 1, the degradation trends of monitored parameters are normalized to support the selection of healthy and faulty baseline engines. The degradation direction of the PHI and P30 parameters is normalized to an upward trend, while the degradation information remains unchanged. Next, all parameter values ​​from the source domain engines in the initial and final cycles are traversed and sorted in ascending and descending order. The top 10 engines in both initial operating state and end-of-life state are selected as healthy and faulty baseline engines. Table 1 shows the selection results.

[0137] Table 1 Healthy and faulty baseline engine screening results

[0138] Engine number Health Baseline Engine 77#、82#、94#、14#、8#、1#、46#、60#、27#、81# Faulty baseline engine 55#、61#、21#、83#、7#、39#、90#、72#、65#、15#

[0139] like Figure 8 As shown, Win d The sliding window width defines the parameter data for the initial and final windows of the healthy and faulty baseline engines as the healthy baseline and faulty baseline, respectively. The parameters enclosed by the two dashed boxes represent the healthy and faulty baselines of the source engine. The 10 healthy engine baselines and 10 faulty engine baselines constitute the healthy baseline set Gh and the faulty baseline set Gf.

[0140] For engine performance status evaluation, Win d is the sliding window width, and the distance between the parameter sequence of the target domain engine in each time window and the healthy / faulty baseline is measured based on the WD algorithm. Figure 9 As shown in , the reverse distribution leads to a certain measurement error at the beginning and end of the healthy and faulty distances. The distance between the initial cycle and the healthy baseline, and the distance between the end cycle and the faulty baseline cannot represent the true performance status of the engine. Figure 9 As shown in Figure 2, the linear fusion algorithm is used to reduce the weights of the initial segment of the healthy distance and the final segment of the fault distance. The linear weight fusion method is used to transform the two WD curves into HI, which effectively solves the endpoint measurement deviation problem caused by the reverse distribution.

[0141] The results show that the HI degradation rate is slow during the initial stages of engine operation and gradually increases with the number of cycles. By identifying the HI inflection point, the time when the target engine begins to significantly degrade can be determined. Table 2 shows the parameter settings for engine health assessment and migration time identification.

[0142] Table 2 Engine health status assessment and migration time identification parameter settings

[0143] Parameter name <![CDATA[Win d ]]> <![CDATA[Win HI ]]> <![CDATA[Thre a ]]> Parameter value 5 10 0.00075

[0144] The linear fitting slope of the HI curve in the sliding window is defined as the degradation sensitive feature. a As a discrimination condition, the time when the degradation sensitive feature exceeds the threshold is the migration time. Taking the target engine No. 3 as an example, the identification results of the degradation stage and migration timing are as follows: Figure 10 shown.

[0145] The results show that as the engine operating time increases, the engine performance decreases, which is manifested by the acceleration of the HI value decrease rate. The degradation sensitive feature decreases and exceeds the degradation threshold Thre a , indicating that the engine has entered the degradation stage. The target engine's health fluctuates slightly but steadily in the early stages of degradation, and the corresponding degradation-sensitive feature fluctuates around 0, ranging from 0.0005 to 0.001. Therefore, Threa is set to 0.00075.

[0146] Part III: Transferability Measurement and Multi-Source Domain Deconstruction

[0147] The eight engine gas path parameters were input into the principal component analysis model. The results showed that the cumulative contribution rate of the first principal component reached 85%. Therefore, the first principal component was selected as the main degradation information to remove redundant information and reduce the calculation cost. The results are shown in Figure 2. Figure 11 As shown in Figure 3, source domain engines differ in terms of lifespan, initial health, degradation rate, and degradation characteristics. In addition to these differences, the operating time of each target engine also varies significantly. Therefore, the transferability of source domains and sequences varies significantly between target engines. The TL-EDM algorithm is used to measure the transferability of target engines and select source domains with high transferability. The parameter settings are shown in Table 3. The integrated distance weight based on WD and cosine distance is set to [1 / 3, 2 / 3], that is, WD accounts for 1 / 3 of the weight, cosine distance accounts for 1 / 3 of the weight, and the remaining 1 / 3 of the weight is determined by the DPC length of the target engine. Figure 12 The transferable distance measurement results between source engine sequence No. 1 and target engine No. 3 are shown.

[0148] In the multi-source domain deconstruction process, the number of pre-screened source domains S is set to 10, that is, the engines with the top 10% transferability are selected. The secondary selection ratio r is set to 1.5, and the source domains with RUL* in the interval [μ-1.5σ, μ+1.5σ] are retained as high transferability source domains. The multi-source domain deconstruction results of target engine No. 3 are shown in the figure below. Figure 13 shown.

[0149] Table 3. Transferability metrics and high-transferability source domain screening parameter settings

[0150] Parameter name <![CDATA[Ther d-up ]]> <![CDATA[Ther d-low ]]> S r Parameter value 1 / 3 2 / 3 10 1.5

[0151] Based on the proposed deconstruction algorithm, the highly transferable source domains of the target engine are screened. The pre-screening results of target engine No. 3 are as follows: Figure 13 -a. Figure 13 -b shows the secondary screening results based on the RUL* label. The results show that as the transferability of the source domain decreases, the distribution of RUL* gradually becomes unstable. Through secondary selection, source domain samples with unstable RUL* distribution are removed. Finally, the highly transferable source domains and sequences of the target engine are determined. The results are shown in Figure 13 -c. The results show that the proposed multi-source domain deconstruction algorithm can screen out source domains and sequences that are similar to the degradation trajectory of the target engine.

[0152] Part 4: RUL Personalized Migration Prediction

[0153] According to the characteristics of CMAPSS data and the model calculation cost, the parameter settings of the life prediction model are shown in Table 4. The number of neurons in the input layer of the DLSTM model is set to 14, which is consistent with the time series window length Win lstm Consistent. Seven hidden layers are used to learn engine degradation information. The number of neurons in each hidden layer is 14, 12, 10, 8, 6, 4, and 2, respectively. The activation function of the LSTM layer is tanh. The number of neurons in the output layer is set to 1. The loss function is the mean squared error function, and the training optimizer is Adam.

[0154] Table 4 Prediction model parameter settings

[0155]

[0156] Table 5 shows the training parameters of the DLSTM model. After pre-training, the weights of the first four LSTM layers were frozen. To reduce overfitting, the L2 regularization coefficient and the denoising rate were set to 0.01 and 0.1, respectively. The training batch size was set to 960, and the number of training iterations was set to 100.

[0157] Table 5 Prediction model training parameter settings

[0158] Parameter name L2 regularization coefficient Number of iterations Training batch size Noise reduction rate Parameter value 0.01 100 960 0.1

[0159] In the first transfer phase, the DPCs of all multi-source domain data are used as training data. The input is the normalized DPCs, and all data training weights are equal. The output is the normalized RUL. After model training, the first four layers of the model are set to the "untrainable" state, and the weights of the remaining layers are initialized. In the second transfer phase, source domains with high transferability are selected as input data for model fine-tuning. The data training weights are determined by the transferability based on the multi-source domain decomposition. Figure 14 The training data and weights for target engine No. 3.

[0160] As the transferability decreases, the training weight of the source domain gradually decreases. In addition, for a source domain, the weights of different position sequences are Gaussian-like, that is, the weights of sequences with higher transferability are higher, and the weights of sequences at other positions gradually decrease. The weights of the remaining layers of the prediction model are determined by fine-tuning, and the training results are as follows: Figure 15 The training results show that the initial training loss is low, indicating that the model has a certain degree of degenerate fitting ability. As the model iterates, the loss decreases and remains stable. The training loss and accuracy results show that the prediction model has stable performance.

[0161] According to the trained model, the prediction results of target engine No. 3 are as follows Figure 16 As shown in the figure, the prediction results for each target domain data window are close to the true RUL. Combining the prediction values ​​for all windows, the predicted RUL for target engine No. 3 is 83.18 cycles, which is very close to the true RUL of 82 cycles. The calculation results verify the accuracy of the model.

[0162] Although the present invention has been described in detail above, it is not limited thereto, and those skilled in the art can make various modifications based on the principles of the present invention. Therefore, any modifications made based on the principles of the present invention should be understood to fall within the scope of protection of the present invention.

Claims

1. A method for predicting the remaining useful life of equipment, characterized in that: include: Acquire operation monitoring data of the target device, and calculate the healthy bulldozer distance and the faulty bulldozer distance between the operation monitoring data of the target device and the screened healthy baseline set and the screened faulty baseline set; Converting the healthy bulldozer distance and the faulty bulldozer distance into the health of the target device using a linear weight fusion algorithm, and determining the migration time of the target device using the health of the target device; Obtaining the length of the target device degradation period according to the migration time of the target device, and calculating the cosine distance weight and the bulldozer distance weight respectively according to the length of the target device degradation period; Calculating the transferable distance between the target device and each device of the same model using the cosine distance weight and the bulldozer distance weight, and screening out multiple devices of the same model that are similar to the target device based on the transferable distance; Obtaining device information of each similar device of the same model, and fine-tuning the pre-trained DLSTM-FC prediction network model using the device information of each similar device of the same model to obtain a fine-tuned DLSTM-FC prediction network model; Inputting the device degradation sensitive parameter information of the target device into the fine-tuned and trained DLSTM-FC prediction network model for prediction processing to obtain the remaining service life information of the target device; Among them, DLSTM-FC refers to a deeper long short-term memory-fully connected layer.

2. The method according to claim 1, characterized in that Calculating the healthy bulldozer distance and the faulty bulldozer distance between the operation monitoring data of the target device and the screened healthy benchmark set and the faulty benchmark set includes: Acquire historical operation monitoring data of the training device, and filter out a healthy baseline set and a fault baseline set from the historical operation monitoring data of the training device; The bulldozer distance measurement method is used to calculate the healthy bulldozer distance between the operation monitoring data of the target device and the healthy reference set, and the faulty bulldozer distance between the operation monitoring data of the target device and the faulty reference set.

3. The method according to claim 2, characterized in that The method of converting the healthy bulldozer distance and the faulty bulldozer distance into the health of the target device by using a linear weight fusion algorithm includes: Among them, HI T is the health of the target device; w f is the fault distance weight; w h is the health distance weight; d f is the faulty bulldozer distance of the faulty benchmark set; d h It is the healthy bulldozer distance.

4. The method according to claim 1, wherein Calculating the cosine distance weight and the bulldozer distance weight according to the degradation period length of the target equipment includes: Among them, W cos is the weight of the cosine distance; t l is the degradation period length of the target device; T up ,T lower are the upper and lower limits of the distance weight respectively, and T up +T lower =1;L max ,L min are the maximum and minimum values ​​of the weight change length range respectively; W EMD is the weight of the bulldozer distance, and W cos +W EMD =1.

5. The method according to claim 4, characterized in that The calculating the migration distance between the target device and each device of the same model by using the cosine distance weight and the bulldozer distance weight includes: Extracting first degradation main information from the multi-dimensional monitoring parameters of the target device by a principal component analysis method, and extracting second degradation main information from the multi-dimensional monitoring parameters of the same model device; Obtaining a target degradation feature sequence from the first degradation main information, and obtaining a degradation feature sequence of the same model from the second degradation main information; Calculating the cosine distance and the bulldozer distance between the target degradation feature sequence and the degradation feature sequence of the same model respectively using the target degradation feature sequence and the degradation feature sequence of the same model; The migration distance between the target device and each device of the same model is calculated according to the cosine distance, the bulldozer distance, the cosine distance weight, and the bulldozer distance weight.

6. The method according to claim 5, characterized in that Calculating the migration distance between the target device and each device of the same model according to the cosine distance, the bulldozer distance, the cosine distance weight, and the bulldozer distance weight includes: d t =W cos *d cos (Z s ,Z t ,t)+W EMD *d EMD (Z s ,Z t ,t) Among them, d t is the migration distance corresponding to the time delay t; d cos is the cosine distance; d EMD is the bulldozer distance; Z s is the degenerate characteristic sequence of the same model; Z t is the target degradation feature sequence; t is the starting point of the delay sliding window.

7. The method according to claim 1, characterized in that The pre-trained DLSTM-FC prediction network model includes: Construct a DLSTM-FC prediction network model that consists of 5 layers of LSTM networks with different model structure parameters and 2 layers of FC networks with different model structure parameters from top to bottom; Device information of multiple devices of the same model is obtained, and the DLSTM-FC prediction network model is pre-trained using the device information of the multiple devices of the same model to obtain a pre-trained DLSTM-FC prediction network model.

8. The method according to claim 7, characterized in that The fine-tuning training of the pre-trained DLSTM-FC prediction network model using the device information of each device of the same model to obtain the fine-tuned DLSTM-FC prediction network model includes: Calculating a similarity value between each of the similar devices of the same model and the target device, and configuring a sample weight value for each of the similar devices of the same model according to each of the similarity values; By using the device degradation sensitive parameter information and device remaining service life information of each similar device of the same model with configured sample weight values ​​as the input and output of the pre-trained DLSTM-FC prediction network model, the weight values ​​of the last three network layers in the pre-trained DLSTM-FC prediction network model are fine-tuned to obtain the fine-tuned DLSTM-FC prediction network model.

9. The method according to claim 7, characterized in that The pre-training of the DLSTM-FC prediction network model by using the device information of the plurality of devices of the same model to obtain the pre-trained DLSTM-FC prediction network model comprises: Pre-training the weight value of each network layer in the DLSTM-FC prediction network model using the device information of the multiple devices of the same model to obtain the weight value of each network layer in the DLSTM-FC prediction network model; The pre-trained DLSTM-FC prediction network model is obtained by randomly initializing the weight values ​​of the last three network layers in the DLSTM-FC prediction network model, setting the training status of the weight values ​​of the first four network layers in the DLSTM-FC prediction network model to a non-trainable state, and setting the training status of the weight values ​​of the last three network layers to a trainable state; The device information includes device degradation sensitive parameter information and device remaining service life information.

10. A device for predicting the remaining useful life of equipment, characterized in that: include: a migration time determination module, configured to obtain operation monitoring data of a target device, and calculate a healthy bulldozer distance and a faulty bulldozer distance between the operation monitoring data of the target device and the screened healthy baseline set and faulty baseline set; Converting the healthy bulldozer distance and the faulty bulldozer distance into the health of the target device using a linear weight fusion algorithm, and determining the migration time of the target device using the health of the target device; a similar device screening module, configured to obtain the length of the degradation period of the target device according to the migration time of the target device, and calculate the cosine distance weight and the bulldozer distance weight respectively according to the length of the degradation period of the target device; Calculating the transferable distance between the target device and each device of the same model using the cosine distance weight and the bulldozer distance weight, and screening out multiple devices of the same model that are similar to the target device based on the transferable distance; The remaining useful life prediction module is used to obtain device information of each similar device of the same model, and use the device information of each similar device of the same model to fine-tune the pre-trained DLSTM-FC prediction network model to obtain a fine-tuned DLSTM-FC prediction network model; input the device degradation sensitive parameter information of the target device into the fine-tuned DLSTM-FC prediction network model for prediction processing to obtain the remaining useful life information of the target device; Among them, DLSTM-FC refers to a deeper long short-term memory-fully connected layer.

Images