Bearing residual service life prediction method based on causal discovery

Abstract

The invention relates to the technical field of bearing health monitoring, in particular to a bearing remaining service life prediction method based on causal discovery. The RUL prediction model provided by the invention comprises a data preprocessing module, a causal discovery module, a feature selection module and a life prediction module, according to the method, a causal discovery idea is introduced, and a multi-stage feature screening and time sequence prediction model is combined, so that an industrial field-oriented bearing life prediction framework is provided. According to the framework, false correlation information can be eliminated at a feature level, and causal driving factors are enhanced, so that prediction robustness and cross-scene applicability are remarkably improved. The maintenance cost can be effectively reduced, the equipment utilization rate is improved, and the method has important theoretical and practical values for promoting intelligent operation and maintenance and predictive maintenance.

Description

Technical Field

[0001] This invention relates to the field of bearing health monitoring technology, and in particular to a method for predicting the remaining service life of bearings based on causal discovery. Background Technology

[0002] With the development of industrial automation and intelligent manufacturing, rolling bearings, as core supporting components in various machines, directly affect the overall machine performance and safe production due to their reliability and operational stability. Bearing failure typically leads to equipment downtime, production interruptions, and even serious safety accidents, causing significant economic losses and safety risks to enterprises. Meanwhile, traditional periodic or reactive maintenance methods fall into two extremes: over-maintenance and delayed maintenance, wasting considerable financial and material resources and failing to detect potential problems in a timely manner. With advancements in sensor technology and the widespread adoption of industrial IoT platforms, data such as vibration signals, acoustic noise, and current can be collected in real time. Based on this data, the reliability of machine operation can be assessed, thus preventing more serious consequences. This provides a foundation for accurate prediction of RUL (Remaining Usage) of bearings. Accurate RUL prediction can issue early warnings before equipment failure, helping maintenance personnel to rationally plan repairs and equipment replacements, reducing unexpected downtime. Statistics show that predictive maintenance based on RUL can reduce maintenance costs by 18% to 25%.

[0003] Currently, bearing life prediction technology based on machine learning and deep learning is becoming increasingly mature, especially recurrent neural networks, which exhibit strong sequence modeling capabilities and can capture complex dynamic features in the bearing degradation process. However, existing data-driven methods generally suffer from two problems: first, redundant information in the high-dimensional feature space can easily lead to model overfitting, affecting prediction performance; second, traditional feature selection methods ignore the causal relationships between variables, resulting in insufficient generalization ability of the model across operating conditions or new equipment, making it difficult to provide reliable decision support for actual production in enterprises. Summary of the Invention

[0004] To overcome the shortcomings in robustness and generalization of existing bearing life prediction technologies, this invention proposes a bearing remaining service life prediction method based on causal discovery, which can effectively improve the generalization ability and performance stability of bearing cross-domain diagnosis.

[0005] The present invention proposes a bearing remaining service prediction method based on causal discovery. First, a RUL prediction model is constructed and trained, and then the trained RUL prediction model is used to predict the remaining service of the target bearing. The RUL prediction model includes: a data preprocessing module, a causal discovery module, a feature selection module, and a life prediction module. After the collected bearing vibration signal is processed by the data preprocessing module, the causal discovery module extracts candidate causal features that are causally dependent on RUL. The feature selection module selects the final causal features from the candidate causal features, and the life prediction module generates the remaining life prediction value based on the final causal features.

[0006] Preferably, the causal discovery module uses the GES algorithm or the PC algorithm.

[0007] Preferably, the feature selection module uses a voting algorithm or a random forest algorithm.

[0008] Preferably, the lifetime prediction module includes a GRU network and a fully connected output layer. The final causal feature set output by the feature selection module is converted into the remaining lifetime by the final hidden state extracted by the GRU network and then by the fully connected output layer.

[0009] Preferably, the fully connected output layer uses a linear activation function.

[0010] Preferably, the data preprocessing module processes the bearing vibration signal in one or more of the following ways: missing value supplementation, outlier removal, and noise addition.

[0011] Preferably, the RUL prediction model is first trained in the source domain, and then RUL prediction is performed in the target domain; the source domain and the target domain are bearings in different positions or under different operating conditions on the same device.

[0012] Preferably, the loss function used in the training of the RUL prediction model is the mean squared error loss.

[0013] The present invention proposes a bearing remaining service life prediction system based on causal discovery, comprising a memory and a processor. The memory stores a computer program, and the processor is connected to the memory. The processor is used to execute the computer program to realize the bearing remaining service life prediction method based on causal discovery.

[0014] The present invention proposes a storage medium storing a computer program, which, when executed, is used to implement the bearing remaining service life prediction method based on causal discovery.

[0015] The advantages of this invention are: (1) This invention proposes a bearing remaining service life prediction method based on causal discovery. It introduces the concept of causal discovery and combines it with a multi-stage feature selection and time-series prediction model to propose a bearing life prediction framework for industrial sites. This framework can eliminate pseudo-related information and strengthen causal driving factors at the feature level, thereby significantly improving prediction robustness and cross-scenario applicability. It can effectively reduce maintenance costs and improve equipment utilization, and has important theoretical and practical value for promoting the implementation of intelligent operation and maintenance and predictive maintenance.

[0016] (2) In predicting bearing remaining life, the present invention effectively utilizes causal discovery technology to uncover the causal driving features that are truly related to RUL, and on this basis, improves the robustness and generalization of the prediction model. Attached Figure Description

[0017] Figure 1 This invention proposes a RUL prediction model architecture; Figure 2 GRU framework diagram; Figure 3 This is a flowchart of a bearing remaining service life prediction method based on causal discovery proposed in this invention. Figure 4 The statistical results are shown for the example. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, the RUL prediction model proposed in this embodiment includes: a data preprocessing module, a causal discovery module, a feature selection module, and a lifetime prediction module.

[0020] The data preprocessing module is used to process the data samples, including missing value imputation, outlier removal, and noise addition. The observation data collected by the system (i.e., the raw data samples) often contains missing values ​​and outliers. Without processing, this can lead to singular covariance matrices and invalidation of independence tests, preventing the causal discovery algorithm from yielding results correctly. Furthermore, causal discovery relies on the accuracy of statistical tests. In this implementation, the data preprocessing module performs anomaly cleaning and adds a small amount of noise to the data samples, which can improve the robustness of the model's independence test. The data sample is a bearing vibration signal.

[0021] The causal discovery module aims to recover the directed acyclic graph between variables based on the data samples output by the data preprocessing module, thereby identifying features that can serve as direct parent nodes of the RUL and which are causally dependent on the RUL.

[0022] The causal discovery module can employ a greedy equivalence search algorithm (hereinafter referred to as the GES algorithm) to construct a directed acyclic graph (DAG) based on a global score function through forward edge addition and backward pruning. Alternatively, the PC algorithm can be used to construct a skeleton graph and determine edge direction constraints through conditional independence checks, thereby constructing a DAG. Considering that the independence check of the PC algorithm is sensitive to noise, and the GES algorithm is more robust with a limited sample size, the GES algorithm is preferred in this implementation.

[0023] Specifically, the GES algorithm uses the Bayesian Information Criterion (BIC) as a metric for graph structure, applying it to each node. Define local score: ; in, For the i-th time-frequency feature extracted by the GES algorithm, It exhibits multi-scale characteristics. for The number of parent nodes that belong to the RUL associated variables. ( ) represents the maximum log-likelihood of a node given its parent set; N represents the number of feature nodes, i.e., the video features extracted for a single data sample. The number of, 1≤i≤N, Extracted by GES algorithm The number of parent nodes that belong to the RUL associated variables.

[0024] The GES algorithm constructs a directed acyclic graph (DAG) through two phases: forward addition and backward pruning. .

[0025] 1) Forward addition: Starting from an empty graph, repeatedly try to add a new edge to any broken edge in the current equivalence class, evaluate the score gain ΔScore, select the step with the largest score gain to add, until there are no more steps to gain or the number of parent nodes of any node reaches the upper limit.

[0026] 2) Backward pruning: Based on the forward result, repeatedly try deleting arbitrary edges and evaluate the score change; if deleting an edge can improve the score, perform the deletion operation until there are no more steps that can improve the score. This ultimately yields a locally optimal sparse DAG. .

[0027] The causal discovery module obtains information from the generated DAG. Extract all incoming edges pointing to RUL nodes, as shown in the following formula. These parent nodes are the candidate causal features, and the set of candidate causal features is... Notation: ; in, This represents the j-th time-frequency feature extracted by the GES algorithm. Indicated in DAG middle Corresponding node (abbreviation) The node points to the corresponding node of RUL (referred to as the RUL node). Let RUL be the set of incoming edges of a node. The incoming edges of a RUL node are the edges from other nodes to the RUL node.

[0028] The feature selection module selects from the candidate causal feature set The final causal features are selected so that the life prediction module can predict the remaining life of the bearing based on the final causal features.

[0029] The feature selection module filters candidate causal features, eliminating those with little or redundant influence, which helps avoid model overfitting and improves model generalization. At the same time, by filtering causal features, the final causal features are ensured to have a high degree of causal dependence on the RUL, thereby improving the robustness of the model.

[0030] The feature selection module can specifically use a random forest regressor. Each candidate causal feature is ranked by importance to achieve secondary screening; specifically, the number of features to be screened can be specified, so that a specified number of final causal features are selected in descending order of importance.

[0031] A random forest consists of T decision trees. Each tree generates different subsamples from the original training set through sampling with replacement. At the same time, m features are selected for optimal splitting when splitting nodes.

[0032] Random forests use a scrambling method to calculate importance; see the formula below for details. ; in, Features Importance Let be the original mean squared error of the m-th tree on the out-of-bag samples. To feature The mean squared error of the feature on samples outside the same bag after random scrambling. The higher the importance value, the greater the contribution of the feature to the prediction of RUL regression.

[0033] The lifetime prediction module specifically includes a GRU network (gated recurrent network) and a fully connected output layer. The GRU network introduces update gates and reset gates to dynamically control the retention and discarding of information in time, maintaining performance comparable to LSTM (Long Short-Term Memory) networks with a simpler gating mechanism and fewer parameters.

[0034] The internal structure of the GRU network is as follows: Figure 2 As shown. Input for time step t. (i.e., the final causal feature set obtained by processing data samples at time step t through the preprocessing module, causal discovery module, and feature selection module) and the hidden layer output at step t-1. (i.e., the input at time step t-1) After processing by the GUR network, the hidden states output by the hidden layers of the GUR network are multiplied by their respective weight matrices and then summed. Finally, these sums are input into the Sigmoid activation function to obtain the update gate. : ; in, and As weight, For bias, This represents the Sigmoid activation function.

[0035] The initial value of t is 1, corresponding to the first observation time point; at t=1, the hidden state... It is a vector of all zeros.

[0036] It is a hidden state A vector of the same dimension, representing the proportion of new information that needs to be updated in the final hidden state. GRU uses this vector to perform the final hidden state update. Perform interpolation to obtain the updated hidden state. : ; The candidate hidden states generated at time step t; when The value of the i-th dimension When the value is close to 0, it indicates that more states are retained, ultimately... ≈ , representing the hidden state at time step t The value of the i-th dimension Hidden state at time step t-1 The value of the i-th dimension ;when When the value is close to 1, it indicates that more new candidate states are introduced, and ultimately... ≈ , express The value of the i-th dimension.

[0037] For the reset door In general, its calculation formula is similar to that of the update gate, but the weight matrices used are different: ; in, and As weight, For bias.

[0038] After obtaining the reset door Then, the candidate hidden states are obtained using the following formula. : ; in, It is the hyperbolic tangent function. and As weight, For brand.

[0039] In bearing RUL prediction, the GRU network can effectively capture the nonlinear dynamic changes of vibration signals over time, mitigating long-term dependence.

[0040] When the GRU network processes the input time-series feature sequence After loop processing, the hidden state H at the final time step T T It contains condensed information about the entire sequence. H T It is passed to the fully connected output layer. This fully connected output layer typically uses a linear activation function to pass H... T Map to bearing remaining life (RUL) and output.

[0041] Reference Figure 3 In this embodiment, the RUL prediction model can be trained on the source domain RUL dataset and then used for RUL prediction in the target domain. The source domain training process includes the following steps: S1. Extract training samples from the source domain RUL dataset and let the RUL prediction model perform machine learning on the training samples; S2. Extract validation samples from the source domain RUL dataset and have the RUL prediction model output the remaining lifetime prediction value of the validation samples. S3. Calculate the model loss on the validation samples and update the RUL prediction model based on backpropagation of the loss; Repeat the above steps until the model converges.

[0042] The model convergence condition can be set to the number of model updates reaching a set value, or the model loss being less than a set loss threshold for three consecutive times. The model loss can specifically be the mean squared error loss, the mean error, or cross-entropy, etc.

[0043] The RUL prediction model described above is illustrated and verified below with specific embodiments using a typical publicly available dataset for bearing life prediction.

[0044] The datasets used in this embodiment include: horizontal vibration signals from the XJTU-SY bearing dataset and horizontal vibration signals from the IEEE PHM Challenge 2012 bearing dataset. The IEEE PHM Challenge 2012 bearing dataset is referred to as the PHM bearing dataset.

[0045] The XJTU-SY bearing dataset, jointly released by Xi'an Jiaotong University and Zhejiang Changjiang Sanyang Technology Co., Ltd., contains vibration data for 15 rolling bearings throughout their entire lifecycle, from initial operation to failure. The bearings used in the experiments were LDK UER204 rolling bearings. Vibration signals were acquired using two PCB 352C33 accelerometers, installed horizontally and vertically on the bearing housing, respectively. The sampling frequency was 25.6 kHz, with a sampling interval of 1 minute, and each sampling lasted 1.28 seconds.

[0046] Three operating conditions were set up for the experiment, as shown in Table 1.

[0047] Table 1 Introduction to the XJTU-SY dataset ; The PHM bearing dataset was provided by the French FEMTO-ST Institute. The bearing used in the experiment was a Rexnord ZA-2115 pillow-type double-row spherical roller bearing. The measurement system used a PCB 353B33 ICP miniature accelerometer, and the vibration signal was acquired at 25.6 kHz, with a single sampling time of 0.1 s and an interval of 10 s.

[0048] Three operating conditions were set up for the experiment, as shown in Table 2.

[0049] Table 2 Introduction to the PHM dataset ; In this embodiment, historical time-series data of bearing horizontal vibration signals are extracted from the dataset. After processing by the preprocessing module, the data samples are divided into windows. The GES algorithm extracts the time-frequency features of each data sample, including mean, variance, skewness, kurtosis, entropy, fractal dimension, peak factor, impulse factor, margin factor, energy ratio, spectral flatness, peak vibration, and RMS vibration.

[0050] In this embodiment, the remaining bearing life is uniformly processed by combining window division to generate normalized RUL labels; thus, the RUL label of the data sample corresponding to the nth window on a certain data line is... for: ; Where N is the total number of bearing acquisition windows.

[0051] In this embodiment, three sets of experiments were conducted; in each set of experiments, the model was trained on the source domain and then the model performance was tested on the target domain.

[0052] Experiment 1: Using the working conditions and bearings of the XJTU-SY dataset to divide the data into domains, a generalization experiment was conducted.

[0053] In this experiment, three RUL predictions were performed on the bearing with different source and target domains. The specific details of the experimental task are shown in Table 3.

[0054] Table 3 XJTU Generalization Experiment Tasks ; Bearing pq represents the qth bearing under operating condition p.

[0055] In Experiment 1, twelve cross-domain RUL predictions were performed for each task, and the average of ten trials for each task was taken to represent the model's average performance. The experimental results were evaluated using RMSE (Root Mean Squared Error). Detailed experimental results are shown in Table 4.

[0056] Table 4. Generalization Experiment Results of XJTU-SY (RMSE) ; Table 4 shows that Task A has the smallest error, followed by Task B, while Task C has the largest error. This indicates that the model has the best generalization ability when bearing 1_1 is used as the source domain, and the average error is the highest when bearing 3_1 is used as the source domain. Bearings 2_5 and 3_4 show relatively high errors in all three tasks, possibly because the vibration data of these two bearings varies significantly, and their data distribution differs considerably from that of the other bearings. In conclusion, the model of this invention demonstrates generalization ability under different operating conditions.

[0057] Experiment 2: A cross-dataset experiment was conducted using the XJTU-SY dataset as the source domain and the PHM dataset as the target domain. The experimental setup is shown in Table 5.

[0058] Table 5 Generalization Experimental Tasks Across PHM ; Task D was conducted 10 times, and the average value was taken to demonstrate the average performance of the model. The experimental results were evaluated using RMSE. The specific experimental results are shown in Table 6.

[0059] Table 6. Results of generalization experiments across PHM ; As can be seen from Table 6, the method of the present invention still performs well on cross-dataset tasks.

[0060] Experiment 3 is an ablation experiment. First, the causal discovery module and feature selection module were removed from the RUL prediction model proposed in this invention (hereinafter referred to as the invention model), forming an ablation model. In the ablation model, data samples are processed by the preprocessing module and then directly enter the GRU network for processing. The remaining lifetime is then output by the fully connected output layer.

[0061] The ablation model was tested on tasks A, B, C, and D, respectively, and RMSE was used as the evaluation metric. The ablation experimental results are compared with the experimental results of the model of this invention. Figure 4 As shown in the figure, it can be seen that the performance of the model of this invention is better than that of the ablation model on the same task. Moreover, the performance of the model of this invention is more stable across different tasks.

[0062] Of course, those skilled in the art will recognize that the present invention is not limited to the details of the exemplary embodiments described above, but also includes the same or similar structures that can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0063] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

[0064] The technologies, shapes, and structures not described in detail in this invention are all known technologies.

Claims

1. A method for predicting the remaining service life of bearings based on causal discovery, characterized in that, First, a RUL prediction model is constructed and trained. Then, the trained RUL prediction model is used to predict the remaining life of the target bearing. The RUL prediction model includes: a data preprocessing module, a causal discovery module, a feature selection module, and a life prediction module. After the collected bearing vibration signal is processed by the data preprocessing module, the causal discovery module extracts candidate causal features that are causally dependent on RUL. The feature selection module selects the final causal features from the candidate causal features, and the life prediction module generates the remaining life prediction value based on the final causal features.

2. The bearing remaining service life prediction method based on causal discovery as described in claim 1, characterized in that, The causal discovery module uses either the GES algorithm or the PC algorithm.

3. The bearing remaining service life prediction method based on causal discovery as described in claim 1, characterized in that, The feature selection module uses either a voting algorithm or a random forest algorithm.

4. The bearing remaining service life prediction method based on causal discovery as described in claim 1, characterized in that, The lifetime prediction module includes a GRU network and a fully connected output layer. The final causal feature set output by the feature selection module is converted into the remaining lifetime by the final hidden state extracted by the GRU network and then by the fully connected output layer.

5. The bearing remaining service life prediction method based on causal discovery as described in claim 4, characterized in that, The fully connected output layer uses a linear activation function.

6. The bearing remaining service life prediction method based on causal discovery as described in claim 1, characterized in that, The data preprocessing module processes bearing vibration signals in one or more of the following ways: missing value supplementation, outlier removal, and noise addition.

7. The bearing remaining service life prediction method based on causal discovery as described in claim 1, characterized in that, First, a RUL prediction model is trained in the source domain, and then RUL prediction is performed in the target domain; the source domain and the target domain are bearings in different positions or under different operating conditions on the same device.

8. The bearing remaining service life prediction method based on causal discovery as described in claim 1, characterized in that, The loss function used in the training of the RUL prediction model is the mean squared error loss.

9. A bearing remaining service life prediction system based on causal discovery, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, the processor is connected to the memory, and the processor is used to execute the computer program to implement the bearing remaining service life prediction method based on causal discovery as described in any one of claims 1-8.

10. A storage medium, characterized in that, The system contains a computer program that, when executed, is used to implement the bearing remaining service life prediction method based on causal discovery as described in any one of claims 1-8.

Images