Incremental scene-based mechanical equipment fault prediction method and system based on meta learning

Abstract

The application provides a mechanical equipment fault prediction method and system based on meta-learning in an incremental scene, which comprises the following steps: constructing an initial prediction model and a memory bank, and collecting full life cycle signals generated by corresponding mechanical equipment; acquiring monitoring signals generated by the mechanical equipment that fails and the remaining time label before failure in the degradation trajectory corresponding to the full life cycle signals according to a preset time interval, and generating time series information according to the data set corresponding to the degradation trajectory; generating a support set according to the time series information, generating a query set according to the data read out from the memory bank, meta-training the initial prediction model through the support set and the query set, performing meta-testing on the initial prediction model through the support set, and generating a corresponding fault prediction model; and predicting the fault time according to the truncated data corresponding to the monitoring signals through the fault prediction model to output a corresponding prediction result. The application can improve the prediction efficiency.

Description

Technical Field

[0001] This invention relates to the field of equipment testing technology, and in particular to a method and system for predicting mechanical equipment faults based on meta-learning in incremental scenarios. Background Technology

[0002] In industrial production, meta-learning-based fault prediction methods for mechanical equipment in incremental scenarios are core technologies for ensuring stable equipment operation, mitigating downtime risks, and optimizing the allocation of operation and maintenance resources. Their core objective is to accurately predict equipment degradation trends and remaining failure time based on equipment operation monitoring data. Currently, data-driven fault prediction methods (especially deep learning-based solutions) have become the mainstream technology in the industry because they do not rely on complex physical models and possess excellent temporal feature extraction and nonlinear modeling capabilities.

[0003] In real-world industrial scenarios, mechanical equipment has long operating cycles and random failures, making it difficult to acquire full-lifecycle degradation data in a short period. Instead, data is often presented as incremental updates that are used sequentially. This necessitates that prediction methods possess the ability to continuously learn from new data. Existing deep learning-based fault prediction methods generally lack this capability. After acquiring new data, they require integrating old and new data to train the model from scratch, which not only consumes significant computational resources and time but also suffers from high historical data storage pressure and low model update efficiency, severely hindering their practical application.

[0004] Furthermore, to address the aforementioned issues, some studies have proposed improved incremental prediction methods based on regularization or replay strategies. However, these methods still have significant technical limitations: the former relies on manually designed heuristic rules to constrain parameter updates in order to reduce the forgetting of historical knowledge, lacks a mechanism for adaptively learning general degradation features, has insufficient generalization ability, and is difficult to adapt to new equipment degradation patterns; the latter, when selecting replay samples in resource-constrained scenarios, does not fully consider inherent characteristics such as differences in equipment degradation rates and trajectory diversity, resulting in samples that cannot truly reflect the dynamic laws of degradation, ultimately affecting prediction accuracy. Both methods are insufficient to meet the actual needs of high-precision fault prediction in industrial scenarios. Summary of the Invention

[0005] Based on this, the purpose of this invention is to provide a mechanical equipment fault prediction method and system based on meta-learning in incremental scenarios, so as to solve the problem that the existing technology does not fully consider the inherent characteristics such as the difference in equipment degradation rate and the diversity of trajectories, which makes it difficult to meet the requirements of high-precision fault prediction.

[0006] The first aspect of the present invention proposes:

[0007] A method for predicting mechanical equipment faults based on meta-learning in incremental scenarios, wherein the method includes:

[0008] An initial prediction model and memory are constructed, and the full life cycle signals generated by the mechanical equipment are collected synchronously during the execution of tasks by the mechanical equipment.

[0009] According to a preset time interval, the monitoring signal generated by the failed mechanical equipment and the remaining time label before the failure are obtained from the degradation trajectory corresponding to the full life cycle signal, and time series information is generated synchronously based on the dataset corresponding to the degradation trajectory.

[0010] A support set is generated based on the time series information, and a query set is generated synchronously based on the data read from the memory bank. The initial prediction model is then meta-trained using the support set and the query set, and the initial prediction model is then meta-tested using the support set, and a corresponding fault prediction model is generated.

[0011] Based on the remaining time stamp, the fault prediction model predicts the fault time according to the truncated data corresponding to the monitoring signal, and outputs the corresponding prediction result.

[0012] The beneficial effects of this invention are as follows: This solution addresses the problem that existing technologies fail to adequately consider the inherent characteristics of equipment degradation rate differences and trajectory diversity, making it difficult to meet the demand for high-precision fault prediction. By constructing an initial prediction model and memory, synchronously collecting signals throughout the entire life cycle of mechanical equipment, accurately extracting monitoring signals, remaining time labels, and time series information of failed equipment, and combining support sets and query sets to complete model meta-training and meta-testing, it effectively adapts to the differences in equipment degradation characteristics, enhances the model's adaptability and generalization ability to diverse degradation trajectories, and achieves accurate fault time prediction based on remaining time labels and monitoring truncation data, truly meeting the high-precision prediction needs in practical applications.

[0013] Furthermore, the steps of generating a support set based on the time series information and simultaneously generating a query set based on the data read from the memory include:

[0014] The monitoring signals are normalized to generate corresponding target monitoring signals;

[0015] A sliding time window strategy is introduced, which uses multi-point information continuously collected within a preset time span as the overall input based on the target monitoring signal to generate the time series information accordingly;

[0016] Based on the time series signal, target data corresponding to the current task under the full life cycle degradation trajectory is collected, and the support set is constructed based on the target data. Simultaneously, the query set is constructed based on the data read from the memory bank.

[0017] Furthermore, the expression for the support set is:

[0018]

[0019] The expression for the query set is:

[0020]

[0021] in, Indicates task Number of samples in the degradation trajectory. This indicates the number of samples in the replay experience. and Indicates task It supports centralized preprocessed input data and corresponding remaining time labels before the fault. and The labels represent the preprocessed input data and the corresponding remaining time before the fault in the replay experience.

[0022] Furthermore, the step of meta-training the initial prediction model using the support set and the query set includes:

[0023] The initial prediction model includes a meta-knowledge encoder, a regressor, a masking function, and a health-aware unit. The meta-training process is divided into an inner loop and an outer loop, and includes the following steps:

[0024] The inner loop uses the support set to update the regressor, the mask function, and the health sensing unit:

[0025]

[0026] in, For the inner loop learning rate, For mean square error loss, This indicates multiplication by channel. Representing tasks The mask function, health sensing unit, and regressor parameters updated in the inner loop. Indicates regular loss. and Indicates task It supports centralized preprocessed input data and corresponding remaining time labels before the fault. The parameter is Meta-knowledge encoder, The parameter is The regressor, The parameter is The mask function.

[0027] Furthermore, the expression for the regularization loss is:

[0028]

[0029] in, This represents sparse regularization. This indicates that health perception is regularized. and represent the weight coefficients of the sparse regularization term and the health perception regularization term, respectively. Indicates potential health factors, Describing the L1 norm, Indicates task Supports centralized preprocessed input data. The parameter is Health sensing unit The parameter is mask function, This represents the mean square error loss.

[0030] Furthermore, the outer loop updates the meta-knowledge encoder using the query set:

[0031]

[0032] in, For the outer loop learning rate, Indicates the parameters of the meta-knowledge encoder. For mean square error loss, The parameter is Meta-knowledge encoder, This indicates multiplication by channel. The parameter is mask function, and The labels represent the preprocessed input data and the corresponding remaining time before the fault in the replay experience. This represents the update coefficient.

[0033] Furthermore, the step of performing a meta-test on the initial prediction model using the support set includes:

[0034] During the meta-test, the regressor, the masking function, and the health-aware unit are updated using the support set to perform task fine-tuning:

[0035]

[0036] in, Representing tasks The mask function, health sensing unit, and regressor parameters after the mid-cycle test. For the inner loop learning rate, For mean square error loss, The parameter is The regressor, The parameter is Meta-knowledge encoder, This indicates multiplication by channel. The parameter is mask function, and Indicates task It supports centralized preprocessed input data and corresponding remaining time labels before the fault. This indicates a regular loss.

[0037] The second aspect of the present invention proposes:

[0038] A mechanical equipment fault prediction system based on meta-learning in incremental scenarios, wherein the system includes:

[0039] The construction module is used to build an initial prediction model and a memory library, and synchronously collect the full life cycle signals generated by the mechanical equipment during the execution of tasks.

[0040] The processing module is used to acquire the monitoring signals generated by the failed mechanical equipment and the remaining time label before the failure in the degradation trajectory corresponding to the full life cycle signal according to a preset time interval, and synchronously generate time series information based on the dataset corresponding to the degradation trajectory.

[0041] The reading module is used to generate a support set based on the time series information, and synchronously generate a query set based on the data read from the memory bank, so as to perform meta-training on the initial prediction model through the support set and the query set, perform meta-testing on the initial prediction model through the support set, and generate a corresponding fault prediction model.

[0042] The prediction module is used to predict the fault time based on the remaining time tag and the fault prediction model according to the truncated data corresponding to the monitoring signal, so as to output the corresponding prediction result.

[0043] Furthermore, the reading module is specifically used for:

[0044] The monitoring signals are normalized to generate corresponding target monitoring signals;

[0045] A sliding time window strategy is introduced, which uses multi-point information continuously collected within a preset time span as the overall input based on the target monitoring signal to generate the time series information accordingly;

[0046] Based on the time series signal, target data corresponding to the current task under the full life cycle degradation trajectory is collected, and the support set is constructed based on the target data. Simultaneously, the query set is constructed based on the data read from the memory bank.

[0047] Furthermore, the expression for the support set is:

[0048]

[0049] The expression for the query set is:

[0050]

[0051] in, Indicates task Number of samples in the degradation trajectory. This indicates the number of samples in the replay experience. and Indicates task It supports centralized preprocessed input data and corresponding remaining time labels before the fault. and The labels represent the preprocessed input data and the corresponding remaining time before the fault in the replay experience.

[0052] Furthermore, the reading module is specifically used for:

[0053] The initial prediction model includes a meta-knowledge encoder, a regressor, a masking function, and a health-aware unit. The meta-training process is divided into an inner loop and an outer loop, and includes the following steps:

[0054] The inner loop uses the support set to update the regressor, the mask function, and the health sensing unit:

[0055]

[0056] in, For the inner loop learning rate, For mean square error loss, This indicates multiplication by channel. Representing tasks The mask function, health sensing unit, and regressor parameters updated in the inner loop. Indicates regular loss. and Indicates task It supports centralized preprocessed input data and corresponding remaining time labels before the fault. The parameter is Meta-knowledge encoder, The parameter is The regressor, The parameter is The mask function.

[0057] Furthermore, the expression for the regularization loss is:

[0058]

[0059] in, This represents sparse regularization. This indicates that health perception is regularized. and represent the weight coefficients of the sparse regularization term and the health perception regularization term, respectively. Indicates potential health factors, Describing the L1 norm, Indicates task Supports centralized preprocessed input data. The parameter is Health sensing unit The parameter is mask function, This represents the mean square error loss.

[0060] Furthermore, the outer loop updates the meta-knowledge encoder using the query set:

[0061]

[0062] in, For the outer loop learning rate, Indicates the parameters of the meta-knowledge encoder. For mean square error loss, The parameter is Meta-knowledge encoder, This indicates multiplication by channel. The parameter is mask function, and The labels represent the preprocessed input data and the corresponding remaining time before the fault in the replay experience. This represents the update coefficient.

[0063] Furthermore, the reading module is specifically used for:

[0064] During the meta-test, the regressor, the masking function, and the health-aware unit are updated using the support set to perform task fine-tuning:

[0065]

[0066] in, Representing tasks The mask function, health sensing unit, and regressor parameters after the mid-cycle test. For the inner loop learning rate, For mean square error loss, The parameter is The regressor, The parameter is Meta-knowledge encoder, This indicates multiplication by channel. The parameter is mask function, and Indicates task It supports centralized preprocessed input data and corresponding remaining time labels before the fault. This indicates a regular loss.

[0067] The third aspect of the present invention proposes:

[0068] A computer includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the incremental scenario-based mechanical equipment fault prediction method as described above.

[0069] The fourth aspect of the present invention proposes:

[0070] A readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the incremental scenario-based mechanical equipment fault prediction method as described above.

[0071] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0072] Figure 1 A schematic diagram illustrating the model structure and the meta-training and meta-testing processes of the initial prediction model provided in the first embodiment of the present invention;

[0073] Figure 2 This is a schematic diagram of the structure of the memory bank provided in the first embodiment of the present invention.

[0074] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0075] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0076] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0077] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0078] Please see Figure 1 The figure shows the mechanical equipment fault prediction method based on meta-learning in incremental scenarios provided by the first embodiment of the present invention. The mechanical equipment fault prediction method based on meta-learning in incremental scenarios provided by this embodiment can accurately identify the degradation trajectory of the equipment, thereby improving the fault prediction efficiency of the mechanical equipment.

[0079] Specifically, this embodiment provides:

[0080] S1. Construct a prediction model, which includes a meta-knowledge encoder, a regressor, a masking function, and a health-aware unit. The meta-knowledge encoder extracts generalized representations of shared knowledge, the regressor learns the decision boundary for fault prediction, and the masking function selectively activates relevant features in the encoder through binary channels to achieve task fine-tuning. The health-aware unit further constrains the channel mask to reflect semantic features related to health status during degradation. Simultaneously, a high-quality memory is constructed: representative experiences are replayed from it before training for each task, and a portion of the current task's samples are written into the memory after training for efficient updates.

[0081] S2. Collect the full lifecycle signals of the mechanical equipment in the current task, and acquire the monitoring signals of the currently failed mechanical equipment throughout its entire lifecycle at certain sampling time intervals. Simultaneously, generate a corresponding remaining time label before failure based on the remaining number of cycles the equipment has undergone from that moment until final failure.

[0082] S3. A normalization strategy is employed to map the measurements from different sensors throughout the entire lifecycle degradation trajectory to a range of -1 to 1, thereby promoting stable convergence during model training. Subsequently, a sliding time window is used to acquire data from all collection points within a given time period, extracting richer temporal features. Based on this, a support set is constructed using the preprocessed degradation data for the current task, and a query set is formed by reading a subset of samples from the memory.

[0083] S4. Perform the meta-training process using the support set and query set. Meta-training consists of an inner loop and an outer loop: the inner loop rapidly updates the regressor, mask function, and health-aware unit on the support set, while preserving its computation graph for the formation of the meta-gradient flow in the subsequent outer loop; the outer loop calculates the meta-gradient based on the data in the query set and updates the meta-knowledge encoder accordingly, enabling it to learn highly generalizable representations.

[0084] S5. Perform meta-testing using the support set: The model first updates the regressor, mask function, and health-aware unit on the support set to fine-tune it for the current task; then the fine-tuned model is applied to the truncated degradation data to be tested to achieve high-quality and efficient prediction of the remaining time before failure.

[0085] S6. After the task is completed, a portion of samples are selected from the full lifecycle degradation trajectory of the current task and written into the memory bank to support the replay of experience in the next task. To this end, a degradation information-enhanced sample selection strategy is introduced: degradation rate information is incorporated into a density-based sampling method to calculate the sampling probability at each time point in the degradation trajectory; at the same time, a selection criterion based on the diversity of degradation trajectories is used to determine the sampling ratio of the current degradation trajectory, thereby preserving more representative degradation experience under limited storage resources.

[0086] Furthermore, in this embodiment, it should be noted that, preferably, in step S1, a prediction model is constructed, which includes a meta-knowledge encoder, a regressor, a masking function, and a health perception unit. The parameters are as follows: Meta-knowledge encoder A generalized representation space for extracting shared knowledge, with parameters as follows: The regressor Used to obtain the decision boundary for fault prediction, with parameters as follows: mask function The generalized representation extracted by the meta-knowledge encoder is selectively activated using a binary channel mask to perform task fine-tuning, with parameters as follows: Health Sensing Unit Further encouragement is given to ensure that channel masks exhibit health-related semantic information during degradation. Simultaneously, a high-quality memory is constructed, providing replay experience before each task training and storing partial data of the current task into the memory after each task training to perform efficient memory updates.

[0087] Preferably, in step S2, the full life cycle signal of the mechanical equipment in the current task is collected, and the monitoring data and the remaining time label before the failure are obtained at each time point in the degradation trajectory according to a certain sampling time interval.

[0088] Preferably, in step S3, the raw monitoring signals from different sensors are normalized to map them all to the range of -1 to 1. For data with multiple operating conditions, the measured values ​​under each condition also need to be normalized in the same way to ensure that the input scale is consistent across different operating conditions, thereby avoiding model convergence difficulties caused by differences in units. Subsequently, a sliding time window strategy is introduced, using multi-point information continuously collected within a certain time span as the overall input, so that it can contain more complete time series features, allowing the model to capture the correlation between adjacent time points and learn the equipment degradation pattern more fully from the time series structure, thereby obtaining more accurate failure time prediction. A support set is constructed using preprocessed data from the current task's full life cycle degradation trajectory, and a query set is constructed by randomly reading a portion of data from the memory. For the current task t, the support set is represented as:

[0089] ,

[0090] The query set is represented as:

[0091] ,

[0092] in, This represents the number of samples in the degenerate trajectory in task t. This indicates the number of samples in the replay experience. and This indicates that task t supports the preprocessed input data and the corresponding remaining time label before the fault. and The labels represent the preprocessed input data and the corresponding remaining time before the fault in the replay experience.

[0093] Preferably, in step S4, the present invention performs a meta-training process using the support set and the query set. The meta-training process is divided into an inner loop and an outer loop, as follows:

[0094] (1) The inner loop uses the support set update regressor, mask function and health sensing unit:

[0095]

[0096]

[0097] in, For the inner loop learning rate, For mean square error loss, This indicates multiplication by channel. These represent the mask function, health sensing unit, and regressor parameters updated in the inner loop of task t, respectively. This represents a regularized loss. The inner loop simulates a temporary update during task fine-tuning, preserving the computation graph to provide the initial gradient for the subsequent outer loop's meta-gradient. Furthermore, to further ensure the interpretability of the binary channel mask, this invention designs a regularized loss. Including sparse regularization and health perception regularization Sparse regularization guides the mask generation process to produce task-specific activations, while health-aware regularization ensures that the binary mask is meaningful in terms of health-related semantics.

[0098]

[0099] in, and represent the weight coefficients of the sparse regularization term and the health perception regularization term, respectively. Indicates potential health factors, This represents the L1 norm.

[0100] (2) The outer loop uses the query set to update the metadata encoder.

[0101]

[0102] in, The learning rate is set to the outer loop learning rate. The meta-knowledge encoder is optimized by taking the generalization performance of the model on the query set after updates in the inner loop as the optimization objective. This optimization obtains the generalization representation through meta-gradient information, where the query set should have sufficient diversity to approximate the generalization evaluation of the test scenario.

[0103] Preferably, in step S5, the present invention utilizes a support set to perform a meta-testing process. The meta-testing process uses a support set to update the regressor, a mask function, and a health-aware unit to perform task fine-tuning.

[0104]

[0105] in, Let represent the mask function, health-aware unit, and regressor parameters after meta-testing in task t, respectively. Based on the meta-knowledge encoder optimized by meta-training, this process performs substantive updates to the regressor, mask function, and health-aware unit for the task, and then performs fault time prediction on the truncated data to be tested.

[0106] Preferably, in step S6, degradation information is incorporated into the sample writing strategy to ensure the quality of experience in the memory bank under limited resources, as follows:

[0107] (1) Degradation rate information is integrated into a density-based sampling strategy to calculate the sampling probability of samples at different time points within the trajectory. Based on constructing a local sample density metric based on a k-nearest neighbor density estimator to guide the sampling probability, a time difference weighting term is introduced to obtain higher temporal relevance. Simultaneously, degradation rate information is integrated to enhance the focus on samples in the rapid degradation phase. The final sampling probability can be expressed as: :

[0108]

[0109]

[0110] in, This represents a measure of information density enhanced by the degradation rate. express The k nearest neighbor contexts, Describing the L2 norm, Indicates the rate of degradation. These are weighting coefficients for degradation rate information. It is a tuning hyperparameter used to control the degree to which density affects the sampling probability. By incorporating degradation rate information, lower density values ​​are assigned to critical stages of rapid degradation, thus giving them higher sampling priority.

[0111] (2) The sampling ratio of the current degenerate trajectory is determined by using a selection criterion based on the diversity of degenerate trajectories. For each degenerate trajectory in a task, in addition to storing representative samples in the memory, the health factor of the degenerate trajectory is also retained. For the degenerate trajectory of task t, the selection criterion based on diversity is defined by calculating the minimum dynamic time warping (DTW) distance between the health factors of each degenerate trajectory in the memory and the health factors of the existing degenerate trajectories, thereby determining the sampling ratio of the current degenerate trajectory. :

[0112]

[0113] in, It is an exponential function. Indicates the dynamic time-warped distance. It is a scaling factor. A diversity-based selection criterion increases the sampling ratio of trajectories with low similarity to those already stored in memory, thereby encouraging broader data distribution coverage. Furthermore, as more trajectories are stored in memory, the sampling ratio of new trajectories tends to decrease, thus reducing memory pressure.

[0114] Furthermore, in this embodiment, it should be noted that the following uses a turbofan engine as a specific example to further explain the incremental scenario-based meta-learning method for predicting mechanical equipment faults based on replay meta-learning of the present invention. The specific steps are as follows:

[0115] (1) The C-MAPSS turbofan engine dataset was used for model training and testing. This dataset contains four subsets (FD001, FD002, FD003, and FD004), and their specific operating conditions and failure modes are given in Table 1. There are differences between the subsets in terms of operating conditions and degradation modes. Each subset consists of multiple multivariate time series degradation trajectories, and each degradation trajectory contains 21 sensor signals and 3 operating parameters. The dataset is divided into training set and test set according to the full life cycle and truncation form: the degradation trajectory in the training set covers the complete degradation process of the engine from initial health to failure, while the degradation trajectory in the test set is terminated in advance before actual failure to evaluate the model's failure prediction capability. To verify our method, we set up four incremental task flows based on these four subsets. Table 2 shows the details of one incremental task flow. At the same time, we constructed a prediction model and a high-quality memory bank including a meta-knowledge encoder, regressor, masking function, and health sensing unit. The structural diagram of the prediction model is shown in the figure. Figure 2 As shown, the specific settings for the model's hyperparameters are given in Table 3.

[0116] Table 1: Details of the C-MAPSS dataset

[0117]

[0118] Table 2: Details of Incremental Task Flow

[0119]

[0120] Table 3: Model Structure Hyperparameter Table

[0121]

[0122] (2) For the current task, the incremental data is preprocessed. First, all monitoring variables are normalized to between -1 and 1. For FD001 and FD003 with single operating conditions, standard min-max normalization is directly applied; while for FD002 and FD004 with variable operating conditions, the unsupervised clustering method K-means is first used to automatically divide the operating conditions, and then min-max normalization is performed independently on the data of each type of operating condition to ensure the consistency of magnitude between different operating conditions. After normalization, the readings of some sensors remain constant throughout the entire operating cycle and do not carry any degradation information. Therefore, these invalid variables are removed, and only the monitoring signals of sensors numbered 2, 3, 4, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20, and 21 are retained as model input. Subsequently, feature segments across multiple time steps were extracted from the time series using sliding windows of lengths 30 / 20 / 30 / 15 for four subsets, forming a multi-dimensional time-series feature map reflecting the dynamic evolution of the equipment, which was then used as input data for the model. For label construction, a piecewise linear degradation model was introduced to generate a corresponding pre-failure remaining time label (unit: cycle) for each window sample. To reduce the adverse impact of the initial health phase on model training, the pre-failure remaining time in the early degradation interval was fixed at 125 cycles to reflect the equipment's health redundancy and enhance the model's focus on the actual degradation segment. After data preprocessing, a support set was constructed using incremental data, and a query set was formed by reading 300 sample points from the memory.

[0123] (3) Perform the meta-training process using the support set and query set, and perform the meta-testing process using the support set. The specific process is as follows: Figure 1 As shown, meta-training consists of an inner loop and an outer loop: the inner loop rapidly updates the regressor, masking function, and health-aware unit on the support set, while preserving their computational graph for the subsequent meta-gradient flow in the outer loop; the outer loop computes meta-gradients based on the query set data and updates the meta-knowledge encoder accordingly, enabling it to learn highly generalizable representations. The meta-testing process updates the regressor, masking function, and health-aware unit on the support set to fine-tune the model for the current task. The fine-tuned model is then applied to the truncated degradation data to predict the remaining time before failure. After the task, a degradation information-enhanced sample selection strategy selects a subset of samples from the full lifecycle degradation trajectory of the current task and writes them into the memory bank to support experience replay for the next task.

[0124] (4) To verify the effectiveness and advantages of the proposed method in fault prediction under incremental scenarios, this invention comprehensively compared the proposed method with several representative baseline models under four incremental task flows. The root mean square error (RMSE) was used as the main evaluation metric for experimental performance. The calculation method of RMSE is as follows:

[0125]

[0126] in, and These represent the predicted and actual values ​​of the remaining time before the failure, respectively. Specific results are shown in Table 4. In all four quantitative task flows, the overall performance of this invention is superior to other methods, demonstrating the superiority of this invention in solving the mechanical failure prediction problem in incremental scenarios.

[0127] Table 4: Comparison of RMSE of prediction results from different methods

[0128]

[0129] The second embodiment of the present invention provides:

[0130] A mechanical equipment fault prediction system based on meta-learning in incremental scenarios, wherein the system includes:

[0131] The construction module is used to build an initial prediction model and a memory library, and synchronously collect the full life cycle signals generated by the mechanical equipment during the execution of tasks.

[0132] The processing module is used to acquire the monitoring signals generated by the failed mechanical equipment and the remaining time label before the failure in the degradation trajectory corresponding to the full life cycle signal according to a preset time interval, and synchronously generate time series information based on the dataset corresponding to the degradation trajectory.

[0133] The reading module is used to generate a support set based on the time series information, and synchronously generate a query set based on the data read from the memory bank, so as to perform meta-training on the initial prediction model through the support set and the query set, perform meta-testing on the initial prediction model through the support set, and generate a corresponding fault prediction model.

[0134] The prediction module is used to predict the fault time based on the remaining time tag and the fault prediction model according to the truncated data corresponding to the monitoring signal, so as to output the corresponding prediction result.

[0135] Furthermore, the reading module is specifically used for:

[0136] The monitoring signals are normalized to generate corresponding target monitoring signals;

[0137] A sliding time window strategy is introduced, which uses multi-point information continuously collected within a preset time span as the overall input based on the target monitoring signal to generate the time series information accordingly;

[0138] Based on the time series signal, target data corresponding to the current task under the full life cycle degradation trajectory is collected, and the support set is constructed based on the target data. Simultaneously, the query set is constructed based on the data read from the memory bank.

[0139] Furthermore, the expression for the support set is:

[0140]

[0141] The expression for the query set is:

[0142]

[0143] in, Indicates task Number of samples in the degradation trajectory. This indicates the number of samples in the replay experience. and Indicates task It supports centralized preprocessed input data and corresponding remaining time labels before the fault. and The labels represent the preprocessed input data and the corresponding remaining time before the fault in the replay experience.

[0144] Furthermore, the reading module is specifically used for:

[0145] The initial prediction model includes a meta-knowledge encoder, a regressor, a masking function, and a health-aware unit. The meta-training process is divided into an inner loop and an outer loop, and includes the following steps:

[0146] The inner loop uses the support set to update the regressor, the mask function, and the health sensing unit:

[0147]

[0148] in, For the inner loop learning rate, For mean square error loss, This indicates multiplication by channel. Representing tasks The mask function, health sensing unit, and regressor parameters updated in the inner loop. Indicates regular loss. and Indicates task It supports centralized preprocessed input data and corresponding remaining time labels before the fault. The parameter is Meta-knowledge encoder, The parameter is The regressor, The parameter is The mask function.

[0149] Furthermore, the expression for the regularization loss is:

[0150]

[0151] in, This represents sparse regularization. This indicates that health perception is regularized. and represent the weight coefficients of the sparse regularization term and the health perception regularization term, respectively. Indicates potential health factors, Describing the L1 norm, Indicates task Supports centralized preprocessed input data. The parameter is Health sensing unit The parameter is mask function, This represents the mean square error loss.

[0152] Furthermore, the outer loop updates the meta-knowledge encoder using the query set:

[0153]

[0154] in, For the outer loop learning rate, Indicates the parameters of the meta-knowledge encoder. For mean square error loss, The parameter is Meta-knowledge encoder, This indicates multiplication by channel. The parameter is mask function, and The labels represent the preprocessed input data and the corresponding remaining time before the fault in the replay experience. This represents the update coefficient.

[0155] Furthermore, the reading module is specifically used for:

[0156] During the meta-test, the regressor, the masking function, and the health-aware unit are updated using the support set to perform task fine-tuning:

[0157]

[0158] in, Representing tasks The mask function, health sensing unit, and regressor parameters after the mid-cycle test. For the inner loop learning rate, For mean square error loss, The parameter is The regressor, The parameter is Meta-knowledge encoder, This indicates multiplication by channel. The parameter is mask function, and Indicates task It supports centralized preprocessed input data and corresponding remaining time labels before the fault. This indicates a regular loss.

[0159] The third embodiment of the present invention provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the mechanical equipment fault prediction method based on meta-learning in incremental scenarios as described above.

[0160] The fourth embodiment of the present invention provides a readable storage medium storing a computer program thereon, wherein when the program is executed by a processor, it implements the mechanical equipment fault prediction method based on meta-learning in incremental scenarios as described above.

[0161] In summary, the incremental scenario-based mechanical equipment fault prediction method and system provided by the above embodiments of the present invention can accurately identify the degradation trajectory of mechanical equipment, thereby improving the efficiency of incremental scenario-based mechanical equipment fault prediction.

[0162] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination.

[0163] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0164] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0165] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0166] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0167] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for predicting mechanical equipment faults based on meta-learning in incremental scenarios, characterized in that, The method includes: An initial prediction model and memory are constructed, and the full life cycle signals generated by the mechanical equipment are collected synchronously during the execution of tasks by the mechanical equipment. According to a preset time interval, the monitoring signal generated by the failed mechanical equipment and the remaining time label before the failure are obtained from the degradation trajectory corresponding to the full life cycle signal, and time series information is generated synchronously based on the dataset corresponding to the degradation trajectory. A support set is generated based on the time series information, and a query set is generated synchronously based on the data read from the memory bank. The initial prediction model is then meta-trained using the support set and the query set, and the initial prediction model is then meta-tested using the support set, and a corresponding fault prediction model is generated. Based on the remaining time stamp, the fault prediction model predicts the fault time according to the truncated data corresponding to the monitoring signal, and outputs the corresponding prediction result. The step of meta-training the initial prediction model using the support set and the query set includes: The initial prediction model includes a meta-knowledge encoder, a regressor, a masking function, and a health-aware unit. The meta-training process is divided into an inner loop and an outer loop, and includes the following steps: The inner loop uses the support set to update the regressor, the mask function, and the health sensing unit: in, For the inner loop learning rate, For mean square error loss, This indicates multiplication by channel. Representing tasks The mask function, health sensing unit, and regressor parameters updated in the inner loop. Indicates regular loss. and Indicates task It supports centralized preprocessed input data and corresponding remaining time labels before the fault. The parameter is Meta-knowledge encoder, The parameter is The regressor, The parameter is mask function, This represents the parameters before the inner loop is updated. Represents the mask coefficient, This represents the regression coefficient.

2. The mechanical equipment fault prediction method based on meta-learning in incremental scenarios according to claim 1, characterized in that, The steps of generating a support set based on the time series information and simultaneously generating a query set based on the data read from the memory include: The monitoring signals are normalized to generate corresponding target monitoring signals; A sliding time window strategy is introduced, which uses multi-point information continuously collected within a preset time span as the overall input based on the target monitoring signal to generate the time series information accordingly; Based on the time series signal, target data corresponding to the current task under the full life cycle degradation trajectory is collected, and the support set is constructed based on the target data. Simultaneously, the query set is constructed based on the data read from the memory bank.

3. The mechanical equipment fault prediction method based on meta-learning in incremental scenarios according to claim 2, characterized in that, The expression for the support set is: The expression for the query set is: in, Indicates task Number of samples in the degradation trajectory. This indicates the number of samples in the replay experience. and Indicates task It supports centralized preprocessed input data and corresponding remaining time labels before the fault. and The labels represent the preprocessed input data and the corresponding remaining time before the fault in the replay experience.

4. The mechanical equipment fault prediction method based on meta-learning in incremental scenarios according to claim 1, characterized in that, The expression for the regularization loss is: in, This represents sparse regularization. This indicates that health perception is regularized. and represent the weight coefficients of the sparse regularization term and the health perception regularization term, respectively. Indicates potential health factors, Describing the L1 norm, Indicates task Supports centralized preprocessed input data. The parameter is Health sensing unit The parameter is mask function, This represents the mean square error loss.

5. The mechanical equipment fault prediction method based on meta-learning in incremental scenarios according to claim 1, characterized in that, The outer loop updates the meta-knowledge encoder using the query set: in, For the outer loop learning rate, Indicates the parameters of the meta-knowledge encoder. For mean square error loss, The parameter is Meta-knowledge encoder, This indicates multiplication by channel. The parameter is mask function, and The labels represent the preprocessed input data and the corresponding remaining time before the fault in the replay experience. This represents the update coefficient.

6. The mechanical equipment fault prediction method based on meta-learning in incremental scenarios according to claim 5, characterized in that, The step of performing a meta-test on the initial prediction model using the support set includes: During the meta-test, the regressor, the masking function, and the health-aware unit are updated using the support set to perform task fine-tuning: in, Representing tasks The mask function, health sensing unit, and regressor parameters after the mid-cycle test. For the inner loop learning rate, For mean square error loss, The parameter is The regressor, The parameter is Meta-knowledge encoder, This indicates multiplication by channel. The parameter is mask function, and Indicates task It supports centralized preprocessed input data and corresponding remaining time labels before the fault. Indicates regular loss. This represents the parameters before the inner loop is updated. Represents the mask coefficient, This represents the regression coefficient.

7. A mechanical equipment fault prediction system based on meta-learning in incremental scenarios, characterized in that, For implementing the incremental scenario-based mechanical equipment fault prediction method as described in any one of claims 1 to 6, the system comprises: The construction module is used to build an initial prediction model and a memory library, and synchronously collect the full life cycle signals generated by the mechanical equipment during the execution of tasks. The processing module is used to acquire the monitoring signals generated by the failed mechanical equipment and the remaining time label before the failure in the degradation trajectory corresponding to the full life cycle signal according to a preset time interval, and synchronously generate time series information based on the dataset corresponding to the degradation trajectory. The reading module is used to generate a support set based on the time series information, and synchronously generate a query set based on the data read from the memory bank, so as to perform meta-training on the initial prediction model through the support set and the query set, perform meta-testing on the initial prediction model through the support set, and generate a corresponding fault prediction model. The prediction module is used to predict the fault time based on the remaining time tag and the fault prediction model according to the truncated data corresponding to the monitoring signal, so as to output the corresponding prediction result.

8. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the mechanical equipment fault prediction method based on meta-learning in incremental scenarios as described in any one of claims 1 to 6.

9. A readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the mechanical equipment fault prediction method based on meta-learning in incremental scenarios as described in any one of claims 1 to 6.

Images