Method for predicting residual service life of equipment based on improved Mama framework

Abstract

The invention discloses a method for predicting the residual service life of equipment based on an improved Mama framework. The method comprises the following steps: step 1, preprocessing multi-dimensional data; step 2, a Mama model is improved; and 3, predicting the remaining service life. According to the method, the unique selective state space mechanism of Mamba is fully utilized, so that the time complexity is controlled to be O (N) when long-sequence data is processed, and the calculation cost which is increased along with the square level of the sequence length in a traditional model is avoided. And meanwhile, an intelligent scaling regulator is constructed in Residual Block of the Mama framework so as to dynamically adjust the weight of residual error connection. A plurality of attention branches which are operated in parallel are creatively integrated in MambaBlock and act on data processed by a convolutional layer, so that the capturing capability of a model on a complex data relationship is enhanced, and the prediction precision is improved.

Description

Technical Field

[0001] The present invention belongs to the field of prediction and health management, and relates to a method for predicting the remaining service life of equipment, and specifically to a method for predicting the remaining service life of equipment based on an improved Mamba framework. Background Art

[0002] In the modern industrial field that is highly dependent on reliability and safety, the remaining useful life (RUL) prediction technology of equipment has become the core link in prediction and health management (PHM), and is most widely used in the fields of aircraft engines, lithium-ion batteries, rolling bearings, and mining machinery and equipment. For example, due to factors such as material fatigue, operational wear and natural aging, accurate prediction of the RUL of aircraft engines can ensure flight safety, improve operational efficiency, and reduce maintenance costs; accurate prediction of the RUL of lithium-ion batteries helps to ensure the safe operation of electric vehicles, energy storage systems, and other fields, and achieve cost control; accurate prediction of the RUL of rolling bearings can ensure the healthy operation of industrial equipment, wind power generation, rail transportation, and CNC machine tools; RUL prediction of key components of mining machinery and equipment can pre-arrange maintenance and replacement plans, avoid production interruptions and safety accidents caused by sudden equipment failures, and achieve efficient, safe, and sustainable development of mining operations.

[0003] At present, RUL prediction methods can be mainly divided into two categories: model-driven prediction methods and data-driven prediction methods. Among them, model-driven prediction methods mainly rely on physical properties and engineering principles, predict the future performance of turbine engines by building mathematical models, and simulate the degradation process of lithium-ion battery performance by establishing corresponding physical failure models or electrochemical models. A major advantage of this method is that it can describe their degradation process relatively accurately, providing strong theoretical support for prediction. However, the numerous and complex structures of aircraft engines and the complex physical and chemical changes inside lithium-ion batteries make the modeling process more difficult, and this method also has relatively high requirements for professional knowledge and high-quality data. Therefore, data-driven prediction methods are more advantageous, focusing on extracting valuable information from historical operating data rather than relying on in-depth understanding of its internal physical mechanisms. However, traditional data-driven machine learning algorithms, such as decision trees (DT), support vector machines (SVM), random forests (RF), extreme learning machines (ELM), and remaining service life prediction methods based on long short-term memory networks (LSTM) and transformers, all have disadvantages such as high computational cost and poor robustness. Summary of the invention

[0004] The present invention provides a method for predicting the remaining useful life of equipment based on an improved Mamba framework. The method analyzes and diagnoses the status of equipment through a large amount of data, predicts potential failures, and adopts corresponding repair and maintenance strategies to improve equipment reliability, reduce repair costs, and avoid production interruptions.

[0005] The objective of the present invention is achieved through the following technical solutions:

[0006] A method for predicting the remaining useful life of equipment based on an improved Mamba framework comprises the following steps:

[0007] Step 1: Preprocessing of multidimensional data:

[0008] Preprocessing of multidimensional data, including data set segmentation, feature cleaning, feature scale consistency, and data smoothing;

[0009] Step 2: Improve the Mamba model:

[0010] Step 21: Build an intelligent scaling regulator to speed up model convergence:

[0011] An intelligent scaling regulator is constructed to act on the ResidualBlock in the Mamba framework, thereby dynamically adjusting the weight of the residual connection. Through the intelligent scaling regulator, the Mamba model learns how to use the residual connection more effectively, thereby optimizing the information flow and feature representation of the network. Specifically: During the training process, the gradient is propagated from the output layer to the input layer through the back-propagation algorithm to update the parameters of the Mamba model. The scaling parameter in the intelligent scaling regulator is used as part of the model parameters and participates in the calculation and update process of the gradient. The initial value of the parameter is 1. During the training process, the optimizer automatically adjusts the value of the parameter according to the gradient of the loss function. When the value is greater than 1, the weight of the residual connection increases; when the value is less than 1, the weight of the residual connection decreases;

[0012] Step 22: Multi-directional collaborative attention focus feature enhancement:

[0013] After the convolutional layer in MambaBlock, multiple parallel attention branches are designed. The data processed by the convolutional layer is input into multiple parallel attention branches to further extract and fuse features to form a richer and more comprehensive feature representation. Through the residual connection mechanism, the outputs of multiple parallel attention branches are residually connected with the inputs and processed through layer normalization.

[0014] Step 3: Remaining useful life prediction:

[0015] The data set filtered in step 1 is input into the improved Mamba model in step 2 to predict the remaining useful life.

[0016] Compared with the prior art, the present invention has the following advantages:

[0017] The present invention makes full use of Mamba's unique selective state space mechanism, so that when processing long sequence data, the time complexity is controlled to O(N), avoiding the computational cost that grows with the square of the sequence length in the traditional model. At the same time, an intelligent scaling regulator is built in the ResidualBlock of the Mamba framework to dynamically adjust the weight of the residual connection. In MambaBlock, multiple parallel attention branches are innovatively integrated to act on the data processed by the convolution layer, which enhances the model's ability to capture complex data relationships and improves prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a multidimensional data processing diagram;

[0019] Figure 2 This is the Mamba network structure diagram;

[0020] Figure 3 is the parameter scaling processing diagram;

[0021] Figure 4 These are Lossp visualization diagrams, (a) Lossp visualization diagram of aircraft engines, (b) Lossp visualization diagram of lithium-ion batteries;

[0022] Figure 5 It is a feature enrichment processing graph;

[0023] Figure 6 It is a model structure diagram based on the improved Mamba framework. DETAILED DESCRIPTION

[0024] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0025] The present invention provides a method for predicting the remaining useful life of equipment based on an improved Mamba framework. First, the multidimensional data is preprocessed, including data set division, feature cleaning, feature scale consistency, data smoothing, etc. Secondly, an intelligent scaling regulator is constructed to act on the ResidualBlock in the Mamba framework, thereby dynamically adjusting the weights of the residual connections; multiple parallel attention branches are innovatively designed in the MambaBlock, and each branch focuses on different aspects or feature subsets of the input data. Finally, the method is compared with the current advanced research results on aircraft engine and lithium-ion battery data sets. The experimental results show that the model is superior to the current advanced methods in the RUL prediction task and has good generalization, and can be used as a general RUL prediction method for RUL prediction of rolling bearings and mining machinery and equipment. Specifically, it includes the following steps:

[0026] Step 1: Preprocessing of multidimensional data:

[0027] Taking the data sets of aircraft engines and lithium-ion batteries as examples, multidimensional data preprocessing is performed, including data set division, feature cleaning, feature scale consistency, and data smoothing. Specifically: the aircraft engine data set is divided into training set, test set, and prediction set, namely train_FD001-train_FD004, test_FD001-test_FD004, RUL_FD001-RUL_FD004; the lithium-ion battery data set is divided according to different temperatures, namely 0℃, 10℃, 25℃, 30℃, 40℃, and 50℃. Then, all data are visualized and feature data with no obvious trend from the healthy stage to the failure stage are deleted, namely data cleaning. Secondly, the data after deleting invalid features is kept consistent in feature scale and dimension differences are eliminated, so that the model can treat each feature equally. Finally, data smoothing technology is further adopted to effectively reduce the random fluctuations and noise components in the data, making the long-term trend and periodic changes of the data more significant.

[0028] Step 2: Improve the Mamba model:

[0029] Step 21: Build an intelligent scaling regulator to speed up model convergence:

[0030] A smart scaling regulator is constructed to act on the ResidualBlock in the Mamba framework, thereby dynamically adjusting the weights of the residual connections. Through the smart scaling regulator, the model can learn how to use the residual connection more effectively, thereby optimizing the information flow and feature representation of the network. Specifically: During the training process, the gradient needs to be propagated from the output layer to the input layer through the back-propagation algorithm to update the parameters of the model. The scaling parameters in the smart scaling regulator, as part of the model parameters, will also participate in the calculation and update process of the gradient. The results show that after 60 epochs, the data tends to be stable, indicating that the model converges quickly, which helps to alleviate the problem of gradient vanishing or gradient exploding, making the gradient more stable during the back-propagation process, converge to the optimal solution faster, and improve the efficiency of the training process.

[0031] Step 22: Multi-directional collaborative attention focus feature enhancement:

[0032] In MambaBlock, multiple parallel attention branches are innovatively designed, each focusing on different aspects or feature subsets of the input data. These branches can be built based on different weights, distance metrics, or feature selection strategies to capture the diverse attention distribution in the input sequence. The outputs of multiple parallel attention branches are residually connected to the input and processed through layer normalization. At the same time, they act on the data processed by the convolutional layer, which enhances the model's ability to capture complex data feature relationships and improves prediction accuracy.

[0033] Step 3: Remaining useful life prediction:

[0034] The useful data set screened in step 1 is input into the improved Mamba model for remaining useful life prediction, and the root mean square error and mean absolute error are introduced as evaluation indicators. The smaller the root mean square error and mean absolute error, the more accurate the remaining useful life prediction of the model is.

[0035] Example:

[0036] The experiments were implemented on an Intel i7-11800H processor, 16GB of memory, and an NVIDIA GeForce RTX3090 graphics card, using Python 3.9 and Pytorch 2.0.1.

[0037] The first step is to process multidimensional data. Although the data sets of aircraft engines, lithium-ion batteries, and other data such as rolling bearings, mining machinery and equipment are all pure data samples, the data feature scales of different data sets are inconsistent, and there are some outliers or noise in the data samples, which will cause certain interference to the prediction of the next model. Figure 1As shown in the figure, after visualizing all the data, the feature data with no obvious trend from the healthy stage to the failed stage are deleted as new data sample input. At the same time, all feature scales are unified, and smoothing technology is used for the unified data features to reduce random fluctuations and noise components in the data, making the long-term trend and periodic changes of the data more significant.

[0038] The second step is to accelerate the convergence of the model. This invention introduces the Mamba model, which can process long sequences more effectively, into the RUL prediction task of aircraft engines for the first time. Figure 2 As shown in the figure, the Mamba model can selectively process input information through a selective mechanism and effectively store results through parallel scanning, kernel fusion, and recalculation. Next, an intelligent scaling regulator is constructed to act on the ResidualBlock in the Mamba framework, thereby dynamically adjusting the weights of the residual connection. Through the intelligent scaling regulator, the model can learn how to use the residual connection more effectively, thereby optimizing the information flow and feature representation of the network.

[0039] Specifically: Figure 3 As shown in the figure, the data tensor X input to the Mamba model is processed with consistent feature scales to make the data more stable, which is helpful for the subsequent feature transformation process. The output data is passed through a mixer containing multiple neural network layers (convolutional layer, fully connected layer, activation layer) to extract useful features, so that the model can learn effective representations from the original data. Then, the output after the mixer is added to X to form a residual connection, which allows the gradient to be directly propagated to the shallow layer through the identity mapping, accelerating the convergence speed and improving the stability of the model. Figure 4 Taking (a) aircraft engines and 4 (b) lithium-ion batteries as examples, the Lossp results show that after 60 epochs, the data tends to be stable, indicating that the model converges quickly, which helps to alleviate the gradient vanishing or gradient exploding problem, making the gradient more stable during back propagation, converging to the optimal solution faster, and improving the efficiency of the training process.

[0040] The third step is to focus on feature enhancement. In the MambaBlock architecture, multiple parallel attention branch designs are creatively integrated. Each branch focuses on a unique dimension or feature subset of the input data, achieving in-depth mining of the breadth of information. The construction of these branches is flexible and diverse, and can be customized according to differentiated weight configurations, distance measurement criteria or feature screening strategies, aiming to accurately capture the rich and varied attention patterns in the input sequence. Furthermore, through the residual connection mechanism, the outputs of these parallel attention branches are cleverly fused with the original input data, which not only retains the integrity of the input information, but also introduces the key features extracted by the attention mechanism. Subsequently, the application of layer normalization effectively balances the data distribution and enhances the stability and generalization ability of the model. In addition, this design is seamlessly integrated into the convolutional layer processing flow, allowing the model to make full use of the local features extracted by the convolution operation, and on this basis, further explore and strengthen the complex feature relationships between the data through parallel attention branches. This innovation not only significantly improves the model's ability to parse complex data structures, but also directly promotes the improvement of prediction accuracy, providing strong support for dealing with complex and changeable practical problems. The specific processing process is as follows Figure 5 shown.

[0041] Step 4: Remaining service life prediction. Input the data processed in the first step into Figure 6 The improved Mamba model is used to predict the remaining service life. As shown in Tables 1 to 4, by comparing the root mean square error and mean absolute error results with the existing advanced methods, the root mean square error and mean absolute error of this method are both smaller than the compared methods, and it has better prediction accuracy. At the same time, this method is not only suitable for the remaining service life prediction of aircraft engines and lithium-ion batteries, but also suitable for the remaining service life prediction tasks such as rolling bearings, mining machinery and equipment, etc.

[0042] Table 1 Comparison of RMSE results with the current mainstream methods for the task of predicting the remaining useful life of aircraft engines

[0043]

[0044]

[0045] Table 2 Comparison of MAE results with the current mainstream methods for the task of predicting the remaining useful life of aircraft engines

[0046]

[0047] Table 3 Comparison of RMSE results with the current mainstream methods at different temperatures for lithium-ion battery tasks

[0048]

[0049] Table 4 Comparison of RMSE results of lithium-ion battery tasks at the same temperature (30°C) with the current mainstream methods

[0050]

Claims

1. A method for predicting the remaining useful life of equipment based on an improved Mamba framework, characterized in that The method comprises the following steps: Step 1: Preprocessing of multidimensional data: Preprocess multidimensional data; Step 2: Improve the Mamba model: Step 21: Build an intelligent scaling regulator to speed up model convergence: An intelligent scaling regulator is constructed to act on the ResidualBlock in the Mamba framework, thereby dynamically adjusting the weight of the residual connection. Through the intelligent scaling regulator, the Mamba model learns how to use the residual connection more effectively, thereby optimizing the information flow and feature representation of the network. Specifically: During the training process, the gradient is propagated from the output layer to the input layer through the back-propagation algorithm to update the parameters of the Mamba model. The scaling parameter in the intelligent scaling regulator is used as part of the model parameters and participates in the calculation and update process of the gradient. The initial value of the parameter is 1. During the training process, the optimizer automatically adjusts the value of the parameter according to the gradient of the loss function. When the value is greater than 1, the weight of the residual connection increases; when the value is less than 1, the weight of the residual connection decreases; Step 22: Multi-directional collaborative attention focus feature enhancement: exist After the convolutional layer in MambaBlock, multiple parallel attention branches are designed. The data processed by the convolutional layer is input into multiple parallel attention branches to further extract and fuse features to form a richer and more comprehensive feature representation. Through the residual connection mechanism, the outputs of multiple parallel attention branches are residually connected with the inputs and processed through layer normalization. Step 3: Remaining useful life prediction: The data set filtered in step 1 is input into the improved Mamba model in step 2 to predict the remaining useful life.

2. The method for predicting the remaining useful life of equipment based on the improved Mamba framework according to claim 1 is characterized in that The preprocessing includes data set division, feature cleaning, feature scale consistency, and data smoothing.

Images