A method and system for predicting the remaining life of a device component

Abstract

The application discloses a kind of equipment component remaining life prediction method and system, and the remaining useful life and health state prediction of industrial equipment component are realized based on FM-Mamba-SE model.Adopting the technical scheme of the application, the degradation characteristics in the vibration signal during the operation of industrial equipment component are accurately extracted to realize the rapid and accurate prediction of the remaining useful life of industrial equipment component.

Description

Technical Field

[0001] This invention belongs to the field of fault prediction and health management technology for industrial equipment components, and specifically relates to a method and system for predicting the remaining service life of equipment components. Background Technology

[0002] In the Industrial Internet field, where reliability and production safety are paramount, predicting the remaining useful life of industrial equipment components has become a crucial aspect of Predictive Health Management (PHM). In modern intelligent industrial production processes, the operating components of industrial equipment are susceptible to cracking, fracture, and corrosion due to factors such as high temperature and pressure, high salt and humidity, and metal fatigue and material wear caused by high-intensity operation. This leads to gradual performance degradation, eventually resulting in failure and complete operational inefficiency. Furthermore, failure of core industrial components can even trigger a cascading failure of other components, threatening the entire production equipment and jeopardizing operational safety.

[0003] To monitor equipment health and ensure safe operation, fault prediction and health management of industrial equipment components, especially critical parts, are necessary. In the field of fault prediction and health management, assessing equipment health, accurately predicting remaining service life, and providing timely warnings of failure risks are key objectives. Currently, there are two feasible technical approaches: one is based on physical models, which predicts remaining service life by constructing physical failure models, using failure mechanism constraints, and combining mathematical calculations. While this approach is highly efficient, it suffers from difficulties in constructing complex models, unclear failure mechanisms, reliance on expert knowledge, large number of computational parameters, high dependence on data accuracy, and poor generalizability. The other approach is based on data-driven machine learning and deep learning methods. This method constructs deep learning models to automatically extract features from large amounts of data, exhibiting good generalization and applicability to complex systems. However, it also suffers from large data requirements, poor interpretability, slow inference speed, and long-term dependence. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a method and system for predicting the remaining service life of equipment components. This method accurately extracts degradation features from vibration signals during the operation of industrial equipment components, thereby enabling rapid and accurate prediction of the remaining service life of industrial equipment components.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for predicting the remaining useful life of a device component, comprising:

[0007] Collect and preprocess horizontal and vertical vibration signal data of industrial equipment components during operation;

[0008] The time domain, frequency domain, and time-frequency domain features of the preprocessed horizontal and vertical vibration signal data of the operating components are extracted, life labels are added, and the data is divided into training set, validation set, and test set.

[0009] Construct an encoder-decoder architecture model; wherein the encoder-decoder architecture model includes: a 2-layer encoder multi-layer neural network module and a 1-layer decoder multi-layer neural network module; the encoder multi-layer neural network module includes: factorization machine layer, multilayer perceptron (MLP) layer, learnable gated fusion layer and Mamba module; the decoder multi-layer neural network module includes: SENet module and dilated convolutional layer;

[0010] Train the FM-Mamba-SE model based on the training dataset;

[0011] The trained FM-Mamba-SE model is used to predict the remaining lifetime on the test dataset, and the remaining lifetime is output for each step.

[0012] The present invention also provides a system for predicting the remaining useful life of equipment components, comprising:

[0013] The first processing unit is used to collect and preprocess horizontal and vertical vibration signal data during the operation of industrial equipment components;

[0014] The second processing unit is used to extract time-domain, frequency-domain, and time-frequency-domain features from the preprocessed horizontal and vertical vibration signal data of the operating components, label them with life tags, and divide them into training set, validation set, and test set.

[0015] The third processing unit is used to construct the encoder-decoder architecture model, which includes a 2-layer encoder multi-layer neural network module and a 1-layer decoder multi-layer neural network module. The encoder multi-layer neural network module includes a factorization machine layer, a multi-layer perceptron (MLP) layer, a learnable gated fusion layer, and a Mamba module. The decoder multi-layer neural network module includes a SENet module and dilated convolutional layers.

[0016] The fourth processing unit is used to train the FM-Mamba-SE model based on the training dataset;

[0017] The fifth processing unit is used to predict the remaining lifetime on the test dataset using the trained FM-Mamba-SE model and output the remaining lifetime for each step.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0019] 1. Three-dimensional feature relationship modeling capability: This invention mines performance degradation information during operation by calculating the time domain, frequency domain, and time-frequency domain features of vibration monitoring signal data of industrial equipment components. It also uses factor decomposition machine combined with multilayer perceptron to extract the relationship between features to clarify the mutual influence between different features. This allows for three-dimensional modeling of the complex changes in data during performance degradation, significantly improving the accuracy of the prediction method.

[0020] 2. Powerful ability to extract long-sequence temporal dependencies: The Mamba model used in this invention accurately models ultra-long-term time series by introducing HiPPO technology, which greatly alleviates the problem of long-term dependencies. Its unique parallel computing mechanism can extract temporal features in conjunction with hardware-aware algorithms without disrupting the cyclic computing process, thus optimizing time complexity and greatly improving computing efficiency and temporal feature extraction performance.

[0021] 3. Accurate remaining useful life prediction capability: The decoder neural network module used in this invention combines SENet and dilated convolutional neural networks. The SENet network assigns weights to different feature channels through a unique squeeze-excitation mechanism, taking into account the weights of key features. At the same time, it uses a dilated convolutional model to expand the receptive field and efficiently capture multi-scale temporal dependencies, effectively fusing feature extraction results and improving prediction accuracy. Attached Figure Description

[0022] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the device component remaining life prediction method according to an embodiment of the present invention;

[0024] Figure 2 This is a diagram of the encoder's multi-layer neural network module structure.

[0025] Figure 3 This is a diagram of the multi-layer neural network module structure for the decoder. Detailed Implementation

[0026] 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 skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0028] Example 1

[0029] like Figure 1 As shown, the present invention provides a method for predicting the remaining life of equipment components, comprising:

[0030] Step S1: Collect and preprocess the horizontal and vertical vibration signal data of the industrial equipment components during operation;

[0031] Step S2: Extract time-domain, frequency-domain, and time-frequency-domain features from the preprocessed horizontal and vertical vibration signal data of the operating components, label them with life tags, and divide them into training set, validation set, and test set;

[0032] Step S3: Construct the encoder-decoder architecture model (FM-Mamba-SE model).

[0033] Step S4: Construct the encoder multilayer neural network module;

[0034] Step S5: Construct the decoder multilayer neural network module;

[0035] Step S6: Train the FM-Mamba-SE model based on the training dataset;

[0036] Step S7: Use the trained FM-Mamba-SE model to predict the remaining lifetime on the test dataset and output the remaining lifetime for each step.

[0037] In one embodiment of the present invention, in step S1, in order to remove the influence of operating noise, it is necessary to use the variational mode decomposition method to perform noise filtering on the vibration monitoring signal data of industrial equipment components, and set the number of decomposition layers K=8.

[0038] In one embodiment of the present invention, in step S2, to extract as much information as possible from the denoised vibration signal data, time-domain features, frequency-domain features, and time-frequency-domain features are calculated and extracted from the denoised vibration signal data. Time-domain features include peak amplitude, peak-to-peak value, peak-to-peak ratio, root mean square value, absolute mean, standard deviation, skewness, kurtosis, margin factor, waveform factor, and impulse factor; frequency-domain features include amplitude mean, amplitude sample standard deviation, amplitude skewness, amplitude kurtosis, weighted average frequency, weighted frequency standard deviation, weighted root mean square frequency, frequency coefficient of variation, and frequency root mean square; time-frequency-domain features include approximate entropy, sample entropy, permutation entropy, and wavelet entropy; RMS features are extracted from the above time-domain features to confirm the degradation path of the running component, and the remaining lifetime label for each time step is calculated; the above time-domain, frequency-domain, and time-frequency-domain features and the remaining lifetime label are combined to construct a dataset; the above dataset is divided into a training dataset, a validation dataset, and a test dataset. The extraction of multi-domain features greatly uncovers the hidden information in the denoised vibration signal, providing an important basis for the subsequent extraction and prediction of the FM-Mamba-SE model.

[0039] In one embodiment of the present invention, the encoder-decoder architecture constructed in step S3 needs to include a two-layer encoder multi-layer neural network module and a one-layer decoder multi-layer neural network module.

[0040] In one embodiment of the present invention, in step S4, the constructed encoder multilayer neural network module needs to include a factorization machine layer, a multilayer perceptron (MLP) layer, a learnable gated fusion layer, and a Mamba module, with the encoder structure as follows: Figure 2As shown, the training dataset is input into the factorization machine layer to extract nonlinear feature relationships. The dimension of its latent vector matrix is ​​set to 8, and the latent vectors are initialized using Xavier to aid convergence. The weights are randomly initialized following a normal distribution with a mean of 0 and a standard deviation of 0.01. The training dataset is input into the multilayer perceptron (MLP) layer to extract linear feature relationships. The MLP dimension is set to 1024. The extracted nonlinear and linear feature relationships are combined and input into a gated fusion layer for weighted fusion. The output is a set of feature relationships strongly correlated with the remaining service life and health status of industrial operating components. The weights α for nonlinear feature relationships and β for linear feature relationships are set as learnable parameters, and the optimal weight ratio is learned autonomously during model training. The above feature relationship set is input into the Mamba module to extract the temporal dependencies of industrial equipment component features. The Mamba model dimension is set to 256, the state parameter is set to 32, the convolution kernel size is 2, and one layer of zero padding is added. The output is the temporal dependencies of vibration signal data. The four-layer neural network in the encoder module is tightly integrated, extracting and organically fusing the relationships between linear features, nonlinear features, and temporal dependencies, effectively improving the feature extraction performance of the encoder module for vibration signal data in complex environments.

[0041] In one embodiment of the present invention, in step S5, the constructed decoder multilayer neural network module needs to include an SENet module and dilated convolutional layers, and the decoder structure is as follows: Figure 3 As shown, the temporal dependency data output from the encoder neural network is input into the SENet module. First, global average pooling is performed to obtain the global feature description of the data. Then, a squeeze-excitation mechanism is used to adaptively and organically weight and fuse different feature channels to obtain single-channel temporal prediction data. This single-channel temporal prediction data is then input into a dilated convolutional layer with a kernel size of 3 and a dilation rate of 2. This effectively increases the receptive field without increasing computational cost. The dilated convolution fuses the temporal prediction data into single-step prediction data, which is then output. This method effectively fuses data from different channels and time steps in both the channel and time dimensions, providing a precise prediction of remaining lifetime as the prediction head. This effectively reduces information loss during feature fusion and improves prediction accuracy.

[0042] In one embodiment of the present invention, in step S6, before training, the lifetime labels in the training dataset need to be removed in advance, and the time stamps of the training dataset are processed into three additional time dimensions. Simultaneously, the data width parameter is set to a time step of 96 units, the maximum number of training epochs is 40, and the maximum number of early stopping patience epochs is 3. The lifetime labels are then combined with the processed training dataset and input into the model for training. The training data needs to pass through two layers of encoding neural network model layers to extract the relationships between features and temporal dependencies, and then through one layer of decoding neural network model layers to fuse the data features and output predicted values. During training, the mean squared error loss (MSE) is calculated based on the predicted values ​​and lifetime labels. The model parameters are then updated using backpropagation with the Adam optimizer, which has a learning rate of 1e-4 and weight decay of 2e-4 to help the model converge faster and more stably. After each training epoch, the model enters the validation phase. The training parameters are frozen, and the remaining lifetime is predicted using the validation dataset. The validation data MSE loss is calculated based on the predicted values ​​and lifetime labels. After each complete validation cycle, the validation data MSE loss is recorded and compared with the historical lowest validation data MSE loss. If the current loss is lower than the historical lowest loss, [further action is taken]. The current loss is set to the historical minimum loss, and the current model parameters are saved as optimal parameters. The patience value is set to zero. If the current loss is higher than the historical minimum loss, the patience value is increased by one point. After completing the loss comparison on the validation dataset, a judgment needs to be made based on the patience value. If the patience value is higher than the early stopping patience limit, training is stopped and the prediction phase begins. If the patience value is lower than the early stopping patience limit, the next training round begins, and the training gradient is updated based on the mean squared error loss of the validation dataset in the current round. When the patience value is higher than the patience limit or the number of training rounds is higher than the set number of rounds, the prediction phase begins. The above encoder-decoder architecture multilayer neural network model needs to be constructed using the optimal model parameters, and the remaining lifetime is predicted on the test dataset.

[0043] Example 2

[0044] The present invention also provides a system for predicting the remaining useful life of equipment components, comprising:

[0045] The first processing unit is used to collect and preprocess horizontal and vertical vibration signal data during the operation of industrial equipment components;

[0046] The second processing unit is used to extract time-domain, frequency-domain, and time-frequency-domain features from the preprocessed horizontal and vertical vibration signal data of the operating components, label them with life tags, and divide them into training set, validation set, and test set.

[0047] The third processing unit is used to construct the encoder-decoder architecture model, which includes a 2-layer encoder multi-layer neural network module and a 1-layer decoder multi-layer neural network module. The encoder multi-layer neural network module includes a factorization machine layer, a multi-layer perceptron (MLP) layer, a learnable gated fusion layer, and a Mamba module. The decoder multi-layer neural network module includes a SENet module and dilated convolutional layers.

[0048] The fourth processing unit is used to train the FM-Mamba-SE model based on the training dataset;

[0049] The fifth processing unit is used to predict the remaining lifetime on the test dataset using the trained FM-Mamba-SE model, and outputs the remaining lifetime for each step.

[0050] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A method for predicting the remaining life of equipment components, characterized in that, The method comprises the following steps: Collecting and preprocessing horizontal and vertical vibration signal data during the operation of an industrial equipment component; Extracting time domain, frequency domain and time-frequency domain features from the preprocessed horizontal and vertical vibration signal data of the operating component, labeling the life label, and dividing it into a training set, a validation set and a test set; Building an encoder-decoder architecture model; wherein the encoder-decoder architecture model comprises: a 2-layer encoder multi-layer neural network module and a 1-layer decoder multi-layer neural network module; the encoder multi-layer neural network module comprises: a factorization machine layer, a multi-layer perceptron MLP layer, a learnable gate fusion layer and a Mamba module; the decoder multi-layer neural network module comprises: a SENet module and a cavity convolution layer; Training the FM-Mamba-SE model according to the training data set; Using the trained FM-Mamba-SE model to predict the remaining useful life on the test data set, and outputting the remaining useful life prediction value of each step; The training data set is input into the factorization machine layer to extract the relationship between nonlinear features, the dimension of the hidden vector matrix is set to 8, the hidden vector is initialized by Xavier to help convergence, and the weight is randomly initialized to follow the normal distribution with a mean of 0 and a standard deviation of 0.01; the training data set is input into the multi-layer perceptron MLP layer to extract the relationship between linear features, and the MLP dimension is set to 1024; the above-mentioned nonlinear feature relationship and linear feature relationship are combined and input into the gate fusion layer for weighted fusion, and the feature relationship set strongly related to the remaining useful life and health status of the industrial operating component is output, wherein the nonlinear feature relationship weight ɑ and the linear feature relationship weight β are set as learnable parameters, and the best weight ratio is learned automatically with the model training; the above-mentioned feature relationship set is input into the Mamba module to extract the time sequence dependence relationship of the industrial equipment component features, the Mamba model dimension is set to 256 dimensions, the state parameter is set to 32, the convolution kernel size is 2, and the expansion is 1 layer of zero padding, and the time sequence dependence relationship of the vibration signal data is output. The time sequence dependence relationship data output from the above encoder neural network is input into the SENet module, first through global average pooling to obtain the global feature description of the data, and then using the compression-excitation mechanism to adaptively and organically fuse different feature channels to obtain single-channel time sequence prediction data; the above-mentioned single-channel time sequence prediction data is input into the cavity convolution layer, and the convolution kernel size is set to 3 and the cavity rate is set to 2.

2. A device component remaining life prediction system that implements the device component remaining life prediction method according to claim 1, characterized by, The method comprises the following steps: The first processing unit is used for collecting and preprocessing horizontal and vertical vibration signal data during the operation of an industrial equipment component; The second processing unit is used for extracting time domain, frequency domain and time-frequency domain features from the preprocessed horizontal and vertical vibration signal data of the operating component, labeling the life label, and dividing it into a training set, a validation set and a test set; The third processing unit is configured to construct an encoder-decoder architecture model, wherein the encoder-decoder architecture model comprises a 2-layer encoder multi-layer neural network module and a 1-layer decoder multi-layer neural network module; the encoder multi-layer neural network module comprises a factorization machine layer, a multi-layer perception (MLP) layer, a learnable gating fusion layer and a Mamba module; and the decoder multi-layer neural network module comprises an SE-Net module and a dilated convolution layer. The fourth processing unit is configured to train the FM-Mamba-SE model according to a training data set. The fifth processing unit is configured to use the trained FM-Mamba-SE model to perform remaining useful life prediction on a test data set, and output a remaining useful life prediction value at each step.

Images