Rotating machine life prediction method based on working condition self-adaption and alternating coupling network
By employing an adaptive operating condition and alternating coupled network approach, the problem of insufficient generalization ability of rotating machinery life prediction models under multiple operating conditions is solved. This approach achieves dynamic fusion of global and local features, thereby improving prediction accuracy and robustness.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-12
- Publication Date
- 2026-04-07
AI Technical Summary
Existing life prediction models for rotating machinery lack generalization ability under multiple operating conditions and are difficult to adapt to dynamic changes in operating conditions. Furthermore, existing hybrid models lack deep interaction and multi-level complementary mechanisms, making it difficult to comprehensively characterize the degradation process.
A method based on adaptive working conditions and alternating coupling networks is adopted. By acquiring vibration signal data under different working conditions, a multi-domain feature set is constructed, and an adaptive working condition embedding vector is introduced to construct a multi-layer network structure containing a global dependency modeling module and a bidirectional temporal convolutional network module, so as to realize the dynamic fusion of global and local features.
The model's robustness and prediction accuracy under multiple operating conditions have been enhanced. It can maintain high prediction accuracy when components are close to failure, adapt to changes in operating conditions, and improve the accuracy of rotating machinery life prediction.
Smart Images
Figure CN121809275A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of mechanical life prediction technology, and more specifically to a method for predicting the life of rotating machinery based on adaptive operating conditions and alternating coupling networks. Background Technology
[0002] Rotating machinery, as core equipment in industrial systems, directly impacts production efficiency, equipment safety, and operating costs through the stable operation of its key components. Accurate prediction of the remaining lifespan of rotating machinery facilitates the implementation of scientific preventative maintenance strategies, thereby reducing the risk of unplanned downtime and extending equipment lifespan. Against the backdrop of rapid development in industrial intelligence and big data applications, data-driven intelligent prediction methods have attracted widespread attention due to their ability to automatically uncover degradation patterns in complex signals.
[0003] However, the monitoring data of rotating machinery is highly heterogeneous and non-stationary due to factors such as changes in operating conditions, load fluctuations, and environmental influences. This results in insufficient generalization ability of traditional data-driven methods under multiple operating conditions, and model training often relies on fixed operating condition data, making it difficult to adapt to the dynamic changes in actual operation. Existing single prediction models, which can only focus on one aspect of the degradation process, have poor predictive performance. To take into account both the global trend and local abrupt changes in rotating machinery degradation, some studies have attempted to combine different models. However, existing hybrid models usually use simple network stacking structures to achieve the fusion of global and local features, lacking deep interaction and multi-level complementary mechanisms, making it difficult to comprehensively represent the degradation process.
[0004] Therefore, how to enhance the model's ability to represent complex degradation processes and improve its prediction accuracy is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] In view of the above problems, the present invention proposes a method for predicting the life of rotating machinery based on adaptive operating conditions and alternating coupling networks, so as to overcome the above problems or at least partially solve them.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for predicting the life of rotating machinery based on adaptive operating conditions and alternating coupling networks, comprising the following steps: The vibration signal data of rotating mechanical components throughout their entire life cycle under different working conditions were acquired, preprocessed, and then used to construct the original dataset. For each original data sample, time-domain features, frequency-domain features, and time-frequency-domain features are extracted to construct a multi-domain feature set; Effective feature sets are selected from multi-domain feature sets based on three dimensions: relevance, monotonicity, and robustness. An adaptive working condition embedding vector is introduced, and each effective feature in the effective feature set is added row by row to the working condition embedding vector to obtain a coupled feature set; A lifetime prediction model is constructed that includes a global dependency modeling module and a bidirectional temporal convolutional network module, wherein the global dependency modeling module and the bidirectional temporal convolutional network module are stacked alternately to form a multi-layer network structure; The lifespan prediction model is trained using the coupled feature set. In each training round, the global dependency modeling module in the first layer of the network structure extracts the global dependency features of the coupled features. The bidirectional temporal convolutional network module performs forward and backward temporal convolution on the global dependency features to obtain local temporal features. The weighted global dependency features and local temporal features of the current layer are used as the input of the next layer of the network structure. Life prediction of rotating machinery components is performed based on a trained life prediction model.
[0007] Furthermore, the process of constructing the original dataset includes: Acquire the original vibration signals in the horizontal and vertical directions of the mechanical rotating component from normal operation to failure under different working conditions. Each sample is represented as ,in, and They represent the first Vibration signals of a sample in the horizontal and vertical directions. The time step for a single sample; Noise suppression is performed on the original vibration signal. The denoised continuous vibration signal is divided into fixed time windows, with each segment containing... The sampling point will be the first sampling point. Each sample is divided into several signal segments. ,in, , Indicates by the first The number of segments obtained from each sample segmentation; the actual remaining lifetime corresponding to each monitoring time is denoted as . , as a tag.
[0008] Furthermore, the initial multi-domain feature vector of a single sample is represented as: Where D is the initial feature dimension. ,, This represents the total number of temporal features extracted from a single sample in the horizontal and vertical directions. This represents the total number of frequency domain features extracted from a single sample in the horizontal and vertical directions. This represents the total number of time-frequency domain features for a single sample; ";" indicates dimension concatenation. This represents the time-domain feature vector of a single sample. This represents the frequency domain feature vector of a single sample. Represents the time-frequency domain feature vector of a single sample; An initial multi-domain feature set is constructed from N samples. ,in, Indicates the current dimension. , For the first The first sample Initial values for the dimensional features; Max-min normalization is performed on all features in the initial multi-domain feature set to map the feature values to the interval [0,1], resulting in the standardized multi-domain feature set. .
[0009] Furthermore, the process of selecting the effective feature set includes: The Pearson correlation coefficient is used to assess the correlation between degradation characteristics and remaining lifespan. The calculation formula is as follows:
[0010] Where Corr(i, d) represents the relevance index of the i-th sample in the d-th feature dimension. For the first The sample at the th Normalized feature values on each feature dimension This represents the actual remaining lifetime of the corresponding sample. Indicates that all samples are in the first... The mean of each feature dimension N represents the mean of the actual remaining lifespan of all samples, and N represents the total number of samples. The Spearman correlation coefficient is used to assess the monotonicity of a feature over time. The calculation formula is as follows:
[0011] Where Mon(i, d) represents the monotonicity index of the i-th sample in the d-th feature dimension. and These are the ranks of the eigenvalues and the remaining lifetime of the samples, respectively. A robustness index is introduced to measure the stability of the characteristic under noise interference and operating condition fluctuations. The calculation formula is as follows:
[0012] Where Rob(i, d) represents the robustness index of the i-th sample in the d-th feature dimension. Indicates the first under the same working condition The standard deviation of the dimensional feature across different samples This represents the standard deviation of the mean of this feature under different operating conditions. To prevent tiny constants with a denominator of zero; Based on correlation, monotonicity, and robustness indicators, a feature comprehensive score function is constructed, and the calculation formula is as follows:
[0013] Where g(i, d) represents the comprehensive score of the i-th sample in the d-th feature dimension. , , For empirical weighting coefficients, satisfying ; Calculate the comprehensive score of each feature dimension on N samples, sort the feature dimensions from highest to lowest comprehensive score, and select the top U feature dimensions to form the final effective feature set. .
[0014] Furthermore, the process of constructing the coupled feature set includes: The various operating condition parameters are subjected to maximum-minimum normalization to obtain the real-time operating condition parameter vector C(t); The real-time operating condition parameter vector C(t) is mapped to an embedding vector with the same effective feature dimension through a two-layer fully connected network. The calculation formula is as follows:
[0015] in, , , This is the weight matrix. , For bias terms; Introducing operating condition-specific bias The final embedding vector is obtained. ; Each effective feature is added row by row to the working condition embedding vector to obtain the coupled features. The calculation formula is as follows:
[0016] in, This represents the coupled feature set, which serves as the input to the lifetime prediction model.
[0017] Furthermore, the processing flow of the input features by the global dependency modeling module in the first-layer network structure includes: Coupling characteristics The query matrix is obtained by inputting the data into the global dependency modeling module in the form of a time series and then performing a linear mapping. Key matrix Sum matrix The attention output is obtained by calculating the correlation weights between different time points in the input sequence using a multi-head attention mechanism. ; Output attention Coupling characteristics with initial input Perform residual connections and layer normalization to obtain normalized features. ; Normalization characteristics Nonlinear global features extracted via feedforward network ; Nonlinear global features With normalization characteristics Perform a second residual connection and layer normalization operation to obtain global dependency features. .
[0018] Furthermore, each bidirectional temporal convolutional network module consists of a forward temporal convolutional network and a backward temporal convolutional network operating in parallel. Its processing flow for input features includes: Perform forward and backward temporal convolution operations on the input feature sequence to obtain forward and backward outputs, respectively; The forward and backward outputs are concatenated to obtain the local fusion features. ; Local fusion features With global dependency features Weighting is performed to obtain new input features. , to new input features Transmitted to the next layer of global dependency modeling module.
[0019] Furthermore, the forward and backward temporal convolution operations are as follows:
[0020]
[0021] in, For activation function, Indicates the first The layer bidirectional temporal convolutional network module in the first layer The input features at the given time and past time steps, where d*k represents the time step offset of the dilated convolution. Indicates the first The layer bidirectional temporal convolutional network module in the first layer Input features of the current time and future time steps; For expansion rate, The kernel size is [size]. This represents the bias term of the forward temporal convolution. This represents the bias term of the inverse temporal convolution. Indicates the first The first layer of forward temporal convolution Each convolutional kernel weight, Indicates the first The first layer of reverse temporal convolution Each convolutional kernel weight, Indicates the forward temporal convolution at the th... Output at any moment Indicates the inverse temporal convolution at the th... Output at any given moment.
[0022] Furthermore, the feature sequence output by the last layer of the network structure is linearly predicted to obtain the lifetime prediction value at each time step, denoted as: ,in, This represents the total number of time steps. The mean absolute error is used as the loss function to measure the deviation between the predicted value and the actual lifetime.
[0023] in, This represents the actual remaining lifetime value at time step t. This is the predicted value at time step t.
[0024] Secondly, the present invention provides a rotating machinery life prediction system based on adaptive operating conditions and alternating coupling networks, which is applicable to the above-mentioned method, including: The data acquisition module is used to acquire vibration signal data of rotating mechanical components throughout their entire life cycle under different working conditions, and after preprocessing, construct the original dataset. The multi-frequency domain feature extraction module is used to extract time-domain features, frequency-domain features, and time-frequency-domain features for each original data sample, and construct a multi-domain feature set. The effective feature selection module is used to select effective feature sets from multi-domain feature sets based on three dimensions: relevance, monotonicity, and robustness. The working condition embedding module is used to introduce an adaptive working condition embedding vector and add each effective feature in the effective feature set to the working condition embedding vector row by row to obtain a coupled feature set. The deep network module is used to build a lifetime prediction model that includes a global dependency modeling module and a bidirectional temporal convolutional network module. The global dependency modeling module and the bidirectional temporal convolutional network module are stacked alternately to form a multi-layer network structure. The training module is used to train the life prediction model using the coupled feature set. In each training round, the global dependency modeling module in the first layer of the network structure extracts the global dependency features of the coupled features. The bidirectional temporal convolutional network module performs forward and backward temporal convolution on the global dependency features to obtain local temporal features. The weighted global dependency features and local temporal features of the current layer are used as the input of the next layer of the network structure. The prediction module is used to predict the lifespan of rotating machinery components based on a trained lifespan prediction model.
[0025] As can be seen from the above technical solution, compared with the prior art, the present invention has the following beneficial effects: 1. This invention uses adaptive embedding based on operating conditions to dynamically adjust the feature space representation according to operating condition parameters such as speed and load, thereby effectively modeling degradation laws under multiple operating conditions and enhancing the robustness and generalization ability of the model under different operating conditions.
[0026] 2. This invention constructs an alternating coupled deep network structure, which dynamically interacts the global dependency modeling module and the bidirectional temporal convolution module at the network level to achieve multi-layer fusion of global temporal trends and local degradation features, taking into account both global dependencies and local mutation features. This breaks through the limitation of traditional stacked models in utilizing only single information, and enables the prediction results to maintain high accuracy even when the component is close to failure. Attached Figure Description
[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0028] Figure 1 This is a flowchart of the rotating machinery life prediction method based on adaptive operating conditions and alternating coupling network provided in the embodiments of the present invention; Figure 2 This is an architecture diagram of the lifetime prediction model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the performance verification results of the present invention; Figure 4 This is a schematic diagram comparing the performance of the present invention with existing methods. Detailed Implementation
[0029] 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.
[0030] like Figures 1-2 As shown in the figure, this invention discloses a method for predicting the life of rotating machinery based on adaptive operating conditions and alternating coupling networks, including the following steps: S1. Obtain vibration signal data of rotating mechanical components throughout their entire life cycle under different working conditions, perform preprocessing, and construct the original dataset; S2. For each original data sample, extract time-domain features, frequency-domain features, and time-frequency-domain features respectively to construct a multi-domain feature set; S3. Select effective feature sets from the multi-domain feature set based on three dimensions: relevance, monotonicity, and robustness. S4. Introduce an adaptive working condition embedding vector, and add each effective feature in the effective feature set to the working condition embedding vector row by row to obtain a coupled feature set; S5. Construct a lifetime prediction model that includes a global dependency modeling module and a bidirectional temporal convolutional network module, wherein the global dependency modeling module and the bidirectional temporal convolutional network module are stacked alternately to form a multi-layer network structure. The lifespan prediction model is trained using the coupled feature set. In each training round, the global dependency modeling module in the first layer of the network structure extracts the global dependency features of the coupled features. The bidirectional temporal convolutional network module performs forward and backward temporal convolution on the global dependency features to obtain local temporal features. The weighted global dependency features and local temporal features of the current layer are used as the input of the next layer of the network structure. S6. Predict the life of rotating machinery components based on the trained life prediction model.
[0031] The following provides further explanation of each of the above steps.
[0032] S1, Data Acquisition and Preprocessing Stage, specifically includes: Using an accelerometer, the original vibration signals in the horizontal and vertical directions of a mechanical rotating component during its normal operation to failure under different operating conditions such as rotational speed and load are acquired. Let the total number of samples be N, and the... Each sample is represented as ,in, and They represent the first Vibration signals of a sample in the horizontal and vertical directions. The time step for a single sample; Noise suppression is performed on the original vibration signal, specifically using wavelet threshold denoising or empirical mode decomposition methods. This decomposes the signal into several components and filters out high-frequency noise terms and low-frequency background interference, retaining only the effective components containing equipment degradation information, denoted as... ,in, For the denoised first The sample at the th Signals in each direction. The denoised continuous vibration signal is divided into fixed time windows, with each segment containing... The sampling point will be the first sampling point. Each sample is divided into several signal segments. ,in, , Indicates by the first The number of segments obtained from each sample segmentation; the actual remaining lifetime corresponding to each monitoring time is denoted as . , as a tag.
[0033] To facilitate feature extraction and subsequent model training, each sample is converted into a tensor form. ,in Indicates the number of channels. This represents the number of sampling points within the time window.
[0034] S2, Multi-domain Feature Extraction Stage, specifically includes: First, input the vibration signal in S1 into the tensor. For the processing object, the vibration signal of each channel of a single sample. ,Right now Features were extracted in the time domain, frequency domain, and time-frequency domain, respectively. For the first The first sample The signal value at each time step.
[0035] Temporal feature extraction: Based on the statistical properties of the signal, the mean, variance, peak factor, skewness, kurtosis, and other temporal features of rotating machinery are extracted from each channel of a single sample to obtain the temporal feature vector of that sample. ,in The number of extracted time-domain features, This represents the total number of temporal features extracted from a single sample in the horizontal and vertical directions.
[0036] Frequency domain feature extraction: The signal is transformed to the frequency domain using Fast Fourier Transform (FFT). The FFT formula is as follows: ,in Sampling frequency, For the first Frequency value at each frequency point , For the first Frequency domain amplitude at each frequency point The imaginary unit is used. Based on the frequency domain amplitude spectrum. Frequency domain features such as spectral centroid, root mean square frequency, spectral skewness, and spectral kurtosis are extracted to obtain the frequency domain feature vector of the sample. ,in The number of frequency domain features extracted, This represents the total number of frequency domain features extracted from a single sample in the horizontal and vertical directions.
[0037] Time-frequency domain feature extraction: Considering the non-stationary nature of the vibration signal, wavelet packet decomposition is used to perform multi-scale decomposition of the vibration signal. Let the number of decomposition levels be... ,get Each one carries a signal. Layer Each sub-band signal is denoted as Its energy characteristics are The time-frequency domain energy feature vector of the sample is obtained. ,in, Indicates the first The energy feature set obtained from the layer decomposition. Further, the feature sets of all decomposed layers are concatenated to form the time-frequency domain energy feature vector of the sample. ,in, This represents the complete time-frequency domain feature vector. Using the first... For a single channel in a layer (final layer) sub-band, the characteristic number is: If the input signal contains two channels, one horizontal and one vertical, then the total number of time-frequency domain features of a single sample is: .
[0038] The time-domain feature vector of a single sample Frequency domain eigenvectors With time-frequency domain feature vectors Horizontal concatenation yields the initial multi-domain feature vector representation of a single sample. Where D is the initial feature dimension. ,, This represents the total number of temporal features extracted from a single sample in the horizontal and vertical directions. This represents the total number of frequency domain features extracted from a single sample in the horizontal and vertical directions. The symbol represents the total number of time-frequency domain features of a single sample, and ";" indicates dimension concatenation.
[0039] An initial multi-domain feature set is constructed from N samples. ,in, Indicates the current dimension. , For the first The first sample Initial values for the dimensional features.
[0040] To avoid slow convergence or prediction bias caused by differences in the magnitude of feature values during model training, max-min normalization is performed on all features in the initial multi-domain feature set, mapping the feature values to the [0,1] interval to obtain the standardized multi-domain feature set. Specifically, for the first feature set... Dimensional features, the first The formula for calculating the normalized eigenvalues of a sample is as follows: ,in, For the first The first sample The original values of the dimensional features; , The first The maximum and minimum values of the dimensional features are processed to obtain a standardized initial multi-domain feature set. .
[0041] S3. Filtering of the effective feature set, specifically including: The Pearson correlation coefficient is used to assess the correlation between degradation characteristics and remaining lifespan. The calculation formula is as follows:
[0042] Where Corr(i, d) represents the relevance index of the i-th sample in the d-th feature dimension. For the first The sample at the th Normalized feature values on each feature dimension This represents the actual remaining lifetime of the corresponding sample. Indicates that all samples are in the first... The mean of each feature dimension N represents the mean of the actual remaining lifespan of all samples, and N represents the total number of samples. The Spearman correlation coefficient is used to assess the monotonicity of a feature over time. The calculation formula is as follows:
[0043] Where Mon(i, d) represents the monotonicity index of the i-th sample in the d-th feature dimension. and These are the ranks of the eigenvalues and the remaining lifetime of the samples, respectively. A robustness index is introduced to measure the stability of the characteristic under noise interference and operating condition fluctuations. The calculation formula is as follows:
[0044] Where Rob(i, d) represents the robustness index of the i-th sample in the d-th feature dimension. Indicates the first under the same working condition The standard deviation of the dimensional feature across different samples This represents the standard deviation of the mean of this feature under different operating conditions. To prevent tiny constants with a denominator of zero; Based on correlation, monotonicity, and robustness indicators, a feature comprehensive score function is constructed, and the calculation formula is as follows:
[0045] Where g(i, d) represents the comprehensive score of the i-th sample in the d-th feature dimension. , , For empirical weighting coefficients, satisfying ; Calculate the comprehensive score of each feature dimension on N samples, sort the feature dimensions from highest to lowest comprehensive score, and select the top U feature dimensions to form the final effective feature set. .
[0046] S4, the adaptive operating condition embedding stage, specifically includes: Multiple operating parameters are subjected to maximum-minimum normalization to obtain the real-time operating parameter vector C(t). Specifically, the rotational speed is... Load The two key operating parameters are normalized using the maximum-minimum method, as shown in the formula. , ,in, Indicates rotational speed, For radial load, and These represent the maximum and minimum rotational speeds in the entire lifecycle dataset, respectively. and These represent the maximum and minimum load values in the entire lifecycle dataset, respectively. After standardization, they form the operating condition parameter vector. The global matrix is If more environmental or operating parameters exist, such as temperature... Lubrication condition parameters etc. can be expanded to: .
[0047] The real-time operating condition parameter vector C(t) is mapped to an embedding vector with the same effective feature dimension through a two-layer fully connected network. The calculation formula is as follows:
[0048] in, , , This is the weight matrix. , For bias terms; Introducing operating condition-specific bias The final embedding vector is obtained. ; Each effective feature is added row by row to the working condition embedding vector to obtain the coupled features. The calculation formula is as follows:
[0049] in, This represents the coupled feature set, which serves as the input to the lifetime prediction model.
[0050] S5. In the construction and training phase of the lifetime prediction model, the coupled feature sequences obtained in S4 are... As model input, a deep network structure is constructed by alternately stacking a global dependency modeling module and a bidirectional temporal convolutional network (BiTCN) module to achieve hierarchical representation and dynamic fusion of global and local features. The global dependency modeling module is designed based on the Transformer encoding structure to capture the global dependencies of features over long time scales. Taking the processing flow of the global dependency modeling module for input features in the first layer of the network structure as an example, the specific processing steps include: Coupling characteristics The query matrix is obtained by inputting the data into the global dependency modeling module in the form of a time series and then performing a linear mapping. Key matrix Sum matrix :
[0051] in, For trainable weights, For the first Layer input feature matrix.
[0052] By employing a multi-head attention mechanism to calculate the correlation weights between different time steps in the input sequence, global feature modeling is achieved, resulting in the attention output. :
[0053] in, The dimension of the key vector. This indicates the operation of a multi-head attention mechanism.
[0054] Output attention Coupling characteristics with initial input Perform residual connections and layer normalization to obtain normalized features. : ; Normalization characteristics Nonlinear global features extracted via feedforward network :
[0055] in, These are the parameters for the feedforward layer.
[0056] Nonlinear global features With normalization characteristics Perform a second residual connection and layer normalization operation to obtain global dependency features. : ; Global dependency features As input to the BiTCN module, it enables the transfer of global features to local feature modeling.
[0057] The BiTCN module consists of parallel forward and backward temporal convolutional networks. Each sub-network employs causal convolution and dilated convolution structures, thereby expanding the receptive field while maintaining temporal causality. The module's processing flow for input features includes: Perform forward and backward temporal convolution operations on the input feature sequence to obtain the forward and backward outputs, respectively:
[0058]
[0059] in, For activation function, Indicates the first The layer bidirectional temporal convolutional network module in the first layer The input features at the given time and past time steps, where d*k represents the time step offset of the dilated convolution. Indicates the first The layer bidirectional temporal convolutional network module in the first layer Input features of the current time and future time steps; For expansion rate, The kernel size is [size]. This represents the bias term of the forward temporal convolution. This represents the bias term of the inverse temporal convolution. Indicates the first The first layer of forward temporal convolution Each convolutional kernel weight, Indicates the first The first layer of reverse temporal convolution Each convolutional kernel weight, Indicates the forward temporal convolution at the th... Output at any moment Indicates the inverse temporal convolution at the th... Output at any given moment.
[0060] The forward and backward outputs are concatenated to obtain the local fusion features. :
[0061] Local fusion features With global dependency features Weighting is performed to obtain new input features. The calculation formula is:
[0062] in, The adaptive fusion coefficients can be learned automatically through training.
[0063] New input features Transmitted to the global dependency modeling module of the next layer.
[0064] The bidirectional temporal convolutional network module introduces residual blocks, leveraging residual connections to enhance model performance and effectively address the vanishing gradient problem, avoiding the increase in prediction error as network depth increases. The residual block consists of two bidirectional causal dilated convolutional layers, a weight normalization layer, an activation function layer, and Dropout regularization. Furthermore, since the input and output of a convolutional network may have different shapes, traditional residual connection networks directly merge the input and output. This invention uses a 1×1 convolutional kernel to ensure consistent input and output dimensions.
[0065] Subsequently, the feature matrix output by the alternating coupling network architecture is... A linear mapping is performed to obtain the predicted lifetime value for each time step. The calculation formula is as follows:
[0066] in, and These are the weight matrix and bias term of the output layer, respectively. For the first The last layer (i.e., the last layer) The feature matrix at each time step.
[0067] The training process for the lifespan prediction model includes: 1) Forward Propagation: The model sequentially passes through alternating stacks of a global dependency modeling module and a bidirectional temporal convolution (BiTCN) module, extracting global and local temporal features layer by layer, and obtaining the lifetime prediction vector through network mapping:
[0068] in, This represents the total number of time steps.
[0069] 2) Loss Calculation: The mean absolute error (MAE) is used as the loss function to measure the deviation between the predicted value and the actual remaining lifetime.
[0070] in, Let be the actual remaining lifetime at time step t. This represents the predicted value corresponding to the model.
[0071] 3) Backpropagation and Parameter Update: The Adam optimization algorithm is used to adaptively update the network parameters to accelerate convergence and prevent gradient oscillations. The parameter update rules are as follows:
[0072] in, For learning rate, For model parameters, This is the gradient of the loss function with respect to the model parameters.
[0073] After training, the independent test set is input into the trained lifetime prediction model for inference, and the predicted remaining lifetime sequence is output. The predictive accuracy of the model is evaluated using metrics such as mean squared error (MSE) and root mean square error (RMSE).
[0074]
[0075] S6. After model training and verification, the vibration signal of the rotating mechanical component to be evaluated is input into the model to predict the life of the rotating mechanical component. Based on the life prediction results, a remaining life curve that changes over time is generated to characterize the degradation process of the equipment from health to failure.
[0076] In other embodiments, the present invention also provides a rotating machinery life prediction system based on adaptive operating conditions and alternating coupling networks, which is applicable to the above-described method and includes: The data acquisition module is used to acquire vibration signal data of rotating mechanical components throughout their entire life cycle under different working conditions, and after preprocessing, construct the original dataset. The multi-frequency domain feature extraction module is used to extract time-domain features, frequency-domain features, and time-frequency-domain features for each original data sample, and construct a multi-domain feature set. The effective feature selection module is used to select effective feature sets from multi-domain feature sets based on three dimensions: relevance, monotonicity, and robustness. The working condition embedding module is used to introduce an adaptive working condition embedding vector and add each effective feature in the effective feature set to the working condition embedding vector row by row to obtain a coupled feature set. The deep network module is used to build a lifetime prediction model that includes a global dependency modeling module and a bidirectional temporal convolutional network module. The global dependency modeling module and the bidirectional temporal convolutional network module are stacked alternately to form a multi-layer network structure. The training module is used to train the life prediction model using the coupled feature set. In each training round, the global dependency modeling module in the first layer of the network structure extracts the global dependency features of the coupled features. The bidirectional temporal convolutional network module performs forward and backward temporal convolution on the global dependency features to obtain local temporal features. The weighted global dependency features and local temporal features of the current layer are used as the input of the next layer of the network structure. The prediction module is used to predict the lifespan of rotating machinery components based on a trained lifespan prediction model.
[0077] Next, using the accelerated degradation experiment of rolling bearings as an example, the run-to-failure data collected in the experiment were used to evaluate the effectiveness of the method of the present invention. The experimental data used an LDK UER204 bearing. The experiment used a PCB 352C33 sensor to record the vibration signals of the bearing in the horizontal and vertical directions. The sampling frequency was set to 25.6 kHz, the sampling time was 1.28 s, and the sampling period was 1 min. Therefore, a total of 32,768 data points were recorded for each sampling. The dataset contains complete run-to-failure data for 15 rolling bearings obtained through multiple accelerated degradation experiments. Three different operating conditions (i.e., different radial forces and rotational speeds) were set up in the accelerated degradation experiment, and five bearings were tested under each condition. The operating conditions are shown in Table 1.
[0078] Table 1 Bearing accelerated life test conditions
[0079] The experiment takes the working condition of 37.5Hz / 11kN as an example. Bearing2_1, Bearing2_2, Bearing2_4 and Bearing2_5 under this working condition are used as the training set, and Bearing2_3 is used as the test set. The model hyperparameter settings are shown in Table 2.
[0080] Table 2 Model Hyperparameter Settings
[0081] The effectiveness of the model is evaluated by comparing it with actual remaining lifespan data, such as... Figure 3 As shown. Furthermore, the model's superiority was evaluated through experimental comparisons with various common deep learning models, such as... Figure 4 As shown in Table 3.
[0082] Table 3 Evaluation results of the performance of the method of the present invention
[0083] The experimental results show that the present invention has good health assessment performance, and the method of the present invention has obvious advantages over other methods.
[0084] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0085] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for predicting the life of rotating machinery based on adaptive operating conditions and alternating coupled networks, characterized in that, Includes the following steps: The vibration signal data of rotating mechanical components throughout their entire life cycle under different working conditions were acquired, preprocessed, and then used to construct the original dataset. For each original data sample, time-domain features, frequency-domain features, and time-frequency-domain features are extracted to construct a multi-domain feature set; Effective feature sets are selected from multi-domain feature sets based on three dimensions: relevance, monotonicity, and robustness. An adaptive working condition embedding vector is introduced, and each effective feature in the effective feature set is added row by row to the working condition embedding vector to obtain a coupled feature set; A lifetime prediction model is constructed that includes a global dependency modeling module and a bidirectional temporal convolutional network module, wherein the global dependency modeling module and the bidirectional temporal convolutional network module are stacked alternately to form a multi-layer network structure; The lifespan prediction model is trained using the coupled feature set. In each training round, the global dependency modeling module in the first layer of the network structure extracts the global dependency features of the coupled features. The bidirectional temporal convolutional network module performs forward and backward temporal convolution on the global dependency features to obtain local temporal features. The weighted global dependency features and local temporal features of the current layer are used as the input of the next layer of the network structure. Life prediction of rotating machinery components is performed based on a trained life prediction model.
2. The method as described in claim 1, characterized in that, The process of constructing the original dataset includes: Acquire the original vibration signals in the horizontal and vertical directions of the mechanical rotating component from normal operation to failure under different working conditions. Each sample is represented as ,in, and They represent the first Vibration signals of a sample in the horizontal and vertical directions. The time step for a single sample; Noise suppression is performed on the original vibration signal. The denoised continuous vibration signal is divided into fixed time windows, with each segment containing... The sampling point will be the first sampling point. Each sample is divided into several signal segments. ,in, , Indicates by the first The number of segments obtained from each sample segmentation; the actual remaining lifetime corresponding to each monitoring time is denoted as . , as a tag.
3. The method as described in claim 1, characterized in that, The initial multi-domain feature vector of a single sample is represented as follows: Where D is the initial feature dimension. ,, This represents the total number of temporal features extracted from a single sample in the horizontal and vertical directions. This represents the total number of frequency domain features extracted from a single sample in the horizontal and vertical directions. This represents the total number of time-frequency domain features for a single sample, with ";" indicating dimension concatenation. This represents the time-domain feature vector of a single sample. This represents the frequency domain feature vector of a single sample. Represents the time-frequency domain feature vector of a single sample; An initial multi-domain feature set is constructed from N samples. ,in, Indicates the current dimension. , For the first The first sample Initial values for the dimensional features; Max-min normalization is performed on all features in the initial multi-domain feature set to map the feature values to the interval [0,1], resulting in the standardized multi-domain feature set. .
4. The method as described in claim 1, characterized in that, The process of selecting an effective feature set includes: The Pearson correlation coefficient is used to assess the correlation between degradation characteristics and remaining lifespan. The calculation formula is as follows: Where Corr(i, d) represents the relevance index of the i-th sample in the d-th feature dimension. For the first The sample at the th Normalized feature values on each feature dimension This represents the actual remaining lifetime of the corresponding sample. Indicates that all samples are in the first... The mean of each feature dimension N represents the mean of the actual remaining lifespan of all samples, and N represents the total number of samples. The Spearman correlation coefficient is used to assess the monotonicity of a feature over time. The calculation formula is as follows: Where Mon(i, d) represents the monotonicity index of the i-th sample in the d-th feature dimension. and These are the ranks of the eigenvalues and the remaining lifetime of the samples, respectively. A robustness index is introduced to measure the stability of the characteristic under noise interference and operating condition fluctuations. The calculation formula is as follows: Where Rob(i, d) represents the robustness index of the i-th sample in the d-th feature dimension. Indicates the first under the same working condition The standard deviation of the dimensional feature across different samples This represents the standard deviation of the mean of this feature under different operating conditions. To prevent tiny constants with a denominator of zero; Based on correlation, monotonicity, and robustness indicators, a feature comprehensive score function is constructed, and the calculation formula is as follows: Where g(i, d) represents the comprehensive score of the i-th sample in the d-th feature dimension. , , For empirical weighting coefficients, satisfying ; Calculate the comprehensive score of each feature dimension across all samples, sort the feature dimensions from highest to lowest comprehensive score, and select the top U feature dimensions to form the final effective feature set. .
5. The method as described in claim 4, characterized in that, The process of constructing the coupled feature set includes: The various operating condition parameters are subjected to maximum-minimum normalization to obtain the real-time operating condition parameter vector C(t); The real-time operating condition parameter vector C(t) is mapped to an embedding vector with the same effective feature dimension through a two-layer fully connected network. The calculation formula is as follows: in, , , This is the weight matrix. , For bias terms; Introducing operating condition-specific bias The final embedding vector is obtained. ; Each effective feature is added row by row to the working condition embedding vector to obtain the coupled features. The calculation formula is as follows: in, This represents the coupled feature set, which serves as the input to the lifetime prediction model.
6. The method as described in claim 1, characterized in that, The processing flow of the input features by the global dependency modeling module in the first layer of the network structure includes: Coupling characteristics The query matrix is obtained by inputting the data into the global dependency modeling module in the form of a time series and then performing a linear mapping. Key matrix Sum matrix The attention output is obtained by calculating the correlation weights between different time points in the input sequence using a multi-head attention mechanism. ; Output attention Coupling characteristics with initial input Perform residual connections and layer normalization to obtain normalized features. ; Normalization characteristics Nonlinear global features extracted via feedforward network ; Nonlinear global features With normalization characteristics Perform a second residual connection and layer normalization operation to obtain global dependency features. .
7. The method as described in claim 1, characterized in that, Each bidirectional temporal convolutional network module consists of a forward temporal convolutional network and a backward temporal convolutional network running in parallel. Its processing flow for input features includes: Perform forward and backward temporal convolution operations on the input feature sequence to obtain forward and backward outputs, respectively; The forward and backward outputs are concatenated to obtain the local fusion features. ; Local fusion features With global dependency features Weighting is performed to obtain new input features. , to new input features Transferred to the next layer of global dependency modeling module.
8. The method as described in claim 7, characterized in that, The operations of forward temporal convolution and backward temporal convolution are as follows: in, For activation function, Indicates the first The layer bidirectional temporal convolutional network module in the first layer The input features at the given time and past time steps, where d*k represents the time step offset of the dilated convolution. Indicates the first The layer bidirectional temporal convolutional network module in the first layer Input features of the current time and future time steps; For expansion rate, The kernel size is [size]. This represents the bias term of the forward temporal convolution. This represents the bias term of the inverse temporal convolution. Indicates the first The first layer of forward temporal convolution Each convolutional kernel weight, Indicates the first The first layer of reverse temporal convolution Each convolutional kernel weight, Indicates the forward temporal convolution at the th... Output at any moment Indicates the inverse temporal convolution at the th... Output at any given moment.
9. The method as described in claim 1, characterized in that, The feature sequence output from the last layer of the network is linearly predicted to obtain the lifetime prediction value at each time step, denoted as: ,in, This represents the total number of time steps. The mean absolute error is used as the loss function to measure the deviation between the predicted value and the actual lifetime. in, This represents the actual remaining lifetime value at time step t. This is the predicted value at time step t.
10. A rotating machinery life prediction system based on adaptive operating condition and alternating coupled network, characterized in that, It is applicable to the method as described in any one of claims 1-9, comprising: The data acquisition module is used to acquire vibration signal data of rotating mechanical components throughout their entire life cycle under different working conditions, and after preprocessing, construct the original dataset. The multi-frequency domain feature extraction module is used to extract time-domain features, frequency-domain features, and time-frequency-domain features for each original data sample, and construct a multi-domain feature set. The effective feature selection module is used to select effective feature sets from multi-domain feature sets based on three dimensions: relevance, monotonicity, and robustness. The working condition embedding module is used to introduce an adaptive working condition embedding vector and add each effective feature in the effective feature set to the working condition embedding vector row by row to obtain a coupled feature set. The deep network module is used to build a lifetime prediction model that includes a global dependency modeling module and a bidirectional temporal convolutional network module. The global dependency modeling module and the bidirectional temporal convolutional network module are stacked alternately to form a multi-layer network structure. The training module is used to train the life prediction model using the coupled feature set. In each training round, the global dependency modeling module in the first layer of the network structure extracts the global dependency features of the coupled features. The bidirectional temporal convolutional network module performs forward and backward temporal convolution on the global dependency features to obtain local temporal features. The weighted global dependency features and local temporal features of the current layer are used as the input of the next layer of the network structure. The prediction module is used to predict the lifespan of rotating machinery components based on a trained lifespan prediction model.