Gear residual life prediction method based on multi-feature fusion and Informer model

Abstract

The invention discloses a gear remaining life prediction method based on multi-feature fusion and an Informer model, and belongs to the technical field of mechanical equipment fault prediction and health management. The method comprises the following steps: firstly, acquiring data to obtain a full-life-cycle vibration signal of a gear; then, extracting an initial feature set capable of representing performance degradation from multiple dimensions of a time domain, a frequency domain and a time-frequency domain; then, a feature fusion method based on kernel principal component analysis is adopted to carry out dimension reduction and fusion on the high-dimensional initial features, and single, sensitive and monotonous health indexes are constructed; and finally, the constructed health index sequence is used as input, and the residual service life of the gear is directly predicted by using the strong long sequence time sequence prediction capability of the Informer model. According to the method, the characterization capability of health indexes is effectively improved through multi-feature fusion, the limitation of a traditional recurrent neural network in long sequence prediction is overcome by using the Informer model, and more accurate and more stable prediction of the gear remaining life is realized.

Description

Technical Field

[0001] This invention belongs to the field of PHM technology for mechanical equipment, and in particular, it is a method for predicting the remaining life of gears based on multi-feature fusion and the Informer model. Background Technology

[0002] As a crucial transmission component in rotating machinery, gears' operating condition directly impacts the reliability and safety of the entire equipment. In the complex operating environment of equipment, gears become one of the most easily damaged and least reliable parts in industry. Gear failure can lead to a series of negative consequences, such as prolonged downtime, major accidents, and even personal injury. Accurate prediction of the RUL (Remaining Useful Life) of gears enables condition-based maintenance, preventing catastrophic failures, thus ensuring production safety and reducing operating costs.

[0003] Existing gear lifespan prediction methods are mainly divided into two categories: physical model-based methods and data-driven methods. Physical model-based methods require precise mathematical models to describe the gear degradation process, resulting in complex modeling and poor generalization ability. Data-driven methods directly learn the degradation patterns of equipment from monitoring data and are currently a research hotspot. Data-driven methods model degradation characteristics based on historical sensor data and then map these degradation features to the remaining service life (RUL) value of the gear. Traditional data-driven methods, such as Support Vector Machines (SVM) and Artificial Neural Networks (ANN), employ shallow machine learning models. These methods require manual feature selection and are highly dependent on expert knowledge. Furthermore, with the advent of the industrial age, the amount of data used in algorithms is increasing dramatically, and shallow machine learning algorithms have limited representational capabilities, making them unsuitable for predicting gear lifespan under complex operating conditions.

[0004] Deep learning can automatically learn hierarchical feature representations from raw or simply preprocessed data, greatly reducing reliance on manual feature engineering. Deep learning-based RUL prediction methods include Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs) and their variants (LSTM, GRU), Transformers and their variants (Informer), etc. Deep learning algorithms can automatically extract deep feature representations from historical data, effectively overcoming the problem of unknown model degradation, and therefore have been widely researched and applied in the field of mechanical prediction. CNNs can automatically learn local degradation features in images, such as the frequency energy changes of gear fault features. RNNs are specifically designed for processing time series data, but they suffer from the problem of vanishing or exploding gradients during long-sequence training, and their inherent sequential computational structure is difficult to parallelize, resulting in limited efficiency and accuracy in processing long-sequence data. LSTMs, through their gating mechanism, effectively capture long-term dependencies in data, making them very suitable for learning the degradation trends of objects, and are one of the most commonly used deep learning models for RUL prediction. The Transformer model based on the self-attention mechanism shows great potential in the field of time series prediction; its parallel processing capability and ability to model ultra-long sequences are expected to solve some of the bottlenecks of RNNs. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the existing technology by providing a gear remaining life prediction method based on multi-feature fusion and the Informer model. This method involves normalizing the acquired raw time-series vibration signal of the rolling bearing and then extracting multi-dimensional features in the time domain, frequency domain, and time-frequency domain. All features are combined to form a high-dimensional feature set. KPCA (Kernel Principal Components Analysis) is then used for nonlinear dimensionality reduction and fusion to construct a better health index. Finally, the powerful long-sequence prediction capability of the Informer model is utilized to achieve a more accurate and stable prediction of the gear's remaining life, thus solving the problem of gear life prediction and intelligent decision-making during operation.

[0006] The technical solution to achieve the objective of this invention is as follows: On the one hand, a method for predicting the remaining life of gears based on multi-feature fusion and the Informer model is provided, the method comprising the following steps:

[0007] Step 1, Data Acquisition: Collect historical gear life monitoring data;

[0008] Step 2, Data Preprocessing: Clean and standardize the collected data;

[0009] Step 3, Multi-dimensional feature extraction: For the preprocessed data, extract multiple features in the time domain, frequency domain, and time-frequency domain respectively, and combine them to obtain a high-dimensional feature set, denoted as the high-dimensional initial feature matrix;

[0010] Step 4, Health Indicator Construction: The initial feature matrix is ​​nonlinearly fused and dimensionality reduced using kernel principal component analysis, and the first principal component is selected as the health indicator HI.

[0011] Step 5, Dataset Construction and Splitting: Construct the health indicator HI sequence into a supervised learning sequence dataset, and split it into training set, validation set and test set;

[0012] Step 6, Gear RUL Prediction Model Establishment: Establish an Infomer model using an Encoder-Decoder architecture to predict gear RUL;

[0013] Step 7, Model Training: Using the training set data, with MSE as the loss function, update the Informer model parameters through the Adam optimizer;

[0014] Step 8, Gear RUL Prediction: For the gear to be predicted, gear RUL prediction is performed through multi-feature fusion and the pre-trained Informer model in the above process.

[0015] Furthermore, the historical gear life monitoring data in step 1 includes time-series vibration signal monitoring data, operating condition record data, and actual gear life; among which, the actual gear life is used as the data label.

[0016] Furthermore, step 2 specifically includes:

[0017] The vibration signal of each time window is read, its effective value is calculated, and extreme anomalies are corrected or eliminated.

[0018] The Z-Score normalization method was used to process all vibration signals within each time window.

[0019] Furthermore, in step 3, features are extracted from the time domain, frequency domain, and time-frequency domain for the vibration signal of each preprocessed time window. The time domain features include: root square, peak value, kurtosis, skewness, waveform factor, and impulse factor. The frequency domain features include: spectral centroid, spectral variance, and spectral kurtosis obtained by fast Fourier transform. The time-frequency domain features include: the energy of each sub-band obtained by wavelet packet transform decomposition, and its energy entropy is calculated.

[0020] Furthermore, step 4 specifically includes the following processes:

[0021] Step 4.1: Standardize the initial feature matrix;

[0022] Step 4.2: Select the radial basis function as the kernel function for kernel principal component analysis, and map the standardized features to a high-dimensional feature space;

[0023] Step 4.3: Perform principal component analysis in the high-dimensional feature space to calculate the eigenvalues ​​and eigenvectors of the kernel matrix;

[0024] Step 4.4: Select the first principal component with the largest contribution rate and use its value as the health indicator HI after fusion.

[0025] Furthermore, in step 5, a supervised learning sequence dataset is constructed using the sliding window method. A continuous sequence of health indicators is used as the input feature, and the health indicator value at a future point in time or a future sequence is used as the prediction target.

[0026] Furthermore, in step 6, the Infomer model introduces a multi-head sparse self-attention mechanism, a distillation structure, and a parallel generative decoder. Specifically, the Infomer model includes:

[0027] The input layer receives feature vectors from the training and test sets and performs input embedding and positional encoding.

[0028] The encoder is used to perform deep feature extraction on the long sequence of input health indicators. It consists of three identical encoder layers stacked together. Each encoder layer contains an attention block and a distillation structure. The distillation structure is a downsampling module composed of a one-dimensional convolutional layer and a max pooling layer.

[0029] The decoder receives the encoder's output and a known partial target sequence to generate the final remaining lifetime prediction sequence; it includes a masked sparse self-attention block and two stacked identical multi-head attention blocks.

[0030] The output layer is a fully connected layer used to output the health index HI representing the gear.

[0031] On the other hand, a gear remaining life prediction system is provided, the system comprising:

[0032] The first module is used to collect historical gear life monitoring data.

[0033] The second module is used to clean and standardize the collected data.

[0034] The third module is used to: extract multiple features in the time domain, frequency domain, and time-frequency domain from the preprocessed data, and combine them to obtain a high-dimensional feature set, denoted as the high-dimensional initial feature matrix;

[0035] The fourth module is used to: perform nonlinear fusion and dimensionality reduction on the initial feature matrix using kernel principal component analysis, and select the first principal component as the health indicator HI;

[0036] The fifth module is used to: construct the health indicator HI sequence into a supervised learning sequence dataset, and divide it into training set, validation set and test set;

[0037] The sixth module is used to implement: building an Informer model that uses an Encoder-Decoder architecture to predict gear RUL;

[0038] The seventh module is used to update the Infomer model parameters using the training set data, with MSE as the loss function, and the Adam optimizer.

[0039] The eighth module is used to perform gear RUL prediction by fusing multiple features from the above process and using the pre-trained Informer model for the gear to be predicted.

[0040] On the other hand, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the gear remaining life prediction method based on multi-feature fusion and the Informer model.

[0041] On the other hand, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the gear remaining life prediction method based on multi-feature fusion and the Informer model.

[0042] Compared with the prior art, the significant advantages of this invention are:

[0043] (1) This invention is true for rolling bearings The label values ​​were normalized, and the bearing was calculated. The percentage of the value is used as the output for model training, with the percentage of degradation serving as the output. Compared to direct calculation, this provides a more accurate representation of the true value of rolling bearings. Normalizing the label values ​​reduces the difficulty of training the network.

[0044] (2) This invention introduces the Informer model into the field of gear RUL prediction for the first time. The model’s unique ProbSparseSelf-Attention mechanism and Distilling structure enable it to efficiently capture long-term dependencies in long-sequence historical data and directly perform long sequence prediction, effectively solving the computational bottleneck and performance degradation problem of traditional RNN / LSTM models in long sequence processing, and improving prediction accuracy and efficiency at the same time.

[0045] (3) This invention provides a complete automated process from raw data to the final RUL prediction value, reducing reliance on expert experience, making it highly replicable and easy to integrate into existing equipment predictive maintenance platforms, thus possessing high engineering application value. In practical applications, it can effectively monitor the working status of gears, predict the RUL value of gears in a timely and accurate manner, ensure that mechanical equipment can operate safely, reliably, and efficiently, and extend the working cycle of the equipment.

[0046] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a gear remaining life prediction method based on multi-feature fusion and the Informer model in one embodiment.

[0048] Figure 2 This is a flowchart illustrating the process of multi-dimensional feature extraction and health indicator construction in one embodiment.

[0049] Figure 3 This is an overall framework diagram of the Informer model in one embodiment.

[0050] Figure 4 This is a schematic diagram illustrating the implementation of the Informer model in one embodiment. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0052] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0053] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0054] In one embodiment, combined Figure 1 This paper provides a method for predicting the remaining life of gears based on multi-feature fusion and the Informer model. The method includes the following steps:

[0055] Step 1, Data Acquisition: Collect historical gear life monitoring data;

[0056] Step 2, Data Preprocessing: Clean and standardize the collected data;

[0057] Step 3, Multi-dimensional feature extraction: For the preprocessed data, extract multiple features in the time domain, frequency domain, and time-frequency domain respectively, and combine them to obtain a high-dimensional feature set, denoted as the high-dimensional initial feature matrix;

[0058] Step 4, Health Indicator Construction: The initial feature matrix is ​​nonlinearly fused and dimensionality reduced using kernel principal component analysis, and the first principal component is selected as the health indicator HI.

[0059] Step 5, Dataset Construction and Splitting: Construct the health indicator HI sequence into a supervised learning sequence dataset, and split it into training set, validation set and test set;

[0060] Step 6, Gear RUL Prediction Model Establishment: Establish an Infomer model using an Encoder-Decoder architecture to predict gear RUL;

[0061] Step 7, Model Training: Using the training set data, with MSE as the loss function, update the Informer model parameters through the Adam optimizer;

[0062] Step 8, Gear RUL Prediction: For the gear to be predicted, gear RUL prediction is performed through multi-feature fusion and the pre-trained Informer model in the above process.

[0063] Furthermore, in one embodiment, the historical gear life monitoring data in step 1 includes time-series vibration signal monitoring data, operating condition record data, and actual gear life; wherein, the actual gear life is used as a data label.

[0064] Among them, the timing vibration signal monitoring data refers to the vibration data of the gear's X, Y, and Z axes; the operating condition record data includes load size, torque, speed, lubrication status, and ambient temperature.

[0065] In step 1, N vibration data files are collected, recording the vibration signals throughout the entire process from normal operation to failure. The recording length (time window) of each file is [length missing]. There are n data points.

[0066] Furthermore, in one embodiment, step 2 specifically includes:

[0067] The vibration signal of each time window is read, its effective value is calculated, and extreme anomalies are corrected or eliminated.

[0068] The Z-Score normalization method was used to process all vibration signals within each time window.

[0069] Preferably, the Z-Score normalization method is used to process the n vibration signals within each time window, and the calculation formula is as follows:

[0070]

[0071] In the formula, For the first A standardized set of data, For the first A vibration signal, This is the mean of all data within that time window. This represents the standard deviation of all data within that time window.

[0072] Furthermore, in one embodiment, combined with Figure 2 In step 3, features are extracted from the time domain, frequency domain, and time-frequency domain for the vibration signal of each preprocessed time window. The time domain features include: root square, peak value, kurtosis, skewness, waveform factor, and impulse factor. The frequency domain features include: spectral centroid, spectral variance, and spectral kurtosis obtained by fast Fourier transform. The time-frequency domain features include: the energy of each sub-band obtained by wavelet packet transform decomposition and its energy entropy calculated.

[0073] The specific process includes:

[0074] Step 3.1, extract time-domain features; calculate features that reflect the statistical characteristics of the vibration signal amplitude. These features are sensitive to wear-related faults and include six features: root mean square (RMS), peak value, kurtosis, skewness, shape factor, and impulse factor. Specifically:

[0075] Root Mean Square Peak value ; kurtosis ,in The standardized mean of the data for each time window. Standard deviation of the standardized data; skewness Waveform factor ; Pulse factor .

[0076] Step 3.2, extract frequency domain features; first, perform a Fast Fourier Transform (FFT) on the vibration signal to convert the signal from the time domain to the frequency domain, then calculate the features that reveal the frequency component changes caused by gear faults, including three features: spectral centroid (FC), spectral variance (VF), and spectral kurtosis (FK). Specifically:

[0077] The signal is transformed to the frequency domain using FFT to obtain the spectrum of the vibration signal for each time window. ,in For frequency lines, It is the first The complex amplitude corresponding to the root frequency line. Its amplitude.

[0078] Then, the frequency domain characteristics of each time window are calculated, namely: spectral centroid ,in For the first The frequency value corresponding to the root frequency line; spectral variance Spectral kurtosis ,in The mean of the spectral amplitude. denoted as the standard deviation of the spectral amplitude.

[0079] Step 3.3: Extract time-frequency domain features; use wavelet packet transform (WPT) to process the vibration signal, select appropriate wavelet basis functions to perform L-level decomposition of the vibration signal, and obtain... Each sub-band. Calculate the energy of each sub-band. Furthermore, the wavelet packet energy entropy (WPEE) is calculated. Specifically:

[0080] The vibration signal for each time window was decomposed into L levels using WPT to obtain... Each frequency band Corresponding to a set of wavelet packet coefficients ( , To calculate the wavelet packet energy entropy for each time window (where the number of coefficients in the frequency band is given), the process is as follows: First, calculate the total energy of each frequency band. Next, calculate the total energy of the entire signal. Then calculate the energy percentage of each frequency band. The calculated wavelet packet energy entropy .

[0081] Furthermore, in one embodiment, step 4 specifically includes the following process:

[0082] Repeat step 3 for all N files collected in step 1 to obtain the feature matrix. .

[0083] Step 4.1: Standardize the initial feature matrix;

[0084] Step 4.2: Select the radial basis function as the kernel function for kernel principal component analysis, and map the standardized features to a high-dimensional feature space;

[0085] Step 4.3: Perform principal component analysis in the high-dimensional feature space to calculate the eigenvalues ​​and eigenvectors of the kernel matrix;

[0086] Step 4.4: Select the first principal component with the largest contribution rate and use its value as the health indicator HI after fusion.

[0087] Furthermore, in one embodiment, step 5 employs a sliding window method to construct a supervised learning sequence dataset, using a continuous sequence of health indicators as input features and a future health indicator value or a future sequence as the prediction target. Specifically:

[0088] Set the input sequence length Given length and predicted sequence length , co-generated The samples were divided into training, validation and test sets in a ratio of 7:2:1.

[0089] Furthermore, in one embodiment, combined with Figure 3 and Figure 4In step 6, the Infomer model introduces a multi-head probSparse self-attention mechanism, a distillation structure, and a parallel generative decoder. Specifically, the Infomer model includes:

[0090] The input layer receives feature vectors from the training and test sets and performs input embedding and positional encoding.

[0091] The encoder is used to perform deep feature extraction on the long sequence of input health indicators. It consists of three identical encoder layers stacked together. Each encoder layer contains an attention block and a distillation structure. The distillation structure is a downsampling module composed of a one-dimensional convolutional layer (Conv1D) and a max pooling layer (MaxPool).

[0092] The decoder receives the encoder's output and a known partial target sequence to generate the final remaining lifetime prediction sequence; it includes a masked probSparse self-attention block and two stacked identical multi-head attention blocks.

[0093] The output layer is a fully connected layer used to output the health index HI representing the gear.

[0094] Step 6: Establishing the gear RUL prediction model. The specific process includes:

[0095] Step 6.1, Input Embedding and Position Encoding. Original Input Data The shape is Historical HI sequences, in which This indicates the number of samples in a batch. =1 indicates the number of HIs. To adapt to the Conv1D input format, first... Transformation dimension is After passing through the Conv1D layer, the output shape is The data is then converted back to time series format. Then, position encoding is performed using standard sine and cosine functions, as shown in the following formula:

[0096]

[0097] in It is the time step position. It's a dimension index. The generated position encoding matrix. The shape is This is then directly added to the input sequence after convolutional embedding to obtain the final encoder input. .

[0098] Step 6.2: The encoder consists of three identical encoder layers stacked together, with each encoder layer distilling the sequence length. Input data The shape is After passing through the ProbSparse Self-Attention block of the first encoder, the output shape remains unchanged. It then undergoes a downsampling module (containing one Conv1D layer and one MaxPool layer) for distillation, halving the length of the output sequence. This is followed by the second and third encoder layers, where the data undergoes the same operation as the first encoder, ultimately resulting in a sequence with the shape... The output data.

[0099] Step 6.3: The decoder consists of one Masked ProbSparse Self-Attention layer and two identical multi-head Attention Blocks stacked together. It receives two inputs: the output of the encoder and the input of the decoder itself.

[0100] Decoder's own input The shape is After passing through the Masked ProbSparse Self-Attention layer, the shape of the output remains unchanged. Then, a two-layer multi-head attention calculation is performed between the Key and Value matrices from the encoder and the Query matrix from the decoder, ultimately yielding a shape of... The output.

[0101] Step 6.4, Output Projection. This involves projecting the output of the decoder. The final predicted value can be obtained by passing each time step through a simple linear layer.

[0102] Furthermore, in one embodiment, step 7, model training, refers to updating model parameters using the training set data and the mean squared error (MSE) as the loss function, performed by the Adam optimizer. During training, a validation set is used to monitor loss changes, early stopping is employed to prevent overfitting, and hyperparameters (such as learning rate, number of attention heads, number of encoder / decoder layers, etc.) are optimized. Specifically:

[0103] Input the HI sequence into the gear RUL prediction model to obtain the predicted value. Then, the loss MSE is calculated to compare the predicted and actual values, and finally, the Adam optimizer is used to update all weight parameters of the model.

[0104] Furthermore, in one embodiment, step 8, gear RUL prediction, refers to inputting the health indicator sequence of the test set into the trained optimal Informer model, and the model directly outputs predictions for the future. A step-by-step HI prediction sequence. A failure threshold for HI is set based on historical data or expert experience. When the predicted HI sequence exceeds the threshold for the first time, the difference between the predicted time point and the current time point is the predicted gear RUL.

[0105] In one embodiment, a gear remaining life prediction system is provided, the system comprising:

[0106] The first module is used to collect historical gear life monitoring data.

[0107] The second module is used to clean and standardize the collected data.

[0108] The third module is used to: extract multiple features in the time domain, frequency domain, and time-frequency domain from the preprocessed data, and combine them to obtain a high-dimensional feature set, denoted as the high-dimensional initial feature matrix;

[0109] The fourth module is used to: perform nonlinear fusion and dimensionality reduction on the initial feature matrix using kernel principal component analysis, and select the first principal component as the health indicator HI;

[0110] The fifth module is used to: construct the health indicator HI sequence into a supervised learning sequence dataset, and divide it into training set, validation set and test set;

[0111] The sixth module is used to implement: building an Informer model that uses an Encoder-Decoder architecture to predict gear RUL;

[0112] The seventh module is used to update the Infomer model parameters using the training set data, with MSE as the loss function, and the Adam optimizer.

[0113] The eighth module is used to perform gear RUL prediction by fusing multiple features from the above process and using the pre-trained Informer model for the gear to be predicted.

[0114] Specific limitations regarding the gear remaining life prediction system based on multi-feature fusion and the Informer model can be found in the limitations of the gear remaining life prediction method based on multi-feature fusion and the Informer model mentioned above, and will not be repeated here. Each module in the aforementioned gear remaining life prediction system based on multi-feature fusion and the Informer model can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0115] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements:

[0116] Step 1, Data Acquisition: Collect historical gear life monitoring data;

[0117] Step 2, Data Preprocessing: Clean and standardize the collected data;

[0118] Step 3, Multi-dimensional feature extraction: For the preprocessed data, extract multiple features in the time domain, frequency domain, and time-frequency domain respectively, and combine them to obtain a high-dimensional feature set, denoted as the high-dimensional initial feature matrix;

[0119] Step 4, Health Indicator Construction: The initial feature matrix is ​​nonlinearly fused and dimensionality reduced using kernel principal component analysis, and the first principal component is selected as the health indicator HI.

[0120] Step 5, Dataset Construction and Splitting: Construct the health indicator HI sequence into a supervised learning sequence dataset, and split it into training set, validation set and test set;

[0121] Step 6, Gear RUL Prediction Model Establishment: Establish an Infomer model using an Encoder-Decoder architecture to predict gear RUL;

[0122] Step 7, Model Training: Using the training set data, with MSE as the loss function, update the Informer model parameters through the Adam optimizer;

[0123] Step 8, Gear RUL Prediction: For the gear to be predicted, gear RUL prediction is performed through multi-feature fusion and the pre-trained Informer model in the above process.

[0124] For specific limitations on each step, please refer to the limitations on the gear remaining life prediction method based on multi-feature fusion and Informer model mentioned above, which will not be repeated here.

[0125] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being implemented when executed by a processor:

[0126] Step 1, Data Acquisition: Collect historical gear life monitoring data;

[0127] Step 2, Data Preprocessing: Clean and standardize the collected data;

[0128] Step 3, Multi-dimensional feature extraction: For the preprocessed data, extract multiple features in the time domain, frequency domain, and time-frequency domain respectively, and combine them to obtain a high-dimensional feature set, denoted as the high-dimensional initial feature matrix;

[0129] Step 4, Health Indicator Construction: The initial feature matrix is ​​nonlinearly fused and dimensionality reduced using kernel principal component analysis, and the first principal component is selected as the health indicator HI.

[0130] Step 5, Dataset Construction and Splitting: Construct the health indicator HI sequence into a supervised learning sequence dataset, and split it into training set, validation set and test set;

[0131] Step 6, Gear RUL Prediction Model Establishment: Establish an Infomer model using an Encoder-Decoder architecture to predict gear RUL;

[0132] Step 7, Model Training: Using the training set data, with MSE as the loss function, update the Informer model parameters through the Adam optimizer;

[0133] Step 8, Gear RUL Prediction: For the gear to be predicted, gear RUL prediction is performed through multi-feature fusion and the pre-trained Informer model in the above process.

[0134] For specific limitations on each step, please refer to the limitations on the gear remaining life prediction method based on multi-feature fusion and Informer model mentioned above, which will not be repeated here.

[0135] Step 1, Data Acquisition: Collect historical gear life monitoring data;

[0136] Step 2, Data Preprocessing: Clean and standardize the collected data;

[0137] Step 3, Multi-dimensional feature extraction: For the preprocessed data, extract multiple features in the time domain, frequency domain, and time-frequency domain respectively, and combine them to obtain a high-dimensional feature set, denoted as the high-dimensional initial feature matrix;

[0138] Step 4, Health Indicator Construction: The initial feature matrix is ​​nonlinearly fused and dimensionality reduced using kernel principal component analysis, and the first principal component is selected as the health indicator HI.

[0139] Step 5, Dataset Construction and Splitting: Construct the health indicator HI sequence into a supervised learning sequence dataset, and split it into training set, validation set and test set;

[0140] Step 6, Gear RUL Prediction Model Establishment: Establish an Infomer model using an Encoder-Decoder architecture to predict gear RUL;

[0141] Step 7, Model Training: Using the training set data, with MSE as the loss function, update the Informer model parameters through the Adam optimizer;

[0142] Step 8, Gear RUL Prediction: For the gear to be predicted, gear RUL prediction is performed through multi-feature fusion and the pre-trained Informer model in the above process.

[0143] For specific limitations on each step, please refer to the limitations on the gear remaining life prediction method based on multi-feature fusion and Informer model mentioned above, which will not be repeated here.

[0144] This invention effectively enhances the representation ability of health indicators through multi-feature fusion and overcomes the limitations of traditional recurrent neural networks in long sequence prediction by utilizing the Informer model, thus achieving more accurate and stable prediction of the remaining life of gears.

[0145] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.

Claims

1. A method for predicting the remaining life of gears based on multi-feature fusion and the Informer model, characterized in that, The method includes the following steps: Step 1, Data Acquisition: Collect historical gear life monitoring data; Step 2, Data Preprocessing: Clean and standardize the collected data; Step 3, Multi-dimensional feature extraction: For the preprocessed data, extract multiple features in the time domain, frequency domain, and time-frequency domain respectively, and combine them to obtain a high-dimensional feature set, denoted as the high-dimensional initial feature matrix; Step 4, Health Indicator Construction: The initial feature matrix is ​​nonlinearly fused and dimensionality reduced using kernel principal component analysis, and the first principal component is selected as the health indicator HI. Step 5, Dataset Construction and Splitting: Construct the health indicator HI sequence into a supervised learning sequence dataset, and split it into training set, validation set and test set; Step 6, Gear RUL Prediction Model Establishment: Establish an Infomer model using an Encoder-Decoder architecture to predict gear RUL; Step 7, Model Training: Using the training set data, with MSE as the loss function, update the Informer model parameters through the Adam optimizer; Step 8, Gear RUL Prediction: For the gear to be predicted, gear RUL prediction is performed through multi-feature fusion and the pre-trained Informer model in the above process.

2. The gear remaining life prediction method based on multi-feature fusion and Informer model according to claim 1, characterized in that, The historical gear life monitoring data in step 1 includes time-series vibration signal monitoring data, operating condition record data, and actual gear life; among them, the actual gear life is used as the data label.

3. The gear remaining life prediction method based on multi-feature fusion and Informer model according to claim 1, characterized in that, Step 2 specifically includes: The vibration signal of each time window is read, its effective value is calculated, and extreme anomalies are corrected or eliminated. The Z-Score normalization method was used to process all vibration signals within each time window.

4. The gear remaining life prediction method based on multi-feature fusion and Informer model according to claim 3, characterized in that, In step 3, features are extracted from the time domain, frequency domain, and time-frequency domain for the vibration signal of each preprocessed time window. The time-domain features include: root square, peak value, kurtosis, skewness, waveform factor, and impulse factor; the frequency-domain features include: spectral centroid, spectral variance, and spectral kurtosis obtained through fast Fourier transform; and the time-frequency-domain features include: obtaining the energy of each sub-band through wavelet packet transform decomposition and calculating its energy entropy.

5. The gear remaining life prediction method based on multi-feature fusion and Informer model according to claim 1, characterized in that, Step 4 includes the following specific steps: Step 4.1: Standardize the initial feature matrix; Step 4.2: Select the radial basis function as the kernel function for kernel principal component analysis, and map the standardized features to a high-dimensional feature space; Step 4.3: Perform principal component analysis in the high-dimensional feature space to calculate the eigenvalues ​​and eigenvectors of the kernel matrix; Step 4.4: Select the first principal component with the largest contribution rate and use its value as the health indicator HI after fusion.

6. The gear remaining life prediction method based on multi-feature fusion and Informer model according to claim 1, characterized in that, In step 5, a supervised learning sequence dataset is constructed using the sliding window method. A continuous sequence of health indicators is used as the input feature, and the health indicator value at a future time point or a future sequence is used as the prediction target.

7. The gear remaining life prediction method based on multi-feature fusion and Informer model according to claim 1, characterized in that, In step 6, the Infomer model introduces a multi-head sparse self-attention mechanism, a distillation structure, and a parallel generative decoder. Specifically, the Infomer model includes: The input layer receives feature vectors from the training and test sets and performs input embedding and positional encoding. The encoder is used to perform deep feature extraction on the long sequence of input health indicators. It consists of three identical encoder layers stacked together. Each encoder layer contains an attention block and a distillation structure. The distillation structure is a downsampling module composed of a one-dimensional convolutional layer and a max pooling layer. The decoder receives the encoder's output and a known partial target sequence to generate the final remaining lifetime prediction sequence; it includes a masked sparse self-attention block and two stacked identical multi-head attention blocks. The output layer is a fully connected layer used to output the health index HI representing the gear.

8. A gear remaining life prediction system based on the method of any one of claims 1 to 7, characterized in that, The system includes: The first module is used to collect historical gear life monitoring data. The second module is used to clean and standardize the collected data. The third module is used to: extract multiple features in the time domain, frequency domain, and time-frequency domain from the preprocessed data, and combine them to obtain a high-dimensional feature set, denoted as the high-dimensional initial feature matrix; The fourth module is used to: perform nonlinear fusion and dimensionality reduction on the initial feature matrix using kernel principal component analysis, and select the first principal component as the health indicator HI; The fifth module is used to: construct the health indicator HI sequence into a supervised learning sequence dataset, and divide it into training set, validation set and test set; The sixth module is used to implement: building an Informer model that uses an Encoder-Decoder architecture to predict gear RUL; The seventh module is used to update the Infomer model parameters using the training set data, with MSE as the loss function, and the Adam optimizer. The eighth module is used to perform gear RUL prediction by fusing multiple features from the above process and using the pre-trained Informer model for the gear to be predicted.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images