Method and system for load bearing nut life prediction based on multi-stage feature extraction

Abstract

The present application relates to a kind of based on multi-stage feature extraction load nut life prediction method and system.The existing prediction method has the problem such as the degradation factor extraction being disturbed.This method obtains load nut full life cycle signal and divides sample point, obtains degradation feature set after time domain feature calculation;Hilbert transform obtains degradation trend curve, calculates demarcation point;Degradation feature set is as the input and output of convolution auto-encoder network, respectively trains according to each degradation stage, obtains multi-stage degradation factor extraction model;The degradation factor and RUL extracted are as the input and output of LSTM network, and load nut life prediction model is obtained.The present application extracts load nut degradation feature in combination with variation modal decomposition and correlation calculation, avoids the degradation factor extraction being disturbed;Division mechanism is combined with convolution auto-encoder, considers the degradation characteristics in different stages, improves the ability of model mining load nut deep degradation feature, and life prediction precision.

Description

Technical Field

[0001] This invention relates to the field of mechanical component condition assessment technology, specifically to a method and system for predicting the life of load-bearing nuts based on multi-stage feature extraction. Background Technology

[0002] Load-bearing nuts are widely used in industrial operations to lift heavy objects through mechanical movement with a lead screw and ensure smooth operation. Nuts are often subjected to long-term heavy loads, which generates fatigue stress during operation. Over time, this stress can lead to damage such as thread wear and breakage. Therefore, predictive maintenance strategies are needed to prevent such failures.

[0003] With the development of artificial intelligence, data-driven methods have been widely applied in industrial production, especially in the field of condition assessment of key components. Such methods do not require in-depth research on mechanisms, have lower costs and stronger transferability. In data processing, there are correlations between features, and feature reduction processing is required to extract degradation factors. After analysis, several aspects affecting degradation factor extraction were found: (1) The features involved in degradation factor extraction determine the upper limit of degradation factor accuracy. Degradation factors are obtained by fusing multi-dimensional features. Among the multi-dimensional features, there are often parts with very low correlation to RUL (Remaining Useful Life). These features are not suitable for degradation factor extraction due to poor monotonicity and other reasons, and will also interfere with the model's extraction of degradation factors. (2) The model is an important factor in the upper limit that degradation factor accuracy can reach. The basic degradation factor extraction model is a shallow learning algorithm or a basic deep learning algorithm. Such methods are limited by their own principles or structures, resulting in unsatisfactory model capabilities. They can only extract shallow features and need to be combined with other algorithms to make up for their defects. (3) The mechanistic characteristics of different degradation stages are different. As the severity of the fault increases, the degradation factors characterizing the degree of degradation exhibit different trends. In cases of complex degradation stages, models extracting a single degradation factor are less effective at describing different stages, making it necessary to construct corresponding models for different degradation stages. Furthermore, the degradation trends of load-bearing nuts differ across operating conditions: bearings also exhibit varying degradation trends under different loads and speeds. Deep learning models trained on data from a single operating condition cannot achieve high RUL prediction accuracy under other operating conditions.

[0004] Therefore, a new method for predicting the life of load-bearing nuts is needed to overcome the problems of the existing methods. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting the life of load-bearing nuts based on multi-stage feature extraction, so as to solve the problems of interference in the extraction of degradation factors, extraction of only shallow features, and neglect of degradation differences at different stages in existing methods.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for predicting the lifespan of load-bearing nuts based on multi-stage feature extraction, the method comprising:

[0008] The vibration signal of the load-bearing nut throughout its entire life cycle is obtained and the sample points are divided to obtain the sample point signal;

[0009] After preprocessing the sample point signals, time-domain feature calculations are performed, and features with high correlation to the degree of degradation are selected to obtain a set of degradation features.

[0010] The degradation trend curve is obtained by performing Hilbert transform on the sample point signal. The boundary points of different degradation stages in the degradation trend curve are calculated. The degradation stages are obtained by dividing the entire life cycle according to the boundary points.

[0011] By fusing convolutional layers and autoencoders, a convolutional autoencoder network is obtained. The set of degradation features is used as the input and output of the convolutional autoencoder network. The network is trained separately according to each degradation stage to obtain a multi-stage degradation factor extraction model.

[0012] The extracted degradation factor and RUL label are used as the input and output of the LSTM network to learn the mapping relationship between the degradation factor and the RUL label, thus obtaining the load-bearing nut life prediction model.

[0013] Furthermore, the vibration signals of the load-bearing nut throughout its entire life cycle are acquired and divided into sample points to obtain sample point signals, including:

[0014] The initial vibration signal of the load-bearing nut is collected by a vibration sensor, and the vibration signal throughout the entire life cycle is divided by setting the number of sample points.

[0015] Define sample point RUL tags, using the start and stop sampling points as the start and complete failure points, respectively, and normalize the RUL tags to the 0-1 range:

[0016]

[0017] In the formula:

[0018] N is the number of sample points;

[0019] i represents the current sample point value;

[0020] y is the current sample's RUL label;

[0021] The sample point signal is obtained.

[0022] Furthermore, the sample point signals are preprocessed, including:

[0023] The signal at each sample point is denoised, including signal decomposition and signal reconstruction.

[0024] Furthermore, temporal feature calculations are performed, and features highly correlated with the degree of degradation are selected to obtain a set of degradation features, including:

[0025] After denoising the sample point signals, the time-domain features of each sample point signal are calculated to obtain the full lifecycle feature set P. N×a :

[0026] P N×a : {p1, p2, ..., p a};

[0027] in:

[0028] p1, p2, ..., p a Features across all time domains throughout the entire lifecycle;

[0029] 'a' represents the number of time-domain features;

[0030] Assessment P N×a The correlation between each temporal feature and the RUL label is used to select the top m temporal features as the filter feature set based on the correlation magnitude.

[0031]

[0032] The selected feature set is used as the degenerate feature set.

[0033] Furthermore, time-domain features include maximum value, minimum value, peak value, peak-to-peak value, absolute mean, root mean square, variance, standard deviation, effective value, kurtosis, skewness, waveform factor, peak factor, impulse factor, boundary factor, and residual factor.

[0034] Furthermore, a Hilbert transform is performed on the sample point signals to obtain a degradation trend curve. The boundary points of different degradation stages in the degradation trend curve are calculated, and the entire life cycle is divided according to the boundary points to obtain each degradation stage, including:

[0035] Perform a Hilbert transform on the sample point signal x(t) to obtain the power spectral density. The set of maximum values ​​Q:

[0036] Q: {q1, q2, ..., q} N};

[0037] in:

[0038] q1, q2, ..., q N This represents the maximum power spectral density of the signal at each sample point;

[0039] Differentiating the fitted curve of Q yields the derivative curve x. psd That is, the degradation trend curve of the load-bearing nut;

[0040] Find x psd The extreme points are used to obtain the boundary points of different degradation stages;

[0041] The entire life cycle is divided into various degradation stages based on the dividing points.

[0042] Furthermore, convolutional layers and autoencoders are fused to obtain a convolutional autoencoder network. The degradation feature set is used as the input and output of the convolutional autoencoder network, and it is trained separately for each degradation stage to obtain a multi-stage degradation factor extraction model, including:

[0043] The degradation feature set data for each degradation stage are enhanced by median interpolation to obtain enhanced data. The degradation feature set data for each degradation stage are time-domain feature values ​​or RUL labels.

[0044] A convolutional autoencoder network is obtained by replacing ordinary neurons in an autoencoder with convolutional layers;

[0045] A multi-stage degradation factor extraction model is obtained by training a convolutional autoencoder network using augmented data.

[0046] Furthermore, the extracted degradation factor and RUL label are used as the input and output of an LSTM network to learn the mapping relationship between the degradation factor and the RUL label, thus obtaining a load-bearing nut life prediction model, including:

[0047] Degradation factors are extracted by using the time-domain features of the test set or real-time signal as input to the multi-stage degradation factor extraction model.

[0048] The extracted degradation factor and RUL label are used as the input and output of the LSTM network for training;

[0049] A bias correction mechanism is introduced during the training process to finally obtain the load-bearing nut life prediction model;

[0050] The degradation factor of the test set or real-time signal is used as input to the load-bearing nut life prediction model to obtain the RUL prediction value.

[0051] On the other hand, a load-bearing nut life prediction system based on multi-stage feature extraction is provided, the system being used to implement the method, including...

[0052] The acquisition module is used to acquire the vibration signal of the load-bearing nut throughout its entire life cycle and divide it into sample points to obtain the sample point signal;

[0053] The degradation feature calculation module is used to preprocess the sample point signals, perform time-domain feature calculation, and filter features that are highly correlated with the degree of degradation to obtain a set of degradation features.

[0054] The stage division module is used to perform Hilbert transform on the sample point signal to obtain the degradation trend curve, calculate the boundary points of different degradation stages in the degradation trend curve, and divide the entire life cycle according to the boundary points to obtain each degradation stage.

[0055] The degradation factor extraction module is used to fuse convolutional layers and autoencoders to obtain a convolutional autoencoder network. The degradation feature set is used as the input and output of the convolutional autoencoder network. The network is trained separately according to each degradation stage to obtain a multi-stage degradation factor extraction model.

[0056] The remaining life prediction module is used to take the extracted degradation factor and RUL label as the input and output of the LSTM network, learn the mapping relationship between the degradation factor and the RUL label, and obtain the load-bearing nut life prediction model.

[0057] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0058] This invention combines variational mode decomposition with correlation calculation to extract the degradation features of load-bearing nuts, avoiding interference with degradation factor extraction, improving the signal-to-noise ratio of vibration signals, and quantifying the degree of degradation of load-bearing nuts by calculating time-domain features. Secondly, it fuses convolutional layers and autoencoder structures to obtain a convolutional autoencoder, and combines the full life cycle partitioning mechanism with the convolutional autoencoder to obtain a multi-stage convolutional autoencoder model, considering the degradation characteristics of different stages, improving the model's ability to mine deep degradation features of load-bearing nuts, and improving the accuracy of degradation factors. In addition, this invention introduces a bias correction mechanism into the loss function of the LSTM network training stage, improving the accuracy of the LSTM network in predicting the RUL of load-bearing nuts under different working conditions. Attached Figure Description

[0059] 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 some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.

[0060] Figure 1 This is a flowchart of the method of the present invention.

[0061] Figure 2This is a system composition diagram of the present invention.

[0062] Figure 3 This is a schematic diagram of the deviation correction mechanism of the present invention.

[0063] Figure 4 This is a comparison chart of the performance of different types of autoencoders in analysis and prediction tasks.

[0064] Figure 5 These are remaining lifetime prediction curves for different types of autoencoders. Detailed Implementation

[0065] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0066] It should be noted that similar reference numerals and letters indicate similar items; therefore, once an item is defined in one embodiment, it does not need to be further defined and explained in subsequent embodiments. Furthermore, the terms "comprising" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0067] It should also be noted that although the order of steps is mentioned in the method description, in some cases, steps may be performed in a different order than that described here, and this should not be interpreted as a restriction on the order of steps.

[0068] This invention provides a method for predicting the remaining service life of load-bearing nuts based on multi-stage feature extraction. Targeting the characteristics of the degradation process of load-bearing nuts, it can effectively predict their remaining service life. A full life-cycle segmentation mechanism is introduced into a convolutional autoencoder to extract degradation factors, and an error correction mechanism is incorporated into the training of the RUL prediction model, thereby improving the prediction accuracy of the remaining service life of load-bearing nuts. The method can be executed on a computer system, such as a set of computer-executable instructions. Figure 1 The method includes:

[0069] S1: Acquire the vibration signal of the load-bearing nut throughout its entire life cycle and divide it into sample points to obtain the sample point signal. This includes:

[0070] S101: The initial vibration signal of the load-bearing nut is collected by a vibration sensor, and the vibration signal of the entire life cycle is divided by setting the number of sample points.

[0071] S102: Define the RUL label for sample points. Take the starting acquisition point and the stopping acquisition point as the starting failure point and the complete failure point respectively, and normalize the RUL label to the range of 0 - 1:

[0072]

[0073] In the formula:

[0074] N is the number of sample points;

[0075] i is the position value of the current sample point;

[0076] y is the RUL label of the current sample. The RUL label represents the specific remaining service life value of the load-bearing nut at a certain sample point;

[0077] S103: Obtain the sample point signals.

[0078] S2: After preprocessing the sample point signals, calculate the time-domain features, screen the features with high correlation with the degradation degree, and obtain the degradation feature set. It includes:

[0079] S201: Aiming at the problem of low signal-to-noise ratio of the signals caused by the poor working environment of the load-bearing nut, adopt a signal reconstruction method based on variational mode decomposition to preprocess each sample point signal, that is, denoising processing, including signal decomposition and signal reconstruction:

[0080] ① Initialize the parameters of the VMD (Variational Modal Decomposition) algorithm λ 1 and the value of n are 0;

[0081] ② Let n = n + 1 and start to execute the loop;

[0082] ③ Let k = k + 1. When k < K, update u using the following formula k and ω k , where K is the number of assumed decomposed IMF (intrinsic mode function) components;

[0083]

[0084]

[0085] ④ Update λ by the following formula:

[0086]

[0087] ⑤ Repeat steps ③ and ④. For a given discrimination value ε, until Pause the iteration and obtain the value of the IMF component K.

[0088] S202: After denoising the sample point signals, calculate the time-domain features of each sample point signal to obtain the full lifecycle feature set P. N×a :

[0089] P N×a : {p1, p2, ..., p a};

[0090] in:

[0091] p1, p2, ..., p a Features across all time domains throughout the entire lifecycle;

[0092] 'a' represents the number of time-domain features.

[0093] The time-domain features include maximum value, minimum value, peak value, peak-to-peak value, absolute mean, root mean square, variance, standard deviation, effective value, kurtosis, skewness, waveform factor, peak factor, impulse factor, boundary factor, and residual factor, as shown in Table 1, where x(i) is the sample point signal.

[0094] Table 1 Formulas for Calculating Time-Domain Features

[0095]

[0096] S203: Evaluate P by calculating the Pearson correlation coefficient between each time-domain feature and the RUL label. N×a The correlation between each temporal feature and the RUL label was analyzed, and the top m temporal features with a correlation coefficient greater than 0.55 were selected as the filter feature set.

[0097]

[0098] The selected feature set is used as the degenerate feature set.

[0099] The formula for calculating the Pearson correlation coefficient is as follows:

[0100]

[0101] In the formula:

[0102] x represents the feature value of the sample, y represents the RUL label of the sample, and σ X and σ Y Let μ represent the standard deviations of X and Y, respectively. X and μ Y Let X and Y represent the average values, respectively. The closer the absolute value of the correlation coefficient is to 1, the stronger the correlation between the two, and vice versa.

[0103] S3: Perform a Hilbert transform on the sample point signals to obtain the degradation trend curve, calculate the boundary points of different degradation stages in the degradation trend curve, and divide the entire life cycle into various degradation stages based on the boundary points, including:

[0104] S301: Perform a Hilbert transform on the sample point signal x(t) to obtain the power spectral density. The set of maximum values ​​Q:

[0105] Q: {q1, q2, ..., q} N};

[0106] in:

[0107] q1, q2, ..., q N This represents the maximum power spectral density of the signal at each sample point;

[0108] The Hilbert transform formula is as follows:

[0109]

[0110] In the formula:

[0111] x(τ) represents the vibration signal.

[0112] S302: Differentiate the fitted curve of Q to obtain the derivative curve x. psd That is, the degradation trend curve of the load-bearing nut;

[0113] S303: Find x psd The extreme points are used to obtain the boundary points of different degradation stages:

[0114]

[0115] In the formula:

[0116] `stage` represents the current stage number;

[0117] x' PSD Representing curve x PSD The derivative value of .

[0118] S304: The entire life cycle is divided into various degradation stages based on the dividing points.

[0119] S4: Fusing convolutional layers and autoencoders yields a convolutional autoencoder network. The degradation feature set is used as the input and output of the convolutional autoencoder network. Training is performed separately for each degradation stage to obtain a multi-stage degradation factor extraction model, including:

[0120] S401: Enhance the degradation feature set data of each degradation stage by means of median interpolation to obtain enhanced data. The degradation feature set data of each degradation stage is time domain feature value or RUL label.

[0121] The specific calculation method for median interpolation is as follows:

[0122]

[0123] In the formula:

[0124] x i x i+1 Let x represent the temporal feature value or RUL label of the i-th and i+1-th sample points of the initial sample, respectively, and the calculated x (i+1) Replace x i+1 As the value of the (i+1)th sample point, the original x i+1 This is used as the value of the (i+2)th sample point, and so on.

[0125] S402: A convolutional autoencoder network is obtained by replacing ordinary neurons in the autoencoder with convolutional layers.

[0126] The activation function of the convolutional autoencoder network is ReLU. The input and output dimensions of the convolutional autoencoder are the same as the dimension m of the degenerate feature set, and the encoder output dimension and decoder input dimension are the same as the dimension n of the degeneracy factor. The number of convolutional layers, l, is tested to reduce the input and output errors of the convolutional autoencoder without blindly increasing the number of layers. The formula for calculating the kernel size of each layer is as follows:

[0127]

[0128] In the formula:

[0129] k represents the size of the convolutional kernel, m represents the input dimension of the convolutional autoencoder, n represents the degradation factor dimension, and l represents the number of convolutional layers.

[0130] S403: Use augmented data to train a convolutional autoencoder network to obtain a multi-stage degradation factor extraction model.

[0131] The convolutional autoencoder is trained using training set samples and the degradation factor of test set samples is obtained. The number of training iterations is the maximum number of times the loss function does not decrease significantly. The loss function is the root mean square error (RMSE).

[0132] S5: Using the extracted degradation factor and RUL label as the input and output of the LSTM network, the mapping relationship between the degradation factor and the RUL label is learned to obtain the load-bearing nut life prediction model, including:

[0133] S501: Extract degradation factors by using the time-domain features of the test set or real-time signal as input to the multi-stage degradation factor extraction model;

[0134] S502: Extract the degradation factor H:{h1,h2,…,h N} and the RUL tag RUL:{r1,r2,…,r N} are used as the input and output of the LSTM network for training;

[0135] S503: A deviation correction mechanism is introduced during the training process to finally obtain the load-bearing nut life prediction model;

[0136] Deviation correction mechanism such as Figure 3 As shown in the figure, the loss function is as follows, which mitigates the impact of different degradation factors on the training process.

[0137]

[0138] In the formula:

[0139] v p Indicates the predicted value;

[0140] v r_t Represents the actual value at the current moment.

[0141] v r_(t-m) and v r_(t-m) These represent the actual values ​​at the previous and next time points, respectively.

[0142] S504: Use the degradation factor of the test set or real-time signal as input to the load-bearing nut life prediction model to obtain the RUL prediction value.

[0143] This invention combines variational mode decomposition with correlation calculation to extract the degradation features of load-bearing nuts, improving the signal-to-noise ratio of vibration signals and quantifying the degree of degradation of load-bearing nuts by calculating time-domain features. Secondly, it fuses convolutional layers and autoencoder structures to obtain a convolutional autoencoder, and combines the full life cycle partitioning mechanism with the convolutional autoencoder to obtain a multi-stage convolutional autoencoder model, improving the model's ability to mine deep degradation features of load-bearing nuts and improving the accuracy of degradation factors. In addition, this invention introduces a bias correction mechanism into the loss function of the LSTM network training stage, improving the accuracy of the LSTM network in predicting the RUL of load-bearing nuts under different working conditions.

[0144] In addition, this invention also provides a load-bearing nut life prediction system based on multi-stage feature extraction, for implementing the above method, such as... Figure 2 The system includes:

[0145] The acquisition module is used to acquire the vibration signal of the load-bearing nut throughout its entire life cycle and divide it into sample points to obtain the sample point signal, which corresponds to S1 of the above method;

[0146] The degradation feature calculation module is used to preprocess the sample point signals, perform time-domain feature calculation, filter features that are highly correlated with the degree of degradation, and obtain a set of degradation features, corresponding to S2 of the above method;

[0147] The stage division module is used to perform Hilbert transform on the sample point signal to obtain the degradation trend curve, calculate the boundary points of different degradation stages in the degradation trend curve, and divide the entire life cycle according to the boundary points to obtain each degradation stage, which corresponds to S3 of the above method.

[0148] The degradation factor extraction module is used to fuse convolutional layers and autoencoders to obtain a convolutional autoencoder network. The degradation feature set is used as the input and output of the convolutional autoencoder network. The network is trained separately according to each degradation stage to obtain a multi-stage degradation factor extraction model, which corresponds to S4 of the above method.

[0149] The remaining life prediction module is used to take the extracted degradation factor and RUL label as the input and output of the LSTM network, learn the mapping relationship between the degradation factor and the RUL label, and obtain the load-bearing nut life prediction model, which corresponds to S5 of the above method.

[0150] Figure 4 The comparison of the performance of different types of autoencoders in analysis and prediction tasks includes autoencoders, convolutional autoencoders, and multi-stage convolutional autoencoders. The multi-stage convolutional autoencoder refers to the multi-stage degradation factor extraction model of this invention. Figure 5 The remaining lifetime prediction curves for different types of autoencoders are shown. After comparison, the prediction accuracy and analytical capabilities of different types of autoencoders show the following differences:

[0151] (1) Autoencoder:

[0152] Prediction performance: Autoencoders reconstruct the input data by compressing it into a low-dimensional representation. They are able to capture the main features of the data, but in complex prediction tasks, they may ignore the details and temporal relationships of the data.

[0153] Accuracy: The accuracy of the prediction results is relatively low because the autoencoder focuses primarily on reconstructing the input and may not be able to fully capture key patterns and variations in the data.

[0154] (2) Convolutional autoencoder:

[0155] Prediction Performance: Convolutional autoencoders, building upon autoencoders, introduce convolutional layers to handle localized features of data. This allows them to better capture spatial relationships and patterns within the data, resulting in better performance for analysis and prediction tasks involving sequence data types.

[0156] Accuracy: The accuracy of the prediction results is relatively improved because the convolutional autoencoder can better handle the spatial features and variation patterns of the data.

[0157] (3) Multi-stage convolutional autoencoder:

[0158] Prediction Performance: Multi-stage convolutional autoencoders further introduce multi-stage processing, progressively extracting higher-level features from the data. This hierarchical feature extraction approach can better capture the abstract patterns and temporal relationships of the data.

[0159] Accuracy: The prediction results are relatively more accurate because the multi-stage convolutional autoencoder processes the data layer by layer in multiple stages, which can more comprehensively analyze the characteristics and changes of the data.

[0160] Experimental analysis shows that multi-stage convolutional autoencoders, compared to traditional autoencoders, exhibit higher accuracy and stronger data analysis capabilities in analytical and predictive tasks. They effectively capture key features, spatial patterns, and temporal relationships in the data, thus providing more reliable prediction results. This hierarchical, convolutional processing approach has advantages in handling complex data and predictive tasks, and can better guide decision-making and resource allocation.

[0161] Those skilled in the art will understand that all or part of the functions of the embodiments of the present invention can be implemented by hardware or by computer program. When all or part of the functions in the above embodiments are implemented by computer program, the program can be stored in a computer-readable storage medium, which may include: read-only memory, random access memory, disk, optical disk, hard disk, etc., and the program is executed by a computer to achieve the above functions. For example, the program can be stored in the memory of a device, and when the program in the memory is executed by the processor, all or part of the above functions can be achieved. In addition, when all or part of the functions in the above embodiments are implemented by computer program, the program can also be stored in a storage medium such as a server, another computer, disk, optical disk, flash drive, or portable hard drive, and can be downloaded or copied to the memory of a local device, or the system of the local device can be updated. When the program in the memory is executed by the processor, all or part of the functions in the above embodiments can be achieved.

[0162] The above examples illustrate the present invention only to aid in understanding it and are not intended to limit the scope of the invention. Those skilled in the art can make various simple deductions, modifications, or substitutions based on the principles of this invention.

Claims

1. A method for predicting the life of load-bearing nuts based on multi-stage feature extraction, characterized in that: The method includes: The vibration signal of the load-bearing nut throughout its entire life cycle is obtained and the sample points are divided to obtain the sample point signal; After preprocessing the sample point signals, time-domain feature calculations are performed, and features with high correlation to the degree of degradation are selected to obtain a set of degradation features. The degradation trend curve is obtained by performing Hilbert transform on the sample point signal. The boundary points of different degradation stages in the degradation trend curve are calculated. The degradation stages are obtained by dividing the entire life cycle according to the boundary points. By fusing convolutional layers and autoencoders, a convolutional autoencoder network is obtained. The set of degradation features is used as the input and output of the convolutional autoencoder network. The network is trained separately according to each degradation stage to obtain a multi-stage degradation factor extraction model. The extracted degradation factor and RUL label are used as the input and output of the LSTM network to learn the mapping relationship between the degradation factor and the RUL label, thus obtaining the load-bearing nut life prediction model. in: Vibration signals of the load-bearing nut throughout its entire life cycle are acquired and divided into sample points to obtain sample point signals, including: The initial vibration signal of the load-bearing nut is collected by a vibration sensor, and the vibration signal throughout the entire life cycle is divided by setting the number of sample points. Define sample point RUL tags, using the start and stop sampling points as the start and complete failure points, respectively, and normalize the RUL tags to the 0-1 range: ； In the formula: This represents the number of sample points. This represents the current sample point value. The current sample's RUL label; The sample point signal is obtained.

2. The method for predicting the life of a load-bearing nut based on multi-stage feature extraction according to claim 1, characterized in that: Preprocessing of the sample point signals includes: The signal at each sample point is denoised, including signal decomposition and signal reconstruction.

3. The method for predicting the life of a load-bearing nut based on multi-stage feature extraction according to claim 2, characterized in that: Temporal feature calculations are performed, and features highly correlated with the degree of degradation are selected to obtain a set of degradation features, including: After denoising the sample point signals, the time-domain features of each sample point signal are calculated to obtain the full lifecycle feature set. : ； in: Features across all time domains throughout the entire lifecycle; The number of time-domain features; Evaluate The correlation between each temporal feature and the RUL label is used to select the top m temporal features as the filter feature set based on the correlation magnitude. : ； The selected feature set is used as the degenerate feature set.

4. The method for predicting the life of a load-bearing nut based on multi-stage feature extraction as described in claim 3, Its features are: Time-domain characteristics include maximum value, minimum value, peak value, peak-to-peak value, absolute mean, root mean square, variance, standard deviation, effective value, kurtosis, skewness, waveform factor, peak factor, impulse factor, boundary factor, and residual factor.

5. The method for predicting the life of a load-bearing nut based on multi-stage feature extraction according to claim 4, characterized in that: The degradation trend curve is obtained by performing a Hilbert transform on the sample point signal. The boundary points of different degradation stages in the degradation trend curve are calculated. Based on the boundary points, the entire life cycle is divided into various degradation stages, including: For sample point signals Perform a Hilbert transform to obtain the power spectral density. The set of maximum values : ； in: This represents the maximum power spectral density of the signal at each sample point; right Differentiating the fitted curve yields the derivative curve. That is, the degradation trend curve of the load-bearing nut; beg The extreme points are used to obtain the boundary points of different degradation stages; The entire life cycle is divided into various degradation stages based on the dividing points.

6. The method for predicting the life of a load-bearing nut based on multi-stage feature extraction according to claim 5, characterized in that: By fusing convolutional layers and autoencoders, a convolutional autoencoder network is obtained. The degradation feature set is used as the input and output of the convolutional autoencoder network, and it is trained separately for each degradation stage to obtain a multi-stage degradation factor extraction model, including: The degradation feature set data for each degradation stage are enhanced by median interpolation to obtain enhanced data. The degradation feature set data for each degradation stage are time-domain feature values ​​or RUL labels. A convolutional autoencoder network is obtained by replacing ordinary neurons in an autoencoder with convolutional layers; A multi-stage degradation factor extraction model is obtained by training a convolutional autoencoder network using augmented data.

7. The method for predicting the life of a load-bearing nut based on multi-stage feature extraction according to claim 6, characterized in that: The extracted degradation factor and RUL label are used as the input and output of an LSTM network to learn the mapping relationship between the degradation factor and the RUL label, thus obtaining a load-bearing nut life prediction model, including: Degradation factors are extracted by using the time-domain features of the test set or real-time signal as input to the multi-stage degradation factor extraction model. The extracted degradation factor and RUL label are used as the input and output of the LSTM network for training; A bias correction mechanism is introduced during the training process to finally obtain the load-bearing nut life prediction model; The degradation factor of the test set or real-time signal is used as input to the load-bearing nut life prediction model to obtain the RUL prediction value.

8. A load-bearing nut life prediction system based on multi-stage feature extraction, characterized in that: The system is used to implement the method according to any one of claims 1-7, including The acquisition module is used to acquire the vibration signal of the load-bearing nut throughout its entire life cycle and divide it into sample points to obtain the sample point signal; The degradation feature calculation module is used to preprocess the sample point signals, perform time-domain feature calculation, and filter features that are highly correlated with the degree of degradation to obtain a set of degradation features. The stage division module is used to perform Hilbert transform on the sample point signal to obtain the degradation trend curve, calculate the boundary points of different degradation stages in the degradation trend curve, and divide the entire life cycle according to the boundary points to obtain each degradation stage. The degradation factor extraction module is used to fuse convolutional layers and autoencoders to obtain a convolutional autoencoder network. The degradation feature set is used as the input and output of the convolutional autoencoder network. The network is trained separately according to each degradation stage to obtain a multi-stage degradation factor extraction model. The remaining life prediction module is used to take the extracted degradation factor and RUL label as the input and output of the LSTM network, learn the mapping relationship between the degradation factor and the RUL label, and obtain the load-bearing nut life prediction model.

Images