A method and device for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding
Through the methods of frequency domain information cropping and multi-scale patch embedding, combined with the neural network model of self-attention mechanism, the problems of complexity and computational complexity of equipment degradation process are solved, and more efficient equipment residual service life prediction is achieved.
Patent Information
- Application Number
- CN202410843154.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-27
- Publication Date
- 2024-12-24
- Estimated Expiration
- 2044-06-27
AI Technical Summary
The existing equipment residual service life prediction methods are difficult to accurately capture the complexity and uncertainty of the equipment degradation process, the calculation complexity is high, and feature engineering depends on expert knowledge, low computing efficiency, and time-frequency domain fusion leads to information redundancy.
Using a method based on frequency domain information clipping and multi-scale patch embedding, the model's flexibility and computing efficiency of feature expression are improved through the random selection and reselecting of Fourier components and the neural network model combined with the self-attention mechanism.
It improves the accuracy and reliability of equipment residual service life prediction, reduces the computational complexity, and makes the method easier to apply in practice.
Smart Images
Figure CN118656623B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method and a device for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding, and belongs to the technical field of equipment health management and remaining useful life prediction. Background Art
[0002] As a core component of the national defense industry, equipment is widely used in many fields such as aviation, navigation, and aerospace. In addition, they also play a vital role in high-tech fields such as electronic warfare, missile defense systems, unmanned vehicles, satellite communications, and deep-sea exploration. The performance stability and reliability of equipment directly affect the operating efficiency and safety of the equipment. With the development of intelligent manufacturing technology, higher requirements are placed on real-time monitoring and predictive maintenance of equipment status. During the operation of equipment, due to factors such as wear, fatigue, and overload, the performance of the equipment gradually degrades, which may eventually lead to failure. Therefore, accurate prediction of the remaining service life of equipment is of great significance for achieving timely maintenance and replacement, reducing unexpected downtime, reducing maintenance costs, and improving production efficiency. Existing remaining service life prediction methods have some limitations. Traditional methods mostly rely on empirical models or simple statistical analysis, which makes it difficult to accurately capture the complexity and uncertainty of the equipment degradation process. In addition, with the continuous changes in the operating conditions of the equipment and the influence of the external environment, the degradation process often presents nonlinear and random characteristics, which further increases the difficulty of remaining service life prediction. The high computational complexity of existing models and the high requirements for equipment have also become obstacles to their widespread application. Therefore, there is an urgent need for a more advanced and effective technical means to improve the accuracy and reliability of equipment remaining service life prediction.
[0003] In recent years, prediction models based on deep learning have gradually become a hot topic of research due to their powerful data processing and modeling capabilities. By analyzing a large amount of historical data, deep learning technology can structure the operating data and facilitate deep knowledge discovery. By using a data model based on a multi-layer neural network to process and learn this data, it can efficiently identify and analyze the complex patterns of equipment health status. This technology can automatically learn multi-level feature representations from vibration signals and other sensor data, thereby accurately capturing key information about the equipment degradation process.
[0004] At present, in the study of applying deep learning technology to predict the remaining service life, researchers generally adopt three main methods according to different processing methods of the original vibration signal: first, the method based on feature engineering, which relies on the knowledge and experience of domain experts to manually extract features that are helpful for prediction from the original data and use these features as the input of the model; second, the end-to-end deep learning method, which directly inputs the original vibration signal into the model, eliminating the tedious feature engineering steps and allowing the model to learn and extract features by itself; finally, the method of fusing time-frequency domain information, which converts the time-domain vibration signal into the frequency domain through Fourier transform and other technologies, and jointly inputs the model with the time-domain information in order to more comprehensively capture the inherent laws and potential characteristics of the data. The method based on feature engineering relies on expert knowledge and manual extraction process, which is not only time-consuming but also may miss subtle features that are crucial to prediction; although the end-to-end deep learning method can automatically extract features, directly processing the original vibration signal often leads to low computational efficiency, and the model performance is limited when the amount of data is insufficient; the time-frequency domain fusion technology is prone to information redundancy and has high computational complexity when processing complex industrial data. To address these issues, further research and improvement of the performance of methods for remaining useful life prediction are needed. Summary of the invention
[0005] The purpose of the present invention is to provide a method and device for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding, which enhances the model's sensitivity to high-frequency features behind specific events by adjusting the high-frequency components of randomly selected Fourier components. At the same time, the introduction of the multi-scale patch embedding method improves the flexibility of the model in expressing features.
[0006] To achieve the above object, the present invention is implemented by adopting the following technical solutions.
[0007] On the one hand, the present invention provides a method for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding, comprising:
[0008] Acquire the original vibration signal generated by the equipment to be studied during its entire life cycle, and use the traditional feature set to perform traditional time domain feature extraction on the original vibration signal to obtain a traditional time domain feature sequence;
[0009] Performing degradation change processing on the extracted traditional time domain features to fit the degradation trend, obtaining features that conform to the degradation trend, and screening the features that conform to the degradation trend based on monotonicity and trend;
[0010] Performing Fourier transform on the screened time domain feature sequence to obtain a frequency domain information matrix, and randomly selecting frequency domain components in the frequency domain information matrix to obtain a randomly selected Fourier component matrix;
[0011] Reselecting the randomly selected Fourier component matrix, and performing inverse Fourier transform on the reselected frequency domain components to obtain a processed time domain feature sequence;
[0012] The patch lengths of patch branches of different scales are calculated, the processed time domain feature sequence is divided into different patch branches according to the patch lengths of different branches, and the processed time domain feature sequence is input into the constructed remaining useful life prediction model to obtain the remaining useful life prediction result.
[0013] Preferably, the traditional time domain features include: mean, root mean square value, skewness, skewness factor, impulse factor, peak factor, margin factor, logarithmic mean absolute value normalization and energy operator.
[0014] Preferably, the extracting the traditional time domain features is subjected to degradation change processing to fit the degradation trend, obtaining features that conform to the degradation trend, and screening the features that conform to the degradation trend based on monotonicity and trend, including:
[0015] The traditional time domain features are processed using the following degradation change function to obtain a feature matrix with monotonicity and trend,
[0016] ;
[0017] ;
[0018] .
[0019] Preferably, the method of performing Fourier transform on the screened time domain feature sequence to obtain a frequency domain information matrix, and randomly selecting frequency domain components in the frequency domain information matrix to obtain a randomly selected Fourier component matrix comprises:
[0020] .
[0021] Preferably, reselecting the randomly selected Fourier component matrix comprises:
[0022] .
[0023] Preferably, the remaining useful life prediction model adopts a neural network model based on a self-attention mechanism.
[0024] The model includes multiple processing layers. Each layer inputs different sequences corresponding to different patch branches into the corresponding branches. After calculation, flattening and connection by the attention mechanism, the calculation results of different branches are fused to obtain the output result of the current layer as the input of the next layer. The output of the last layer is the final remaining service life prediction result.
[0025] Preferably, the calculating patch lengths of patch branches of different scales includes:
[0026] .
[0027] Preferably, in each processing layer of the neural network model based on the self-attention mechanism,
[0028] Linear element-by-element multiplication is used for encoding to capture the global context, and a linear layer output is used.
[0029] Preferably, in each processing layer of the neural network model based on the self-attention mechanism,
[0030] Flatten the output of each branch into a length of The vector of all branches is merged to obtain the output of this processing layer.
[0031] In a second aspect, the present invention provides a device for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding, which is used to implement the above-mentioned method for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding, and the device comprises:
[0032] The original feature extraction module is used to obtain the original vibration signal generated by the equipment to be studied during its entire life cycle, and perform traditional time domain feature extraction on the original vibration signal using a traditional feature set to obtain a traditional time domain feature sequence;
[0033] A feature screening module, used for performing degradation change processing on the extracted traditional time domain features to fit the degradation trend, obtaining features that conform to the degradation trend, and screening the features that conform to the degradation trend based on monotonicity and trend;
[0034] A feature selection module is used to perform Fourier transform on the screened time domain feature sequence to obtain a frequency domain information matrix, and to randomly select frequency domain components in the frequency domain information matrix to obtain a randomly selected Fourier component matrix;
[0035] A feature reselection module, used for reselecting the randomly selected Fourier component matrix and performing inverse Fourier transform on the reselected frequency domain components to obtain a processed time domain feature sequence;
[0036] The prediction module is used to calculate the patch lengths of patch branches of different scales, divide the processed time domain feature sequence into different patch branches according to the patch lengths of different branches, input them into the constructed remaining service life prediction model, and obtain the remaining service life prediction result.
[0037] Compared with the prior art, the present invention has the following beneficial effects:
[0038] The present invention proposes a method for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding. The method introduces a step of reselecting randomly selected Fourier components to enhance the ability to identify specific events. The multi-scale patch embedding strategy effectively makes up for the shortcomings of traditional time series prediction methods in dealing with emergencies. In addition, on the basis of frequency domain information clipping, reselection and multi-scale patch embedding, the present invention further combines a computationally efficient attention mechanism, which not only optimizes the performance of the model, but also reduces the computational complexity to linear time, making the method easier to apply in practical engineering. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 The figure shows a flow chart of a method for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding provided in an embodiment of the present invention;
[0040] Figure 2 FIG. 1 is a schematic diagram of a process of processing and reselecting traditional time domain features in an embodiment of the present invention;
[0041] Figure 3 FIG. 1 is a schematic diagram of a multi-scale patch embedding and model prediction process in an embodiment of the present invention;
[0042] Figure 4 Shown is a schematic diagram of an improved neural network based on the self-attention mechanism in an embodiment of the present invention. DETAILED DESCRIPTION
[0043] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. The embodiments of the present invention and the technical features in the embodiments may be combined with each other unless there is a conflict.
[0044] The term "and / or" is only a description of the association relationship between related objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " generally indicates that the related objects are in an "or" relationship.
[0045] Example 1
[0046] This embodiment introduces a method for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding. Figure 1 ,include:
[0047] Obtain the original vibration signal generated by the equipment to be studied during its entire life cycle, and use the traditional feature set to perform traditional time domain feature extraction on the original vibration signal;
[0048] Perform degradation change processing on the extracted traditional time domain features to fit the degradation trend, obtain features that meet the degradation trend, and screen the features that meet the degradation trend based on monotonicity and trend;
[0049] Performing Fourier transform on the screened time domain feature sequence to obtain a frequency domain information matrix, and performing random selection on the frequency domain components to obtain a randomly selected Fourier component matrix;
[0050] Reselect the randomly selected Fourier component matrix, and perform inverse Fourier transform on the reselected frequency domain components to obtain the processed time domain feature sequence;
[0051] The patch lengths of patch branches of different scales are calculated, and the processed time domain feature sequence is divided into different patch branches according to the patch lengths of different branches. The processed time domain feature sequence is input into the constructed remaining useful life prediction model to obtain the remaining useful life prediction result.
[0052] In this embodiment, multiple sensors located at different positions of the equipment during its entire life cycle collect vibration signals continuously generated by the equipment. These signals contain rich equipment degradation information, and the rationality of using this information is related to the accuracy of the remaining service life prediction. In this embodiment, the vibration signal collected by one sensor is used as the original vibration signal for analysis and processing.
[0053] In this embodiment, trigonometric function changes or statistical features are extracted from the irregular original vibration signal to obtain traditional time domain features of the vibration signal including mean, root mean square value, skewness, skewness factor, pulse factor, peak factor, margin factor, logarithmic mean absolute value normalization and energy operator.
[0054] It should be noted that the above-mentioned traditional time domain feature extraction is implemented using existing methods, and the specific extraction process will not be repeated here.
[0055] In this embodiment, the extracted traditional time domain features are subjected to degradation change processing to fit the degradation trend, and features that conform to the degradation trend are obtained. Figure 2 ,include:
[0056] The degradation change function converts all features into corresponding forms through a function, and uses this function to make the features themselves monotonic and trendy to characterize the degradation process. The degradation function is shown in formula (1).
[0057] (1)
[0058] .
[0059] After the extracted traditional time domain features are processed by degradation change function, the monotonicity and trend of the features are estimated to screen the features that meet the degradation trend, as follows:
[0060] Monotonicity reflects the rising and falling trend of the feature within the period, which is expressed as follows:
[0061] (2)
[0062] .
[0063] The relationship between the trend description feature and the running time is expressed as follows:
[0064] (3)
[0065] .
[0066] Monotonicity reflects the increase or decrease trend of the feature. M=1 and M=0 respectively indicate a highly monotonic trend and a non-monotonic trend of the feature. Trendiness indicates that the degradation state changes with the operation of the event. -1 is strictly decreasing, and 1 is strictly increasing.
[0067] In this embodiment, the screening process actually eliminates the features where M=0 is within 0.2 and T=0 is within 0.1. Here, 0.2 and 0.1 are different empirical values in different experimental scenarios and are not limited to this. Overall, the features where M approaches 0 and T approaches 0 need to be eliminated.
[0068] In this embodiment, the selected traditional time domain feature sequence is Fourier transformed to obtain a frequency domain information matrix, the frequency domain components are randomly selected to obtain a randomly selected Fourier component matrix, and the randomly selected Fourier component matrix is reselected, and the reselected frequency domain components are inversely Fourier transformed to obtain a processed time domain time series; see Figure 2 The specific implementation process is as follows:
[0069] .
[0070] In this embodiment, a reselection mechanism based on frequency clipping is adopted to randomly select a fixed number of Fourier components including low-frequency and high-frequency components and bias towards high-frequency components, as follows:
[0071] Sort the components in the frequency domain information matrix A;
[0072] Calculate according to the following formula , remove the most frequent ones from the selected set and the original set After components, re-add the previous The high frequency components,
[0073] ;
[0074] in is a hyperparameter, indicating the high frequency ratio, and its range is .
[0075] It should be known to those skilled in the art that the matrix after fast Fourier transform processing has a frequency order from low to high; and the original set is the frequency domain matrix formed after fast Fourier transform, that is, matrix A; the selected set is the set obtained after random selection in the previous step, that is, frequency domain components; the candidate set refers to the part removed from the original set. For example, if the number of the original set is 10, 5 need to be selected from the selection set, and 5 have been selected by random selection in the previous step. If the number of reselections is 2, the highest two are discarded from the already selected ones, and then the highest two are taken from the original set to be the candidate set, and then the 2 candidates are added to the selection set.
[0076] It should be noted that reselection is not added during the random selection process, but removed during the re-screening process. The components act as mid-frequency components in the final random result. Compared with high-frequency and low-frequency components, mid-frequency components have less information, which can further shield the interference of this part and reduce the amount of calculation. If the reselection process is combined with random selection, the ability to capture periodic and specific events may be reduced due to the large number of mid-frequency components and insufficient low-frequency and high-frequency components.
[0077] In this embodiment, the patch lengths of patch branches of different scales are calculated and determined, and the processed time domain time series are divided into different branch inputs according to the patch lengths of different branches, and input into the remaining service life prediction model, see Figure 3 The specific implementation process is as follows:
[0078] make Indicates the patch branch The patch size is the length of the time series encoded at one time, represents the step length between patches, that is, the length of the non-overlapping area between two adjacent patch sequences. Indicates the patch branch The number of patches generated, using the sliding window method, the number of patches generated by any patch branch is shown in formula (4),
[0079] (4)
[0080] in, To represent the input sequence length of the previous processing layer, Indicates rounding down;
[0081] After multi-scale segmentation, different patch branch inputs of the neural network model based on the self-attention mechanism are obtained, which can be expressed as:
[0082] (5)
[0083] .
[0084] It is also the first The output of the patch is expressed as follows:
[0085] (6)
[0086] .
[0087] In this embodiment, the remaining useful life prediction model adopts a neural network model based on a self-attention mechanism.
[0088] See also Figure 3 ,In this embodiment, the neural network model based on the ,selfattention mechanism includes multiple processing layers, each layer inputs different sequences corresponding to different ,patch branches into the corresponding branches, and the calculation results of different branches are ,fused through calculation, flattening, and connection by a computationally efficient ,attention mechanism, to obtain the output result of the current layer, which is used as the input of the next layer, and the output of the last layer obtains the final ,remaining service life prediction result.
[0089] .
[0090] The final prediction value is the processing layer The output value of is formula (9), where represents the predicted value of each processing layer, is the predicted length.
[0091] (7)
[0092] (8)
[0093] (9)
[0094] As a preferred embodiment, see Figure 4 In one embodiment of the present invention, an improved neural network model based on the self-attention mechanism is adopted to effectively replace the traditional quadratic matrix multiplication operation by linear element-by-element multiplication, and to use linear layers to replace key-value interactions, while ensuring the accuracy of prediction. Figure 4 The triangle symbol represents element-by-element multiplication, which is implemented as follows:
[0095] ,
[0096] (10)
[0097] and The attention scores between each pair of tokens in are calculated using the dot product operation, and each score is normalized and finally combined with the matrix The multiplication operation finally gives the weighted output. In this process, it is not difficult to obtain the event complexity of the self-attention mechanism as ,in is the number of tags. As the number of tags increases, The computation and memory requirements grow quadratically, making it impractical to process long sequences in real time.
[0098] The additive attention mechanism used in natural language processing captures the global context by exploiting pairwise interactions between tokens through element-wise multiplication rather than dot product operations.
[0099] .
[0100] (11)
[0101] (12)
[0102] (13)
[0103] .
[0104] Furthermore, the present invention realizes branch fusion by flattening, splicing, and linear transformation of each branch, as follows:
[0105] At the final output of each layer of the model, the labels of all branches will be fused into a single input, which is conducive to sharing information at different patch scales and is beneficial for learning time series. The output of each branch will be flattened to a length of Then all branches will be merged into the vector as shown in equation (14):
[0106] (14)
[0107] ,
[0108] (15),
[0109] .
[0110] Example 2
[0111] Based on the same inventive concept, embodiment 2 of the present invention provides an equipment remaining service life prediction device based on frequency domain information clipping and multi-scale patch embedding, which is used to implement the method of embodiment 1 above, and the device includes:
[0112] The original feature extraction module is used to obtain the original vibration signal generated by the equipment to be studied during its entire life cycle, and perform traditional time domain feature extraction on the original vibration signal using a traditional feature set to obtain a traditional time domain feature sequence;
[0113] A feature screening module, used for performing degradation change processing on the extracted traditional time domain features to fit the degradation trend, obtaining features that conform to the degradation trend, and screening the features that conform to the degradation trend based on monotonicity and trend;
[0114] A feature selection module is used to perform Fourier transform on the screened time domain feature sequence to obtain a frequency domain information matrix, and to randomly select frequency domain components in the frequency domain information matrix to obtain a randomly selected Fourier component matrix;
[0115] A feature reselection module, used for reselecting the randomly selected Fourier component matrix and performing inverse Fourier transform on the reselected frequency domain components to obtain a processed time domain feature sequence;
[0116] The prediction module is used to calculate the patch lengths of patch branches of different scales, divide the processed time domain feature sequence into different patch branches according to the patch lengths of different branches, input them into the constructed remaining service life prediction model, and obtain the remaining service life prediction result.
[0117] It is worth pointing out that the device embodiment corresponds to the above-mentioned method embodiment, and the implementation methods of the above-mentioned method embodiments are all applicable to the device embodiment and can achieve the same or similar technical effects, so they will not be repeated here.
[0118] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0119] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0120] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0121] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0122] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the enlightenment of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the purpose of the present invention and the claims, which all fall within the protection of the present invention.
Claims
1. A method for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding, characterized in that: include: The original vibration signal generated by the equipment to be studied during its entire life cycle is obtained, and a traditional time domain feature extraction is performed on the original vibration signal using a traditional feature set to obtain a traditional time domain feature sequence; the traditional time domain features include: mean value, root mean square value, skewness, skewness factor, pulse factor, peak factor, margin factor, logarithmic mean absolute value normalization, and energy operator; Performing degradation change processing on the extracted traditional time domain features to fit the degradation trend, obtaining features that conform to the degradation trend, and screening the features that conform to the degradation trend based on monotonicity and trend; Performing Fourier transform on the screened time domain feature sequence to obtain a frequency domain information matrix, and randomly selecting frequency domain components in the frequency domain information matrix to obtain a randomly selected Fourier component matrix; Reselecting the randomly selected Fourier component matrix, and performing inverse Fourier transform on the reselected frequency domain components to obtain a processed time domain feature sequence; reselecting the randomly selected Fourier component matrix, including: Sort the frequency domain components in the frequency domain information matrix A; Calculate R according to the following formula, remove the R components with the highest frequency from the s components, and select the R components with the highest frequency from the frequency domain information matrix A and put them into the s components. R = λ·S; Where λ is a hyperparameter, indicating the high-frequency ratio, and its range is [0.0, 1.0]; Calculate the patch lengths of patch branches of different scales, divide the processed time domain feature sequence into different patch branches according to the patch lengths of different branches, input them into the constructed remaining service life prediction model, and obtain the remaining service life prediction result. The calculating patch lengths of patch branches of different scales includes: Among them, J b represents the number of patches generated by patch branch b, d n-1 To represent the input sequence length of the previous processing layer, Indicates rounding down, P b represents the patch size of patch branch b, S b represents the step size between patches; The remaining useful life prediction model adopts a neural network model based on the self-attention mechanism. The model includes multiple processing layers. Each layer inputs different sequences corresponding to different patch branches into the corresponding branches. After calculation, flattening and connection by the attention mechanism, the calculation results of different branches are fused to obtain the output result of the current layer as the input of the next layer. The output of the last layer is the final remaining useful life prediction result.
2. According to claim 1, a method for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding is characterized in that: The step of performing degradation change processing on the extracted traditional time domain features to fit the degradation trend, obtaining features that conform to the degradation trend, and screening the features that conform to the degradation trend based on monotonicity and trend, includes: The traditional time domain features are processed using the following degradation change function to obtain a feature matrix with monotonicity and trend, The degradation change function is: Among them, n is the calculation window length, f j (i) represents the i-th feature value in the j-th traditional time-domain feature, j = 1, 2, ..., J, i = 1, 2, ..., n, J is the number of feature types, d j (i) is f j (i) The corresponding difference, f n (X) is the result of the traditional time domain feature after being processed by the degradation change function, and X represents the matrix processed under the calculation window of n; Monotonicity is calculated as follows: Where M(X) is the monotonicity of the matrix X, Indicates the changing trend of the difference value. Indicates that the difference value is greater than 0, which is the number of ascending sequence pairs in the adjacent sequence. Indicates that the difference value is less than 0, which is the number of descending sequence pairs in the adjacent sequence; The trend is calculated as follows: Among them, x n represents the eigenvalue after being processed by the degenerate change function, y n Indicates time value, T is trend, the range is [-1,1], -1 means the eigenvalue is strictly decreasing, 1 means the eigenvalue is strictly increasing; Eliminate the eigenvalues where M(X)=0+floating value and T=0+floating value, where the floating value is selected based on experience.
3. The method for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding according to claim 2, characterized in that: The method of performing Fourier transform on the screened time domain feature sequence to obtain a frequency domain information matrix, and randomly selecting frequency domain components in the frequency domain information matrix to obtain a randomly selected Fourier component matrix includes: The filtered time domain feature sequence is described as X1(t),…,X m (t), m is the dimension of the filtered time domain feature sequence; Perform a fast Fourier transform on one of the sequences to get a Fourier vector: a i,s ,s=1,2,…,d is the Fourier component, d is the number of Fourier components; All time domain feature sequences are subjected to fast Fourier transform to obtain a frequency domain information matrix A: By randomly selecting s components from the frequency domain information matrix, a randomly selected Fourier component matrix is obtained.
4. The method for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding according to claim 3 is characterized in that: In each processing layer of the neural network model based on the self-attention mechanism, Linear element-by-element multiplication is used for encoding to capture the global context, and a linear layer output is used.
5. The method for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding according to claim 4, characterized in that: In each processing layer of the neural network model based on the self-attention mechanism, Flatten the output of each branch into a length J b ·P b The vector of all branches is merged to obtain the output of this processing layer.
6. A device for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding, characterized in that: The device is used to implement the method for predicting the remaining useful life of equipment based on frequency domain information clipping and multi-scale patch embedding as described in any one of claims 1 to 5, the device comprising: The original feature extraction module is used to obtain the original vibration signal generated by the equipment to be studied during its entire life cycle, and perform traditional time domain feature extraction on the original vibration signal using a traditional feature set to obtain a traditional time domain feature sequence; A feature screening module, used for performing degradation change processing on the extracted traditional time domain features to fit the degradation trend, obtaining features that conform to the degradation trend, and screening the features that conform to the degradation trend based on monotonicity and trend; A feature selection module is used to perform Fourier transform on the screened time domain feature sequence to obtain a frequency domain information matrix, and to randomly select frequency domain components in the frequency domain information matrix to obtain a randomly selected Fourier component matrix; A feature reselection module, used to reselect the randomly selected Fourier component matrix, and perform inverse Fourier transform on the reselected frequency domain components to obtain a processed time domain feature sequence; The prediction module is used to calculate the patch lengths of patch branches of different scales, divide the processed time domain feature sequence into different patch branches according to the patch lengths of different branches, input them into the constructed remaining service life prediction model, and obtain the remaining service life prediction result.