Method for predicting residual life of mechanical equipment based on CDDKD framework
Through the method based on the CDDKD framework and information decoupling network model, the problems of distribution differences and real-time monitoring in mechanical equipment life prediction are solved, real-time accurate residual life prediction of mechanical equipment is achieved, and prediction accuracy and robustness are improved.
Patent Information
- Application Number
- CN202411808345.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-10
- Publication Date
- 2025-04-25
AI Technical Summary
When the service life prediction method of existing mechanical equipment is changed, the difference in the distribution of test data and training data leads to a decrease in accuracy, and real-time online monitoring cannot be achieved.
The residual life prediction method of mechanical equipment based on the CDDKD framework is adopted, and the residual life of mechanical equipment is predicted by constructing a common distribution difference knowledge distillation framework and information decoupling network model, and using convolutional neural network and high-order moment distance, real-time accurate prediction of the residual life of mechanical equipment is achieved.
This method can realize the residual life prediction of cross-working conditions of mechanical equipment under complex variable working conditions, improve prediction accuracy and robustness, and meet the needs of real-time online monitoring in industrial real-time.
Smart Images
Figure CN119884733A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of mechanical equipment life prediction, and relates to a method for predicting the remaining life of mechanical equipment based on a CDDKD framework. Background Art
[0002] Mechanical equipment plays a vital role in industry and is the core foundation of productivity. The safe operation of mechanical equipment can significantly increase the speed of industrial production and reduce labor costs. However, when mechanical equipment fails, it will not only cause production line shutdowns and economic losses, but may also cause safety accidents and pose a threat to personnel safety. By predicting the remaining service life of mechanical equipment in advance, maintenance and replacement can be carried out in a timely manner to prevent unexpected failures, ensure the continuity and safety of production, and extend the life of the equipment, thereby improving overall production efficiency. Due to changes in the service conditions of mechanical equipment, there will be distribution differences between test data and training data, resulting in a decrease in the accuracy of existing mechanical equipment life prediction methods. In addition, considering the actual industrial requirements of real-time online monitoring, existing prediction methods cannot be achieved. Therefore, a new mechanical equipment life prediction method is urgently needed to solve the above problems. Summary of the invention
[0003] In view of this, the purpose of the present invention is to provide a method for predicting the remaining life of mechanical equipment based on the CDDKD framework, so as to realize real-time and accurate prediction of the remaining life of mechanical equipment.
[0004] In order to achieve the above object, the present invention provides the following technical solutions:
[0005] A method for predicting the remaining life of mechanical equipment based on the CDDKD framework specifically includes the following steps:
[0006] S1: Collect the original vibration signals of mechanical equipment under different working conditions, and then form the full life cycle samples based on the collected original vibration signals;
[0007] S2: Construct a common distribution discrepancy knowledge distilling (CDDKD) framework based on the task to be implemented;
[0008] S3: Based on the CDDKD framework and convolutional neural network, an information decoupling network (IDN) model is constructed;
[0009] S4: Input sufficient sample data from heterogeneous mechanical equipment into the IDN model, and then use the prediction loss L P , noise constraint loss L o, distribution difference loss L based on high-order moment distance (HMD) distance dd , reconstruction loss L r and the domain confusion loss L based on HMD distance dc Iteratively train the weight parameters of the constructed IDN model;
[0010] S5: Inherit the weight parameters of the feature extractor in step S4, and then use the prediction loss L P , noise constraint loss L o , distribution difference loss L based on high-order moment (HMD) distance dd and the reconstruction loss L r Input the restricted sample data of the target mechanical equipment into the IDN model for fine-tuning training;
[0011] S6: After multiple iterations of training, observe whether the loss curve tends to converge smoothly. If the curve converges, the IDN model training is completed. The trained IDN model will be used to predict the remaining life of mechanical equipment under cross-operating condition migration under monitoring data.
[0012] Furthermore, in step S2, the constructed CDDKD framework is specifically: first, the distribution difference knowledge in heterogeneous mechanical equipment with sufficient sample data is distilled and extracted, and then according to the similar action mechanism of distribution differences in different mechanical systems or components (load changes, speed changes, etc.), the obtained distribution difference knowledge is used for the remaining life prediction of target mechanical equipment with scarce data, thereby enhancing the domain invariance of the trend degradation characteristics of the target mechanical equipment under complex and variable working conditions.
[0013] Further, in step S3, the constructed IDN model includes a feature extractor G fe (θ fe ), feature decoupler G fd (θ fd ), Remaining useful life predictor G p (θ p ) and feature reconstructor G fr (θ fr ), where θ fe ,θ fd ,θ p and θ fr Respectively represent the corresponding network weight parameters; feature extractor G fe Used to obtain the most critical relationship representation and extract coupling features F in the CDDKD framework c ; Feature decoupler G fd Used to decouple the degradation trend feature F dt and the degradation noise feature F dn ; Feature Reconstructor Gfr For the degenerate trend feature F dt and the degradation noise feature F dn Recovering the extracted coupling features; Remaining useful life predictor G p For outputting the final remaining service life.
[0014] Furthermore, in step S3, the optimization objectives of the IDN model include the following three parts: 1) learning a feature decoupled representation to extract fe The obtained coupling characteristics F c Separate the degradation trend feature F dt and the degradation noise feature F dn ; 2) By jointly training the feature decoupler G fd and feature reconstructor G fr , maintain the degradation trend characteristic F dt and the degradation noise feature F dn semantic consistency to avoid feature information distortion; 3) Degenerate trend feature F from any two domains dt Distribution alignment of,learning differential relation representations.
[0015] Furthermore, in step S3, feature decoupling specifically includes: first, by predicting the loss L P Train the feature extractor G with mean square error fe , feature decoupler G fd and the remaining life predictor G p , in order to achieve accurate predictions:
[0016]
[0017] Where Y i Represents the original data sample X i The actual remaining useful life label; represents the predicted value of the remaining life of the i-th sample, which can be calculated as follows:
[0018]
[0019] Then, fix G fe and G p The weight parameter of the noise constraint loss L o Train the feature disentangler G separately fd :
[0020]
[0021] In the formula, represents the degradation noise characteristics of the i-th sample, u dn represents the feature mean, and udn It can be calculated as follows:
[0022]
[0023] In the formula, n represents the number of samples;
[0024] Finally, the degradation trend characteristic F is obtained through equations (1) and (3): dt and the degradation noise feature F dn .
[0025] Further, in step S3, maintaining semantic consistency specifically includes: using a distribution difference loss L based on high-order moment (HMD) distance dd To increase the degradation trend characteristic F dt and the degradation noise feature F dn The difference between:
[0026]
[0027] In the formula, represents the probability expectation of the degradation trend characteristics, represents the probability expectation of the degraded noise feature; φ(·) is the kernel function in RKHS, RKHS represents the reproducing kernel Hilbert space, and the mapping function represents the tensor product operation; ⊙ represents the Hadamard product operation; F dt and F dn It can be calculated by the following equation:
[0028]
[0029] In the formula, and Respectively represent the degradation trend characteristics and degradation noise characteristics of the i-th sample;
[0030] On the other hand, in order to prevent and The semantic information of fd and feature reconstructor G fr Considered as an autoencoder structure, the reconstruction loss L of the mean square error is r To restore the coupling feature:
[0031]
[0032] in, represents the expected probability of the sample.
[0033] Furthermore, in step S3, learning the difference relationship specifically includes: striving to achieve distribution alignment between the inherent degradation trend features from any two domains, using a domain confusion loss L based on the high-order moment (HMD) distance dc :
[0034]
[0035] In the formula, and Indicates that they come from the i-th domain and the jth domain The degradation trend characteristics of
[0036] In Q domains, K domain confusion losses will be obtained
[0037]
[0038] In the formula, represents the number of combinations of taking 2 from Q;
[0039] In the distribution alignment process, larger domain confusion losses are assigned greater weights; the weight assignment strategy is constructed as follows:
[0040]
[0041] In the formula, τ is the temperature factor; the smaller the τ value, the closer the The larger the difference is; finally, the weight w j Act on the corresponding Update network parameters:
[0042]
[0043] In the formula, denotes the corrected domain confusion loss.
[0044] The beneficial effects of the present invention are as follows: the present invention proposes a method for predicting the remaining life of mechanical equipment based on the common distribution difference knowledge distillation (CDDKD) framework, which can realize real-time and accurate prediction of the remaining life of mechanical equipment. The method includes a high-order moment (HMD) distance for improving the sensitivity of the distribution difference metric; an information decoupled network (IDN) basic network for improving the model's learning ability for the invariant features of the true fault domain.
[0045] Other advantages, objectives and features of the present invention will be described in the following description to some extent, and to some extent, will be obvious to those skilled in the art based on the following examination and study, or can be taught from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below in conjunction with the accompanying drawings, wherein:
[0047] Figure 1 This is a schematic diagram of the CDDKD framework principle;
[0048] Figure 2 This is the IDN model structure diagram. DETAILED DESCRIPTION
[0049] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.
[0050] See also Figure 1-2 , this paper proposes a method for predicting the remaining life of mechanical equipment based on the CDDKD framework. The method includes a HMD distance for improving the sensitivity of the distribution difference metric; in order to improve the model's ability to learn the invariant features of the true fault domain, an IDN model is proposed.
[0051] 1. CDDKD framework
[0052] The schematic diagram of the CDDKD framework is as follows: Figure 1 As shown in the figure, the framework first distills and extracts the distribution difference knowledge from heterogeneous mechanical equipment with sufficient sample data, and then uses the obtained distribution difference knowledge to predict the remaining life of target mechanical equipment with scarce data based on the similar action mechanism of distribution difference in different mechanical systems or components (load change, speed change, etc.), thereby strengthening the domain invariance of the trend degradation characteristics of target mechanical equipment under complex variable working conditions. This framework can effectively solve the current problems caused by the lack of full life cycle data samples, incomplete working condition information, and high heterogeneity of monitoring data.
[0053] 2. HMD distance
[0054] The distribution difference metric is the basis of the entire knowledge transfer framework. It not only guides the formulation and optimization of the transfer strategy, but also ensures the effective transfer of knowledge in the target domain and the improvement of model performance. The maximum mean difference (MMD) utilizes the excellent representation ability of distribution difference in the reproducing kernel Hilbert space (RKHS) and has become the most commonly used distance metric in the distribution alignment mechanism. Its mathematical expression is:
[0055]
[0056] Where φ(·) is the kernel function in RKHS, which can map the data samples in low-dimensional space to high-dimensional space and enhance the difference representation between the source domain and the target domain; and Represent the source domain and target domain data samples respectively; n S and n T Represent the number of source domain and target domain samples respectively.
[0057] Although MMD has been verified in many scenarios, its essence still relies on mean statistics to measure the differences in data distribution across domains. However, the distribution characteristics of actual monitored vibration signals are symmetrical along the x-axis, that is, the mean is always equal to zero. Therefore, the difference representation capability of MMD needs to be further improved. One intuition is to take the absolute value of the vibration signal. To ensure the feasibility of optimization, square operations are used instead of absolute value operations; then considering the rich statistical information contained in the high-order moments, a p-order HMD distance is defined:
[0058]
[0059] It should be noted that in order to reduce the amount of calculation, the Hadamard product ⊙ operation is used instead of the ordinary inner product operation. However, formula (2) only improves the ability to express the distribution difference in the original sample space, ignoring the high-dimensional reproducing kernel Hilbert space (RKHS). In order to ensure the regeneration characteristics of RKHS, the tensor product operation is performed in RKHS. Redefine equation (2) as follows:
[0060]
[0061] According to the tensor product theorem of Hilbert space, we know that the mapping function The computational properties of the kernel function are still satisfied. Here, the reason why the mapping feature of RKHS does not perform high-order moment operations is that the computational complexity of the tensor product is much higher than that of the Hadamard product.
[0062] 3. IDN Model
[0063] In engineering practice, vibration monitoring signals collected from mechanical equipment often contain a large number of interference components from external factors such as environmental noise and operating conditions. This makes the original data complex and difficult to analyze. In severe cases, it may even mask the inherent degradation characteristics, posing a huge challenge to the subsequent remaining life prediction. To this end, an IDN model with strong information decoupling capability is proposed to separate and filter out interference features in heterogeneous mechanical equipment and extract common distribution difference knowledge. Figure 2 As shown in Figure 2, the architecture of the IDN model consists of four modules: feature extractor G fe (θ fe ), feature decoupler G fd (θ fd ), Remaining useful life predictor G p (θ p ) and feature reconstructor G fr (θ fr ), where θ fe ,θ fd ,θ p and θ fr Respectively represent the corresponding network weight parameters. Feature extractor G fe It is used to obtain the most critical relationship representation and extract coupling features F in the CDDKD framework. c ; Feature decoupler G fd Used to decouple the degradation trend feature F dt and the degradation noise feature F dn ; Feature Reconstructor G fr For the degenerate trend feature F dt and the degradation noise feature F dn Recovering the extracted coupling features; Remaining useful life predictor G p For outputting the final remaining service life.
[0064] The detailed optimization objectives of the IDN model include the following three parts: 1) learning a feature disentangled representation to extract fe The obtained coupling characteristics F c Separate the degradation trend feature F dt and the degradation noise feature F dn . 2) By jointly training the feature decoupler G fd and feature reconstructor G fr , maintain the degradation trend characteristic F dt and the degradation noise feature F dn 3) Realize the degradation trend feature F from any two domains dt The distribution of is aligned to learn the difference relation representation. The following will describe these points in detail:
[0065] 1) Feature decoupling:
[0066] This part aims to use the feature decoupler G fd To separate the degraded noise features, we can obtain the decoupled representation. First, we predict the loss L P Train the feature extractor G with mean square error fe , feature decoupler G fd and the remaining life predictor G p , in order to achieve accurate predictions:
[0067]
[0068] Where Y i Represents the original data sample X i The real remaining useful life label, It can be calculated as follows:
[0069]
[0070] Then fix G fe and G p The weight parameter is L. o Train the feature disentangler G separately fd :
[0071]
[0072] In the formula, and the characteristic mean u dn It can be calculated as follows:
[0073]
[0074] It can be seen that formula (6) aims to suppress the variance fluctuation in all feature dimensions. Finally, the degradation trend feature F is obtained through formula (4) and formula (6): dt and the degradation noise feature F dn .
[0075] 2) Semantic consistency:
[0076] In order to further improve the characteristic variables ( and ), the present invention uses the distribution difference loss L based on HMD distance dd To increase and The difference between:
[0077]
[0078] In the formula, and It can be calculated by the following equation:
[0079]
[0080] On the other hand, in order to prevent and The semantic information of G is lost. fd and G fr Considered as an autoencoder structure, the reconstruction loss L of the mean square error is r To restore the coupling feature:
[0081]
[0082] 2) Difference relationship:
[0083] This part is dedicated to achieving distribution alignment between the inherent degradation trend features from any two domains to extract common distribution difference knowledge. The domain confusion loss L based on HMD distance is adopted. dc :
[0084]
[0085] In the formula, and From the i-th domain and the jth domain
[0086] In Q domains, K domain confusion losses will be obtained
[0087]
[0088] According to the transfer learning theory, The larger the value of G fe The less focused on the distribution confusion of the two domains. However, the proposed CDDKD framework aims to make G fe Obtain a universal difference relationship representation that is insensitive to the size of the distribution difference. Therefore, in the distribution alignment process, we should assign greater weights to larger domain confusion losses. The weight assignment strategy is constructed as follows:
[0089]
[0090] In the formula, τ is the temperature factor. The smaller the τ value, the closer the Finally, the weight w j Act on the corresponding Update network parameters:
[0091]
[0092] The embodiment of the present invention provides a method for predicting the remaining life of mechanical equipment based on the CDDKD framework, and the process is as follows:
[0093] Step 1: Install an accelerometer on the mechanical equipment to collect the original vibration signals under different working conditions. Then, form a full life cycle sample based on the collected original vibration signals.
[0094] Step 2: Construct the CDDKD framework according to the task to be implemented, such as Figure 1 shown.
[0095] Step 3: Then, based on the CDDKD framework and convolutional neural network constructed above, an information decoupling network (IDN) model is constructed, as shown in Figure 2 shown.
[0096] Step 4: Input sufficient sample data from heterogeneous mechanical equipment into the IDN model, and then use equations (4, 6, 8, 10, 14) to iteratively train the weight parameters of the constructed network.
[0097] Step 5: Inherit the weight parameters of the feature extractor in the above step, and then use equations (4, 6, 8, 10) to input the restricted sample data in the target mechanical equipment into the IDN model for fine-tuning training.
[0098] Step 6: After multiple iterations of training, observe whether the loss curve tends to converge smoothly. If the curve converges, the network model training is completed. The trained network will be used to predict the remaining life of cross-condition migration under heterogeneous monitoring data of bearings and gears.
[0099] The above are the process steps of the proposed method for predicting the remaining life of mechanical equipment based on the CDDKD framework. The following experimental results have demonstrated the effectiveness of the method.
[0100] Verification experiment:
[0101] 1) The gear life cycle dataset was collected by the gear contact fatigue test rig at Newcastle University. The test rig uses two accelerometers mounted on the test gearbox to acquire vibration monitoring signals at a sampling frequency of 25kHz. The sampling duration is set to 0.2 seconds, followed by a one-minute interval. The criterion for gear failure is that the pitting area on the tooth surface exceeds 4%. Fatigue life tests were carried out under three different operating conditions, with load and speed information of 4400N·m&1100rpm(G1_1), 4070N·m&800rpm(G1_2), and 3750N·m&800rpm(G1_3).
[0102] 2) The bearing life cycle data set comes from Xi'an Jiaotong University. The sampling frequency is 25.6kHz, the sampling interval is 1 minute, and the sampling duration is 1.28 seconds. The bearing data set includes three working conditions, and its load and speed information are 12kN&2100rpm, 11kN&2250rpm and 10kN&2400rpm respectively. Five operation failure monitoring vibration signals were collected under each working condition. The present invention will select the working condition data numbered B1_1, B1_2, and B1_3 for the remaining life migration prediction experiment.
[0103] Through the above-collected full life cycle data sets of bearings and gears, a total of two cases can be constructed, each of which contains six cross-condition migration remaining life prediction tasks, as shown in Tables 1 and 2.
[0104] Comparative experiment:
[0105] In order to prove the superiority of the remaining life prediction method proposed in the present invention, the present invention uses Score and Mean Absolute Error (MAE) as evaluation indicators. Compared with the current typical remaining life prediction network model, the experimental results are shown in Table 1 and Table 2. According to the data results in Table 1 and Table 2, it can be concluded that the mechanical remaining life prediction method based on the CDDKD framework of the present invention has higher migration prediction accuracy, stronger robustness and generalization ability.
[0106] Table 1 Experimental results of gear life cycle dataset
[0107]
[0108] Table 2 Experimental results of bearing life cycle data set
[0109]
[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.
Claims
1. A method for predicting the remaining life of mechanical equipment based on the CDDKD framework, characterized in that: The method specifically comprises the following steps: S1: Collect the original vibration signals of mechanical equipment under different working conditions, and then form the full life cycle samples based on the collected original vibration signals; S2: According to the task to be implemented, a CDDKD framework is constructed, where CDDKD stands for Common Distribution Difference Knowledge Distillation; S3: Based on the CDDKD framework and convolutional neural network, an IDN model is constructed, where IDN stands for information decoupling network; S4: Input sufficient sample data from heterogeneous mechanical equipment into the IDN model, and then use the prediction loss L P , noise constraint loss L o , distribution difference loss L based on high-order moment distance dd , reconstruction loss L r and the domain confusion loss L based on HMD distance dc Iteratively train the weight parameters of the constructed IDN model; S5: Inherit the weight parameters of the feature extractor in step S4, and then use the prediction loss L P , noise constraint loss L o , distribution difference loss L based on high-order moment distance dd and the reconstruction loss L r Input the restricted sample data of the target mechanical equipment into the IDN model for fine-tuning training; S6: After multiple iterations of training, observe whether the loss curve tends to converge smoothly. If the curve converges, the IDN model training is completed. The trained IDN model will be used to predict the remaining life of mechanical equipment under cross-operating condition migration under monitoring data.
2. The method for predicting the remaining service life of mechanical equipment according to claim 1, characterized in that: In step S2, the constructed CDDKD framework is specifically as follows: first, the distribution difference knowledge in heterogeneous mechanical equipment with sufficient sample data is distilled and extracted, and then, based on the similar action mechanism of distribution differences in different mechanical systems or components, the obtained distribution difference knowledge is used for the remaining life prediction of target mechanical equipment with scarce data, thereby enhancing the domain invariance of the trend degradation characteristics of the target mechanical equipment under complex and variable working conditions.
3. The method for predicting the remaining service life of mechanical equipment according to claim 1, characterized in that: In step S3, the constructed IDN model includes a feature extractor G fe (θ fe ), feature decoupler G fd (θ fd ), Remaining useful life predictor G p (θ p ) and feature reconstructor G fr (θ fr ), where θ fe ,θ fd ,θ p and θ fr Respectively represent the corresponding network weight parameters; feature extractor G fe Used to obtain the most critical relationship representation and extract coupling features F in the CDDKD framework c ; Feature decoupler G fd Used to decouple the degradation trend feature F dt and the degradation noise feature F dn ; Feature Reconstructor G fr For the degenerate trend feature F dt and the degradation noise feature F dn Recovering the extracted coupling features; Remaining useful life predictor G p For outputting the final remaining service life.
4. The method for predicting the remaining service life of mechanical equipment according to claim 3, characterized in that: In step S3, the optimization objectives of the IDN model include the following three parts: 1) learning a feature decoupled representation to extract fe The obtained coupling characteristics F c Separate the degradation trend feature F dt and the degradation noise feature F dn ; 2) By jointly training the feature decoupler G fd and feature reconstructor G fr , maintain the degradation trend characteristic F dt and the degradation noise feature F dn semantic consistency to avoid feature information distortion; 3) Degenerate trend feature F from any two domains dt Distribution alignment of,learning differential relation representations.
5. The method for predicting the remaining service life of mechanical equipment according to claim 3, characterized in that: In step S3, feature decoupling specifically includes: first, by predicting the loss L P Train the feature extractor G with mean square error fe , feature decoupler G fd and the remaining life predictor G p , in order to achieve accurate predictions: Where Y i Represents the original data sample X i The actual remaining useful life label; represents the predicted value of the remaining life of the i-th sample, which is calculated as follows: Then, fix G fe and G p The weight parameter of the noise constraint loss L o Train the feature disentangler G separately fd : In the formula, represents the degradation noise characteristics of the i-th sample, u dn represents the feature mean, and u dn It is calculated as follows: In the formula, n represents the number of samples; Finally, the degradation trend characteristic F is obtained through equations (1) and (3): dt and the degradation noise feature F dn .
6. The method for predicting the remaining service life of mechanical equipment according to claim 5, characterized in that: In step S3, maintaining semantic consistency specifically includes: Use the distribution difference loss L based on the high-order moment distance dd To increase the degradation trend characteristic F dt and the degradation noise feature F dn The difference between: In the formula, represents the probability expectation of the degradation trend characteristics, represents the probability expectation of the degraded noise feature; φ(·) is the kernel function in RKHS, RKHS represents the reproducing kernel Hilbert space, and the mapping function represents the tensor product operation; ⊙ represents the Hadamard product operation; F dt and F dn Calculated by the following equation: In the formula, and Respectively represent the degradation trend characteristics and degradation noise characteristics of the i-th sample; On the other hand, in order to prevent and The semantic information of fd and feature reconstructor G fr Considered as an autoencoder structure, the reconstruction loss L of the mean square error is r To restore the coupling feature: in, represents the expected probability of the sample.
7. The method for predicting the remaining service life of mechanical equipment according to claim 6, characterized in that: In step S3, learning the difference relationship specifically includes: striving to achieve distribution alignment between the inherent degradation trend features from any two domains, using a domain confusion loss L based on high-order moment distance dc : In the formula, and Indicates that they come from the i-th domain and the jth domain The degradation trend characteristics of In Q domains, K domain confusion losses will be obtained In the formula, represents the number of combinations of taking 2 from Q; In the distribution alignment process, large weights are assigned to large domain confusion losses; the weight assignment strategy is constructed as follows: In the formula, τ is the temperature factor; the smaller the τ value, the closer the The larger the difference is; finally, the weight w j Act on the corresponding Update network parameters: In the formula, denotes the corrected domain confusion loss.