Rotary mechanical equipment health monitoring method based on depth residual error alignment

Abstract

The invention provides a time-varying rotating speed equipment health monitoring method based on depth residual error alignment. The method comprises the specific steps that complete and sufficient health state vibration signals and corresponding rotating speed signals are collected; inputting all health enhancement samples into a feature extractor for self-supervised contrast learning training, and extracting feature manifolds of the vibration signals; constructing nonlinear mapping from the rotating speed to the characteristic trend manifold by using a regression device; constructing residual vectors of manifold feature vectors and rotating speed regression feature vectors of all fault type vibration data; constructing nonlinear mapping from the rotating speed to the residual vector by using a second regression device; calculating a secondary residual vector according to the extractor, the regression device and the second regression device; inputting the secondary residual vector into a detector to perform a reconstruction task, and obtaining a health index according to reconstruction loss; and real-time vibration signals and rotating speed signals are obtained, and health indexes are calculated to realize real-time online monitoring of the rotating machinery. The method has the advantages that meaningless fluctuation and interference brought to health state monitoring by continuous and great change of the rotating speed under the time-varying rotating speed working condition are effectively eliminated, and the method is more sensitive and effective to various faults.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of mechanical intelligent maintenance and abnormality detection, and particularly relates to a rotating mechanical equipment health monitoring method based on deep residual alignment. BACKGROUND

[0002] In recent years, with the rise of intelligent manufacturing and industrial internet technologies, the design of mechanical equipment is increasingly integrated, the structure is increasingly coupled, and the working conditions are increasingly complex. The health monitoring technology of rotating mechanical equipment has long been hindered by time-varying working conditions and has not made much progress. At the same time, it has greatly hindered the landing and application of related scientific and technological. Among them, the problem of mechanical equipment monitoring under time-varying speed conditions is the most prominent. The substantial change of speed will cause the observation to change dramatically, resulting in a visible transition in amplitude and energy of collected quantities such as vibration and sound, causing the model of the monitoring site to fail, and causing frequent false negatives and false positives. Therefore, how to effectively realize the self-adaptive health monitoring of mechanical equipment under time-varying speed conditions is a scientific and application problem that needs to be solved in the field of intelligent maintenance. The conventional processing method usually obtains health indicators based on static feature representation. In this process, the operation condition of the equipment is usually fixed or limited.

[0003] As an effective way to realize deep health monitoring of machinery, the method based on deep learning has been widely applied in the field of equipment intelligent maintenance. The monitoring deep learning model usually realizes systematic learning of input distribution through the structure of auto-encoder to obtain good reconstruction ability of the data in the department. Therefore, when the reconstruction error is lower than a certain threshold, it can be considered that the mechanical equipment is in a relatively healthy state. However, in the actual industrial scene, machinery usually needs to run safely and reliably in changing working conditions. Continuous and large amplitude speed changes will bring fluctuations unrelated to fault information to machine learning, deep learning and a series of monitoring methods based on indicators, resulting in frequent false negatives and false positives in industrial monitoring. The risk of false alarms will cause shutdown and production stoppage, thereby causing a series of economic losses. How to effectively reduce the impact of speed on health indicators so that the indicators are only sensitive to fault information is a technical problem that needs to be solved by technical personnel in the field.

[0004] Self-supervised learning, as an emerging deep learning paradigm in recent years, has seen initial applications in fields such as visual classification and natural language processing. Self-supervised learning guides models to automatically summarize and learn features from completely unlabeled data, enabling them to perform generalization training and adapt broadly to various downstream tasks, achieving good results. In the field of fault diagnosis, self-supervised learning is often used to extract inherent distinguishable features between various fault classes. However, in variable speed environments, there is coupling interference between changes in operating conditions and fault components, causing many monitoring models to be unable to effectively distinguish fault component information under varying speeds. Summary of the Invention

[0005] To address the aforementioned issues, this application's technical solution utilizes a mechanism of constructing multi-level residual vectors through multi-level regression. This achieves explicit operating condition alignment in the representation space and effectively solves the health monitoring problem under time-varying speed conditions using dynamic thinking. It effectively reduces the impact of speed on health indicators, improves the sensitivity of indicators to fault information, and provides a time-varying speed equipment health monitoring method based on deep residual alignment, specifically:

[0006] This invention provides a health monitoring method for rotating machinery based on deep residual alignment, for adaptive health monitoring under time-varying speed conditions, comprising the following steps:

[0007] Step 1) Collect vibration signals of rotating machinery in a healthy state and corresponding rotational speed signals;

[0008] Step 2) Perform self-supervised contrastive learning training based on the healthy vibration signal to obtain an extractor for extracting the feature manifold of the vibration signal;

[0009] Step 3) Input the vibration signal of the healthy state into the trained extractor to obtain the characterization features, and then input the current rotation speed value into the set regressor to obtain the regression features. Train to obtain a regressor used to construct a nonlinear mapping from transient rotation speed to characterization features.

[0010] Step 4) Input all vibration signals and corresponding rotation speed signals of healthy states into the trained extractor and regressor respectively, and subtract the corresponding vibration characterization features and rotation speed regression features to form a residual vector, thereby achieving the first deep residual alignment of different characterization features under all rotation speed conditions.

[0011] Step 5) Construct a second regressor that maps the time-varying transient speed values ​​to the corresponding residual vectors in a nonlinear regression, and train the second regressor to obtain the regression residual features;

[0012] Step 6) Based on the trained extractor, regressor, and second regressor, the vibration signal and corresponding rotation speed signal of the healthy state are input, and the residual vector and the regression residual features are subtracted to form a second-level residual vector, so as to realize the second deep residual alignment of different characterization features of all rotation speed conditions.

[0013] Step 7) Input the secondary residual vector into the set detector to perform the vector reconstruction task, and obtain the health index from the reconstruction loss;

[0014] Step 8) Calculate health indicators based on the real-time acquired vibration and rotation speed signals to achieve real-time online monitoring of rotating machinery.

[0015] A further design of the deep residual aligned health monitoring method for rotating machinery is that the training process of the extractor in step 2) is as follows: First, the vibration signals in the healthy state are flipped and negatively charged in sequence to form enhanced positive sample pairs, which are then input into the extractor to generate the representation features of the enhanced samples in the feature space. Then, based on the cosine similarity metric, a contrastive learning loss and the final cross-entropy loss are constructed. By aggregating similar vibration samples, the automatic summarization and extraction of the feature manifold of the vibration signal under time-varying rotation speed is completed.

[0016] A further design of the depth residual aligned rotating machinery health monitoring method is that, in step 2), the vibration signal of the health state is flipped and negativeed according to equations (1) and (2) respectively.

[0017] In equations (1) and (2), x IO The vibration signal after flipping, x NO The negative vibration signal, Let be the data points in a vibration signal of length n.

[0018] The further design of the deep residual aligned health monitoring method for rotating machinery is that, in step 2), a contrastive learning loss and a final cross-entropy loss are constructed based on the cosine similarity metric. Specifically, by aggregating similar vibration samples, the obtained representation features are input to a set projector, and the representation features of the feature space are projected to the latent space through nonlinear mapping. In the latent space, the contrastive learning loss is used to calculate the representation similarity between positive and negative latent features, and the extraction and projector training are guided by the mode of maximizing the similarity of negative latent features and minimizing the similarity of positive latent features. The representation features are extracted according to equation (3), and the latent features are represented according to equation (4).

[0019] z i =Pro(Ext(IO(x) k))) (3)

[0020] z j =Pro(Ext(NO(x) k (4)

[0021] In equations (3) and (4), z i With z j These represent the latent feature vectors of the two augmented samples, where Ext represents the extractor and Pro represents the projector.

[0022] The extreme cosine similarity measure is calculated according to equation (5);

[0023] Sim i,j =z i T z j / (‖z i |||z j ||) (5)

[0024] Calculate the contrastive learning loss according to equation (6);

[0025] In equation (6), The indicator function is defined if and only if k≠i and its value is 0; τ is the built-in temperature coefficient for self-supervised contrastive learning, which is usually taken as 0.07.

[0026] The final cross-entropy loss is calculated according to equation (7);

[0027]

[0028] A further design of the deep residual aligned rotating machinery health monitoring method is that the extractor itself is an arbitrary form of coding structure network or model, and the projector is a fully connected neural network.

[0029] A further design of the deep residual aligned health monitoring method for rotating machinery is that the training process of the regressor in step 3) is specifically as follows: the training of the regressor is guided by calculating the mean square error between the regression features and the extractor representation features, and the loss function of the regressor is established according to equation (8).

[0030] In equation (8), x k For the k-th vibration signal in a single training batch, s k Let Ext be the k-th rotational speed signal in a single training batch, and Spy be the extractor and regressor, respectively. N is the number of samples in a single training batch.

[0031] A further design of the deep residual aligned health monitoring method for rotating machinery is that the training process of the second regressor in step 5) is specifically as follows: the training of the second regressor is guided by calculating the mean square error between the secondary regression features and the residual vector, and the loss function of the second regressor is established according to equation (9).

[0032] In equation (9), x k For the k-th vibration signal in a single training batch, s k Let represent the k-th rotational speed signal in a single training batch. Ext, Spy, and Spyt are the extractor, regressor, and second regressor, respectively, and N is the number of samples in a single training batch.

[0033] A further design of the deep residual aligned health monitoring method for rotating machinery is that the health index is obtained from the reconstruction loss specifically by establishing the training loss function of the detector according to equation (10).

[0034] In equation (10), Det refers to the detector, SRes k This represents the second-order residual vector. The health index value is the reconstruction loss value corresponding to the reconstructor after a single vibration signal and rotational speed input.

[0035] A further design of the deep residual aligned rotating machinery health monitoring method is that the regressor and the second regressor are fully connected neural networks FCNN with a single hidden layer.

[0036] A further design of the deep residual aligned rotating machinery health monitoring method is that the detector is a convolutional autoencoder with skip connections.

[0037] Beneficial effects:

[0038] The health monitoring method for rotating machinery based on deep residual alignment proposed in this invention proposes a mechanism for constructing multi-level residual vectors through multi-level regression, which realizes explicit working condition alignment in the representation space. It effectively solves the health monitoring problem under time-varying speed conditions by using dynamic thinking, effectively eliminating the meaningless fluctuations and interferences brought to health status monitoring by the continuous and large changes in speed under time-varying speed conditions, and is more sensitive and effective for various faults.

[0039] The method of this invention far surpasses existing comparative health monitoring methods in terms of stability and reliability, achieving the best monitoring performance. Attached Figure Description

[0040] Figure 1This is a flowchart of a time-varying speed device health monitoring method based on deep residual alignment according to an embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram of the extractor characterization feature extraction according to an embodiment of the present invention.

[0042] Figure 3 This is a schematic diagram of latent feature extraction and comparative learning training of the projector according to an embodiment of the present invention.

[0043] Figure 4 This is a schematic diagram illustrating the alignment of deep regression features between the regressor and the second regressor in an embodiment of the present invention.

[0044] Figure 5 This is a schematic diagram of the detector health index generation process based on multiple depth residual alignment in an embodiment of the present invention.

[0045] Figure 6 This is a schematic diagram of the vibration signal manifold trend obtained by the health monitoring method for rotating machinery based on deep residual alignment according to the present invention.

[0046] Figure 7 This is a comparison of the results obtained by the health monitoring method for rotating machinery based on deep residual alignment in this application with the results obtained by other methods. Detailed Implementation

[0047] The specific embodiments of the present invention will be described below with reference to the accompanying drawings and examples.

[0048] Example 1:

[0049] This invention discloses a health monitoring method for rotating machinery based on deep residual alignment, used for adaptive health monitoring under time-varying speed conditions, specifically including the following steps:

[0050] Step 1) Collect the vibration signal of the rotating machinery in a healthy state and the corresponding rotation speed signal.

[0051] Step 2) Perform self-supervised contrastive learning training based on the healthy vibration signal to obtain an extractor for extracting the feature manifold of the vibration signal;

[0052] Step 3) Input the vibration signal of the healthy state into the trained extractor to obtain the characterization features, and then input the current rotation speed value into the set regressor to obtain the regression features. Train to obtain a regressor used to construct a nonlinear mapping from transient rotation speed to characterization features.

[0053] Step 4) Input all vibration signals and corresponding rotation speed signals of healthy states into the trained extractor and regressor respectively, and subtract the corresponding vibration characterization features and rotation speed regression features to form a residual vector, thereby achieving the first deep residual alignment of different characterization features under all rotation speed conditions.

[0054] Step 5) Construct a second regressor that maps the time-varying transient speed values ​​to the corresponding residual vectors in a nonlinear regression, and train the second regressor to obtain the regression residual features;

[0055] Step 6) Based on the trained extractor, regressor, and second regressor, the vibration signal and corresponding rotation speed signal of the healthy state are input, and the residual vector and the regression residual features are subtracted to form a second-level residual vector, so as to realize the second deep residual alignment of different characterization features of all rotation speed conditions.

[0056] Step 7) Input the secondary residual vector into the set detector to perform the vector reconstruction task, and obtain the health index from the reconstruction loss;

[0057] Step 8) Calculate health indicators based on the real-time acquired vibration and rotation speed signals to achieve real-time online monitoring of rotating machinery.

[0058] Example 2:

[0059] This embodiment provides a method for health monitoring of rotating machinery with deep residual alignment, used for online adaptive health monitoring of rotating machinery under time-varying speed conditions. The following is combined with... Figure 1 The health monitoring method for time-varying speed equipment based on deep residual alignment in this embodiment will be described.

[0060] Step 1) Obtain vibration and rotational speed data of the mechanical equipment under various speed conditions to assess its health status. The richness of the rotational and vibrational signals under different healthy speed conditions greatly affects and determines the effectiveness of this monitoring method.

[0061] Step 2) Segment the vibration data of all healthy states to form a one-dimensional vibration signal sample set. For each individual sample in the sample set, perform a negativeing ​​and flipping operation sequentially to obtain two positive enhanced samples corresponding to each sample. The enhanced samples of the original sample and the enhanced samples of other original samples are each other's negative enhanced samples. The negativeing ​​and flipping operations in this embodiment can effectively preserve the inherent frequency, period, amplitude, and other characteristics of the vibration signal itself, and are easy to implement. The enhanced sample pairs generated by the negativeing ​​and flipping operations allow the contrastive learning loss function (see Equation 6) to learn the samples based on feature similarity in the representation feature space and the latent feature space. In the feature space, similar samples (corresponding to vibration signals under similar speed conditions) are clustered, while different samples (corresponding to vibration signals with large differences in speed conditions) are separated, see [link to relevant documentation]. Figure 6 ,in Figure 6 'a' represents the characteristic of the original variable speed vibration signal. Figure 6 b represents the vibration signal characteristics after self-supervised comparative learning.

[0062] Let the original vibration signal be The negativeing ​​and flipping operations of the vibration signal are shown in Equation (1) and Equation (2), respectively.

[0063]

[0064] Where, x IO The vibration signal after flipping, x NO The negative vibration signal, Let n be the data point in the vibration signal, and n be the length of the vibration signal.

[0065] All positive and negative sample pairs are input into the extractor to generate enhanced representation features in the feature space. Further, the obtained representation features are input into the projector, which projects the representation features from the feature space onto the latent space using a nonlinear mapping. In the latent space, contrastive learning loss is used to calculate the representation similarity between positive and negative latent features; see [link to relevant documentation]. Figure 3 The training of the extractor and projector is guided by the pattern of maximizing the similarity of latent features of negative examples and minimizing the similarity of latent features of positive examples. Representational features and latent features are extracted according to equations (3) and (4), respectively.

[0066] z i =Pro(Ext(IO(x) k ))) (3)

[0067] z j =Pro(Ext(NO(x) k (4)

[0068] Where z i With z j Let represent the latent feature vectors of the two enhanced samples respectively, Ext represent the extractor, Pro represent the projector, and the cosine similarity measure is given by equation (5).

[0069] Sim i,j =z i T z j / (‖z i |||z j ||) (5)

[0070] Contrastive learning loss is:

[0071] in, The indicator function is defined if and only if k ≠ i and its value is 0. τ is the built-in temperature coefficient for self-supervised contrastive learning, typically taken as 0.07. The final cross-entropy loss is given by equation (7).

[0072]

[0073] By clustering similar vibration samples and separating dissimilar vibration samples, the system achieves automatic summarization and extraction of the characteristic manifold of vibration signals under time-varying rotational speeds without any labels. The extractor itself can be any form of encoded network or model, and the projector is a fully connected neural network.

[0074] Step 3) as Figure 4 The vibration signals at various rotational speeds under healthy conditions are input into the trained extractor to obtain representational features, and the current rotational speed value is input into the regressor to obtain regression features. The mean square error between the regression features and the extractor representational features is calculated to guide the training of the regressor, enabling the regressor to become a nonlinear mapper between rotational speed and representational features. When a new rotational speed value is generated, the core position of the current rotational speed value in the representational space can be quickly calculated. The loss function of the regressor is given by equation (8). The regressor in this embodiment is a fully connected neural network FCNN with a single hidden layer.

[0075] Where, x k For the k-th vibration signal in a single training batch, s k Let Ext be the k-th rotational speed signal in a single training batch, and Spy be the extractor and regressor, respectively. N is the number of samples in a single training batch.

[0076] Step 4) The successfully trained extractor and the trained first regressor are input into the trained extractor and first regressor respectively for all healthy vibration signals and speed signals to obtain vibration characterization features and speed regression features. The difference between the two types of features is calculated to form a residual vector, realizing the first deep residual alignment of different characterization features for all speed conditions.

[0077] Res k =Ext(x k )-Spy(s k (9)

[0078] Step 5) Figure 4 A nonlinear regression mapping from rotational speed values ​​to residual vectors is constructed using a second regressor.

[0079] Throughout the training process of the second regressor, the parameters of both the extractor and the first regressor are in the state after training.

[0080] Where x k For the k-th vibration signal in a single training batch, s k Let represent the k-th rotational speed signal in a single training batch. Ext, Spy, and Spyt are the extractor, regressor, and second regressor, respectively, and N is the number of samples in a single training batch.

[0081] Step 6) Based on the trained extractor, regressor, and second regressor, the vibration signal and corresponding rotational speed signal of the healthy state are input. The difference between the obtained residual vector and the regression residual features is used to form a second-level residual vector, achieving a second-level deep residual alignment of different representational features for all rotational speed conditions. Second-level residual vector SRes k The calculation formula is shown in Equation 11. The second regressor in this implementation is a fully connected neural network FCNN with a single hidden layer.

[0082] SRes k =Ext(x k )-Spy(s k )-Spyt(s k (11)

[0083] Step 7) Input the secondary residual vector into the detector for reconstruction and calculate the reconstruction loss. The training loss function of the detector is shown in Equation 12. In this embodiment, the detector uses a convolutional autoencoder with skip connections. These skip connections are used to implement the network layers corresponding to the encoder and decoder. The health index value is the reconstruction loss value corresponding to the reconstructor after a single vibration signal and rotational speed input.

[0084]

[0085] Here, Det stands for detector, and the final health index value is obtained from the reconstruction loss.

[0086] Step 8) Based on the real-time acquired vibration and rotational speed signals, calculate health indicators to achieve real-time online monitoring of rotating machinery. The health indicators in this embodiment are based on conventional technologies used in actual industrial applications, are highly condensed, and can intuitively reflect the real-time status of the equipment. Furthermore, these scalar health indicators can be visualized through time series analysis, thereby reflecting the health change trend of the equipment within a certain time period, such as... Figure 7 As shown.

[0087] Using a time-varying speed bearing test bench as the pre-verification object, the results obtained by the time-varying speed equipment health monitoring method based on deep residual alignment in this embodiment are compared with the results obtained by other methods. Figure 7 As shown: From Figure 7 The results show that the method of the present invention effectively eliminates the meaningless fluctuations and interferences caused by the continuous and large changes in speed under time-varying speed conditions to health status monitoring, and is more sensitive and effective for various faults (including bearing inner ring faults, outer ring faults, ball faults, combined inner and outer ring faults, and combined outer ring and ball faults). The method of this embodiment far surpasses the other four comparative health monitoring methods (including conventional deep learning methods such as Convolutional Auto-encoder (CAE), Long-Short Term Memory Auto-encoder (LSTMAE), Densely-connected U-Net (DenseUNet), and 1D Residual Convolutional Auto-encoder (1DRCAE)) in terms of stability and reliability, achieving the best monitoring performance.

[0088] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A health monitoring method for rotating machinery based on deep residual alignment, used for adaptive health monitoring under time-varying speed conditions, characterized in that, Includes the following steps: Step 1) Collect the vibration signal of the rotating machinery in a healthy state and the corresponding rotational speed signal; Step 2) Perform self-supervised contrastive learning training based on the healthy vibration signal to obtain an extractor for extracting the feature manifold of the vibration signal; Step 3) Input the vibration signal of the healthy state into the trained extractor to obtain the characterization features, and then input the current rotation speed value into the set regressor to obtain the regression features. Train to obtain a regressor used to construct a nonlinear mapping from transient rotation speed to characterization features. Step 4) Input all vibration signals and corresponding rotation speed signals of healthy states into the trained extractor and regressor respectively, and subtract the corresponding vibration characterization features and rotation speed regression features to form a residual vector, thereby achieving the first deep residual alignment of different characterization features under all rotation speed conditions. Step 5) Construct a second regressor that maps the time-varying transient speed values ​​to the corresponding residual vectors in a nonlinear regression, and train the second regressor to obtain the regression residual features; Step 6) Based on the trained extractor, regressor, and second regressor, the vibration signal and corresponding rotation speed signal of the healthy state are input, and the residual vector and the regression residual features are subtracted to form a second-level residual vector, so as to realize the second deep residual alignment of different characterization features of all rotation speed conditions. Step 7) Input the secondary residual vector into the set detector to perform the vector reconstruction task, and obtain the health index from the reconstruction loss; Step 8) Calculate health indicators based on the real-time acquired vibration and rotation speed signals to achieve real-time online monitoring of rotating machinery.

2. The method for health monitoring of rotating machinery with depth residual alignment according to claim 1, characterized in that, The training process of the extractor in step 2) is as follows: First, the vibration signals in the healthy state are flipped and negative in sequence to form enhanced positive sample pairs and input into the extractor to generate the representation features of the enhanced samples in the feature space. Then, based on the cosine similarity measure, the contrastive learning loss and the final cross-entropy loss are constructed. By aggregating similar vibration samples, the automatic summarization and extraction of the feature manifold of the vibration signal under time-varying speed is completed.

3. The method for health monitoring of rotating machinery with depth residual alignment according to claim 2, characterized in that, In step 2), the vibration signal of the healthy state is flipped and negativeed according to equations (1) and (2) respectively. In equations (1) and (2), x IO The vibration signal after flipping, x NO The negative vibration signal, These are data points with a vibration signal length of n in the vibration signal.

4. The method for health monitoring of rotating machinery with depth residual alignment according to claim 3, characterized in that, In step 2), the contrastive learning loss and the final cross-entropy loss are constructed based on the cosine similarity metric. Specifically, by clustering similar vibration samples, the obtained representation features are input into a set projector, and the representation features of the feature space are projected into the latent space through nonlinear mapping. In the latent space, the contrastive learning loss is used to calculate the representation similarity between positive and negative latent features, and the extraction and projector training is guided by the pattern of maximizing the similarity of negative latent features and minimizing the similarity of positive latent features. Representation features are extracted according to Equation (3), and latent features are represented according to Equation (4). z i =Pro(Ext(IO(x k ))) (3) z j =Pro(Ext(NO(x k ))) (4) In equations (3) and (4), z i With z j These represent the latent feature vectors of the two augmented samples, where Ext represents the extractor and Pro represents the projector. The extreme cosine similarity measure is calculated according to equation (5); Sim i,j =z i T With j / (‖With i ‖||of j ||) (5) Calculate the contrastive learning loss according to equation (6); In equation (6), The indicator function is defined if and only if k≠i and its value is 0; τ is the built-in temperature coefficient for self-supervised contrastive learning, which is usually taken as 0.

07. The final cross-entropy loss is calculated according to equation (7); 5. The method for health monitoring of rotating machinery with depth residual alignment according to claim 4, characterized in that, The extractor itself can be any form of encoded network or model, and the projector is a fully connected neural network.

6. The method for health monitoring of rotating machinery with depth residual alignment according to claim 1, characterized in that, The training process of the regressor in step 3) is specifically as follows: the training of the regressor is guided by calculating the mean square error between the regression features and the extractor representation features, and the loss function of the regressor is established according to equation (8). In equation (8), x k For the k-th vibration signal in a single training batch, s k Let Ext be the k-th rotational speed signal in a single training batch, and Spy be the extractor and regressor, respectively. N is the number of samples in a single training batch.

7. The method for health monitoring of rotating machinery with depth residual alignment according to claim 1, characterized in that, The training process of the second regressor in step 5) is as follows: the training of the second regressor is guided by calculating the mean square error between the secondary regression features and the residual vector, and the loss function of the second regressor is established according to equation (9). In equation (9), x k For the k-th vibration signal in a single training batch, s k Let represent the k-th rotational speed signal in a single training batch. Ext, Spy, and Spyt are the extractor, regressor, and second regressor, respectively, and N is the number of samples in a single training batch.

8. The method for health monitoring of rotating machinery with depth residual alignment according to claim 1, characterized in that, The health index obtained from the reconstruction loss is specifically as follows: The training loss function of the detector is established according to equation (10). In equation (10), Det refers to the detector, SRes k This represents the second-order residual vector. The health index value is the reconstruction loss value corresponding to the reconstructor after a single vibration signal and rotational speed input. N is the number of samples in a single training batch.

9. The method for health monitoring of rotating machinery with depth residual alignment according to claim 7 or 8, characterized in that, The regressor and the second regressor are both fully connected neural networks (FCNNs) with a single hidden layer.

10. The method for health monitoring of rotating machinery with depth residual alignment according to claim 1, characterized in that, The detector is a convolutional autoencoder with skip connections.

Images