A method and system for predicting the remaining life of a rotating machine

Abstract

The application relates to a rotating machine residual life prediction method and system, and belongs to the field of mechanical life prediction. The method comprises the following steps: collecting signals capable of reflecting mechanical life of a rotating machine to be measured at least at two positions, extracting one-dimensional time characteristic data from the signals capable of reflecting mechanical life by using principal component analysis pooling, training a related vector machine by using the one-dimensional time characteristic data, so that the full life cycle of the rotating machine is divided into a period without obvious failure trend and a failure tendency period; training a deep separable convolution gate recurrent unit network by using data of the failure tendency period, and predicting the residual life of the rotating machine in the failure tendency period. The application can improve the prediction accuracy when the residual service life of the rotating machine is predicted.

Description

Technical Field

[0001] This invention relates to a method and system for predicting the remaining life of rotating machinery, belonging to the field of machinery life prediction. Background Technology

[0002] Currently, in the fault diagnosis and remaining life prediction of rotating machinery, traditional methods mostly employ physical models and data-driven models. Physical models rely on complex mathematical equations, making it difficult to cope with the varied operating conditions in real-world applications. Data-driven models, such as Support Vector Machines (SVM) and Random Forests (RF), make predictions through feature extraction and model training. However, because they require manually designing the feature extraction process, they are easily limited by sensor data under specific conditions, and processing time-series data under different states is quite difficult.

[0003] Chinese invention patent application CN113051689A discloses a method for predicting the remaining service life of bearings based on convolutionally gated recurrent neural networks (GNRN). This method is based on: collecting vibration acceleration signals throughout the bearing's entire lifespan; integrating and preprocessing the vibration acceleration data throughout the bearing's lifespan, dividing the processed data into corresponding training and testing sets; designing a network structure that fuses a GNRN neural network and an attention mechanism for predicting the remaining service life of bearings; feeding the training set into the GNRN neural network and attention mechanism fusion network structure for automatic feature extraction; and then feeding the extracted features into a fully connected layer to obtain the predicted remaining service life. The bearing remaining service life predicted by this method is more accurate, which can prevent major accidents and provide reference for predictive maintenance.

[0004] However, this method of predicting the remaining service life of a bearing using full-life bearing data has the problem of large errors. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting the remaining service life of rotating machinery, in order to solve the problem of low accuracy in predicting the remaining service life of rotating machinery in the prior art.

[0006] To achieve the above objectives, the present invention includes:

[0007] The present invention provides a method for predicting the remaining life of rotating machinery, comprising the following steps:

[0008] Signals reflecting the lifespan of the rotating machinery under test are collected at at least two locations. One-dimensional time feature data is extracted from the signals reflecting the lifespan using principal component analysis pooling. The one-dimensional time feature data is then used to train a correlation vector machine to divide the entire lifespan of the rotating machinery into periods with no obvious failure trend and periods with failure tendency.

[0009] By training a deep separable convolutional gated recurrent unit network using data from the fault-prone period, the remaining life of rotating machinery during the fault-prone period can be predicted.

[0010] The depthwise separable convolutional gated recurrent unit network includes alternating separable convolutional building blocks and a single-layer gated recurrent unit network. The separable convolutional building blocks are used to model the spatial features in multi-channel sequences, while the single-layer gated recurrent unit network is used to mine the temporal dependencies of data according to the acquisition order.

[0011] Furthermore, signals that can reflect mechanical life include vibration signals, temperature signals, and pressure signals.

[0012] Furthermore, acquiring signals reflecting the mechanical life of the rotating machinery under test at two locations includes acquiring signals reflecting the mechanical life of the rotating machinery under test in a horizontal direction perpendicular to the rotating machinery and in a vertical direction perpendicular to the rotating machinery.

[0013] Furthermore, before using principal component analysis pooling to extract one-dimensional time feature data from the signal that reflects mechanical life, the signal that reflects mechanical life is preprocessed.

[0014] Preprocessing includes at least normalization.

[0015] Furthermore, the relevant vector machine is iteratively trained on the pooled dataset using an iterative reweighted least squares method.

[0016] Furthermore, it also includes: verifying the effectiveness and reliability of the prediction accuracy of depth-separable convolutional gated recurrent unit networks using a scoring function and root mean square error.

[0017] The present invention provides a system for predicting the remaining life of rotating machinery, comprising a processor that executes a computer program to perform the following steps:

[0018] Signals reflecting the lifespan of the rotating machinery under test are collected at at least two locations. One-dimensional time feature data is extracted from the signals reflecting the lifespan using principal component analysis pooling. The one-dimensional time feature data is then used to train a correlation vector machine to divide the entire lifespan of the rotating machinery into periods with no obvious failure trend and periods with failure tendency.

[0019] By training a deep separable convolutional gated recurrent unit network using data from the fault-prone period, the remaining life of rotating machinery during the fault-prone period can be predicted.

[0020] The depthwise separable convolutional gated recurrent unit network includes alternating separable convolutional building blocks and a single-layer gated recurrent unit network. The separable convolutional building blocks are used to model the spatial features in multi-channel sequences, while the single-layer gated recurrent unit network is used to mine the temporal dependencies of data according to the acquisition order.

[0021] Furthermore, signals that can reflect mechanical life include vibration signals, temperature signals, and pressure signals;

[0022] The signals that reflect the mechanical life of the rotating machinery under test are collected at two locations: in the horizontal direction perpendicular to the rotating machinery and in the vertical direction perpendicular to the rotating machinery.

[0023] Furthermore, before extracting one-dimensional time feature data from the signal that reflects mechanical life using principal component analysis pooling, the signal that reflects mechanical life is preprocessed; the preprocessing includes at least normalization.

[0024] The Relevant Vector Machine is trained iteratively on the pooled dataset using an iterative reweighted least squares method.

[0025] Furthermore, the effectiveness and reliability of the prediction accuracy of the depthwise separable convolutional gated recurrent unit network are verified by using the scoring function and root mean square error.

[0026] The beneficial effects of this invention are as follows: As a pioneering invention, this invention provides a method and system for predicting the remaining life of rotating machinery. By collecting signals reflecting the life of the rotating machinery under test at at least two locations, principal component analysis pooling is used to extract one-dimensional time feature data from the signals reflecting the life of the machinery. The one-dimensional time feature data is then used to train a correlation vector machine to divide the entire life cycle of the rotating machinery into a period without obvious failure trend and a period of failure tendency, thereby achieving data dimensionality reduction and simplifying the dataset. Then, the data from the period of failure tendency is used to train a depthwise separable convolutional gated recurrent unit network to predict the remaining life of the rotating machinery during the period of failure tendency, which can improve the accuracy of predicting the remaining life of rotating machinery. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating a method for predicting the remaining life of rotating machinery provided in an embodiment of the present invention;

[0028] Figure 2This is a schematic diagram of the structure of a device for collecting signals that reflect the lifespan of rotating machinery under test, according to an embodiment of the present invention.

[0029] Figure 3 This is a flowchart illustrating how a correlation vector machine, according to an embodiment of the present invention, divides the entire lifecycle of rotating machinery into periods with no obvious failure trend and periods with failure tendency.

[0030] Figure 4 This is a schematic diagram illustrating the comparison of data processed using different pooling methods, as provided in an embodiment of the present invention.

[0031] Figure 5 This is a schematic diagram of the structure of a depth-separable convolutional gated recurrent network provided in an embodiment of the present invention;

[0032] Figure 6 This is a graph showing the relationship between metrics for evaluating depth-separable convolutional gated recurrent networks, as provided in an embodiment of the present invention.

[0033] Figure 7 This is a schematic diagram of an evaluation result provided by an embodiment of the present invention;

[0034] Figure 8 This is a schematic diagram illustrating a prediction of remaining useful life provided by an embodiment of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0036] The concept of this invention is to use a correlation vector machine to divide the entire life cycle of the rotating machinery under test into a fault-free trend and a fault-prone period, and then use the data from the fault-prone period to train a deep separable convolutional gated recurrent network to predict the remaining service life of the rotating machinery.

[0037] An embodiment of a method for predicting the remaining life of rotating machinery:

[0038] Figure 1 This is a flowchart illustrating a method for predicting the remaining life of rotating machinery according to an embodiment of the present invention, as shown below. Figure 1 As shown, the prediction method includes the following steps S101 to S103.

[0039] Specifically, S101, signals reflecting the mechanical life of the rotating machinery under test are collected at at least two locations. One-dimensional time feature data is extracted from the signals reflecting the mechanical life using principal component analysis pooling. The one-dimensional time feature data is used to train a correlation vector machine to divide the entire life cycle of the rotating machinery into periods with no obvious failure trend and periods with failure tendency.

[0040] Among them, the two positions are in the horizontal direction perpendicular to the rotating machinery and in the vertical direction perpendicular to the rotating machinery.

[0041] The rotating machinery can be a bearing or other types of machinery. This invention does not impose any particular limitation on this type of machinery. The following description will use the example of a bearing as the rotating machinery.

[0042] The signals that can reflect the mechanical life include at least vibration signals, temperature signals, and pressure signals, and may also include other signals, etc. This invention does not make any special limitations on this. The following description will take the example of signals that can reflect the mechanical life including at least vibration signals, temperature signals, and pressure signals.

[0043] For devices that collect signals that reflect mechanical lifespan, reference can be made to... Figure 2 . Figure 2 This is a schematic diagram of the structure of a device for acquiring signals reflecting the lifespan of rotating machinery under test, as provided in an embodiment of the present invention. Figure 2 As shown, vibration signals from fifteen rolling element bearings (model LDKUER204) were collected under three test conditions, representing two full lifecycle events, with the fault location clearly marked for each test bearing. Under all these variable test conditions, radial force was applied to the bearing housing by a hydraulic loading system, and the rotational speed was determined by a speed controller of an AC motor. Vibration signals from the fault were directly collected by two accelerometers (PCB 352C33 type, vertical and horizontal) with a sampling period of 1 minute and a sampling frequency of 25.6 kHz. Since the sampling time was set to 1.28 seconds, a total of 32,768 samples were recorded per operation. For each operating condition, five datasets were divided into training and test sets; the first four datasets were used as training datasets, and the remaining datasets were used as test datasets. Multi-sensor data from the rotating machinery, including vibration, temperature, and pressure signals, were collected.

[0044] The three different experimental conditions are shown in Table 1 below.

[0045] Table 1 Three different experimental conditions

[0046] Operating condition number 1 2 3 Rotational speed (r / min) 2100 2250 2400 Radial force / kN 12 11 10

[0047] The three different experimental conditions shown in Table 1 are: Condition 1, with an experimental speed of 2100 rpm and a radial force of 12 kN; Condition 2, with an experimental speed of 2250 rpm and a radial force of 11 kN; and Condition 3, with an experimental speed of 2400 rpm and a radial force of 10 kN.

[0048] The relevant parameters of the bearing model LDKUER204 can be found in Table 2 below.

[0049] As shown in Table 2, the bearing of model LDKUER204 has an inner ring raceway diameter of 29.30 mm, an outer ring raceway diameter of 39.80 mm, a bearing pitch diameter of 34.55 mm, a basic rated dynamic load of 12820 N, a ball diameter of 7.92 mm, 8 balls, a contact angle of 0°, and a basic rated static load of 6.65 kN.

[0050] Table 2 shows the relevant parameters of the LDKUER204 bearing.

[0051] Parameter name numerical values Parameter name numerical values Inner ring raceway diameter / mm 29.30 Ball diameter / mm 7.92 Outer ring raceway diameter / mm 39.80 Number of balls 8 Bearing mean diameter / mm 34.55 Contact angle / (°) 0 Basic rated dynamic load / N 12820 Basic rated static load / kN 6.65

[0052] To make the prediction more accurate, as an optional implementation, the signal reflecting mechanical life can be preprocessed before extracting one-dimensional time feature data from the signal reflecting mechanical life using principal component analysis pooling.

[0053] The preprocessing includes at least normalization, which converts signal data that reflects mechanical life into a dimensionless form.

[0054] As an alternative implementation, preprocessing also includes noise reduction to make the resulting signal that reflects mechanical life cleaner and improve the accuracy of prediction.

[0055] The following is combined with Figure 3 Detailed instructions for step S101.

[0056] Figure 3 This is a flowchart illustrating how a correlation vector machine, according to an embodiment of the present invention, divides the entire lifecycle of rotating machinery into periods with no obvious failure trend and periods with a failure tendency. Figure 3 As shown, the data 1 collected by the horizontal sensor and the data 2 collected by the vertical sensor are normalized respectively. Then, Principal Component Analysis (PCA) pooling technology (hereinafter referred to as PCA-Pooling) is used to extract key features from the normalized data 1 collected by the horizontal sensor and the data 2 collected by the vertical sensor to obtain a one-dimensional time series (i.e., 1-D data). Then, the one-dimensional time series is used to train the Relevance Vector Machine (RVM).

[0057] The specific process of 1-D PCA-Pooling includes the following steps S101A to S101F.

[0058] S101A, Sampling operation.

[0059] Specifically, 1) Define the 1-D sampling kernel S K (l, s). Where l represents the length of the sampling kernel and s is the step size.

[0060] 2) The result of the p-th sampling is represented as Z. p ∈R 1×l Among them, R 1×l This represents a matrix with 1 row and 1 column.

[0061] 3) Sampling matrix Z = {Z p |1≤ p ≤P}∈R P×l ,in This represents the maximum number of samples. Where L represents the length of the input channel.

[0062] S101B, Constrain the length of the sampling kernel.

[0063] The length l of the sampling kernel cannot shrink indefinitely, nor can it be too large; the constraint is set as follows:

[0064] S101C, Perform singular decomposition on the sampling matrix Z.

[0065] Perform singular value decomposition (SVD) on the sampling matrix Z: Z = UΛV T .

[0066] Where, U∈R l×l It is an orthogonal matrix composed of left singular vectors; Λ∈R l×P It is a diagonal matrix, with singular values ​​on the diagonal; V∈R P×P It is an orthogonal matrix composed of right singular vectors.

[0067] S101D, dimensionality reduction processing is performed.

[0068] Choose the r largest singular values ​​to approximate the original matrix:

[0069] Where Z′∈R l×P ,

[0070] S101E, further optimized.

[0071] First pass Simplify the time-domain subdivision features. Since each vector is independent, to further optimize the data scale, the mean and variance of each vector need to be consistent.

[0072] Taking the u-th vector as an example, its mini-batch mean is μ. u The variance is σ u Then the new vector should be reshaped as:

[0073]

[0074] S101F, recover the 1-D sequence after dimensionality reduction.

[0075] The dimensionality-reduced 1-D sequence is recovered by stacking the row vectors of the compressed and dimensionality-reduced matrix P.

[0076] Figure 4 This is a schematic diagram illustrating the comparison of data processed using different pooling methods, as provided in an embodiment of the present invention. Figure 4 As shown, Figure 4 In the diagram, (a) represents the original data signal acquired by the sensor. Figure 4 (b) indicates that PCA-Pooling with a step size s of 16 and a maximum number of singular values ​​of 2 is used to process the raw sensor data signal. Figure 4 (c) in the text indicates that the original sensor data signal is processed using Average-Pooling with a step size S of 8.

[0077] Through observation Figure 4 (a) Figure 4 (b) and Figure 4 As can be seen from (c), the pooling result after PCA dimensionality reduction can reduce the amount of data and computational complexity while preserving key features compared to the original data.

[0078] The RVM training using 1-D sequences includes: using a Relevance Vector Machine (RVM) classifier to distinguish mechanical degradation states, classifying them into two states: No Significant Fault Trend (NOFT) and Fault Trend (FP). The RVM classifier is trained within a Bayesian framework, iteratively training the pooled dataset using iterative reweighted least squares, updating the model parameters using Automatic Relevance Determination (ARD) theory, and providing probabilistic classification results.

[0079] To find the optimal hyperparameter combination, grid search and cross-validation were used. Specific parameters included: sampling step size *s*, varying from 2 to 496; and the retained dimension *r*, varying from 1 to 23. A total of 3718 combinations were tested (only valid when *r* ≤ *s*). The XJTU-SY dataset was extracted and divided into three batches, each containing different bearing data. Training and testing: The RVM classifier was independently trained and tested on each batch of data, and training accuracy, testing accuracy, and other evaluation metrics were recorded. Through the above optimization and training process, the performance of the RVM classifier was significantly improved. Specifically, the accuracy of the RVM classifier varied with parameter settings, with the highest accuracy achieved under the optimal parameter combination; its performance was consistent across different datasets, demonstrating good generalization ability and high robustness.

[0080] S102. Use data from the fault-prone period to train a deep separable convolutional gated recurrent unit network to predict the remaining life of rotating machinery during the fault-prone period.

[0081] The Depth Separable Convolutional Gated Recurrent Unit Network (DSCGRN) includes alternating separable convolutional building blocks (SCN) and single-layer gated recurrent unit networks (GRU). The separable convolutional building blocks are used to model the spatial features in multi-channel sequences, while the single-layer gated recurrent unit network is used to mine the temporal dependencies of data according to the acquisition order.

[0082] Specifically, the process of training a deep separable convolutional gated recurrent unit network using data from the fault-prone period to predict the remaining life of rotating machinery during the fault-prone period can be referenced. Figure 5 .

[0083] Figure 5 This is a schematic diagram of the structure of a depth-separable convolutional gated recurrent network provided in an embodiment of the present invention, as shown below. Figure 5 As shown, a depthwise separable convolutional gated recurrent unit network (DSCGRN) is constructed, which consists of alternating separable convolutional networks (SCN) and gated recurrent units (GRU), aiming to learn high-level feature representations from raw multi-sensor data and capture long-term dependencies in time-series data.

[0084] DSCGRN is trained using sample data labeled as FP states in the first stage to assess the mechanical degradation process and predict remaining lifetime. DSCGRN achieves accurate prediction of remaining lifetime through optimized hyperparameter settings (such as feature filtering preservation dimension, sequence length, etc.) and efficient feature extraction capabilities.

[0085] Specifically, it includes the following steps S102A to S102C.

[0086] S102A, PCA-Pooling: Employs PCA-Pooling technology to achieve feature extraction with controllable granularity. This technique can reduce the dimensionality of features while maintaining high accuracy, thereby building lightweight networks and accelerating model training.

[0087] S102B, Feature Filtering Unit: A feature filtering unit is constructed to recalibrate the multi-channel feature representations learned from separable convolutions. In this process, information-rich channel data is enhanced, while useless channel data is filtered out.

[0088] S102C, Alternating Structure Design: An alternating arrangement of DSC and GRU was designed to simultaneously learn high-level feature representations and temporal dependencies. This structure helps to better capture complex patterns in multi-sensor data.

[0089] As an optional implementation, after completing step S102, step S103 can be executed to improve the accuracy of the prediction results.

[0090] S103. The effectiveness and reliability of the prediction accuracy of the depth-separable convolutional gated recurrent unit network are verified by using the scoring function and root mean square error.

[0091] The trained model was applied to predict the remaining useful life (RUL) of actual rotating machinery to verify its effectiveness and reliability. Two quantitative indicators for RUL prediction were evaluated: the adapted RUL scoring function (SCORE) and the root mean square error (RMSE). Given the monitoring samples, both SCORE and RMSE used absolute precision. (See reference...) Figure 6 .

[0092] Figure 6 This is a graph showing the relationship between metrics for evaluating depthwise separable convolutional gated recurrent networks, as provided in an embodiment of the present invention. Figure 6 As shown, the horizontal axis represents the error distance, and the vertical axis represents performance metrics, including SCORE and RMES. The blue line represents SCORE, and the orange line represents RMES. Lower SCORE and RMES scores indicate better RUL predictions. Another metric is C. M It is a convergence rate metric that measures how quickly any test metric (such as accuracy or precision) improves over time. M Defined as the Euclidean distance between the origin of the initial observation and the centroid of the region under the time and prediction error curve.

[0093] Figure 7 This is a schematic diagram of an evaluation result provided by an embodiment of the present invention, such as... Figure 7 As shown, Figure 7In the diagram, (a) represents the SCORE results for the DSCGRN model using different pooling methods. Figure 7 (b) in the figure represents the RMSE results of the DSCGRN model using different pooling methods.

[0094] like Figure 7 (a) and Figure 7 As shown in (b), there are four different DSCGRN models: the DSCGRN model with PCA-Pooling (represented in dark blue), the DSCGRN model without PCA-Pooling (represented in light blue), the DSCGRN model with Average-Pooling (represented in orange), and the DSCGRN model with Max-Pooling (represented in aquamarine).

[0095] Observations show that for the first bearing (Bearing1_5), the second bearing (Bearing2_5), and the third bearing (Bearing3_5), the SCORE and RMSE are lowest when using the DSCGRN model with PCA-Pooling to predict their remaining useful lives. Therefore, the DSCGRN model with PCA-Pooling has the highest accuracy.

[0096] Figure 8 This is a schematic diagram illustrating a prediction of remaining useful life provided by an embodiment of the present invention, as shown below. Figure 8 As shown, Figure 8 In the diagram, (a) represents the prediction result of the DSCGRN model with PCA-Pooling for the first bearing, Bearing1_5. Figure 8 In the diagram, (b) represents the prediction result of the DSCGRN model with PCA-Pooling for the second bearing, Bearing2_5. Figure 8 In the figure, (c) represents the prediction result of the DSCGRN model with PCA-Pooling for the third bearing, Bearing3_5.

[0097] like Figure 8 (a) Figure 8 (b) and Figure 8 As shown in (c), the horizontal axis represents the sample number, the vertical axis represents the remaining useful life (min), the red dashed line represents the ground truth RUL, the blue dots represent the point estimation of RUL from the DSCGRN model, and the pink area represents the 95% confidence interval (Estimation band).

[0098] like Figure 8 As shown in (a), for the first bearing Bearing1_5, the failure threshold occurs near sample number 34, predicting that the remaining service life of the first bearing Bearing1_5 is about 18 minutes.

[0099] like Figure 8 As shown in (b), for the second bearing Bearing2_5, the failure threshold occurs near sample number 184, predicting that the remaining service life of the second bearing Bearing2_5 is about 152 minutes.

[0100] like Figure 8 As shown in (c), for the third bearing Bearing3_5, the failure threshold occurs near sample number 11, predicting that the remaining service life of the third bearing Bearing3_5 is about 100 minutes.

[0101] This invention provides a method and system for predicting the remaining life of rotating machinery. It collects signals reflecting the machinery's lifespan from at least two locations, extracts one-dimensional time feature data from these signals using principal component analysis and pooling, and trains a correlation vector machine using this one-dimensional time feature data to divide the entire lifespan of the rotating machinery into periods with no obvious failure trend and periods with failure tendency, thus reducing the dimensionality of the data and simplifying the dataset. Then, it uses data from the failure tendency period to train a depthwise separable convolutional gated recurrent unit network to predict the remaining lifespan of the rotating machinery during the failure tendency period. This method improves the accuracy of predicting the remaining lifespan of rotating machinery and has significant application value in practical machinery health management, helping to reduce equipment downtime and improve production efficiency and equipment reliability.

[0102] An embodiment of a system for predicting the remaining life of rotating machinery:

[0103] This invention provides a rotating machinery remaining life prediction system, including a processor that executes a computer program to implement the method steps of a rotating machinery remaining life prediction method as described above.

[0104] For the "method steps of a method for predicting the remaining life of rotating machinery", please refer to the relevant description in the aforementioned "Example of a method for predicting the remaining life of rotating machinery", which will not be repeated here.

[0105] The present invention provides a rotating machinery remaining life prediction system that can achieve the same beneficial effects as the aforementioned rotating machinery remaining life prediction method, which will not be repeated here.

Claims

1. A method for predicting the remaining life of rotating machinery, characterized in that, Includes the following steps: Signals reflecting the lifespan of the rotating machinery under test are collected at at least two locations. One-dimensional time feature data is extracted from the signals reflecting the lifespan using principal component analysis pooling. The one-dimensional time feature data is used to train a correlation vector machine to divide the entire lifespan of the rotating machinery into periods with no obvious failure trend and periods with failure tendency. The depthwise separable convolutional gated recurrent unit network is trained using the data from the fault-prone period to predict the remaining life of the rotating machinery during the fault-prone period. The depthwise separable convolutional gated recurrent unit network includes alternating separable convolutional building blocks and a single-layer gated recurrent unit network. The separable convolutional building blocks are used to model the spatial features in multi-channel sequences, and the single-layer gated recurrent unit network is used to mine the temporal dependencies of data according to the acquisition order.

2. The method for predicting the remaining life of rotating machinery according to claim 1, characterized in that, The signals that can reflect the mechanical life include vibration signals, temperature signals, and pressure signals.

3. The method for predicting the remaining life of rotating machinery according to claim 2, characterized in that, The method of acquiring signals reflecting the mechanical life of the rotating machinery under test at two locations includes: acquiring signals reflecting the mechanical life of the rotating machinery under test in a horizontal direction perpendicular to the rotating machinery and in a vertical direction perpendicular to the rotating machinery.

4. The method for predicting the remaining life of rotating machinery according to claim 3, characterized in that, Before using principal component analysis pooling to extract one-dimensional time feature data from the signal that reflects mechanical life, the signal that reflects mechanical life is preprocessed. The preprocessing includes at least: normalization.

5. The method for predicting the remaining life of rotating machinery according to claim 4, characterized in that, The related vector machine is trained iteratively on the pooled dataset using an iterative reweighted least squares method.

6. The method for predicting the remaining life of rotating machinery according to any one of claims 1-5, characterized in that, It also includes: verifying the effectiveness and reliability of the prediction accuracy of the depth-separable convolutional gated recurrent unit network using a scoring function and root mean square error.

7. A system for predicting the remaining life of rotating machinery, comprising a processor, characterized in that, The processor executes a computer program to perform the following steps: Signals reflecting the lifespan of the rotating machinery under test are collected at at least two locations. One-dimensional time feature data is extracted from the signals reflecting the lifespan using principal component analysis pooling. The one-dimensional time feature data is used to train a correlation vector machine to divide the entire lifespan of the rotating machinery into periods with no obvious failure trend and periods with failure tendency. The depthwise separable convolutional gated recurrent unit network is trained using the data from the fault-prone period to predict the remaining life of the rotating machinery during the fault-prone period. The depthwise separable convolutional gated recurrent unit network includes alternating separable convolutional building blocks and a single-layer gated recurrent unit network. The separable convolutional building blocks are used to model the spatial features in multi-channel sequences, and the single-layer gated recurrent unit network is used to mine the temporal dependencies of data according to the acquisition order.

8. The rotating machinery remaining life prediction system according to claim 7, characterized in that, The signals that can reflect the mechanical life include vibration signals, temperature signals, and pressure signals; The method of acquiring signals reflecting the mechanical life of the rotating machinery under test at two locations includes: acquiring signals reflecting the mechanical life of the rotating machinery under test in a horizontal direction perpendicular to the rotating machinery and in a vertical direction perpendicular to the rotating machinery.

9. The rotating machinery remaining life prediction system according to claim 8, characterized in that, Before using principal component analysis pooling to extract one-dimensional time feature data from the signal that reflects mechanical life, the signal that reflects mechanical life is preprocessed. The preprocessing includes at least: normalization; The related vector machine is trained iteratively on the pooled dataset using an iterative reweighted least squares method.

10. The rotating machinery remaining life prediction system according to any one of claims 7-9, characterized in that, The effectiveness and reliability of the prediction accuracy of the depthwise separable convolutional gated recurrent unit network are verified by using a scoring function and root mean square error.

Images