Continuous health prediction method and system for rotating machinery

Abstract

The invention relates to the technical field of mechanical health management, and discloses a continuous health prediction method and system for a rotating machine, and the method comprises the steps: collecting the full-life-cycle operation data of a training rotating part and a target rotating part in the rotating machine, and setting the target rotating part and a satisfaction condition; inputting the initial stage training data into the network model for learning and training, and after training is completed, inputting a target prediction sample for preliminary health prediction; in the second stage, collecting uncollected full life cycle operation data of the training rotating part of the rotating machine, inputting the data into the network model trained in the initial stage for training again, and inputting a target prediction sample for continuous health prediction after the training is completed; the above process is repeated, and data collection and health prediction are carried out continuously in several stages until the prediction performance meets a set condition. The method can gradually improve the health prediction precision of the rotating machine, and provides a new application scene for the residual life prediction of the rotating machine.

Description

Technical Field

[0001] The present invention belongs to the technical field of mechanical health management, and relates to a method and system for continuous health prediction of rotating machinery. Background Art

[0002] Health prediction of rotating machinery is of great significance for improving the operation reliability of industrial equipment. The main goal of rotating machinery health prediction is to estimate its remaining useful life (RUL). Due to its strong non-linear feature capture ability, deep learning (DL) methods can establish an end-to-end mapping from operation data to remaining life without time-consuming feature screening and physical knowledge storage. So far, convolutional neural network (CNN), recurrent neural network (RNN), graph neural network (GNN), Transformer, attention mechanism, etc. have achieved relatively prominent performance in remaining life prediction.

[0003] However, in real industrial scenarios, due to limitations such as storage resources and full-life test cycles, it is often difficult to collect all at once a data set that can characterize the extremely long operation-to-failure process and multiple failure modes. In other words, operation data is usually gradually collected in multiple stages, and the device data collected at one time is limited, resulting in the difficulty for a deep model trained with limited data to represent the complete data distribution and to effectively perform continuous health prediction, thus affecting the prediction accuracy. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for continuous health prediction of rotating machinery, which can utilize data collected in multiple stages to gradually improve the accuracy of remaining life prediction, realize continuous prediction of the health state of rotating machinery, and provide a new application scenario for the remaining life prediction of rotating machinery.

[0005] To solve the above technical problems, the present invention is implemented by adopting the following technical solutions.

[0006] In a first aspect, the present invention proposes a method for continuous health prediction of rotating machinery, including the following steps:

[0007] Set conditions for stopping data collection and continuous health prediction;

[0008] In the initial stage:

[0009] Collect vibration signals of the entire life cycle of the training rotating component and the target rotating component in the rotating machinery until they fail, and obtain the vibration signal data of their entire life cycles;

[0010] Preprocess the vibration signal data of the entire life cycle of the training rotating component collected to obtain training samples in the initial stage;

[0011] Preprocess the vibration signal data of the entire life cycle of the target rotating component collected to obtain target prediction samples;

[0012] Input the obtained training samples in the initial stage into an untrained network model for training to obtain a trained network model. Input the target prediction samples into the trained network model for health prediction to obtain the predicted remaining service life value of the target rotating component in the rotating machinery in the initial stage;

[0013] In the second stage:

[0014] Collect the vibration signals of the entire life cycle of the training rotating component in the rotating machinery that has not been collected before until it fails to obtain the vibration signal data of its entire life cycle;

[0015] Preprocess the vibration signal data of the entire life cycle of the collected training rotating component to obtain training samples in the second stage;

[0016] Input the training samples in the second stage into the network model trained in the initial stage for retraining to obtain a network model trained in this stage. Input the target prediction samples into the network model trained in this stage for health prediction to obtain the predicted remaining service life value of the target rotating component in the rotating machinery in the second stage;

[0017] Repeat the steps of the previous stage until the prediction performance meets the conditions for stopping data collection and continuous health prediction.

[0018] Combined with the first aspect, further, the method for preprocessing the vibration signal data set to obtain samples includes:

[0019] Conduct statistical feature analysis on the vibration signal data of the entire life cycle collected to obtain the frequency domain features and time domain features of the vibration signal of the entire life cycle; Combine the time domain features and frequency domain features through tensor concatenation operations to obtain preprocessed samples.

[0020] It should be noted that: the target prediction samples remain unchanged during the multi-stage continuous health prediction process.

[0021] Combined with the first aspect, further, the method for obtaining the time domain features is: extract features from the time domain signal within the time window to obtain time domain features; The time domain features include maximum value, minimum value, standard deviation, root mean square value, mean value, peak-to-peak value, variance, entropy, arcsine standard deviation, arctangent standard deviation, kurtosis, and skewness.

[0022] In combination with the first aspect, further, the method for obtaining the frequency-domain features is as follows: performing a fast Fourier transform on the time-domain signal within a time window to obtain its frequency-domain signal. By statistically analyzing the frequency spectrum, the frequency-domain features are obtained. The frequency-domain features include: average frequency, median frequency, band power, occupied bandwidth, power bandwidth, maximum power spectral density, maximum amplitude, and the frequency corresponding to the maximum amplitude.

[0023] In combination with the first aspect, further, the conditions for setting to satisfy stopping data collection and continuous health prediction include:

[0024] When the RMSE of a certain stage is less than a and the AF is less than b, the termination condition is satisfied, and data collection and continuous health prediction end;

[0025] Wherein, RMSE is the root mean square error, AF is the average forgetting, and a and b are respectively the set first threshold and second threshold.

[0026] When the RMSE and AF values meet the set conditions, end data collection and continuous health prediction, and output the network model completed in this stage for real-time health prediction of rotating machinery; otherwise, enter the next stage, continue to collect data and perform training, prediction, and evaluation until the set conditions are met.

[0027] In combination with the first aspect, further, using the root mean square error RMSE and the average forgetting AF as prediction performance evaluation indicators, RMSE represents the accuracy of the remaining life. The smaller the RMSE value, the better the prediction performance. The expression is:

[0028]

[0029] In the formula, and respectively represent the true value and the predicted value of the th remaining life, is the total number of remaining lives;

[0030] Set the average forgetting AF as the evaluation indicator for continuous prediction knowledge retention. The lower the AF, the stronger the ability of the network model to retain learning knowledge. The calculation formula is as follows:

[0031] ;

[0032] In the formula, represents the average forgetting of all previous learning knowledge by the network model after completing the training of the th stage; is the final stage; represents the forgetting of the network model of the previous th learning after completing the training of the th stage, which is expressed as:

[0033]

[0034] Wherein, is the RMSE value calculated by inputting the training sample used in the th learning after the network model completes the training of the th stage; represents the RMSE value calculated by inputting the current training sample after the th learning of the network model.

[0035] Combined with the first aspect, further, the continuous health prediction method of the present invention further includes the following steps: In each stage, specify a training rotating part as a memory test rotating part in the training rotating parts. After the network model in each stage is trained, the training samples of this memory test rotating part need to be input into the network model trained in this stage to calculate its remaining life, which is used to calculate AF.

[0036] Combined with the first aspect, further, in the initial stage, the specific method for training the untrained network model is:

[0037] During the training process, it is necessary to minimize the prediction loss to optimize the parameters of the network model in the initial stage, and the training data set in the initial stage is given , , are the th training sample and the remaining life label respectively, and the th remaining life label is the true value of the th remaining life. The learning process is:

[0038]

[0039] Wherein, represents the prediction loss; is the regularization loss, and its function is to ensure that the network model can learn in future stages, is the weight parameter that the network model can learn.

[0040] Combined with the first aspect, further, input the target prediction sample into the trained network model to obtain the predicted value of the remaining life of the target rotating part. The expression is:

[0041] ;

[0042] Wherein, is the th remaining life prediction value; is a network model; is the th prediction sample; is the target prediction sample set.

[0043] Combined with the first aspect, further, in the second stage, the specific method for retraining the network model trained in the previous stage is:

[0044] Collect the training data set of the stage as , and through the introduction of the memory loss term, the learning and training process of the network model is:

[0045] ;

[0046] In the formula, is the memory loss term, including the weight preservation loss and the memory replay loss ; is the replay buffer for storing samples from the previous initial stage to the stage; represents the prediction loss; is the regularization loss.

[0047] In the second aspect, the present invention proposes a rotating machinery continuous health prediction system, including:

[0048] A condition setting module configured to set conditions for satisfying the stop of data collection and continuous health prediction;

[0049] A prediction module configured to, in the initial stage:

[0050] Collect the vibration signals of the training rotating components and the target rotating components in the rotating machinery during the entire life cycle until failure, and obtain the vibration signal data of the entire life cycle of both;

[0051] Perform data preprocessing on the vibration signal data of the entire life cycle of the collected training rotating components to obtain the training samples in the initial stage;

[0052] Perform data preprocessing on the vibration signal data of the entire life cycle of the collected target rotating components to obtain the target prediction samples;

[0053] Input the obtained training samples in the initial stage into the untrained network model for training to obtain the trained network model, input the target prediction samples into the trained network model for health prediction, and obtain the predicted remaining service life value of the target rotating components in the rotating machinery in the initial stage;

[0054] In the second stage:

[0055] Collect the vibration signals of the training rotating components in the rotating machinery that have not been collected during their entire life cycle until failure, and obtain the vibration signal data of their entire life cycle;

[0056] Perform data preprocessing on the vibration signal data of the entire life cycle of the collected training rotating components to obtain the training samples in the second stage;

[0057] Input the training samples in the second stage into the network model that has been trained in the initial stage for retraining, obtain the network model that has been trained in this stage, input the target prediction samples into the network model that has been trained in this stage for health prediction, and obtain the predicted remaining service life value of the target rotating component in the rotating machinery in the second stage;

[0058] Repeat the steps of the previous stage until the prediction performance meets the conditions for stopping data collection and continuous health prediction.

[0059] In a third aspect, the present invention proposes a computer-readable storage medium, on which a computer program is stored, characterized in that: when the computer program is executed by a processor, the steps of the above-mentioned continuous health prediction method for rotating machinery are implemented.

[0060] In a fourth aspect, the present invention proposes a computer device, characterized by comprising:

[0061] A memory for storing a computer program;

[0062] A processor for executing the computer program to implement the steps of the above-mentioned continuous health prediction method for rotating machinery.

[0063] In a fifth aspect, the present invention proposes a computer program product, comprising a computer program, characterized in that: when the computer program is executed by a processor, the steps of the above-mentioned continuous health prediction method for rotating machinery are implemented.

[0064] Compared with the prior art, the beneficial effects achieved by the present invention:

[0065] The present invention can process the operation data collected gradually; through multi-stage data, the prediction performance can be gradually enhanced; a memory module is introduced to complete the continuous accumulation of multi-stage knowledge; the continuous prediction of the remaining life is realized, providing a new method and application scenario for the health prediction of rotating machinery. Description of the Drawings

[0066] Figure 1 It is a schematic flowchart of the continuous health prediction of rotating machinery in Embodiment 1 of the present invention;

[0067] Figure 2 It is a graph of the predicted remaining life in Embodiment 1 of the present invention;

[0068] Figure 3 This is the memory effect diagram for continuous remaining life prediction in Embodiment 1 of the present invention. Detailed implementation manners

[0069] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. Without conflict, the technical features in the embodiments of the present invention and the embodiments can be combined with each other.

[0070] The term "and / or" is merely a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone. In addition, the character " / " generally represents an "or" relationship between the associated objects before and after.

[0071] Embodiment 1

[0072] As Figure 1 shown, a method for continuous health prediction of a rotating machine in this embodiment includes the following steps:

[0073] S1. Collect the vibration signals of the whole life cycle of several rotating components of the rotating machine through an acceleration sensor;

[0074] Using the sensor system of the rotating machine, collect and record the vibration signals of the characteristic factor j (j = 1, 2,..., J) of the rotating components (such as bearings, gears, pumps, etc.) i (i = 1, 2,..., I) within the time period [0, T] (T is the fault occurrence time) through the acceleration sensor, and construct time series data {x ij (1), x ij (2), …,x ij (n),…,x ij (N)} based on the vibration signals. Wherein, I represents the total number of rotating components of the rotating machine, J represents the number of characteristic factors of faults occurring within the time period T, N represents the number of constructed time series (corresponding to the number of remaining lives described below), and x ij (n) represents the operation time series data of the i-th key component with the characteristic factor j of the fault occurring within the time period T, and n ∈ [1, N].

[0075] It should be noted that: In the initial stage, vibration signal data of the entire life cycle of at least two rotating components are collected. The vibration signal data of one rotating component (referred to as the training rotating component) of the entire life cycle are used as training samples, and the vibration signal data of one rotating component (referred to as the target rotating component) of the entire life cycle are used as target prediction samples. In each stage after the initial stage, only the vibration signal data of the training rotating component can be collected and used as training sample data. The target prediction samples in the initial stage are used for the prediction performance evaluation after the network model training in each subsequent stage, that is, the prediction performance evaluation process in each stage after the initial stage shares the target prediction samples in the initial stage.

[0076] S2. Conduct statistical feature analysis on the vibration signal to obtain its time-domain and frequency-domain features, and divide the preprocessed data;

[0077] S21. Extract features from the time-domain signal within the time window to obtain time-domain features, mainly including the following features:

[0078] Maximum value, minimum value, standard deviation, root mean square value, mean value, peak-to-peak value, variance, entropy, arcsine standard deviation, arctangent standard deviation, kurtosis, skewness;

[0079] S22. Perform fast Fourier transform on the time-domain signal within the time window to obtain its frequency-domain signal. Through statistical analysis of the frequency spectrum, the following frequency-domain features are obtained:

[0080] Average frequency, median frequency, band power, occupied bandwidth, power bandwidth, maximum power spectral density, maximum amplitude, frequency corresponding to the maximum amplitude.

[0081] S23. Combine the time-domain features and frequency-domain features to obtain the preprocessed data for comprehensively characterizing the statistical characteristics of the signal.

[0082] According to the different rotating components, the preprocessed data samples are divided into training samples and target prediction samples, and the target prediction samples remain unchanged during the multi-stage continuous health prediction process.

[0083] Furthermore, step S2 also includes the step of making remaining life labels. After constructing the time series data {x ij (1), x ij (2), …, x ij (n), …, x ij (N)} based on the collected vibration signals, the number of remaining life labels can be expressed as (N - n) / N, that is, the total number of time series (remaining life) minus the number of a certain time series (remaining life), and then divided by the total number of time series (remaining life). There is a relationship: the number of time series = the number of training samples = the number of remaining life labels (the number of remaining life).

[0084] S3. Set the satisfaction conditions for continuous health prediction;

[0085] Taking the root mean square error (RMSE) and average forgetting (AF) as performance evaluation indicators, the satisfaction conditions are set as:

[0086] When the RMSE of a certain stage is less than a and the AF is less than b, the termination condition is satisfied, and data collection and continuous health prediction end.

[0087] Preferably, a and b are set thresholds, which are set according to expert experience or the requirements of technicians.

[0088] Specifically, the calculation processes of RMSE and AF are as follows.

[0089] S31. RMSE represents the accuracy of the remaining life. The smaller the RMSE value, the better the prediction performance. The expression is:

[0090]

[0091] In the formula, and respectively represent the true value and predicted value of the th remaining life. is the total number of remaining lives. Additionally, , that is, one time series corresponds to one remaining life value; the true value of the remaining life is obtained by subtracting the actual running time from the running-to-failure time (total life).

[0092] S32. While continuous health prediction realizes the prediction of the data life in the current stage, it is necessary to retain the prediction ability for the data in previous stages. Therefore, the average forgetting AF is set as an evaluation index for the retention of continuous prediction knowledge. The lower the AF, the stronger the ability of the network model to retain learned knowledge. The calculation formula is as follows:

[0093]

[0094] In the formula, represents the average forgetting of all previous learned knowledge by the network model after completing the training of the th stage. is the final stage. represents the forgetting of the th previous learning by the network model after completing the training of the th stage, which can be expressed as:

[0095]

[0096] In the formula, is the input of the network model after completing the training of the th stage, and the RMSE value calculated from the training samples of the memory test rotating component used in the current learning; Indicates that after the th learning, the RMSE value calculated by inputting the current training sample into the network model that has completed training in the current stage exists .

[0097] In the present invention, the true remaining life value and the predicted value for calculating RMSE can be the target predicted rotating component or the memory test rotating component in the training rotating component; calculating the RMSE of the target rotating component is used for evaluating the prediction performance of the network model; after calculating the RMSE of the memory test rotating component, AF is calculated, which is used for evaluating the memory performance of the network model.

[0098] S4. Input the training samples in the initial stage into the network model for training and learning. After the training is completed, input the target prediction samples for health prediction to obtain the remaining useful life (RUL) of the target rotating component;

[0099] During the training process, it is necessary to minimize the prediction loss to optimize the network model parameters in the initial stage, and a given initial stage training data set , , are respectively the th training sample and the remaining life label (since one sample input into the network model corresponds to an output of one remaining life value, the remaining life label is the true remaining life value). The learning process is as follows:

[0100]

[0101] In the formula, represents the prediction loss; is the regularization loss, and its function is to ensure that the network model is learnable in the future stage, is the learnable weight parameter of the network model.

[0102] After that, input the target prediction samples into the trained network model to obtain the predicted value of the remaining life of the target rotating component. The expression is:

[0103] ;

[0104] In the formula, is the th remaining life prediction value of the target rotating component in the initial stage; is the target prediction sample set. After obtaining the remaining life value of the target rotating component in the initial stage, the RMSE value in the initial stage can be calculated through step S31.

[0105] S5. Next, collect the vibration signal data of the training rotating components that have not been collected in the previous stage until failure (because the training rotating components in the previous stage have failed and cannot be used normally), and conduct continuous health prediction;

[0106] S51. Retrain the network model. Retrain the network model trained in the initial stage using the data in the second stage. While minimizing the prediction loss of the current data, the prediction ability for the data in the initial stage needs to be retained. From the second stage, introduce a memory module, including two strategies: weight preservation and memory replay. Collect the training dataset for the stage as , and obtain its time-frequency domain features through data preprocessing. Then, by introducing a memory loss term, the learning and training process of the network model is:

[0107]

[0108] where is the memory loss term, including the weight preservation loss and the memory replay loss ; is the replay buffer for storing samples from the previous initial stage to the stage.

[0109] In the above formula, the memory loss term includes the weight preservation loss and the memory replay loss , which are calculated respectively as:

[0110] S511. Weight preservation loss . The learnable parameters of the network model play a decisive role in prediction. To maintain the prediction ability for the data in the previous stage, when learning the data of the current th stage, the update of the important parameters for the prediction of the previous stages can be restricted. The important parameters are those that have a greater impact on the prediction performance of the previous stages, which can be represented by the importance of the parameters, that is, the following . Specifically, by penalizing the change of important parameters, the update of the important parameters of old knowledge is restricted. The weight preservation loss for learning the current th stage is:

[0111]

[0112] where represents the element-wise multiplication operation; is the The multi-dimensional gradient vector of the importance of the network model parameters in a stage, where the parameter gradient is obtained by back extrapolation from the calculated prediction loss, and the higher the importance of the parameter, the slower the update of the parameter; For the optimal parameters of the model trained in the previous stage.

[0113] S512, Memory replay loss . The memory replay strategy establishes a buffer for storing samples of the previous dataset and performs sample replay training when learning a new task. Therefore, the loss of memory replay can be expressed as the training loss of the samples in the buffer:

[0114]

[0115] S52, Health prediction. In the stage, when the network model training is completed, the target prediction sample set is input into the network model for prediction, and the remaining life prediction value of the target rotating component in the stage is:

[0116]

[0117] where, is the th remaining life prediction value of the target rotating component in the stage; is the th prediction sample. After obtaining the remaining life value of the target rotating component in the stage, the RMSE value of the current stage is calculated through step S31.

[0118] In addition, the training dataset from the previous initial stage to the th stage (i.e., the data of the memory test rotating component) is sequentially input into the current network model, and the remaining life prediction values of the training rotating components from the initial stage to the th stage are:

[0119]

[0120] After obtaining the remaining life prediction values of the training rotating components from the initial stage to the th stage , the is calculated through S31, that is, the calculation process corresponding to in S32, and finally the value of the performance evaluation index AF is obtained.

[0121] S6, Health prediction performance calculation and evaluation;

[0122] After obtaining the true value of the remaining life in the current stage the RMSE value of the network model is calculated through the performance evaluation index formula in the conditions set in step S3.

[0123] After calculating the predicted values of the remaining life in each previous stage the RMSE values in each previous stage are first calculated through the performance evaluation index formula in the conditions set in step S3, and finally the AF value of the network model can be calculated.

[0124] When both the RMSE and AF values meet the conditions set in step S3, the data collection and health prediction are terminated, and the network model trained in this stage is output for real-time health prediction of rotating machinery; otherwise, enter the next stage, continue to collect data and perform training, prediction, and evaluation until the conditions set in step S3 are met.

[0125] In a specific implementation manner of this embodiment, the network model of the present invention can be a traditional deep learning neural network, such as a convolutional neural network, a long short-term memory network, a graph neural network, etc., or other improved effective network models can also be used.

[0126] It should be noted that: in the present invention, the initial stage, the second stage... the stage is divided according to the time axis of the operation of the rotating machinery. The training rotating components in each stage can be different types of rotating components (for example, the training rotating component in the initial stage is a bearing, the target rotating component is another bearing, and the training rotating component in the second stage is a gear...). Because the method of the present invention can continuously learn the degradation process of rotating components with different data distributions and is not easy to forget the degradation knowledge of the previously learned rotating components. However, it should be noted that: compared with the same type of rotating components, the data distribution differences between different types of rotating components are relatively large. It may be necessary to use a network model with a larger capacity to increase the fault tolerance of learning knowledge, or to use transfer learning technology to solve the problem of large data distribution differences. However, after learning the degradation modes of multiple types of rotating components, the network model can be more widely used in actual applications. For example, using this network model in industry can predict the remaining life of multiple types of rotating components.

[0127] Preferably, the rotating components in each stage are of the same type. For example, the training rotating components in each stage are all bearings, and the target rotating component is also a bearing. The reason is that: compared with different types of rotating components, the data distribution differences of the same type of rotating components are relatively small, and the method of the present invention can form good knowledge accumulation for the degradation processes of multiple same type of rotating components.

[0128] The following further illustrates the solution and effect of the present invention through specific application examples.

[0129] In this embodiment, the full life cycle dataset of rolling bearings collected by a certain dynamics and mechatronics (LDM) laboratory is used as the research object to illustrate the effectiveness of the method of the present invention. The full life cycle dataset of rolling bearings records the data of 17 rolling bearings running to failure under time-varying working conditions, and the numbers of the rolling bearings are B01 - B17. Using the leave-one-out method, one rolling bearing is set as the target prediction bearing, and the remaining rolling bearings are used as training bearings. Based on this, three groups of prediction tasks are set, denoted as A, B, and C respectively. In order to simulate the scenario of continuous health prediction, in the ascending order of the rolling bearing numbers, each group of tasks is sequentially set with 8 stages of incremental training sample data and common target prediction sample data. In addition, a training bearing is set as a memory test bearing in each stage. The memory test bearing is used to test the memory performance of the network model, that is, to calculate AF (the memory test bearing is taken out from the training bearings in each stage. After the network model is trained in each subsequent stage, the data of this memory test bearing needs to be input to calculate its remaining life, aiming to see the difference in its remaining life value from the previous stage, so as to judge the degree of forgetting of the network model for it). While adopting the method of the present invention, the fine-tuning method and the joint training method are also used for comparative illustration. Among them, the fine-tuning method is to only use the deep network model and fine-tune the model with the newly obtained training sample data in each stage, which is usually used as a performance lower limit for reference; joint training is to add the newly added training sample data in each stage to the existing data for joint training, which is usually used as a performance upper limit for reference. During the prediction process, the data sample values and the remaining life values are normalized between 0 and 1, so the obtained remaining life result is the percentage of the remaining life (remaining life value / total life value). Table 1 shows the specific information for arranging the continuous health prediction tasks, as follows.

[0130]

[0131] Table 2 shows the calculation metrics of the prediction results, as follows.

[0132]

[0133] In this embodiment, in order to simulate the continuous health prediction scenario, multi-stage incremental learning tasks are arranged on the entire full life cycle dataset of rolling bearings. A total of three groups of prediction tasks are set, and each group includes eight stages. The specific task information is shown in Table 1. The remaining life prediction results in several stages during the continuous health prediction process are as Figure 2As shown, Figures (2a) to (2d) are the remaining life prediction results of Bearing B04 in the second, fifth, sixth, and seventh stages respectively, Figures (2e) to (2h) are the remaining life prediction results of Bearing B08 in the second, fourth, fifth, and eighth stages, and Figures (2i) to (2l) are the remaining life prediction results of Bearing B10 in the second, fourth, fifth, and sixth stages. From Figure 2 it can be seen that as multi-stage incremental data is used for training, the remaining useful life (RUL) prediction curve gradually approaches the actual RUL curve, indicating that the method of the present invention for continuous health prediction is feasible and effective. The calculated continuous remaining life prediction performance indicators are shown in Table 2. From Table 2, it can be seen that the method of the present invention effectively improves the accuracy of continuous remaining life prediction, and the prediction performance is reasonably between the upper and lower limits. Further, the continuous health prediction memory performance of the method of the present invention is as Figure 3 shown, where Figures (3a) to (3d) are the health prediction memory performance of Bearing B01 in the first (i.e., initial), second, third, and fourth stages respectively, Figures (3e) to (3h) are the health prediction memory performance of Bearing B06 in the third, sixth, seventh, and eighth stages, and Figures (3i) to (3l) are the health prediction memory performance of Bearing B04 in the second, fourth, sixth, and eighth stages. Observing Figure 3 it can be seen that the method of the present invention still has a high accuracy in predicting the remaining life of previous stages, and has strong memory ability and knowledge accumulation ability.

[0134] Embodiment 2

[0135] Based on the same inventive concept as Embodiment 1, this embodiment introduces a continuous health prediction system for rotating machinery, including:

[0136] A condition setting module configured to set conditions for stopping data collection and continuous health prediction;

[0137] A prediction module configured to, in the initial stage:

[0138] Collect vibration signals of the entire life cycle of the training rotating component and the target rotating component in the rotating machinery until failure, and obtain the vibration signal data of their entire life cycles;

[0139] Perform data preprocessing on the vibration signal data of the entire life cycle of the collected training rotating component to obtain training samples in the initial stage;

[0140] Perform data preprocessing on the vibration signal data of the entire life cycle of the collected target rotating component to obtain target prediction samples;

[0141] Input the obtained training samples in the initial stage into an untrained network model for training to obtain a trained network model. Input the target prediction samples into the trained network model for health prediction to obtain the predicted remaining service life value of the target rotating component in the rotating machinery at the initial stage.

[0142] In the second stage:

[0143] Collect the vibration signals of the entire life cycle of the uncollected training rotating components in the rotating machinery until they fail to obtain the vibration signal data of their entire life cycle.

[0144] Perform data preprocessing on the vibration signal data of the entire life cycle of the collected training rotating components to obtain the training samples in the second stage.

[0145] Input the training samples in the second stage into the network model trained in the initial stage for retraining to obtain the network model trained in this stage. Input the target prediction samples into the network model trained in this stage for health prediction to obtain the predicted remaining service life value of the target rotating component in the rotating machinery at the second stage.

[0146] Repeat the steps of the previous stage until the prediction performance meets the conditions for stopping data collection and continuous health prediction.

[0147] Embodiment 3

[0148] Based on the same inventive concept as other embodiments, this embodiment introduces a computer-readable storage medium with a computer program stored thereon. When the computer program is executed by a processor, the steps of the above-mentioned continuous health prediction method for rotating machinery are implemented.

[0149] Embodiment 4

[0150] Based on the same inventive concept as other embodiments, this embodiment introduces a computer device, including: a memory for storing a computer program; a processor for executing the computer program to implement the steps of the above-mentioned continuous health prediction method for rotating machinery.

[0151] Embodiment 5

[0152] Based on the same inventive concept as other embodiments, this embodiment introduces a computer program product, including a computer program. When the computer program is executed by a processor, the steps of the above-mentioned continuous health prediction method for rotating machinery are implemented.

[0153] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0154] The present invention is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present invention. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices produce means for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0155] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means that implement the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0156] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0157] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms, and these all fall within the protection scope of the present invention.

Claims

1. A method for predicting the continuous health of rotating machinery, characterized in that: The following steps are involved: Setting the conditions for stopping data collection and continuing health prediction; In the initial stage: Collect vibration signals of the training rotating parts and the target rotating parts in the rotating machinery during their entire life cycle until failure, and obtain vibration signal data of the entire life cycle of both; Perform data preprocessing on the collected vibration signal data of the training rotating parts throughout their life cycle to obtain training samples in the initial stage; Perform data preprocessing on the collected vibration signal data of the target rotating parts throughout their life cycle to obtain target prediction samples; Input the obtained training samples in the initial stage into the untrained network model for training to obtain a trained network model, input the target prediction samples into the trained network model for health prediction, and obtain the remaining service life prediction value of the target rotating part in the rotating machinery in the initial stage; In the second stage: Collect the vibration signals of the training rotating parts in the rotating machinery that have not been collected before, and obtain the vibration signal data of the whole life cycle; Perform data preprocessing on the collected vibration signal data of the training rotating parts during the entire life cycle to obtain training samples for the second stage; The training samples of the second stage are input into the network model trained in the initial stage for retraining to obtain the network model trained in this stage, and the target prediction samples are input into the network model trained in this stage for health prediction to obtain the remaining service life prediction value of the target rotating parts in the rotating machinery in the second stage; Repeat the steps of the previous stage until the prediction performance meets the conditions for stopping data collection and continuing healthy prediction.

2. The method for predicting the continuous health of rotating machinery according to claim 1, characterized in that: The settings satisfy the conditions for stopping data collection and continuing health prediction, including: When the RMSE of a certain stage is less than a and the AF is less than b, the termination condition is met, and data collection and continuous health prediction are completed; Among them, RMSE is the root mean square error, AF is the average forgetting, a and b are the set first threshold and second threshold.

3. The method for predicting the continuous health of rotating machinery according to claim 2, characterized in that: The root mean square error RMSE and average forgotten AF are used as prediction performance evaluation indicators. RMSE represents the accuracy of the remaining life span. The smaller the RMSE value, the better the prediction performance. The expression is: ； In the formula, and Respectively represent The true and predicted values ​​of the remaining life, is the total number of remaining lifespan; The average forgotten AF is set as the evaluation index for continuous prediction of knowledge retention. The lower the AF, the stronger the ability of the network model to retain learning knowledge. The calculation formula is as follows: ； In the formula, Indicates that the network model is completed The average forgetting of all previously learned knowledge after a phase of training; For the final stage; Indicates that the network model is completed After the first stage of training, The forgetting of learning is expressed as: ； In the formula, For the network model, complete the After the first stage of training, input The RMSE value calculated from the training samples used in the learning; Indicates that the network model is After learning, input the RMSE value calculated from the current training sample.

4. The method for predicting the continuous health of rotating machinery according to claim 1, characterized in that: In the initial stage, the specific method for training an untrained network model is: During the training process, it is necessary to minimize the prediction loss to optimize the network model Parameters in the initial stage, given the initial stage training data set , , Respectively training samples and remaining lifetime labels, The remaining life label is The true value of the remaining life, the learning process is: ； In the formula, represents the predicted loss; is the regularization loss, which ensures that the network model In the future stage, it is possible to learn. is the weight parameter that the network model can learn.

5. The method for predicting the continuous health of rotating machinery according to claim 1, characterized in that: The target prediction sample is input into the trained network model to obtain the predicted value of the remaining life of the target rotating component, which is expressed as: ； In the formula, For the The remaining life prediction value; For the network model; For the prediction samples; Predict the sample set for the target.

6. The method for predicting the continuous health of rotating machinery according to claim 4, characterized in that: In the second stage, the specific method for retraining the network model trained in the previous stage is: Collect The training data set is , by introducing the memory loss term, the learning and training process of the network model is: ； In the formula, is the memory loss term, including weight preservation loss and memory replay loss ; To store the previous initial stage to the Playback buffer for stage samples; represents the predicted loss; is the regularization loss.

7. A method system for predicting the continuous health of rotating machinery, characterized in that: include: A condition setting module, configured to set conditions for stopping data collection and continuing health prediction; The prediction module is configured to: Collect vibration signals of the training rotating parts and the target rotating parts in the rotating machinery during their entire life cycle until failure, and obtain vibration signal data of the entire life cycle of both; Perform data preprocessing on the collected vibration signal data of the training rotating parts throughout their life cycle to obtain training samples in the initial stage; Perform data preprocessing on the collected vibration signal data of the target rotating parts throughout their life cycle to obtain target prediction samples; Input the obtained training samples in the initial stage into the untrained network model for training to obtain a trained network model, input the target prediction samples into the trained network model for health prediction, and obtain the remaining service life prediction value of the target rotating part in the rotating machinery in the initial stage; In the second stage: Collect the vibration signals of the training rotating parts in the rotating machinery that have not been collected before, and obtain the vibration signal data of the whole life cycle; Perform data preprocessing on the collected vibration signal data of the training rotating parts during the entire life cycle to obtain training samples for the second stage; The training samples of the second stage are input into the network model trained in the initial stage for retraining to obtain the network model trained in this stage, and the target prediction samples are input into the network model trained in this stage for health prediction to obtain the remaining service life prediction value of the target rotating parts in the rotating machinery in the second stage; Repeat the steps of the previous stage until the prediction performance meets the conditions for stopping data collection and continuing healthy prediction.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for predicting the continuous health of a rotating machinery described in any one of claims 1 to 6 are implemented.

9. A computer device, characterized in that: include: Memory for storing computer programs; A processor is used to execute the computer program to implement the steps of the rotating machinery continuous health prediction method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for predicting the continuous health of a rotating machine according to any one of claims 1 to 6 are implemented.

Images