Data and model hybrid driven mechanical component remaining useful life prediction method

Abstract

The application discloses a data and model hybrid driving mechanical component residual service life prediction method, and relates to the field of intelligent manufacturing and health management of equipment. The method adopts extended Kalman filtering to calibrate parameters of an exponential random model, learns position information of input embedding through an adaptive coding layer of a hybrid driving prediction model, and then models a mapping relationship between input data and residual service life through a multi-head self-attention mechanism. The application combines the calibrated exponential random model and the multi-head attention neural network structure, simultaneously retains accuracy of a model-based method and generalization ability of a data-driven method, can improve the accuracy of residual service life prediction of mechanical components, and has important significance for application of the data / model hybrid driving method in the field of intelligent manufacturing and health management of mechanical equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment intelligent manufacturing and health management, and particularly relates to a data and model hybrid driving mechanical component remaining useful life prediction method. BACKGROUND

[0002] In the field of mechanical component state prediction and health management, the future degradation process and the remaining useful life (RUL) of a system are estimated based on state monitoring data. However, the degradation process of a mechanical component contains various uncertainties, and different devices may have different degradation trajectories. Different operating conditions and environmental conditions can also lead to different degradation processes. For example, in the degradation process of a rolling bearing, the type of bearing, the running speed and load, and the type of bearing failure can all cause differences in the degradation process, bringing greater uncertainty to the prediction of the degradation process of the bearing. Therefore, it is necessary to study a remaining useful life prediction method that has high prediction accuracy and is less affected by uncertainties. At present, remaining useful life prediction methods are mainly divided into two categories: data-driven and model-driven methods.

[0003] On the one hand, traditional model-based remaining useful life prediction methods use mathematical derivation or rely on fault domain knowledge to establish a model of the degradation process of a system, such as the Paris law of crack propagation. Then, statistical signal processing techniques such as Kalman filtering, extended Kalman filtering (EKF), and particle filtering are used to identify and calibrate the model parameters, and then the remaining useful life is predicted. However, in practical applications, it is difficult to obtain the degradation and failure mechanisms of different systems, and the use of model-based methods may face certain difficulties. However, the physical understanding in the model-based method can improve the interpretability of the prediction research and improve the accuracy of the prediction.

[0004] On the other hand, the data-driven method uses state monitoring data combined with machine learning techniques to train a prediction model, and the trained model can be used for the estimation of the remaining useful life. At present, with the emergence of large-scale sensor data and system health monitoring data, the method of neural network and deep learning is hot, such as convolutional neural network, recurrent neural network, long short-term memory (LSTM) and transformer network model are applied in the field of life prediction. Data-driven methods have many advantages, such as data-driven methods can directly model the complex relationship between the degradation process and the historical observation data, without applying much domain knowledge, which can reduce the cost of application; and the learning-based artificial intelligence method can greatly reduce the model error and improve the accuracy of model prediction. However, at present, the data-driven method still has some shortcomings, such as the lack of physical understanding of the degradation process leads to the lack of interpretability of the data-driven model; the measurement error and noise in the original sensor data lead to poor prediction accuracy of the model. SUMMARY

[0005] In view of the problems in the above background art, the present application provides a data and model hybrid driven mechanical component remaining useful life prediction method to retain the generalization ability of the data-driven method and the fidelity of the model-driven method at the same time, thereby improving the accuracy of the mechanical component remaining useful life prediction.

[0006] To achieve the above object, the present application provides the following scheme:

[0007] In one aspect, the present application provides a data and model hybrid driven mechanical component remaining useful life prediction method, comprising:

[0008] An exponential random model is used to model the degradation process of the mechanical component and establish a system state space equation;

[0009] Based on the system state space equation, the exponential random model is parameter estimated by an extended Kalman filter to obtain an optimal state estimation;

[0010] The state monitoring data of the degradation stage of the mechanical component is obtained based on the first prediction time, and the fast Fourier transform is used to obtain the frequency domain data corresponding to the degradation stage state monitoring data;

[0011] The neural network training data set of all mechanical components is constructed according to the optimal state estimation and the frequency domain data;

[0012] A hybrid driven prediction model is established, including a full connection layer, a one-dimensional convolution long short-term memory network adaptive coding layer, a multi-head attention mechanism module, a feedforward module and a full connection regression layer;

[0013] The hybrid driving prediction model is trained and tested by using the neural network training data set, and a trained hybrid driving prediction model is obtained.

[0014] The trained hybrid driving prediction model is used to predict the remaining useful life of the mechanical component.

[0015] Optionally, the exponential random model is used to model the degradation process of the mechanical component and establish a system state space equation, specifically comprising:

[0016] The exponential random model is used Model the degradation process of the mechanical component; wherein a k ,b k is a parameter related to the health state of the mechanical component in the degradation process;

[0017] Based on the exponential random model Establish a system state space equation Wherein the state vector at time k f and h are nonlinear functions; x k-1 is the state vector at time k-1; u k-1 is the system input at time k-1; w k-1 is a random 0-mean error at time k-1; z k is the measurement at time k; v k is the measurement error at time k.

[0018] Optionally, based on the system state space equation, the parameters of the exponential random model are estimated by an extended Kalman filter to obtain the optimal state estimation, specifically comprising:

[0019] Based on the system state space equation, the nonlinear functions f k and h k about the state prior estimation Local linearization is performed to obtain the corresponding Jacobian matrices F k and H k ;

[0020] The prediction and update equations of the extended Kalman filter are established according to the Jacobian matrices F k and H k ;

[0021] Based on the prediction and update equations, the prediction and update processes of the extended Kalman filter are alternately performed to continuously update the predicted state vector and obtain the optimal state estimation.

[0022] Optionally, the state monitoring data of the degradation stage of the mechanical component is obtained based on the first prediction time, and the frequency domain data corresponding to the degradation stage state monitoring data is obtained by using fast Fourier transform, specifically comprising:

[0023] determining a first prediction time according to the original condition monitoring data of the mechanical components collected by the sensor;

[0024] extracting condition monitoring data of a degradation stage of the mechanical components based on the first prediction time;

[0025] extracting frequency domain information of the condition monitoring data of the degradation stage by using fast Fourier transform to obtain frequency domain data of the degradation stage.

[0026] Optionally, the neural network training data set of all mechanical components is constructed according to the optimal state estimation and the frequency domain data, and specifically includes:

[0027] optimal state estimation of the degradation stage of the mechanical components and frequency domain data constructing a neural network training data set of all mechanical components wherein t i is the first prediction time; n i is the variance feature sequence length; Q is the number of mechanical components i; is the remaining service life of the mechanical component i at time k.

[0028] In another aspect, the present application provides a data and model hybrid driven mechanical component remaining service life prediction system, comprising:

[0029] a degradation model establishing module for modeling the degradation process of the mechanical components by using an exponential random model and establishing a system state space equation;

[0030] an extended Kalman filter module for parameter estimation of the exponential random model by using an extended Kalman filter based on the system state space equation to obtain optimal state estimation;

[0031] an FFT feature extraction module for obtaining condition monitoring data of a degradation stage of the mechanical components based on a first prediction time, and obtaining frequency domain data corresponding to the condition monitoring data of the degradation stage by using fast Fourier transform;

[0032] a training data set construction module for constructing a neural network training data set of all mechanical components according to the optimal state estimation and the frequency domain data;

[0033] a hybrid driven prediction model establishing module for establishing a hybrid driven prediction model, including a full connection layer, a one-dimensional convolution long short memory network adaptive coding layer, a multi-head attention mechanism module, a feedforward module and a full connection regression layer;

[0034] The hybrid driving prediction model training module is configured to train and test the hybrid driving prediction model by using the neural network training data set, and obtain a trained hybrid driving prediction model.

[0035] The remaining useful life prediction module is configured to predict the remaining useful life of the mechanical component by using the trained hybrid driving prediction model.

[0036] In another aspect, the present application also provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the data and model hybrid driving mechanical component remaining useful life prediction method when executing the computer program.

[0037] In another aspect, the present application also provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executed to implement the data and model hybrid driving mechanical component remaining useful life prediction method.

[0038] According to the specific embodiments of the present application, the following technical effects are provided:

[0039] The data and model hybrid driving mechanical component remaining useful life prediction method provided by the present application retains the accuracy of the model-based method and the generalization ability of the data-driven method, calibrates the parameters of the exponential random model by using the extended Kalman filter, automatically learns the position information of the input embedding through the one-dimensional convolution long short memory network adaptive coding layer, then models the mapping relationship between the input data and the remaining useful life by using the multi-head self-attention mechanism, improves the prediction accuracy of the remaining useful life, and has important significance for the application of the data / model hybrid driving method in the fields of prediction and health management. BRIEF DESCRIPTION OF DRAWINGS

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0041] Figure 1 A flowchart of the data and model hybrid driving mechanical component remaining useful life prediction method provided by the present application;

[0042] Figure 2 A network structure diagram of the hybrid driving prediction model established in the method of the present application;

[0043] Figure 3The figure shows the actual operation effect of the method of the application on the FEMTO bearing dataset. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the application will be apparently and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. All other embodiments obtained by a person of ordinary skill in the art without creative work belong to the protection scope of the application.

[0045] The object of the application is to provide a data and model hybrid driven mechanical component residual service life prediction method, which can introduce relevant field knowledge based on a model driven method, calibrate the original input data, and model the complex relationship between the input data and the degradation process by using a data driven method, so as to retain the generalization ability of the data driven method and the fidelity of the model driven method at the same time, thereby improving the accuracy of the mechanical component residual service life prediction.

[0046] In order to make the above object, characteristics and advantages of the application more apparent and easy to understand, the application will be further described in detail below with reference to the drawings and specific embodiments.

[0047] Referring to Figure 1 A data and model hybrid driven mechanical component residual service life prediction method comprises the following steps.

[0048] Step 1: An exponential random model is used to model the degradation process of the mechanical component and establish a system state space equation.

[0049] The time series signal measured by the sensor is expressed as a vector is a vector containing the sensor signals from time k to k+p. i=1,...,Q indicates that the sensor signals are sampled from Q different mechanical components of the same kind, such as bearings. i The service life of the i th component is represented by The variance feature of the time series signal is extracted to obtain a variance feature sequence wherein represents the variance of the sensor signals from time k to k+p of the i th component, n i = m i / p is the length of the variance feature sequence of the i th component.

[0050] The method of the application considers using an exponential random model to fit the variance feature curve, that is, to express the degradation process of the mechanical component by an exponential model, that is, to use the model as the degradation model of the mechanical component, and the unknown parameters a k ,bk is a parameter related to the health state of the component during degradation. The parameters of the exponential random model are estimated by an extended Kalman filter. Generally, the model parameter estimation process needs to establish the system state space equation, which is generally in the form of:

[0051]

[0052] where the state vector at time k is composed of the exponential function and the parameters a k and b k of the exponential model. f and h are nonlinear functions. u k is the system input at time k, which is 0 in this method. w k is a random 0-mean error, and its error covariance matrix is Q k . z k is the measurement value, including the variance characteristics of the sensor signal of the mechanical component i v k is the measurement error, and its error covariance matrix is R k .

[0053] Step 2: Based on the system state space equation, the parameters of the exponential random model are estimated by an extended Kalman filter to obtain the optimal state estimation.

[0054] Step 2 estimates the parameters of the exponential random model by an extended Kalman filter. Step 2 specifically includes:

[0055] Step 2.1: Based on the system state space equation, the nonlinear functions f k and h k at time k are locally linearized with respect to the state prior estimate to obtain the corresponding Jacobian matrices F k and H k .

[0056] In the EKF, the nonlinear functions f and h in the system state space equation (1) need to be locally linearized at the estimated state by calculating the respective Jacobian matrices F and H, respectively, to generate the matrices F and H.

[0057] In the Kalman filter iteration process, the parameters a k and b k of the exponential model change constantly, so the nonlinear function f k composed of the two also changes at different times k. The nonlinear functions f k and h k at time k are locally linearized with respect to the state prior estimate to obtain the corresponding Jacobian matrices Fk and H k :

[0058]

[0059]

[0060] where the state prior estimate refers to the state update value obtained by Kalman filtering, i.e., the prior estimate of the state x k .

[0061] The Jacobian matrix F k and H k are obtained, the nonlinear system in the method can be approximated as a linear system, so that the prediction and update processes of Kalman filtering can be performed to estimate the state x k .

[0062] Step 2.2: Establish the prediction and update equations of the extended Kalman filter according to the Jacobian matrix F k and H k .

[0063] The prediction and update equations of the extended Kalman filter established according to the Jacobian matrix F k and H k in the method of the application are

[0064]

[0065] where x k (1) = H k x k is the first term of the state x k .

[0066] Step 2.3: Alternately perform the prediction and update processes of the extended Kalman filter based on the prediction and update equations, constantly update the predicted state vector, and obtain the optimal state estimate.

[0067] Based on the degenerate model (i.e., the exponential random model ) and the state space equation (1) given in step 1, the degenerate model parameters are estimated by using the extended Kalman filter. The state vector x0 and the covariance matrix P0 are initialized, the prediction is performed based on the prediction and update equations (4) provided in step 2.2, and the prediction process is shown in equation (5):

[0068]

[0069] P k-1 represents the error covariance matrix at time k-1. Q k-1 represents the process error w k-1The error covariance matrix M. k Let represent the prior estimate of the covariance matrix at time k.

[0070] Next, equation (6) uses the Kalman gain K. k Update the predicted state vector The covariance matrix P k :

[0071]

[0072] Among them, z k The measured value is the variance characteristic var of the sensor signal. k The prediction and update steps are continuously repeated until the variable k reaches the maximum number of iterations, i.e., the length n of the sensor signal variance feature. After extended Kalman filtering, the optimal estimated state variable time series can be obtained. Simply put, this is the optimal state estimate. This time series contains the changing trends of sensor signals during the degradation process described by the model method. This information will be used in subsequent steps along with the original sensor signal data. The data is then fused and ultimately input into a hybrid-driven prediction model for remaining useful life prediction. This data will increase the interpretability and accuracy of the remaining useful life prediction method; the specific fusion method will be described in later steps.

[0073] Step 3: Obtain the condition monitoring data of the mechanical component during the degradation stage based on the first prediction time, and use Fast Fourier Transform to obtain the frequency domain data corresponding to the condition monitoring data of the degradation stage.

[0074] Step 3 determines the First Prediction Time (FPT) based on the collected state monitoring data, using the kurtosis of the signal as an indicator. When the kurtosis exceeds the 3σ range, it indicates the onset of degradation. A Fast Fourier Transform (FFT) is used to perform preliminary feature extraction on the target, providing time-domain and frequency-domain information. The remaining useful life is then labeled as a percentage. Specifically, Step 3 includes:

[0075] Step 3.1: Determine the initial prediction time based on the original state monitoring data of the mechanical parts collected by the sensors.

[0076] Step 2 yielded the optimal state estimate by estimating the parameters of the exponential model. To achieve hybrid data model driving, it is also necessary to process the original sensor time-series signal data. Perform feature pre-extraction. This refers to the raw state monitoring data of mechanical parts collected by sensors, which are usually vibration signal data in the horizontal or vertical direction, collected using accelerometers.

[0077] First, the first predicting time (FPT) is determined, which divides the whole life cycle of the mechanical component into a healthy stage and a degradation stage, and only the data of the degradation stage is used for the remaining useful life prediction, which can reduce the influence of other irrelevant data other than the degradation data and improve the accuracy of the prediction.

[0078] The kurtosis of the sensor signal is extracted based on formula (7) for determining the FPT:

[0079] |kurtosis k+j -u|>3σ s j=0,1,2 (7)

[0080] wherein kurtosis k+j represents the kurtosis at the k+j time, and u and σ s represent the mean and standard deviation of the kurtosis, respectively. Formula (7) indicates that when the kurtosis of the sensor signal continuously exceeds 3 times the standard deviation σ i for three consecutive times t i ,t i +1,t i +2 from the time t s , then the time t i is the first predicting time, and the sensor data i starting from the time t will be used for subsequent prediction.

[0081] Step 3.2: Extracting the condition monitoring data of the degradation stage of the mechanical component based on the first predicting time.

[0082] The sensor data i starting from the time t is the condition monitoring data of the degradation stage of the mechanical component.

[0083] Step 3.3: Extracting the frequency domain information of the condition monitoring data of the degradation stage using fast Fourier transform to obtain the frequency domain data of the degradation stage.

[0084] After the FPT, the sensor data of the degradation stage is obtained, and then the frequency domain information of the time series sensor signal is extracted using fast Fourier transform (FFT). After FFT, the frequency domain data is obtained, wherein represents the kth time dimension q frequency domain vector of the ith component.

[0085] For subsequent long short-term memory network training, let​ i.e. the time step for the long short-term memory network is 3. Frequency domain data The remaining useful life label corresponding to each time instance k is determined by the following equation (8):

[0086]

[0087] wherein, represents the RUL value of the i-th component at the k-th time instance, is a ratio between 0 and 1, n i -3-t i is the total length of the degradation phase.

[0088] Step 4: Construct the neural network training dataset of all mechanical components according to the optimal state estimation and the frequency domain data.

[0089] Through FPT and FFT, the frequency domain data of the degradation phase is obtained in step 3 together with the optimal state estimation of the degradation phase obtained by the degradation model to form the neural network training dataset of all mechanical components This dataset will be used for subsequent hybrid driving prediction model training.

[0090] Step 5: Establish a hybrid driving prediction model, including a fully connected layer, a one-dimensional convolution long short-term memory network adaptive coding layer, a multi-head attention mechanism module, a feedforward module, and a fully connected regression layer.

[0091] Referring to Figure 2 , the hybrid driving prediction model established by the present application comprises one layer of fully connected layer, one layer of one-dimensional convolution long short-term memory network adaptive coding layer, one layer of multi-head attention mechanism module, one layer of feedforward module, and one layer of fully connected regression layer. In the network structure of the hybrid driving prediction model established by the present application, the outputs of the fully connected layer and the one-dimensional convolution long short-term memory network adaptive coding layer are both taken as the inputs of the multi-head attention mechanism module, the multi-head attention mechanism module and the feedforward module are connected through a residual manner, and the output of the feedforward module is also connected with the fully connected regression layer through a residual manner.

[0092] The hybrid driving prediction model adopts one-dimensional convolution layer to extract shallow features, and then inputs the shallow features into the convolution long short-term memory network to adaptively encode the inputs. The state vector learned in step 2 The mapped features are inputted into the multi-head attention mechanism module and the feedforward module together with the RegressionToken and the shallow features after adaptive coding to mine the category-related information from the input embedding sequence. The output of the feedforward module is inputted into the input full connection regression layer to map the fused feature embedding to the corresponding RUL. Since the RUL is expressed in the form of percentage in the present application, a RUL mapping layer composed of a single full connection layer and a Sigmoid activation function is used to achieve the above-mentioned target.

[0093] Specifically, the neural network training data set is obtained after step 4 is performed The data set of the lth component is selected as the test set The other Q-1 data sets are selected as the training set

[0094] At the beginning of training, a batch of data {F, X, Y} is randomly selected, wherein Wherein, batch is the number of data in each batch. First, X is mapped to The full connection layer network can be represented as:

[0095]

[0096] Wherein, And are network parameters, and r represents the number of output layer neurons of the full connection layer network, i.e. the output layer dimension of the full connection layer.

[0097] Then, the tensor is inputted into the one-dimensional convolution long short memory network adaptive coding layer. The data F is adaptively positionally coded by the convolution long short memory network (CLSTM), and the role of this step is to add position information to the input data. As an extension of the LSTM network, the CLSTM adopts a similar structure to the LSTM and uses convolution operation instead of full connection operation in the LSTM. The related formulas of the CLSTM are as follows:

[0098]

[0099] Wherein, i k , e k , o k and c k represent the input gate, the forget gate, the output gate and the cell state respectively, and the information at each time k is transmitted in the CLSTM calculation unit. Different types of gates can selectively add new information to the cell state. represents the input vector at time k, and h k-1 represents the hidden layer output at time k-1. W fi , Wfe W fo W fc W hi W he W ho W hc ,b i ,b e ,b o ,b c For network parameters. σ(·) represents the Sigmoid activation function, and * and These represent convolution and Hamiltonian multiplication, respectively.

[0100] The output of the CLSTM network is Where kernel is the number of convolution kernels in CLSTM. This is the output of the CLSTM hidden layer.

[0101] The output of the fully connected layer Concatenate the H output from the CLSTM adaptive coding layer to obtain in The splicing operation is the process of integrating extended Kalman filter parameter estimation and neural network methods. This operation expands the input feature space of the neural network, which can improve the accuracy of model prediction.

[0102] See Figure 2 tensor The input multi-head self-attention module and feedforward module are used to extract high-level features related to the degradation process from the data. It is assumed that the multi-head self-attention mechanism uses fully connected layers on the input tensor. A certain matrix in (For ease of reference, the following abbreviations are...) By performing a linear transformation, the query can be obtained. key Sum in Here are the network parameters, H represents the number of heads in the multi-head attention mechanism, and d represents the dimension. k =r / H. The single-head attention mechanism calculates the dot product of all queries and keys and divides it by the scaling factor. And input the softmax function to obtain the corresponding attention weight matrix.

[0103]

[0104] Then the multi-head attention mechanism output

[0105] Multihead = Concat(A 1 ,...,A H W o (12)

[0106] where Concat is a column-wise concatenation operation, are network parameters.

[0107] The input tensor The input multi-head attention module obtains The multi-head attention module and the feed-forward module are connected through a residual manner, so there is

[0108]

[0109] where LayerNorm represents a layer normalization operation, is the output of the residual connection mode, which will be passed into the feed-forward module. The feed-forward module is composed of two fully connected layers, which can be specifically represented as

[0110]

[0111] where W1, W2, b1, b2 are network parameters, and ReLU is an activation function. Z FF is the output of the feed-forward module. Similarly, the output of the feed-forward module is also connected with the next module through a residual manner, so there is

[0112]

[0113] where is the output of the final feed-forward module.

[0114] Referring to Figure 2 the output of the final feed-forward module is input into a fully connected regression layer, which can be specifically represented as:

[0115]

[0116] where W reg ,b reg are network parameters, and Sigmoid is an activation function. The obtained output is the RUL prediction value corresponding to the input {F, X} The obtained prediction value is used to calculate the loss value L using a mean square error loss function:

[0117]

[0118] The network training adopts the Adam algorithm to update the parameters of the entire neural network.

[0119] Step 6: The hybrid driving prediction model is trained and tested using the neural network training data set, and a trained hybrid driving prediction model is obtained.

[0120] The neural network training training set obtained from step 4 A batch of data {F, X, Y} is randomly selected to train the hybrid driving prediction model, and the neural network parameters are updated during the training process until the training period reaches the maximum training period limit. The effect of the network model is verified on the test set, and the neural network model with the optimal effect on the test set is reserved as the trained hybrid driving prediction model

[0121] Step 7: The trained hybrid driving prediction model is used to predict the remaining useful life of the mechanical part.

[0122] When new sensor data s k * , s k+1 * , s k+2 * arrives at k+2 time, first extract the variance feature var k * at k time, obtain the optimal state estimation through the extended Kalman filter Then, the frequency domain vector is constructed by FFT The vectors and are input into the trained hybrid driving prediction model

[0123] The application provides a data / model hybrid driving mechanical part remaining useful life prediction method, which combines a calibrated exponential random model and a multi-head attention neural network structure, and a hybrid driving prediction model is constructed, which can combine the advantages of data driving and model-based methods to accurately predict the remaining useful life.

[0124] Figure 3A running effect diagram of an instance of a data and model hybrid driving mechanical component residual useful life prediction method provided by the application on a FEMTO bearing dataset. The FEMTO bearing dataset (i.e., a condition monitoring dataset) is composed of data collected from a PRONOSTIA platform and contains vibration signals monitored by sensors in horizontal and vertical directions. The dataset contains three operating conditions, and the rotating speed and load of each condition are different. In this embodiment, the number of iterations of the training process is 120, the number of batch samples extracted each time is 16, the learning rate l is 0.0002, the hybrid driving prediction model contains 1 full connection layer, 1 one-dimensional convolution long short memory network adaptive coding layer, 1 multi-head attention mechanism module, 1 feedforward module and 1 full connection regression layer. In the running effect diagram, the actual value of the residual useful life of the mechanical component is represented as actual RUL, and the data and model hybrid driving residual useful life prediction method provided by the application (represented as Hybrid Transformer[proposed]in the method) and the other two methods (ConvLSTM and LSTMTransformer) are compared in terms of effect. From a visual point of view, the prediction effect of the method of the application is closer to the actual value and more accurate, and is superior to the other two models. Figure 3

[0125] As shown in Table 1, in terms of evaluation indexes, the root mean square error (RMSE), the mean absolute error (MAE) and the mean absolute percentage error (MAPE) of the three methods on the test set are compared.

[0126] Table 1

[0127]

[0128] As can be seen from the data in Table 1, the three error indexes of the method of the application are lower than those of the other two methods. Thus, the data and model hybrid driving residual useful life prediction method provided by the application has an advantage in the accuracy of the trained model.

[0129] Based on the method provided by the application, the application further provides a data and model hybrid driving residual useful life prediction system of a mechanical component, comprising:

[0130] A degradation model establishing module is configured to model a mechanical component degradation process by using an exponential random model and establish a system state space equation;

[0131] An extended Kalman filter module is configured to perform parameter estimation on the exponential random model by using an extended Kalman filter based on the system state space equation to obtain an optimal state estimation;

[0132] ​An FFT feature extraction module is configured to acquire state monitoring data of a degradation stage of a mechanical component based on a first prediction time, and to obtain frequency domain data corresponding to the state monitoring data of the degradation stage by using a fast Fourier transform (FFT);

[0133] A training data set construction module is configured to construct a neural network training data set of all mechanical components based on the optimal state estimation and the frequency domain data;

[0134] A hybrid driving prediction model establishment module is configured to establish a hybrid driving prediction model, which includes a full connection layer, an adaptive coding layer of a one-dimensional convolution long short memory network, a multi-head attention mechanism module, a feedforward module, and a full connection regression layer.

[0135] A hybrid driving prediction model training module is configured to train and test the hybrid driving prediction model by using the neural network training data set, so as to obtain a trained hybrid driving prediction model.

[0136] A remaining useful life prediction module is configured to predict the remaining useful life of the mechanical component by using the trained hybrid driving prediction model.

[0137] Further, the present application also provides an electronic device, which can include a processor, a communication interface, a memory and a communication bus. Wherein the processor, the communication interface and the memory complete mutual communication through the communication bus. The processor can invoke a computer program in the memory to execute the data and model hybrid driving mechanical component remaining useful life prediction method.

[0138] In addition, the computer program in the memory described above is implemented in the form of a software function unit and sold or used as an independent product. Based on such understanding, the technical solutions of the present application or parts of the technical solutions that essentially contribute to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk or an optical disk, and various media that can store program codes.

[0139] Further, the present application also provides a non-transitory computer readable storage medium having a computer program stored thereon, wherein the computer program is executed to implement the data and model hybrid driving mechanical component remaining useful life prediction method.

[0140] The application provides a data / model hybrid driving mechanical component residual service life prediction method, which combines a calibrated exponential random model and a multi-head attention neural network structure to realize accurate prediction of the residual service life. The method calibrates the exponential random model by extending a Kalman filter, iteratively learns the exponential function value and the exponential model parameter; a neural network module based on a multi-head attention mechanism is combined to establish a prediction model, the position information of input embedding is automatically learned through a self-adaptive position coding layer, and then the mapping relationship between the input data and the residual service life is modeled through a multi-head self-attention mechanism; therefore, the accuracy of the model-based method and the generalization ability of the data-driven method are retained, the prediction accuracy of the residual service life is greatly improved, the application of the data / model hybrid driving method in the field of intelligent manufacturing and mechanical equipment health management has important significance, and the method has a wide application prospect.

[0141] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the difference from other embodiments, and the same or similar parts between the embodiments can be referred to each other. For the system disclosed by the embodiments, since it corresponds to the method disclosed by the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0142] The principles and implementation manners of the application are described by applying specific examples in the specification, and the above embodiment description is only used to help understand the method of the application and its core idea; meanwhile, for the general technical personnel in the field, the specific implementation manner and application range of the application will be changed according to the idea of the application. In conclusion, the content of the specification should not be understood as the limitation of the application.

Claims

1. A data and model hybrid-driven method for predicting the residual useful life of a mechanical component, characterized in that, The method comprises the following steps: obtaining original condition monitoring data of the mechanical component, wherein the original condition monitoring data comprises vibration signals monitored by sensors in horizontal and vertical directions; modeling the degradation process of the mechanical component by using an exponential random model and establishing a system state space equation; based on the system state space equation, estimating parameters of the exponential random model by using an extended Kalman filter to obtain optimal state estimation; based on the first prediction time, obtaining condition monitoring data of the degradation stage of the mechanical component, and using fast Fourier transform to obtain frequency domain data corresponding to the condition monitoring data of the degradation stage; constructing a neural network training data set of all mechanical components according to the optimal state estimation and the frequency domain data; establishing a hybrid driving prediction model, which comprises a fully connected layer, an adaptive encoding layer of one-dimensional convolutional long short memory network, a multi-head attention mechanism module, a feedforward module and a fully connected regression layer; wherein the fully connected layer is used to map the optimal state estimation into a feature vector; the adaptive encoding layer of the one-dimensional convolutional long short memory network is used to perform adaptive position encoding on the frequency domain data; the feature vector output by the fully connected layer and the encoded features output by the adaptive encoding layer of the one-dimensional convolutional long short memory network are input into the multi-head attention mechanism module together; the output of the multi-head attention mechanism module is input into the fully connected regression layer through the feedforward module, and the fused feature embedding is mapped to the remaining useful life; training and testing the hybrid driving prediction model by using the neural network training data set to obtain a trained hybrid driving prediction model; using the trained hybrid driving prediction model to predict the remaining useful life of the mechanical component.

2. The data and model hybrid-driven mechanical component remaining useful life prediction method of claim 1, wherein, The modeling of the degradation process of the mechanical component by using the exponential random model and the establishment of the system state space equation specifically comprise: using an exponential random model modeling a degradation process of a mechanical component; wherein is a parameter related to a health state of the mechanical component during the degradation process; based on the exponential stochastic model establishing system state space equations ; wherein time state vector ; and is a nonlinear function; is a state vector at time ; is a system input at time ; is a random 0-mean error at time ; is a measurement at time ; is a measurement error at time .

3. The data and model hybrid-driven mechanical component remaining useful life prediction method of claim 2, wherein, The parameter estimation of the exponential random model based on the system state space equation by using the extended Kalman filter to obtain the optimal state estimation specifically comprises: Based on the system state-space equation, at time... nonlinear functions and Regarding state prior estimation Local linearization is performed to obtain the corresponding Jacobian matrix. and ; According to the Jacobian matrix and establishes the prediction and update equations of the extended Kalman filter; Based on the prediction and update equations, the prediction and update processes of the extended Kalman filter are alternately executed, and the predicted state vector is constantly updated to obtain the optimal state estimation.

4. The data and model hybrid-driven mechanical component remaining useful life prediction method of claim 3, wherein, The condition monitoring data of the degradation stage of the mechanical component is obtained based on the first prediction time, and the frequency domain data corresponding to the condition monitoring data of the degradation stage is obtained by using fast Fourier transform, specifically comprising: determining the first prediction time according to the original condition monitoring data of the mechanical component collected by the sensor; extracting the condition monitoring data of the degradation stage of the mechanical component based on the first prediction time; extracting the frequency domain information of the condition monitoring data of the degradation stage by using fast Fourier transform to obtain the frequency domain data of the degradation stage.

5. The data and model hybrid-driven mechanical component remaining useful life prediction method of claim 4, wherein, The neural network training data set of all mechanical components is constructed according to the optimal state estimation and the frequency domain data, specifically comprising: Optimal state estimation according to degradation phase of mechanical components and frequency domain data constructing a neural network training dataset for all mechanical components wherein time is a first prediction time; is a variance feature sequence length; is a number of mechanical components ; and is a remaining useful life of the mechanical component at time .

6. A data and model hybrid-driven mechanical component remaining useful life prediction system, characterized in that, The method comprises the following steps: an original data acquisition module is configured to obtain original condition monitoring data of the mechanical component, wherein the original condition monitoring data comprises vibration signals monitored by sensors in horizontal and vertical directions; a degradation model establishment module is configured to model the degradation process of the mechanical component by using an exponential random model and establish a system state space equation; An extended Kalman filter module is configured to perform parameter estimation on the exponential random model based on the system state space equation by using an extended Kalman filter to obtain optimal state estimation. An FFT feature extraction module is configured to obtain state monitoring data of a degradation stage of a mechanical component based on a first prediction time and to obtain frequency domain data corresponding to the state monitoring data of the degradation stage by using fast Fourier transform. A training data set construction module is configured to construct a neural network training data set of all mechanical components based on the optimal state estimation and the frequency domain data. A hybrid driving prediction model establishment module is configured to establish a hybrid driving prediction model, which includes a full connection layer, an adaptive encoding layer of one-dimensional convolutional long short memory network, a multi-head attention mechanism module, a feedforward module, and a full connection regression layer. The full connection layer is configured to map the optimal state estimation into a feature vector. The adaptive encoding layer of one-dimensional convolutional long short memory network is configured to perform adaptive position encoding on the frequency domain data. The feature vector output by the full connection layer and the encoded features output by the adaptive encoding layer of one-dimensional convolutional long short memory network are input into the multi-head attention mechanism module. The output of the multi-head attention mechanism module is input into the full connection regression layer after being processed by the feedforward module, so as to map the fused feature embedding to the remaining useful life. A hybrid driving prediction model training module is configured to train and test the hybrid driving prediction model by using the neural network training data set, so as to obtain a trained hybrid driving prediction model. A remaining useful life prediction module is configured to predict the remaining useful life of the mechanical component by using the trained hybrid driving prediction model.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the data and model hybrid driving mechanical component remaining useful life prediction method as claimed in any one of claims 1 to 5 when executing the computer program.

8. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program implements the data and model hybrid driving mechanical component remaining useful life prediction method as claimed in any one of claims 1 to 5 when being executed.

Images