A method for predicting the remaining service life of rolling bearings based on residual convolutional neural network and LSTM

Abstract

The present invention discloses a method for predicting the remaining useful life of a rolling bearing based on a residual convolutional neural network and a long-term memory (LSTM). The method comprises the following steps: obtaining a time-series vibration signal over the full life cycle of the bearing and normalizing it; using the preprocessed signal as input to the residual convolutional neural network; automatically extracting local abstract information from the data using a one-dimensional deep residual convolutional neural network to mine deep features of the bearing's time-series vibration signal; inputting the extracted deep features into a deep long-term memory (LSTM) network to construct bearing trend features; optimizing all network hyperparameters using an optimization algorithm to complete the prediction of the remaining useful life of the bearing and obtain a rolling bearing life prediction model based on the residual convolutional neural network and the deep long-term memory (LSTM) network; and inputting a time-series vibration signal over a non-full life cycle of a test bearing into the model to obtain the remaining useful life of the test bearing. The present invention utilizes a deep neural network model to effectively improve the accuracy of bearing life prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of predicting the remaining service life of a rolling bearing, and relates to a method for predicting the remaining service life of a rolling bearing based on a residual convolutional neural network and an LSTM (Long Short-Term Memory) network. Background Art

[0002] In recent years, rolling bearings, one of the most common and crucial core components in rotating machinery, have become increasingly important. Their health impacts the precision and safety of the entire mechanical system. In complex operating environments, rolling bearings are among the most vulnerable and least reliable components in industry. Failure of a rolling bearing can lead to a range of negative consequences, including extended downtime, major accidents, and even casualties. Therefore, accurately predicting the remaining useful life (RUL) of rolling bearings is crucial for preventive maintenance decisions in rotating machinery.

[0003] Existing RUL prediction methods can generally be categorized into four main categories: physical model-based methods, signal processing-based methods, data-driven methods, and hybrid methods. Traditional algorithms require a certain level of understanding of the physical characteristics of the equipment before performing rolling bearing life prediction. This requires prior knowledge, complex signal filtering and noise reduction, and feature design and selection. However, with the advent of Industry 4.0 and the big data era, traditional algorithms are becoming increasingly infeasible in the face of massive amounts of data. In recent years, with the advancement of information technology, data-driven rolling bearing life prediction methods have become a hot topic of research. Among these data-driven methods, degradation characteristics are modeled based on historical sensor data, and then the degradation characteristics of the rolling bearing are mapped to the RUL value of the rolling bearing. Traditional data-driven methods, such as support vector machines (SVMs), artificial neural networks (ANNs), and naive Bayesian methods, all employ shallow machine learning models. However, these shallow machine learning methods rely heavily on expert knowledge and manual feature selection. Furthermore, these shallow methods have limited representational capabilities and cannot fully utilize massive data to learn sufficient features. Therefore, rolling bearing life prediction under complex operating conditions remains to be improved.

[0004] Deep learning, a popular branch of data-driven methods, has developed rapidly and has been widely applied in various fields due to its advantages such as automatic feature extraction and strong model generalization. Deep learning-based RUL prediction methods can automatically extract deep feature representations of historical data, effectively overcoming the problem of unknown model degradation. Therefore, they have been widely studied and applied in the field of mechanical prediction. Deep learning models, represented by convolutional neural networks (CNNs), have gradually been applied to rolling bearing RUL prediction. A small number of convolution operations cannot accurately predict the RUL value of rolling bearings. However, as the number of deep CNN layers increases to a certain level, the model may degrade and the prediction accuracy of the model will decrease. In addition, using a simple CNN to perform adaptive feature extraction on raw time series data only considers the multi-dimensional characteristics of the data, not the time series characteristics of the data, which will result in the loss of characteristic information of the original data sequence. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems that the existing rolling bearing vibration signal feature extraction method relies too much on expert experience and the prediction accuracy of the existing model is low. A rolling bearing RUL prediction method based on residual convolutional neural network and LSTM is provided. The original time series vibration signal of the rolling bearing is obtained and normalized and used as the input of a one-dimensional deep residual convolutional neural network to preliminarily extract the deep degradation characteristics of the rolling bearing. The trend characteristics of the rolling bearing are then constructed using the deep LSTM network. Finally, the rolling bearing RUL prediction is realized through a fully connected neural network, thereby solving the life prediction and intelligent decision-making problems of the rolling bearing during operation.

[0006] In order to solve the above problems, the present invention adopts the following technical solutions.

[0007] The present invention provides a method for predicting the remaining service life of a rolling bearing based on a residual convolutional neural network and LSTM, comprising the following steps:

[0008] Step 1: Data preprocessing: Obtain the time series vibration signal of the bearing throughout its life cycle and construct the training set D under different working conditions. train ={x t ,y t} T , D train is the full life cycle time series vibration signal of a bearing under a certain working condition, x t ∈R P×2 The feature input of the bearing is P×2, P=2560, y t∈[0,1] is the output of the prediction model, and T is the failure time of the bearing during operation, that is, the full life value of the bearing. First, the full life cycle time series vibration signal is normalized and used as the feature input of the model. To facilitate network training, and based on the fact that the health status of the bearing degrades linearly with operating time, the percentage of the full life value of the bearing is calculated and used as the output to train the prediction model.

[0009] Step 2: Preliminary feature extraction: Set the structure of each residual block and the size of the convolution kernel in the one-dimensional residual convolutional neural network; input the normalized bearing time series vibration signal into the one-dimensional residual convolutional neural network, and use the residual blocks of different structures composed of Conv1D layer, pooling layer, BN layer, and Relu layer in the network to further extract the deep features of the bearing time series vibration signal;

[0010] Step 3: Trend feature construction: The output of the one-dimensional residual convolutional neural network is input into the deep LSTM. The LSTM basic memory unit in the deep LSTM has the advantage of long-term and short-term memory for time series data, and the trend feature information of the rolling bearing is further constructed.

[0011] Step 4: Model training: The training data set constructed in step 1 is input into steps 2 and 3 to construct quantitative features of bearing trends. The linear layer of the fully connected neural network is then used to fit the performance degradation trend of the rolling bearing. The Adam optimization algorithm is used to optimize all hyperparameters to complete the self-extraction of deep-level bearing features and RUL prediction, thereby obtaining a bearing RUL prediction model based on residual convolutional neural network and LSTM.

[0012] Step 5: Model testing: Select non-full-life cycle time-series vibration signals under different operating conditions to construct a test set for the model; perform normalization processing on them, adaptively extract deep features of the vibration data through step 2, and combine them with the deep LSTM model in step 3 to obtain trend quantitative features of the test set bearings; finally, use a fully connected neural network to fit the degradation trend of the bearing life value and predict the RUL of the test set bearings.

[0013] Specifically, the data preprocessing process in step 1 includes:

[0014] For the full-cycle time series vibration signals of the rolling bearing under different working conditions obtained in step 1, min-max normalization processing is performed, and the conversion function is as follows:

[0015]

[0016] Among them, x is the time series vibration signal before normalization, x t is the normalized time series vibration signal, x min is the minimum value of the time series vibration signal, xmax is the maximum value of the time series vibration signal;

[0017] The normalized vibration signal is used as the input of the overall prediction model. After continuous iterative training of the neural network to optimize its network parameters, the network finally outputs the RUL prediction value corresponding to the data set. Assuming that each training sample is set to {(x i ,y i )}, where x i is the vibration data feature of the i-th sample; y i is the label value of the i-th sample, that is, the real rolling bearing RUL corresponding to the collection time; the health status of the rolling bearing degrades linearly with the running time. Based on this condition, y i The calculation formula is as follows:

[0018]

[0019] Where t is the failure time of the bearing, that is, the full life value of the bearing; tt i is the RUL value of the i-th sampling point, t0 is the starting time of the bearing operation, and N is the total sampling points of the bearing.

[0020] Specifically, the preliminary feature extraction in step 2 includes the following steps:

[0021] 2.1. Build a one-dimensional deep residual convolutional neural network model: The preliminary feature extraction model based on the one-dimensional deep residual convolutional neural network includes a post-activation residual block and a pre-activation residual block; the residual blocks of different structures in the one-dimensional residual convolutional neural network are used to process the entire input vibration signal sequence, and the deep degradation characteristics of the vibration signal are preliminarily adaptively mined; the x obtained after normalization of the time series vibration signal is used to extract the t Input into the one-dimensional deep residual convolutional neural network model G;

[0022] y=G(x t )=G(x t1 ,x t2 ) (3)

[0023] Among them, x t1 , x t2 The normalized vibration signal of the rolling bearing, y is the bearing degradation eigenvalue extracted by the one-dimensional residual convolutional network.

[0024] 2.2、Normalized bearing time series vibration signal x tPerform the first convolution operation. The stride size of the first Conv layer is s1, the convolution kernel uses a size of k1×1, and the number of convolution kernels is m1. Then use the BN layer to accelerate the calculation and optimization of the network. Use the ReLU layer for nonlinear operation and output the first layer signal features.

[0025] 2.3. The features output by the first Conv layer are input into the post-activation residual block. The post-activation residual block starts with a Conv layer, followed by a BN layer, a Relu layer, and finally ends with a Conv layer. A Dropout layer is added after the Relu layer. To reduce computation time and eliminate redundant features, a MaxPool layer is added to the residual learning framework. The structures of the two Conv layers and the MaxPool layer both use a convolution kernel with a kernel size of k2×1 and a stride size of s2. In the post-activation residual block, the number of kernels of the two convolution kernels is m2 and m3 respectively. Finally, the feature output of the post-activation residual block is output.

[0026] 2.4. The features output by the post-activation residual block are input into a deep residual module stacked by multiple one-dimensional residual modules, namely pre-activation residual blocks; a pre-activation residual block starts with a BN layer, followed by a Relu layer, a Conv layer, a BN layer, a Relu layer, and a Conv layer, and the MaxPool layer is also used in the residual structure; in the pre-activation residual block, the first Conv layer uses a convolution kernel size of k3×1 and a stride size of s3; the second Conv layer uses a convolution kernel size of k4×1 and a stride size of s4; the MaxPool layer uses a convolution kernel size of k5×1 and a stride size of s5; this structure is repeated in each pre-activation residual block; finally, after several pre-activation residual blocks, the deep features of the bearing are output, and finally the BN layer and the Relu layer are connected to complete the preliminary deep feature extraction process of the rolling bearing based on the residual convolutional neural network.

[0027] Specifically, the process of step 3, constructing trend features, includes:

[0028] Step 3.1. Calculate the value f of the forget gate in the LSTM basic unit t , f t The state value c of the memory unit at the previous moment t-1 t-1 The historical information in the memory is selectively forgotten, so that the state value c of the memory unit at the current time t is tTo have an impact is to filter the information in long-term memory;

[0029] f t =σ[W f (h t-1 ,x t )+b f ] (4)

[0030] Step 3.2, calculate the candidate memory unit value k t , in the update c t Before, k will be generated first t ; Input x at the current time t t And the output h of the network at time t-1 t-1 Jointly decide t Value; by h t-1 and x t Perform corresponding calculations to obtain the k at the current time t t , update c t Information:

[0031] k t =tanh[W k (h t-1 ,x t )+b k ] (5)

[0032] Step 3.3, calculate the value of the input gate i t ,i t Selectively memorize x t C t The influence of , stores the information in the input data into the current state;

[0033] i t =σ[W i (h t-1 ,x t )+b i ] (6)

[0034] Among them, i t and k t Both are to update the information in short-term memory, combined with h t-1 , from x t Extracting important information from memory is to filter the information in short-term memory;

[0035] Step 3.4, calculate c at the current time t t value,

[0036]

[0037] Where: represents the point-by-point product. From formula (7), we can see that c t-1 and k t The value of c t The LSTM basic unit structure is updated by f t and i t The two door controllers are respectively t-1 and k t These two parts are selectively memorized and forgotten, so as to obtain c at the current time t t value;

[0038] Step 3.5, calculate the value of the output gate o t , o t Control the current moment c t The output,

[0039] o t =σ[W o (h t-1 ,x t )+b o ] (8)

[0040] Step 3.6. Calculate the output h of the LSTM basic unit t , through o t Pass the information to the next neuron,

[0041] h t =o t tanh(c t ) (9)

[0042] In formulas (4)-(9), W f 、W k 、W i and W o f and k at time t respectively t , the weights of i and o, b f 、b k 、b i and b o f and k respectively t , i and o bias, x t is the input at the current time t, h t-1 is the output of the network at time t-1, σ is the sigmoid function, the value is (0,1), and the tanh function, the value is (-1,1); in formula (9), the output h of the LSTM basic unit is t That is the trend bearing feature information extracted by the deep LSTM network.

[0043] Specifically, the model training in step 4 includes:

[0044] Step 4.1. Build a bearing RUL prediction model based on the PyTorch deep learning framework. Create a model that specifies the hyperparameters of the one-dimensional deep residual convolutional neural network, the number of network layers and hidden layer nodes of the deep LSTM network, the number of network layers and network nodes of the fully connected neural network, the network learning rate, and the number of iterations during network training.

[0045] Step 4.2: Train the rolling bearing RUL prediction network model on the training set: Input the rolling bearing training set constructed in step 1 into step 2 to preliminarily extract the deep features of the rolling bearings. Input the deep features of the rolling bearings extracted by the one-dimensional deep residual convolutional neural network in step 2 into step 3 to construct the trend features of the rolling bearings. Finally, use the linear layer of the fully connected neural network to fit the performance degradation trend of the rolling bearings. Use the Adam optimization algorithm to optimize the hyperparameters in the one-dimensional deep residual convolutional network, deep LSTM network, and fully connected neural network to complete the life prediction of the training set. Finally, obtain the rolling bearing RUL prediction model based on the residual convolutional neural network and LSTM:

[0046]

[0047] Where M is the total number of data sample points; d i =RUL' i -RUL i , which represents the difference between the predicted value and the true value; the smaller the value of the root mean square error, the higher the prediction accuracy of the model.

[0048] Furthermore, the model test in step 5 includes:

[0049] The non-full life cycle time series vibration signals under different working conditions of rolling bearings are selected to construct a test set, and the accuracy and reliability of the model are tested on the test set; the non-full life cycle time series vibration signals under different working conditions in the test set are normalized and input into the rolling bearing RUL prediction network model based on residual convolutional neural network and LSTM obtained in the model training phase in step 4, completing the model testing phase and realizing the prediction of the RUL value of the rolling bearings in the test set.

[0050] Specifically, in the model training of step 4 and the model testing of step 5, rolling bearings 1_1 and 1_2 under working condition 1, rolling bearings 2_1 and 2_2 under working condition 2, and rolling bearings 3_1 and 3_2 under working condition 3 are selected as training sets, and rolling bearings 1_5, rolling bearing 2_7, and rolling bearing 3_3 are selected as test sets; wherein the training set is the time series vibration signal of the rolling bearing over its entire life cycle, and the test set is the time series vibration signal of the rolling bearing over a non-full life cycle.

[0051] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0052] 1. This method normalizes the actual RUL label values ​​of rolling bearings, calculates the percentage of the bearing RUL value, and uses the degradation percentage as output for model training. Compared to direct calculation, normalizing the actual RUL label values ​​of rolling bearings reduces the difficulty of network training.

[0053] 2. This invention introduces the basic residual unit structure into a one-dimensional convolutional neural network to construct a one-dimensional deep residual convolutional neural network, enabling deep-level feature extraction of rolling bearing degradation vibration signals. Compared to conventional deep convolutional neural networks, this method can operate with more convolutional layers and avoids the gradient vanishing and network degradation problems caused by too many convolutional layers. It also solves the problem of existing rolling bearing vibration signal feature extraction methods being overly dependent on expert experience, thereby automatically extracting local abstract information from the data to mine deep-level features.

[0054] 3. The present invention uses the maximum output pooling layer at the jump connections of different residual learning frameworks in the one-dimensional deep residual convolutional neural network, thereby reducing the calculation time of the network and eliminating the redundant features of the network.

[0055] 4. This invention introduces a deep LSTM network into the field of rolling bearing life prediction, further extracting the temporal characteristics of rolling bearing degradation vibration signals. Compared with other deep learning methods, this network based on the trend characteristics of deep LSTM is superior for processing time series data. It not only solves the long-term dependency problem in time series but also improves the accuracy of rolling bearing life prediction.

[0056] 5. The present invention effectively addresses the following issues: traditional feature extraction methods are overly dependent on expert knowledge; gradient vanishing and model degradation occur in deep convolutional neural networks; and simple convolutional neural networks cause the loss of time series feature information in the original data. The present invention implements end-to-end rolling bearing RUL prediction based on deep learning, without relying on a large number of manual feature extraction methods, which can greatly reduce the consumption of manpower and material resources. Effectively monitoring the working status of the rolling bearing and predicting its RUL value in a timely and accurate manner can ensure the safe, reliable and efficient operation of mechanical equipment and extend the equipment's working cycle. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flowchart of a rolling bearing RUL prediction method based on residual convolutional neural network and LSTM according to an embodiment of the present invention.

[0058] Figure 2 This is an overall framework diagram of a one-dimensional deep residual convolutional neural network according to an embodiment of the present invention;

[0059] Figure 3a-Figure 3b This is a specific implementation diagram of the residual block in a one-dimensional deep residual convolutional neural network according to an embodiment of the present invention. Figure 3a Implementation diagram for the post-activation residual block, Figure 3b Implementation diagram for the pre-activation residual block;

[0060] Figure 4 This is a basic structural diagram of an LSTM memory unit according to an embodiment of the present invention;

[0061] Figure 5 This is a specific implementation diagram of a rolling bearing RUL prediction method according to an embodiment of the present invention;

[0062] Figure 6 A data collection device used in an embodiment of the present invention;

[0063] Figure 7a-7b The following is the prediction result diagram of rolling bearing RUL using different methods. Figure 7a This is a result diagram of a rolling bearing RUL prediction method based on a residual convolutional neural network and LSTM according to an embodiment of the present invention. Figure 7b The figure shows the rolling bearing RUL prediction results based on CNN and LSTM. DETAILED DESCRIPTION

[0064] The present invention proposes a rolling bearing RUL prediction method based on a residual convolutional neural network and LSTM. First, a one-dimensional deep residual convolutional neural network based on the residual principle is used to extract preliminary features from the rolling bearing's raw vibration signal. A deep LSTM network is then used to extract the rolling bearing's temporal features, constructing trend features. Finally, a fully connected regression layer is used to map these features to the rolling bearing's lifespan. This method accurately predicts the RUL value of rolling bearing components, addressing the issue of intelligent decision-making for components during the operation of rotating machinery.

[0065] The present invention is described in further detail below with reference to the accompanying drawings.

[0066] Figure 1 This is an overall flow chart of an embodiment of the present invention. Figure 1 As shown, a rolling bearing RUL prediction method based on residual convolutional neural network and LSTM of the present invention includes the following steps:

[0067] Step 1. Data preprocessing: Select the full life cycle time series vibration signals under different working conditions of the rolling bearing as the training set, normalize the full life cycle time series vibration signals as feature input, and output the percentage of the true RUL value of the rolling bearing as the label value, as shown in D train={x train ,y train} T , where x train ∈R P×2 represents the P×2-dimensional input feature of a rolling bearing at time t under a certain working condition, P=2560 represents the 2560 sampling points of the rolling bearing, y train ∈[0,1] represents the life degradation percentage output of the rolling bearing at time t; D train It represents the full life cycle time series vibration signal of a rolling bearing under a certain working condition in the training set; T is the full life operation time of the rolling bearing;

[0068] Step 2, preliminary feature extraction: set the number of residual block layers and convolution kernel size of the one-dimensional deep residual convolutional neural network, use the normalized rolling bearing full life cycle time series vibration signal as the input of the one-dimensional deep residual convolutional neural network, and use the one-dimensional convolution (Conv1D) layer, pooling layer, batch normalization (BN) layer, and nonlinear activation function (ReLU) layer in the one-dimensional deep residual convolutional neural network to traverse the entire life cycle time series vibration signal, mine the local information in the full life cycle time series vibration signal, and preliminarily extract the deep-level features of the rolling bearing;

[0069] Step 3: Trend feature construction: The deep-level features of the rolling bearings extracted by the one-dimensional deep residual convolutional neural network are input into the deep LSTM network. The LSTM memory units in the deep LSTM network have the advantage of long-term and short-term memory for the full life cycle time series vibration signals, and the trend quantitative features of the rolling bearings are further constructed.

[0070] Step 4: Model training: The rolling bearing training set constructed in Step 1 is input into Step 2 to initially extract the deep features of the rolling bearings. The deep features of the rolling bearings extracted by the one-dimensional deep residual convolutional neural network in Step 2 are input into Step 3 to construct the rolling bearing trend features. Finally, the linear layer of the fully connected neural network is used to fit the performance degradation trend of the rolling bearings. The Adam optimization algorithm is used to optimize the hyperparameters of the one-dimensional deep residual convolutional network, the deep LSTM network, and the fully connected neural network to complete the model training phase, thereby obtaining a rolling bearing RUL prediction model based on the residual convolutional neural network and LSTM.

[0071] Step 5: Model testing: Select the non-full life cycle time series vibration signals under different working conditions of the rolling bearing to construct a test set, normalize the non-full life cycle time series vibration signals in the test set, and input them into the network model obtained in the model training phase in step 4. Complete the model testing phase and realize the prediction of the RUL value of the rolling bearing in the test set.

[0072] Specifically, the data processing in step 1 includes:

[0073] Step 1.1: Perform minimum-maximum (min-max) normalization processing on the rolling bearing full life cycle time series vibration signal under different working conditions obtained in step 1, that is, perform linear change on the rolling bearing full life cycle time series vibration signal so that the final result falls within the interval [0,1]. The conversion function example is as follows:

[0074]

[0075] Where x is the full life cycle time series vibration signal before normalization, x t is the normalized full life cycle time series vibration signal, x min is the minimum value of the full life cycle time series vibration signal, x max is the maximum value of the full life cycle time series vibration signal;

[0076] Step 1.2: Take the min-max normalized full life cycle time series vibration signal as the input of the overall prediction model, and optimize the network parameters through continuous iterative training of the neural network, and finally output the corresponding RUL prediction value. Assume that each training sample is set as {(x i ,y i )}, where x i is the characteristic value of the rolling bearing vibration data of the i-th sample; y i is the label value of the i-th sample, that is, the real rolling bearing RUL value corresponding to the collection. It is generally believed that the health of the rolling bearing degrades linearly with the running time. Based on this condition, y i The calculation formula is as follows:

[0077]

[0078] Where t is the failure time of the rolling bearing, that is, the full life value of the rolling bearing; tt i is the RUL value of the i-th sampling point, t0 is the starting time of the rolling bearing operation, and N is the total number of sampling points of the rolling bearing. For example, in the PHM2012 dataset used in this invention, there are 2803 samples of the rolling bearing 1_1 in the working condition. Since the sampling interval is 10s, the RUL of the rolling bearing is considered to be 28030s. If x i For the 1000th sample, the corresponding real RUL is 18030s, that is, x at this time i is the signal feature of the 1000th sample, and the sample label y i It is 0.64324.

[0079] Specifically, the step 2 of preliminary feature extraction includes:

[0080] 2.1. Build a one-dimensional deep residual convolutional neural network model, such as Figure 2 As shown in Figure 3, the preliminary feature extraction model based on the one-dimensional deep residual convolutional neural network includes a convolution (Conv) layer, a BN layer, a Relu layer, a post-activation residual block, 5 pre-activation residual blocks, and the final BN layer and Relu layer. In Formula 3, the input of the one-dimensional deep residual convolutional neural network at time t is the normalized full life cycle time series vibration signal of the rolling bearing, and the output is the deep-level features extracted by the one-dimensional deep residual convolutional neural network. The purpose of the one-dimensional deep residual convolutional neural network model is to learn the mapping from the full life cycle time series vibration signal segment to the deep-level features, and use the different residual blocks in the one-dimensional deep residual convolutional neural network to traverse the entire input, preliminarily extract the local information of the full life cycle time series vibration signal, and mine the deep-level features of the rolling bearing data. The x obtained after the normalization of the full life cycle time series vibration signal of the rolling bearing is converted to the deep-level features. t Input into the one-dimensional deep residual convolutional neural network model G;

[0081] y=G(x t )=G(x t1 ,x t2 ) (3)

[0082] Among them, x t1 , x t2 is the normalized full life cycle time series vibration signal of the rolling bearing, and y is the deep degradation eigenvalue of the rolling bearing initially adaptively extracted by the one-dimensional deep residual convolutional neural network.

[0083] 2.2 The specific process of the one-dimensional deep residual convolutional neural network includes: first, the normalized full life cycle time series vibration signal x t The first Conv layer is operated with a stride of 2, a 3×1 convolution kernel, and 64 convolution kernels. A BN layer is then used to accelerate the calculation and optimization of the one-dimensional deep residual convolutional neural network. A ReLU layer is used for nonlinear operations and outputs signal features.

[0084] 2.3. Input the signal features output by the first Conv layer, BN layer and Relu layer into the post-activation residual block. The specific implementation diagram of the post-activation residual block is as follows: Figure 3aAs shown in the figure, the post-activation residual block starts with a Conv layer, followed by a BN layer, a Relu layer, a Dropout layer, and finally ends with a Conv layer. Adding a Dropout layer after the Relu layer reduces the overfitting phenomenon in the one-dimensional deep residual convolutional neural network during training and improves the generalization ability of the model. In order to reduce computing time and eliminate redundant features, a maximum output pooling (MaxPool) layer is added to the jump connection of the residual learning framework. The structures of the two Conv layers and the MaxPool layer in the post-activation residual block both use convolution kernels with a kernel size of 3×1 and a stride of 1. The number of convolution kernels of the two Conv layers in the post-activation residual block are 128 and 64 respectively. Finally, the output signal features are obtained after a post-activation residual block.

[0085] 2.4. Input the signal features output by the post-activation residual block into the deep residual module stacked by 5 pre-activation residual blocks. The specific implementation diagram of each pre-activation residual block is as follows: Figure 3b As shown in the figure, each pre-activation residual block starts with a BN layer, followed by a Relu layer, a Conv layer, a BN layer, a Relu layer, and a Conv layer. Similarly, a MaxPool layer is used at the skip connection of the residual learning framework. The first Conv layer in each pre-activation residual block uses a convolution kernel with a kernel size of 3×1 and a stride of 1; the second Conv layer uses a convolution kernel with a kernel size of 3×1 and a stride of 2; the MaxPool layer uses a convolution kernel with a kernel size of 1×1 and a stride of 2. After five pre-activation residual blocks with the same structure are stacked and the output signal features are finally connected to the BN layer and the Relu layer to complete the preliminary deep feature extraction process of rolling bearings based on the one-dimensional deep residual convolutional neural network.

[0086] Specifically, the step 3 of trend-based feature construction includes:

[0087] For the deep features obtained by the one-dimensional deep residual convolutional neural network, a deep LSTM network is used to control the flow of information in the time series to capture the long-term and short-term dependency information in the time series, thereby constructing effective trend features for rolling bearings. The LSTM memory unit selectively memorizes and forgets the current input information and the previous memory state through three gate controllers: input gate i, forget gate f, and output gate o; the three gates can control the flow of information in the time series, capture the long-term and short-term dependency information in the time series data, thereby constructing effective trend features for rolling bearings. The structure of the LSTM basic unit at time t is as follows: Figure 4 shown.

[0088] Step 3.1. Calculate the value f of the forget gate in the LSTM basic unit t , the forget gate selectively forgets the memory unit state value c at the previous moment t-1 The historical information in the current memory unit state value c t The effect is to filter the information in long-term memory;

[0089] f t =σ[W f (h t-1 ,x t )+b f ] (4)

[0090] Step 3.2, calculate the candidate memory unit value k t , in updating memory cell c t Before, candidate memory unit k will be generated t , k t is the deep feature x input at the current time t t And the output h of the hidden layer unit at the previous time t-1 t-1 Jointly decide and perform linear combination of weight matrices to obtain the candidate memory unit value k at the current moment t , update the memory cell state c t Information:

[0091] k t =tanh[W k (h t-1 ,x t )+b k ] (5)

[0092] Step 3.3, calculate the value of the input gate i t , the input gate selectively memorizes the current deep feature data input x t For the memory cell state value c t The influence of , stores the information in the input sample into the current state;

[0093] it =σ[W i (h t-1 ,x t )+b i ] (6)

[0094] Among them, the value of the input gate i t and candidate memory value k t It is to update the information in short-term memory, combined with h t-1 , from the input x t Extract important information from

[0095] Step 3.4: Calculate the memory cell state value c at the current time t t ,

[0096]

[0097] Where: Represents the point-by-point product. From formula (7), we can see that the memory unit state c t The update depends on the memory cell value c at the previous moment t-1 and the current candidate memory cell value k t , the LSTM basic unit is through the forget gate f t and input gate i t The two door controllers are respectively t-1 and k t These two parts perform selective memory and forgetting;

[0098] Step 3.5, calculate the value of the output gate o t , the output gate controls the output of the memory unit state value,

[0099] o t =σ[W o (h t-1 ,x t )+b o ] (8)

[0100] Step 3.6. Calculate the output h of the LSTM memory unit t , through the output gate o t Pass the information to the next neuron,

[0101] h t =o t tanh(c t ) (9)

[0102] In formulas (4)-(9), W f 、W k 、W i 、W o They are respectively the forget gate f and the candidate memory unit k at time tt , the weight values ​​of input gate i and output gate o, b f 、b k 、b i 、b o They are respectively the forget gate f and the candidate memory unit k t , bias of input gate i, output gate o, x t is the input at the current time t, h t-1 is the output of the network at the previous time t-1, σ is the sigmoid function, the value is (0,1), and the tanh function, the value is (-1,1).

[0103] Specifically, the model training phase in step 4 includes:

[0104] Step 4.1, the specific implementation diagram of the rolling bearing RUL prediction method of the present invention is as follows Figure 5 As shown in the figure, a one-dimensional deep residual convolutional neural network and a deep LSTM network model were built based on the PyTorch deep learning framework. When creating the network model, the network hyperparameters for the one-dimensional residual convolutional neural network, the number of network layers and hidden nodes for the deep LSTM network, the number of network layers and nodes for the fully connected neural network, the learning rate for the RUL prediction network, and the number of iterations for network training were specified.

[0105] Step 4.2, train the rolling bearing RUL prediction network model on the training set. Input the rolling bearing training set constructed in step 1 into step 2 to preliminarily extract the deep features of the rolling bearing, input the deep features of the rolling bearing extracted by the one-dimensional deep residual convolutional neural network in step 2 into step 3 to construct the trend features of the rolling bearing, and finally use the linear layer of the fully connected neural network to fit the performance degradation trend of the rolling bearing. Use the Adam optimization algorithm to optimize the hyperparameters in the one-dimensional deep residual convolutional network, the deep LSTM network and the fully connected neural network to complete the life prediction of the training set. In order to evaluate the accuracy of the model prediction, the present invention quotes the performance evaluation index in the International Conference on Fault Prediction and Health Management data competition, namely the root mean square error (RMSE), as shown in formula (10). Finally, the model training phase is completed, and a rolling bearing RUL prediction model based on residual convolutional neural network and LSTM is obtained;

[0106]

[0107] Where M is the total number of data sample points; d i =RUL' i -RUL i , which represents the difference between the predicted value and the true value. The smaller the value of the root mean square error, the higher the prediction accuracy of the model.

[0108] Specifically, the model testing phase in step 5 includes:

[0109] A test set was constructed using non-full-lifecycle time-series vibration signals from rolling bearings under different operating conditions. The model's accuracy and reliability were tested on this test set. The non-full-lifecycle time-series vibration signals from the test set under different operating conditions were normalized and input into the rolling bearing RUL prediction network model based on the residual convolutional neural network and LSTM, obtained during the model training phase in Step 4. This completed the model testing phase and predicted the RUL values ​​of the rolling bearings in the test set.

[0110] The vibration time domain signal used in this embodiment comes from the PRONOSTIA test bench. Figure 6 As shown. The data is collected by two acceleration sensors in the horizontal and vertical directions respectively. The data is recorded every 10s, the sampling time is 0.1s, and the sampling frequency is 25.6kHz, that is, the vibration data is 2560 points each time. The data collected contains the full life cycle time series vibration signals of 15 rolling bearings from operation to failure obtained in the accelerated degradation experiment. The experiment includes three working conditions: working condition 1 (1800rpm and 4000N), working condition 2 (1650rpm and 4200N) and working condition 3 (1500rpm and 5000N). The test rolling bearings under the three working conditions are rolling bearings 1_1 to rolling bearings 1_7, a total of 7 groups, rolling bearings 2_1 to rolling bearings 2_7, a total of 7 groups, and rolling bearings 3_1 to rolling bearings 3_3, a total of 3 groups.

[0111] In this example, rolling bearings 1_1 and 1_2 under operating condition 1, rolling bearings 2_1 and 2_2 under operating condition 2, and rolling bearings 3_1 and 3_2 under operating condition 3 are selected as the training set, and rolling bearings 1_5, 2_7, and 3_3 are selected as the test set. The training set consists of time-series vibration signals of rolling bearings over their entire lifecycle, while the test set consists of time-series vibration signals of rolling bearings over a period of part of their lifecycle.

[0112] The present invention was experimentally verified on the constructed rolling bearing training set and test set, and a one-dimensional deep residual convolutional neural network and a three-layer deep LSTM network were used to predict the RUL value of rolling bearings.

[0113] The verification results are shown in the following table. The left column in the table shows the experimental results of the rolling bearing RUL prediction method based on residual convolutional neural network and LSTM of the present invention, and the right column shows the experimental results of the rolling bearing RUL prediction method based on CNN and LSTM.

[0114]

[0115] The experimental verification results are shown in the figure Figure 7a-7b As shown, the verification result diagram of the present invention is shown in FIG. Figure 7a As shown, Figure 7b This is a verification result diagram of the rolling bearing RUL prediction method based on CNN and LSTM. Combining the above table and the result diagram, it can be seen that the rolling bearing RUL prediction method based on residual convolutional neural network and LSTM can well predict the RUL value of the rolling bearing. The present invention realizes the intelligent life prediction of the rolling bearing, solves the life prediction and intelligent decision-making problems of the rolling bearing parts during operation, and improves the prediction accuracy. At the same time, the present invention solves the problem that the existing rolling bearing vibration signal feature extraction method is too dependent on expert experience, and proposes a one-dimensional deep residual convolutional neural network that can adaptively extract the deep-level features of the rolling bearing vibration signal, realize end-to-end rolling bearing life prediction, solve the life prediction and intelligent decision-making problems of the rolling bearing during operation, and replace the equipment in time before the failure occurs to avoid huge economic losses.

Claims

1. A method for predicting the remaining service life of rolling bearings based on residual convolutional neural network and LSTM, characterized in that: The following steps are involved: Step 1: Data preprocessing: Obtain the time series vibration signal of the bearing throughout its life cycle and construct the training set D under different working conditions. train ={x t ,y t } T , D train is the full life cycle time series vibration signal of a bearing under a certain working condition, x t ∈R P×2 The feature input of the bearing is P×2, P=2560, y t ∈[0,1] is the output of the prediction model, T is the failure time of the bearing during operation, that is, the full life value of the bearing; Step 2: Preliminary feature extraction. The specific process includes: 2.

1. Build a one-dimensional deep residual convolutional neural network model: including a post-activation residual block and a pre-activation residual block; use the residual blocks of different structures in the one-dimensional residual convolutional neural network to process the entire input vibration signal sequence, and preliminarily adaptively mine the deep degradation characteristics of the vibration signal; the x obtained after normalization of the time series vibration signal is converted into t Input into the one-dimensional deep residual convolutional neural network model G; 2.2、Normalized bearing time series vibration signal x t Perform the first convolution operation. The stride size of the first Conv layer is s1, the convolution kernel uses a size of k1×1, and the number of convolution kernels is m1. Then, a BN layer is used to accelerate the calculation and optimization of the network. The nonlinear activation function ReLU is used for nonlinear operation to output the first layer of signal features. 2.

3. The features output by the first Conv layer are input into the post-activation residual block. The post-activation residual block starts with the Conv layer, followed by the BN layer, the Relu layer, and finally the Conv layer. A Dropout layer is added after the Relu layer. A MaxPool layer is added to the residual learning framework. The structures of the two Conv layers and the MaxPool layer both use a convolution kernel of size k2×1 and a stride size of s2. In the post-activation residual block, the number of kernels of the two convolution kernels is m2 and m3 respectively, and finally the features are output. 2.

4. The features output by the post-activation residual block are input into a deep residual module stacked by multiple one-dimensional residual modules, namely pre-activation residual blocks; the pre-activation residual block starts with the BN layer, followed by the Relu layer, Conv layer, BN layer, Relu layer, and Conv layer, and the MaxPool layer is used in the residual structure; the first Conv layer uses a convolution kernel size of k3×1 and a stride size of s3; the second Conv layer uses a convolution kernel size of k4×1 and a stride size of s4; the MaxPool layer uses a convolution kernel size of k5×1 and a stride size of s5; after several pre-activation residual blocks, the deep features of the bearing are output, and finally the BN layer and the Relu layer are connected to complete the preliminary deep feature extraction process of the rolling bearing based on the residual convolutional neural network; Step 3: Trend feature construction: Input the output of the one-dimensional residual convolutional neural network into the deep LSTM to further construct the trend feature information of the rolling bearing; Step 4: Model training: Steps 2 and 3 construct quantitative features of bearing trends, and then use the linear layer of a fully connected neural network to fit the performance degradation trend of the rolling bearing. Use the Adam optimization algorithm to optimize all hyperparameters, complete the self-extraction of deep-level bearing features and RUL prediction, and thus obtain a bearing RUL prediction model based on a residual convolutional neural network and LSTM. Step 5: Model testing.

2. The method for predicting the remaining service life of a rolling bearing based on a residual convolutional neural network and LSTM according to claim 1, characterized in that: The data preprocessing process in step 1 includes: For the full-cycle time series vibration signals of the rolling bearing under different working conditions obtained in step 1, min-max normalization processing is performed, and the conversion function is as follows: Among them, x is the time series vibration signal before normalization, x t is the normalized time series vibration signal, x min is the minimum value of the time series vibration signal, x max is the maximum value of the time series vibration signal; The normalized vibration signal is used as the input of the overall prediction model. After continuous iterative training of the neural network to optimize its network parameters, the network finally outputs the RUL prediction value corresponding to the data set. Assuming that each training sample is {(x i ,y i )}, where x i is the vibration data feature of the i-th sample; y i is the label value of the i-th sample, that is, the corresponding real rolling bearing RUL; the health status of the rolling bearing degrades linearly with the running time. Based on this condition, y i The calculation formula is as follows: Where t is the failure time of the bearing, that is, the full life value of the bearing; tt i is the RUL value of the i-th sampling point, t0 is the starting time of the bearing operation, and N is the total sampling points of the bearing.

3. The method for predicting the remaining service life of a rolling bearing based on a residual convolutional neural network and LSTM according to claim 1, characterized in that: The process of step 3, trend feature construction, includes: Step 3.

1. Calculate the value f of the forget gate in the LSTM basic unit t , f t The state value c of the memory unit at the previous moment t-1 t-1 The historical information in the memory is selectively forgotten, so that the state value c of the memory unit at the current time t is t To have an impact is to filter the information in long-term memory; f t =σ[W f (h t-1 ,x t )+b f ] (4) Step 3.2, calculate the candidate memory unit value k t , in the update c t Before, k will be generated first t ; Input x at the current time t t And the output h of the network at time t-1 t-1 Jointly decide t Value; by h t-1 and x t Perform corresponding calculations to obtain the k at the current time t t , update c t Information: k t =tanh[W k (h t-1 ,x t )+b k ] (5) Step 3.3, calculate the value of the input gate i t ,i t Selectively memorize x t C t The influence of , stores the information in the input data into the current state; i t =σ[W i (h t-1 ,x t )+b i ] (6) Among them, i t and k t Both are to update the information in short-term memory, combined with h t-1 , from x t Extracting important information from memory is to filter the information in short-term memory; Step 3.4, calculate c at the current time t t value, Where: represents the point-by-point product. From formula (7), we can see that c t-1 and k t The value of c t The LSTM basic unit structure is updated by f t and i t The two door controllers are respectively t-1 and k t These two parts are selectively memorized and forgotten, so as to obtain c at the current time t t value; Step 3.5, calculate the value of the output gate o t , o t Control the current moment c t The output, the t =σ[W o (h t-1 ,x t )+b o ] (8) Step 3.

6. Calculate the output h of the LSTM basic unit t , through o t Pass the information to the next neuron, h t =o t fishy t ) (9) In formulas (4)-(9), W f 、W k 、W i and W o are f at time t respectively t 、k t 、i t and o t The weight value, b f 、b k 、b i and b o f t 、k t 、i t and o t Bias, x t is the input at the current time t, h t-1 is the output of the network at time t-1, σ is the sigmoid function, the value is (0,1), and the tanh function, the value is (-1,1); in formula (9), the output h of the LSTM basic unit is t That is the trend bearing feature information extracted by the deep LSTM network.

4. The method for predicting the remaining service life of a rolling bearing based on a residual convolutional neural network and LSTM according to claim 1, characterized in that: The step 4, model training, includes: Step 4.1: Build a bearing RUL prediction model based on the PyTorch deep learning framework. Specify the hyperparameters of the one-dimensional deep residual convolutional neural network, the number of network layers and hidden layer nodes of the deep LSTM network, the number of network layers and network nodes of the fully connected neural network, the network learning rate, and the number of iterations during network training. Step 4.2: Train the rolling bearing RUL prediction network model on the training set: Input the rolling bearing training set constructed in step 1 into step 2 to preliminarily extract the deep features of the rolling bearings. Input the deep features of the rolling bearings extracted by the one-dimensional deep residual convolutional neural network in step 2 into step 3 to construct the trend features of the rolling bearings. Finally, use the linear layer of the fully connected neural network to fit the performance degradation trend of the rolling bearings. Use the Adam optimization algorithm to optimize the hyperparameters in the one-dimensional deep residual convolutional network, deep LSTM network, and fully connected neural network to complete the life prediction of the training set. Finally, obtain the rolling bearing RUL prediction model based on the residual convolutional neural network and LSTM: Where M is the total number of data sample points; d i =RUL' i -RUL i , which represents the difference between the predicted value and the true value; the smaller the value of the root mean square error, the higher the prediction accuracy of the model.

5. The method for predicting the remaining service life of a rolling bearing based on a residual convolutional neural network and LSTM according to claim 1, characterized in that: The step 5 model testing includes: The non-full life cycle time series vibration signals under different working conditions of rolling bearings are selected to construct a test set, and the accuracy and reliability of the model are tested on the test set; the non-full life cycle time series vibration signals under different working conditions in the test set are normalized and input into the rolling bearing RUL prediction network model based on residual convolutional neural network and LSTM obtained in the model training phase in step 4, completing the model testing phase and realizing the prediction of the RUL value of the rolling bearings in the test set.

6. The method for predicting the remaining service life of a rolling bearing based on a residual convolutional neural network and LSTM according to claim 5, characterized in that: In the model training of step 4 and the model testing of step 5, rolling bearings 1_1 and 1_2 under working condition 1, rolling bearings 2_1 and 2_2 under working condition 2, and rolling bearings 3_1 and 3_2 under working condition 3 are selected as training sets, and rolling bearings 1_5, rolling bearing 2_7, and rolling bearing 3_3 are selected as test sets; wherein the training set is the time series vibration signal of the rolling bearing over its entire life cycle, and the test set is the time series vibration signal of the rolling bearing over a non-full life cycle.

Images