Method and system for predicting remaining useful life of rolling bearing

Abstract

The present invention relates to a method and system for predicting the remaining useful life of a rolling bearing. The method comprises: acquiring sample data to be predicted, and inputting same into a trained remaining useful life prediction neural network to obtain an output prediction result, wherein the remaining useful life prediction neural network comprises a feature encoder and a regression predictor, and the training process comprises using the feature encoder to preliminarily extract features from source domain sample data; using a temporal mixed contrastive domain adaptation training module to calculate contrastive loss, and using the contrastive loss to iteratively train the feature encoder, so as to further extract mutual information from target domain sample data features as a high-level feature; and using a fine-grained structural domain adaptation training module to calculate domain discrimination loss and a fine-grained matching degree between the source domain sample data and target domain sample data, and using the domain discrimination loss and the fine-grained matching degree to iteratively train the feature encoder, so as to further extract domain-invariant features between a source domain and a target domain. The system comprises an input interface, an output interface, a processor, a computer readable storage medium and stored program instructions, wherein the processor calls the program instruction to train the remaining useful life prediction neural network, calls the program instruction of the trained remaining useful life prediction neural network to instruct the feature extractor to perform feature extraction on rolling bearing vibration data to be detected, and inputs the extracted features into the regression predictor for prediction processing to obtain a prediction result. The present invention effectively improves the accuracy of the prediction result of the remaining useful life of rolling bearings.

Description

Rolling bearing remaining life prediction method and system TECHNICAL FIELD

[0001] The present application belongs to the technical field of mechanical resource prediction, and particularly relates to a rolling bearing remaining life prediction method and system. BACKGROUND

[0002] As an indispensable component in industrial equipment, the health status of rolling bearings is directly related to the entire mechanical system. Therefore, accurately predicting the remaining useful life (RUL) of bearings is crucial for implementing preventive maintenance of equipment and improving production efficiency. In recent years, although scholars at home and abroad have carried out a large amount of research on bearing remaining useful life prediction based on machine learning, deep learning and other artificial intelligence, in actual engineering, due to the influence of factors such as variable speed and variable load, the working conditions of each bearing are different, which leads to differences in bearing data distribution, resulting in a sharp decline in the performance of the prediction model, limiting the application of deep learning in bearing life prediction.

[0003] In order to solve the problem that the traditional machine learning method is seriously degraded due to the inconsistency between the training samples and the test samples, many scholars use the domain adaptation method under the transfer learning framework to extract domain-invariant features, which effectively alleviates the differences between data distributions. However, these domain adaptation methods mainly emphasize the similarity between the target domain and the source domain, while ignoring the mutual information of the target-specific data. Ignoring the mutual information between the current data and other data to construct the prediction task may lead to limitations in prediction. At the same time, ignoring the unknown distribution distance in the transmission task under multiple working conditions, the commonly used distribution distance measurement methods have the problem of weak gradient information or gradient disappearance, resulting in poor accuracy of rolling bearing remaining life prediction under variable working conditions and poor model transferability.

[0004] SUMMARY

[0005] To solve the above technical problems and improve the accuracy of rolling bearing remaining life prediction under variable working conditions and the transferability of the prediction model, the present application proposes a rolling bearing remaining life prediction method and system.

[0006] The first aspect of the present application proposes a rolling bearing remaining life prediction method, which comprises: obtaining sample data to be predicted and inputting it into a pre-trained remaining life prediction neural network to obtain an output prediction result;

[0007] The remaining life prediction neural network includes a feature encoder and a regression predictor, and the training process of the remaining life prediction neural network includes:

[0008] Obtaining source domain sample data and target domain sample data;

[0009] inputting the source domain sample data and the target domain sample data into a time mixed contrast domain adaptation training module, calculating a contrast loss, and iteratively training the feature encoder using the contrast loss;

[0010] inputting the source domain sample data and the target domain sample data into a fine-grained structure domain adaptation training module, respectively calculating a domain discrimination loss and a fine-grained matching degree between the source domain sample data and the target domain sample data, and iteratively training the feature encoder using the domain discrimination loss and the fine-grained matching degree, respectively;

[0011] inputting the source domain sample data into the feature encoder to extract source domain features, inputting the source domain features into a regression predictor to perform prediction, calculating a regression prediction loss, and iteratively training the regression predictor using the regression prediction loss;

[0012] stopping the iteration when the number of iterations reaches a pre-set number of iterations, and obtaining a trained remaining life prediction neural network.

[0013] Further, the time mixed contrast domain adaptation training module includes a feature encoder, a cross-domain time mixed generation module, a momentum encoder, and a projector, and the training process includes: inputting the source domain sample data and the target domain sample data into the cross-domain time mixed generation module, generating target dominant domain sample data based on a cross-domain time mixed strategy; extracting target dominant domain features of the target dominant domain sample data using the momentum encoder; extracting target domain features of the target domain sample data using the feature encoder, embedding the target domain features using the projector to obtain embedded features; calculating a contrast loss between the target dominant domain features and the embedded features, and iteratively training the projector, the feature encoder, and the momentum encoder in sequence according to the contrast loss.

[0014] Further, the process of generating target dominant domain sample data includes: cyclic shift operation to generate aggregated virtual sample data from a plurality of source domain samples, and generating target dominant domain sample data based on a cross-domain time mixed strategy according to the virtual sample data and the target domain sample data.

[0015] Further, the fine-grained domain adaptation module comprises a feature encoder, a fine-grained calculation module and a domain discriminator, and the training process comprises: extracting source domain features and target domain features from the source domain sample data and the target domain sample data by using the feature encoder; inputting the source domain features and the target domain features into the domain discriminator to perform adversarial training, confusing the source domain features and the target domain features, calculating a domain classification loss, and respectively performing back propagation and iterative training of the domain discriminator and the feature encoder according to the domain classification loss; and calculating a Sinkhorn divergence between the source domain features and the target domain features by using the fine-grained calculation module, wherein a result measured by the Sinkhorn divergence is a fine-grained matching degree between the source domain features and the target domain features, and the feature encoder is iteratively trained according to the fine-grained matching degree.

[0016] The second aspect of the application provides a rolling bearing residual life prediction system, comprising: an input interface, an output interface, a processor, a computer readable storage medium and stored program instructions, wherein:

[0017] The input interface is used to obtain training samples or samples to be detected;

[0018] The stored program instructions comprise program instructions programmed to obtain training samples or samples to be detected;

[0019] The program instructions are programmed to train a residual life prediction neural network to obtain a trained residual life prediction neural network, the residual life prediction neural network comprising a feature extractor and a regression predictor;

[0020] The program instructions are programmed to utilize the trained residual life prediction neural network to perform prediction and output a prediction result;

[0021] The processor: calls the program instructions to train the residual life prediction neural network to obtain a trained residual prediction neural network;

[0022] The program instructions of the trained residual prediction neural network are called, and the feature extractor extracts features from the vibration data of the rolling bearing to be detected, and the extracted features are input into the regression predictor for prediction processing to obtain a prediction result;

[0023] The output interface is used to output a residual life prediction value.

[0024] Further, the processor calls the program instructions to train the residual life prediction neural network, and the training process comprises:

[0025] Initialize system parameters and set the number of iterative training times;

[0026] Obtain training sample data through the input interface;

[0027] The processor calls program instructions of the time mixed contrast domain adaptation training module, and the instructions of the time mixed contrast domain adaptation module calculate a contrast loss, and the feature encoder is iteratively trained by using the contrast loss;

[0028] The processor calls program instructions of the fine-grained domain adaptation training module, and the instructions of the fine-grained domain adaptation module respectively calculate a fine-grained matching degree and a domain classification loss between the source domain sample data and the target domain sample data, and the feature encoder is iteratively trained by using the fine-grained matching degree and the domain classification loss respectively;

[0029] The processor calls program instructions of the feature encoder, and the instructions of the feature encoder extract source domain features, call program instructions of the regression predictor, and the instructions of the regression predictor input the source domain features for prediction, calculate a regression prediction loss, and iteratively train the regression predictor by using the regression prediction loss;

[0030] The iteration is stopped when the number of iterations reaches the pre-set number of iterations, and a trained remaining life prediction neural network is obtained.

[0031] The first aspect and the second aspect of the present application are based on the same inventive concept, and the technical features can be mutually referred to, and the same technical features will not be described again.

[0032] The present application has the following advantages: first, the present application constructs a remaining life prediction neural network, and adopts a time mixed contrast domain adaptation training module and a fine-grained domain adaptation training module for training during network training; second, the time mixed contrast domain adaptation training module in the present application uses a cyclic shift method to aggregate virtual samples obtained from source domain samples, uses a fixed proportion of Mixup data enhancement method, performs cross-domain time mixing on the source domain and the target domain, obtains target dominant domain samples, extracts target dominant domain features through a momentum encoder, calculates a contrast loss between the target domain features and the target dominant domain features through contrast momentum learning, iteratively trains the feature encoder through the contrast loss back propagation, and further retains the mutual information of the target domain; third, the fine-grained domain adaptation training module is used to realize the adversarial training between the source domain features and the target domain features, calculate the fine-grained alignment between the target domain features and the source domain features through Sinkhorn divergence, and iteratively train the feature encoder according to the result of back propagation, and further extract the domain invariant features between the source domain features and the target domain features; in this way, the remaining life prediction neural network constructed by the present application can accurately extract the domain invariant features between the source domain features and the target domain features, and can fully utilize the mutual information (i.e. specific information) of the target domain, improve the accuracy of the remaining life prediction result, and improve the transferability of the remaining life prediction model (i.e. the remaining life prediction neural network). BRIEF DESCRIPTION OF DRAWINGS

[0033] Fig. 1 is a flow chart of the training steps of the remaining life prediction neural network in the present application;

[0034] Fig. 2 is a schematic diagram of the training and testing processes of the remaining life prediction neural network in the present application;

[0035] Fig. 3 is a schematic diagram of the training process of the time-mixing contrastive domain adaptation training module in the present application;

[0036] Fig. 4 is a schematic diagram of the training process of the fine-grained domain adaptation training module in the present application;

[0037] Fig. 5 is a schematic diagram of the network structure of the regression predictor, domain discriminator, and projector in the present application;

[0038] Fig. 6 is a schematic diagram of the network structure of the feature encoder and momentum encoder in the present application;

[0039] Fig. 7 is a schematic diagram of the system structure of the rolling bearing remaining life prediction system in the present application;

[0040] Fig. 8 is a flow chart of the training steps of the processor calling program instructions for training the remaining life prediction neural network of the rolling bearing remaining life prediction system in the present application;

[0041] Fig. 9 is a comparison chart of the experimental results of each task under different conditions. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0043] Fig. 1 is a flow chart of the training steps of the remaining life prediction neural network in the present application. In Fig. 1, steps A, B, C, D, and E are only codes and do not represent a specific execution order. Steps B, C, and D are parallel steps.

[0044] FIG. 2 is a schematic diagram of the training process and the testing process of the remaining useful life prediction neural network. In FIG. 2, the remaining useful life prediction neural network includes a feature encoder and a regression predictor. The purpose of network training is to iteratively train the feature encoder and the regression predictor respectively to optimize their performance, and the process includes: obtaining labeled source domain sample data and unlabeled target domain sample data; inputting the source domain sample data and the target domain sample data into a time-mixed contrastive domain adaptation training module to calculate a contrastive loss, and iteratively training the feature encoder using the contrastive loss; inputting the source domain sample data and the target domain sample data into a fine-grained structure domain adaptation training module to respectively calculate a domain discrimination loss and a fine-grained matching degree between the target domain sample data, and iteratively training the feature encoder using the domain discrimination loss and the fine-grained matching degree respectively; inputting the source domain sample data into the feature encoder to extract source domain features, inputting the source domain features into the regression predictor for prediction, calculating a regression prediction loss, and iteratively training the regression predictor using the regression prediction loss. Network testing is to test the performance of the trained remaining useful life prediction neural network for performance evaluation, and the process includes: inputting the labeled source domain sample data into the trained feature encoder to extract source domain features, inputting the source domain features into the trained regression predictor for regression prediction, and obtaining a prediction result. Comparing the prediction result with the source domain sample data label to obtain a performance evaluation result.

[0045] FIG. 3 is a schematic diagram of the training process of the time-mixed contrastive domain adaptation training module. In FIG. 3, the time-mixed contrastive domain adaptation training module includes a feature encoder, a cross-domain time-mixed generation module, a momentum encoder, and a projector, and the purpose is to train the feature encoder by contrastive learning to further extract the mutual information of the target domain features. The training process is as follows: inputting the source domain sample data and the target domain sample data into the cross-domain time-mixed generation module to generate target dominant domain sample data based on a cross-domain time-mixed strategy, and the target dominant domain sample is the positive sample in the contrastive learning. The feature encoder (i.e. the positive encoder in momentum learning) extracts the target domain features, and the projector is used to embed the target domain features to obtain embedded features; the momentum encoder extracts the target dominant domain features from the target dominant domain sample data; the network structure of the momentum encoder is the same as that of the feature encoder, and the parameters of the momentum encoder are updated based on the momentum update principle according to the parameters of the feature encoder; calculating the contrastive loss between the target dominant domain features and the target domain features, and iteratively updating the projector and the feature encoder by backpropagation through the contrastive loss.

[0046] FIG. 4 is a schematic diagram of the training process of the fine-grained domain adaptation training module. In FIG. 4, the fine-grained domain adaptation module includes a feature encoder, a fine-grained calculation module, and a domain discriminator. The purpose is: 1. Adversarial training to reduce the difference between the source domain features and the target domain features and extract domain-invariant features; 2. To achieve fine-grained alignment of the source domain features and the target domain features. The training process includes: extracting source domain features and target domain features from the source domain sample data and the target domain sample data using the feature encoder; inputting the source domain features and the target domain features into the domain discriminator to perform adversarial training, confusing the source domain features and the target domain features, calculating the domain classification loss, and iteratively training the domain discriminator and the feature encoder according to the domain classification loss; calculating the Sinkhorn divergence between the source domain features and the target domain features through the fine-grained calculation module. The result measured by the Sinkhorn divergence is the fine-grained matching degree between the source domain features and the target domain features. The feature encoder is iteratively trained according to the fine-grained matching degree.

[0047] It should be noted that each feature encoder in FIGS. 2-4 is a feature encoder of the remaining life prediction neural network, which has the same network structure, and the parameters are shared in the training process of each network.

[0048] The embodiment of the present application proposes a rolling bearing remaining life prediction method. Referring to FIGS. 1 and 2, the method comprises: obtaining sample data to be predicted and inputting it into a pre-trained remaining life prediction neural network to obtain an output prediction result.

[0049] The remaining life prediction neural network comprises a feature encoder and a regression predictor. The training process of the remaining life prediction neural network comprises:

[0050] Step A: obtaining source domain sample data and target domain sample data;

[0051] Step B: inputting the source domain sample data and the target domain sample data into the time-mixed contrast domain adaptation training module to calculate a contrast loss, and iteratively training the feature encoder using the contrast loss;

[0052] Step C: inputting the source domain sample data and the target domain sample data into the fine-grained domain adaptation training module to calculate a domain discrimination loss and a fine-grained matching degree between the source domain sample data and the target domain sample data, respectively, and iteratively training the feature encoder using the domain discrimination loss and the fine-grained matching degree;

[0053] Step D: inputting the source domain sample data into the feature encoder to extract source domain features, inputting the source domain features into the regression predictor for prediction, calculating a regression prediction loss, and iteratively training the regression predictor using the regression prediction loss;

[0054] Step E: stopping the iteration when the number of iterations reaches the pre-set number of iterations, to obtain the trained remaining life prediction neural network.

[0055] Before step A, further comprising collecting vibration data of the rolling bearing by using a sensor, the vibration data comprising a rotating part, a degradation generating part, a vibration part and a temperature; dividing the collected vibration data into a source domain sample set of a full life cycle and a target domain sample set of incomplete life cycle data, and taking the ratio of the remaining service life of the rolling bearing to the total service life as the label of the source domain sample. Specifically:

[0056] The data acquisition part of the rolling bearing mainly consists of three parts: a rotating part, a degradation generating part and a measuring part. The rotating part uses an asynchronous motor and two shafts. The motor drives the gear box to transfer the rotating motion, so as to keep the speed of the secondary shaft at a certain speed and provide a rated torque below 2000 revolutions per minute. The degradation generating part mainly consists of a special aluminum plate, which supports the pneumatic jack, the vertical shaft and its lever arm, the force sensor, the test bearing clamping ring, the support test bearing shaft, the two shaft tables and their large super bearings, thereby generating a radial force to shorten the service life of the bearing. The measuring part mainly consists of vibration and temperature sensors. The vibration sensor consists of two micro accelerometers at 90° to each other. The first one is placed on the vertical shaft, and the second one is placed on the horizontal shaft. The two accelerometers are placed radially on the outer ring of the bearing. The temperature sensor is an RTD (Resistance Temperature Detector) platinum PT100 (1 / 3 DIN level) probe placed in a hole close to the outer bearing ring. The temperature measurement is sampled at 0.1 Hz. By installing acceleration sensors with a sampling frequency of 25.6 kHz in the vertical and horizontal directions on different types of rolling bearings, vibration data is collected with a sampling interval of 10 seconds and a sampling duration of 0.1 seconds. The collected vibration data is divided into a source domain vibration data sample set of a full life cycle and a target domain vibration data sample.

[0057] The basic processing method for the original collected vibration data set is: collecting the time domain vibration signal of the bearing Converting the time domain signal to a frequency domain signal using fast Fourier transform as input. The fast Fourier transform is performed on the vibration data sample of the mechanical part to obtain the Fourier phase spectrum of the vibration data sample, and the labeled data is divided into a source domain and the unlabeled data is divided into a target domain.

[0058] The ratio of the remaining life of the source domain bearing to the full life is taken as the label of the source domain sample, which is specifically represented as:

[0059] where y i represents the label of the source domain sample, t iis the remaining useful life at the i-th time, T is the time used by the bearing from the start of operation until failure. Since the data is collected at an average interval, y i is uniformly distributed between 0 and 1.

[0060] Fig. 5 is a schematic diagram of the network structure of the regression predictor, domain discriminator and projector in the embodiment of the present application. In the embodiment of the present application, the regression predictor is composed of three fully connected layers, and the hidden neurons are 128, 32 and 1, respectively. The domain discriminator is composed of three fully connected layers, each of which is composed of a fully connected layer and a Dropout layer, and the hidden neurons are 480, 480 and 1, respectively; the proportion of the dropout regularization used is 0.5. The projector is composed of three fully connected layers, and the hidden neurons are 128, 32 and 960, respectively.

[0061] Fig. 6 is a schematic diagram of the structure of the feature encoder and momentum encoder in the embodiment of the present application. The network structure of the feature extractor and the momentum encoder is the same, and each is composed of three one-dimensional convolutional layers, three pooling layers and a Gate Recurrent Unit (GRU) layer, the convolution kernel of the convolutional layer is 7, 5 and 3, respectively, and the output neuron number is 80, 160 and 320, respectively, and the output dimension of the GRU is 1440. The GRU is a kind of recurrent neural network (RNN).

[0062] While considering obtaining the mutual information of the target domain specific data, the difference between the source domain feature distribution and the target domain feature distribution cannot be ignored. In the contrast learning implemented in the embodiment of the present application, enhanced samples are needed to construct positive and negative sample pairs, therefore, a fixed proportion Mixup technology is applied to construct a cross-domain time mixing strategy to obtain target dominant domain data to provide positive samples for contrast learning.

[0063] The existing Mixup technology generates additional virtual samples by linearly interpolating the labels of randomly selected samples, and in the training process, the neural network learns a function with lower complexity to reduce the generalization error without increasing the model complexity. The mathematical expression of Mixup is:

[0064] In the formula, x i represents the feature vector and label value of any sample i in the sample set, x j represents the feature vector and label value of any sample j in the sample set, represents the feature vector of the newly generated sample, and λ represents the fusion coefficient.

[0065] Further, the time-mixed contrastive domain adaptation training module comprises a feature encoder, a cross-domain time-mixed generation module, a momentum encoder and a projector, and the step B specifically comprises:

[0066] S201: input the source domain sample data x s and the target domain sample data x t into the cross-domain time-mixed generation module to generate the target dominant domain sample data x td based on a cross-domain time-mixed strategy.

[0067] S202: extract the target dominant domain feature of the target dominant domain sample data by using the momentum encoder .

[0068] S203: extract the target domain feature of the target domain sample data by using the feature encoder F(·) .

[0069] S204: calculate the contrastive loss between the target dominant domain feature and the embedded feature, and iteratively train the projector, the feature encoder and the momentum encoder in sequence according to the contrastive loss.

[0070] Further, in S201, the process of generating the target dominant domain sample data comprises: cyclically shifting the aggregated virtual sample data from a plurality of source domain samples, and generating the target dominant domain sample data based on the cross-domain time-mixed strategy according to the virtual sample data and the target domain sample data.

[0071] S201 specifically comprises:

[0072] The embodiment uses a fixed Mixup ratio 0.5<λ<1 on the basis of the existing Mixup technology, and learns the time information of the source domain data by aggregating a plurality of forward and backward time steps of the source domain, mixes with one time step of the target domain, generates the target dominant sample by mixing each sample, takes the aggregated value (T / 2 backward time steps and T / 2 forward time steps) from T time steps of the source domain, wherein the target domain proportion is λ (i.e. the fusion coefficient in the original Mixup), and the source domain proportion is 1-λ; and the cyclic shift operation generates the aggregated virtual sample.

[0073] The source domain sample data and the target domain sample data are input into the time-mixed contrastive domain adaptation module, and the target dominant domain sample data is generated according to the cross-domain time-mixed strategy, and the process comprises:

[0074] S201a: cyclically shift the aggregated virtual sample from a plurality of source domain samples:

[0075] wherein, M represents the virtual sample data, x represents the source domain sample data, which can be a one-dimensional or higher-dimensional data structure, h represents the position of a certain element in the source domain sample data, i represents the center position of the cyclic shift operation, T represents the window size of the aggregated data, N represents the length of the array, (h)modN represents the modulo operation, and when the index range exceeds the array length, the modulo operation will start from the beginning.

[0076] S201b: generating a target dominant domain sample based on a cross-domain time mixing strategy according to the virtual sample and the target sample, and the process includes:

[0077] wherein, represents the generated target dominant domain sample, i represents the serial number, and λ represents the target domain proportion, represents a target domain sample, 1-λ represents the source domain proportion, and T represents the mixing window length, represents a source domain sample, j represents the integer serial number after the modulo operation, and the value range of j is [i-T / 2, i+T / 2].

[0078] In this method, the target domain sample data is the anchor point in the contrast learning, and the constructed target dominant domain sample is the positive sample in the contrast learning, which includes the features of the target domain sample, i.e. the unique information of the target domain.

[0079] To further retain the unique information of the target domain, the embodiment of the application implements momentum contrast learning between the target domain and the target dominant domain to capture the mutual information of the target specific data, so as to retain the unique information of the target domain.

[0080] Further, the feature extractor includes a convolutional neural network and a gated recurrent unit, and the network structure of the feature extractor is the same as that of the momentum encoder.

[0081] In S202, the parameters θ q of the feature extractor are updated based on the momentum update principle to update the parameters θ k of the momentum encoder, and the momentum update principle is specifically: θ k ←αθ k +(1-α)θ q

[0082] wherein, α represents a momentum factor, α∈[0,1), and α usually takes a value very close to 1, and the initial parameters of the momentum encoder are derived from the feature encoder.

[0083] S204 is specifically: using the target dominant domain features of the early batch Construct a dynamic dictionary J with a queue structure. Input the encoded features of the current batch at the head of the queue and pop the oldest batch features from the tail. After each training iteration, input the current target dominant domain features. The target dominant domain features will be calculated and stored in queue J. and embedded features The contrast loss is used, and the loss function for the contrast loss is InfoNCE loss.

[0084] The specific formula for calculating the InfoNCE loss is as follows:

[0085] Among them, L CL Let represent the InfoNCE loss, N represent the total number of samples in the target domain, τ represent the temperature coefficient (τ > 0), J represent the dynamic dictionary, and Q(·) represent the projector. This represents the target domain features extracted from the target domain sample data after passing through the feature encoder (i.e., the positive encoder in contrastive learning). This represents the target dominant domain features extracted after the target dominant domain sample passes through the momentum encoder.

[0086] The contrastive loss updates the parameters θ of the feature encoder through backpropagation. q Since the feature encoder and momentum encoder have the same network structure, the feature encoder is essentially the positive encoder in momentum contrastive learning. Each time the feature encoder updates its parameters, the momentum encoder's parameters θ are updated based on the momentum update principle. k .

[0087] By using contrastive learning, we can further extract mutual information from specific data in the target domain, so that during temporal mixed contrastive domain adaptation training, we can both retain domain-invariant features and consider the unique features of the target domain.

[0088] In this invention, the fine-grained structural domain adaptation training module is used to extract domain-invariant features between the target domain and the source domain. This module extracts domain-invariant features from two aspects: confusing the features of the source domain and the target domain, and fine-grained features between them.

[0089] Furthermore, the fine-grained structural domain adaptation training module includes a feature encoder, a fine-grained computation module, and a domain discriminator. Step C specifically includes:

[0090] S301: Extract source domain features and target domain features from the source domain sample data and target domain sample data using the feature encoder F(·).

[0091] S302: Input the source domain features and target domain features into the domain discriminator D(·) for adversarial training to confuse the source domain features and target domain features, and calculate the domain classification loss L.disc The domain discriminator and the feature encoder are iteratively trained respectively according to the domain classification loss.

[0092] Specifically, the output of the domain discriminator is passed through a gradient reversal layer to maximize the domain classification loss, and the domain discriminator realizes game learning of domain-invariant features between the features extracted from the source domain and the target domain. The domain classification loss L disc The calculation formula of the domain classification loss L

[0093] Wherein, L disc represents the domain classification loss, N s represents the total number of source domain samples, represents the source domain sample, represents the domain label of the source domain, N t represents the total number of target domain samples, represents the target domain sample, represents the domain label of the target domain, L BCE (·, ·) represents the binary cross-entropy loss.

[0094] S303: Calculate the Sinkhorn divergence between the source domain sample data and the target domain sample data through the fine-grained calculation module, and the result measured by the Sinkhorn divergence is the fine-grained matching degree between the source domain sample data and the target domain sample data. The feature encoder is iteratively trained according to the fine-grained matching degree.

[0095] The features of the source domain sample data and the target domain sample data are extracted to obtain the source domain features and the target domain features, and the fine-grained alignment is performed on them. Specifically, the Sinkhorn divergence is used to measure the distribution difference between the source domain features and the target domain features, and the parameters of the feature encoder are optimized according to the Sinkhorn divergence. Specifically:

[0096] The Sinkhorn divergence is an entropy-regularized Wasserstein distance that can support non-intersecting feature distributions.

[0097] The Sinkhorn divergence between the source domain features and the target domain features is calculated, and the specific calculation formula is:

[0098] In the formula, S ηdenotes Sinkhorn divergence, denotes regularization parameter, η > 0, C denotes cost matrix, P is a parameter for controlling the distance measure (e.g. Manhattan distance for p = 1, Euclidean distance for p = 2), H(P) denotes optimal coupling, Π(μ,ν) denotes probability coupling representation, μ denotes probability divergence measure of source domain, ν denotes probability divergence measure of target domain, μ is a non-negative vector of n, v is a non-negative vector of m, μ + v = 1, n and m represent the number of discrete points of two probability distributions respectively. i denotes feature i, a j denotes feature j, the probability coupling matrix Π is composed of n + m equality constraints, and the cost matrix C represents the transmission cost from position i to position j.

[0099] The higher the regularization parameter η, the smoother the coupling matrix; when η tends to zero, the coupling matrix is closer to the optimal transport solution.

[0100] The regularization optimal transport distance problem is introduced as:

[0101] where OT c,η (μ,ν) denotes the regularization optimal transport distance, min() denotes the minimum operation, c(a i ,a j ) denotes the transport cost required to move a unit mass from a i to a j , π(a i ,a j ) denotes the transmission probability, and in the regularization OT problem denoted as H(P) is the optimal coupling, and η denotes the regularization parameter, η > 0.

[0102] The Sinkhorn-Knopp method is used to solve the Sinkhorn iteration to the optimal solution.

[0103] The feature encoder is continuously optimized through the Sinkhorn divergence back propagation, so that the distribution difference between the source domain features and the target domain features is continuously reduced, thereby the domain-invariant features between the source domain features and the target domain features are further preserved by the contrast domain adaptive prediction network.

[0104] The step D can be specifically:

[0105] The source domain sample data is input into the feature encoder to extract the source domain features, the source domain features are input into the regression predictor, the remaining life of the source domain data is predicted, the mean square error loss is used to calculate the loss between the real label domain labels, the regression predictor of the model is continuously optimized through back propagation, and the calculation formula of the regression prediction loss is specifically:

[0106] wherein, L c represents a regression prediction loss, N s represents the total number of source domain samples, L pred represents a mean squared error loss, F(·) represents a feature extractor, C(·) represents a regression predictor, represents a source domain sample, represents a label of the source domain sample.

[0107] The embodiment of the present application provides a computer program product, characterized by comprising a computer readable storage medium, and program instructions for executing the rolling bearing residual life prediction method are stored on the computer readable storage medium.

[0108] The software and hardware environment of the present application is implemented as follows:

[0109] The present application can be a system, a method, and / or a computer program product. The computer program product can include a computer readable storage medium (or media) having computer readable program instructions stored therein for causing a processor to perform aspects of the present application.

[0110] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium can be, for example but not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. A computer readable storage medium includes a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical hard disk encoding device, such as a punch card or a raised structure in a groove on which instructions are recorded, and any suitable combination of the foregoing.

[0111] The computer readable program instructions can be downloaded from the computer readable storage medium to a corresponding computing / processing device, or to an external computer or an external storage device via a network / wireless network (for example, the Internet, a local area network, a wide area network, etc.). The network can include copper transmission cables, optical transmission fibers, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer readable program instructions from the network and forwards the computer readable program instructions to the computer readable storage medium for storage in the corresponding computing / processing device.

[0112] Computer readable program instructions for carrying out operations of the present application can be in assembly, instruction set architecture (ISA), machine instructions, machine dependent instructions, microcode, firmware instructions, state-setting data, or code, written in any combination of one or more programming languages, including an object oriented programming language such as Smalltalk, C++ or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The computer readable program instructions can execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer can be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computer (for example, through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, programmable logic circuitry, field-programmable gate array (FPGA), or programmable logic array (PLA) can execute the computer readable program instructions by utilizing state information of the computer readable program instructions to personalize the electronic circuitry, in order to perform aspects of the present application.

[0113] The methods, systems, and computer program products of embodiments of the application are described by flowcharts and / or block diagrams in the figures. It will be understood that each block of the flowchart and / or block diagrams, and combinations of blocks in the flowchart and / or block diagrams, can be implemented by computer readable program instructions. The computer readable program instructions can be provided to a computer, other programmable data processing apparatus, or other device to produce a machine, such that the instructions, which execute via the computer, other programmable data processing apparatus, or other device, implement the functions / acts specified in the flowchart and / or block diagram block or blocks. Each block in the flowchart and / or block diagram can represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logic function(s). It will also be understood that each block of the flowchart and / or block diagrams, and combinations thereof, can be implemented by computer readable program instructions.

[0114] Based on the same inventive concept, the embodiments of the present application provide a rolling bearing residual life prediction system, in which the technical features (especially the same or similar technical features) in the rolling bearing residual life prediction method can be mutually referred to, and hereinafter the same or similar technical features will not be described again.

[0115] Fig. 7 is a schematic diagram of the system structure of the rolling bearing residual life prediction system. In Fig. 7, the input interface is connected with the peripheral device to obtain the input data, and the peripheral device in the embodiment of the present application includes the vibration sensor. The peripheral device generally includes a device such as a keyboard, a keypad, a touch screen, and / or some other suitable input device for inputting data. The output interface is connected with the display device, and the display device provides a mechanism for displaying data to the user, such as a computer monitor, a display screen (including a display screen of a smart phone / tablet computer), and a display, etc. The computer readable program instructions can implement the "software function", which is a set of machine readable instructions and / or data that can be used to create, manage and control certain software functions. The processor mainly includes a controller and an arithmetic unit, and the main functions are to process instructions, execute operations, control time, and process data. The computer storage medium is mainly used to store readable program instructions and various data.

[0116] In the embodiment of the present application, according to the rolling bearing residual life prediction method, various program instructions are obtained after programming by a computer programming language, and are stored in a computer readable storage medium for the processor to allocate instructions. The sample database is used to store various sample data and pre-process the sample data. The pre-processing includes capturing (obtaining) relevant features (for example, related to residual service life prediction, etc.). The relevant features can be captured / obtained based on sensor measurements, key operation events, operating conditions, laboratory tests / calculations, etc. The pre-processing can extract features using summary statistics, events of interest, etc. The pre-processing can include scaling selected features, for example, by normalization, standardization, etc.

[0117] The embodiment of the present application proposes a rolling bearing residual life prediction system, which, with reference to Fig. 7, includes an input interface, an output interface, a processor, a computer readable storage medium, and stored program instructions, wherein:

[0118] The input interface is used to obtain the training sample or the sample to be detected, and the training sample includes the source domain sample with a label and the target domain sample without a label. The sample in the embodiment of the present application refers to the vibration data of the rolling bearing after pre-processing;

[0119] The stored program instructions include program instructions programmed to obtain the training sample or the sample to be detected;

[0120] The program instructions are programmed to train the residual life prediction neural network to obtain the trained residual life prediction neural network;

[0121] The program instructions are programmed to use the trained residual life prediction neural network to perform prediction processing on the training sample or the sample to be detected and output the prediction result.

[0122] Processor: call program instructions for remaining life prediction neural network training;

[0123] Call program instructions of the trained remaining prediction neural network, and the instruction feature extractor extracts features from the to-be-detected rolling bearing vibration data, and inputs the extracted features into the regression predictor for prediction processing to obtain a prediction result.

[0124] Output interface: for outputting the remaining life prediction value.

[0125] Further, the computer readable storage medium in the system further stores program instructions programmed as a feature encoder for extracting source domain features and target domain features;

[0126] The program instructions are programmed as a time-mixed contrast domain adaptation training module for calculating a contrast loss;

[0127] The program instructions are programmed as a fine-grained domain adaptation training module for calculating a fine-grained matching degree between the source domain features and the target domain features and a domain classification loss;

[0128] The program instructions are programmed as a regression predictor for regression prediction and calculating a regression prediction loss.

[0129] The data processing flow in the program instructions of the time-mixed contrast domain adaptation training module is shown in FIG. 3, and the data processing flow in the program instructions of the fine-grained domain adaptation training module is shown in FIG. 4.

[0130] FIG. 8 is a training step flowchart of the processor calling program instructions for remaining life prediction neural network training in the rolling bearing remaining life prediction system.

[0131] Further, referring to FIG. 8, in the system, the processor calls program instructions for remaining life prediction neural network training, and the training process includes:

[0132] Step a: initialize system parameters, set the number of iteration training times, and the system initialization parameters include: initial parameters of the feature encoder, momentum encoder, domain discriminator, projector, training sample batch, etc.

[0133] Step b: acquire labeled source domain samples and unlabeled target domain samples through the input interface, and store them in the sample database;

[0134] Step c: the processor calls program instructions of the time-mixed contrast domain adaptation training module, the time-mixed contrast domain adaptation module calculates a contrast loss, and the feature encoder is iteratively trained using the contrast loss.

[0135] Step d: the processor calls the program instructions of the fine-grained domain adaptation training module, instructs the fine-grained domain adaptation module to calculate the fine-grained matching degree and domain classification loss between the source domain sample data and the target domain sample data, and iteratively trains the feature encoder using the fine-grained matching degree and the domain classification loss.

[0136] Step e: the processor calls the program instructions of the feature encoder, instructs the feature encoder to extract the source domain features, calls the program instructions of the regression predictor, instructs the regression predictor to input the source domain features for prediction, calculates the regression prediction loss, and iteratively trains the regression predictor through the regression prediction loss.

[0137] Step f: stop iteration when the number of iterations reaches the pre-set number of iterations, and obtain the trained remaining life prediction neural network.

[0138] The steps a, b, c, d, e, and f are only codes of steps and do not represent a specific order. The rolling bearing remaining life prediction system in the embodiment of the application adopts a parallel architecture, and steps c, d, and e are parallel computing, so that the training efficiency can be improved and the training time can be saved.

[0139] Further, the system further comprises a vibration sensor for collecting rolling bearing vibration data and program instructions programmed to pre-process the collected rolling bearing vibration data, and the pre-processing includes:

[0140] The collected vibration data is divided into a source domain sample set of the whole life cycle and a target domain sample set of incomplete life cycle data, and the ratio of the remaining service life of the rolling bearing to the total service life is taken as the label of the source domain sample. The vibration sensor used in the application is DYTRAN 3035B, and the temperature sensor is paltinum RTD PT100PROSENSOR.

[0141] Test effect verification:

[0142] In order to verify the prediction effect of the method proposed in the application on the remaining service life of the key mechanical parts, first, an accelerated degradation experiment is performed on the PRONOSTIA to collect vibration signals, and when the amplitude of the vibration signal exceeds 20g, the bearing is considered to be invalid.

[0143] The collected vibration signals are divided into horizontal and vertical directions, the sampling frequency of the data is 25.6kHz, the sampling interval is 10s, the length of each sampling is 0.1s, and there are 2560 vibration data collected each time. The data set collected is shown in Table 1:

[0144] Table 1 Data set description

[0145] In Table 1, the first working condition (OC1) is that the motor speed is 1800r / min and the load is 4000N; the second working condition (OC2) is that the motor speed is 1650r / min and the load is 4200N; and the third working condition (OC3) is that the motor speed is 1500r / min and the load is 5000N; wherein the working condition 1 and the working condition 2 comprise 7 different bearings, and the working condition 3 comprises 3 different bearings. 1_1-1_7 represent the 7 different bearings contained in the working condition 1; 2_1-2_7 represent the 7 different bearings contained in the working condition 2; and 3_1-3_3 represent the 3 different bearings contained in the working condition 3.

[0146] In order to further compare and analyze the effectiveness of the method of the application, the application and a plurality of existing methods are compared and analyzed:

[0147] (Domain Adaptive Neural Network; DANN) extracts domain invariant features as a baseline method for comparative analysis;

[0148] Deep domain confusion (DDC) measures the distance between features by a maximum mean difference method and learns domain invariant features by reducing the difference in feature distribution between domains as a loss;

[0149] Contrastive adversarial domain adaptation (CADA) introduces a contrastive loss InfoNCE to retain the mutual information of the target domain based on adversarial training, so that the model can consider target-specific information when learning domain-invariant features.

[0150] Fig. 9 shows the residual life prediction curves of the application in six cross-condition tasks under three conditions of the rolling bearing dataset. From (a) to (f), they are the migration tasks from working condition 1 to working condition 2, the migration tasks from working condition 1 to working condition 3, the migration tasks from working condition 2 to working condition 1, the migration tasks from working condition 2 to working condition 3, the migration tasks from working condition 3 to working condition 1, and the migration tasks from working condition 3 to working condition 2. The red curve represents the residual life curve of the bearing under different prediction tasks of the application, and the other curves represent the prediction curves under the methods of DDC, DANN and CADA. As can be seen from Fig. 7, the prediction curve of the application is closer to the actual RUL curve, which verifies the effectiveness of the application.

[0151] The application adopts two commonly used indicators for predicting RUL to evaluate the pros and cons of the model: root mean square error (RMSE) and mean absolute error (MAE):

[0152] wherein a i represents the RUL label, RUL prediction results of the model, N is the number of test set samples. The experimental comparison results are shown in Table 2. Wherein DANN is used as a baseline method.

[0153] Table 2 Comparison of bearing residual life prediction indexes of different domain adaptation tasks

[0154] As can be seen from Table 2, in the experimental results, the method proposed in the present application is significantly better than other methods, and the average RMSE and average MAE values under 6 tasks are 0.142 and 0.165 respectively, thereby proving that the rolling bearing residual life prediction method and system of the present application can better predict the residual service life of the unlabeled target domain bearing, greatly improve the accuracy of the rolling bearing residual life prediction, and also improve the migration of the rolling bearing residual life prediction model (i.e. the comparative domain adaptation prediction network) proposed in the present application.

[0155] A person of ordinary skill in the art can understand that all or part of the steps in the above-mentioned embodiments can be completed by a program instructing the relevant hardware, and the program can be stored in a computer readable storage medium, which can include ROM, RAM, magnetic disk or optical disk, etc.

[0156] Although the embodiments of the present application have been shown and described, it can be understood by those of ordinary skill in the art that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for predicting the remaining life of a rolling bearing, characterized in that, include: Obtain the sample data to be predicted and input it into a pre-trained remaining life prediction neural network to obtain the output prediction result; The remaining lifetime prediction neural network includes a feature encoder and a regression predictor. The training process of the remaining lifetime prediction neural network includes: Obtain source domain sample data and target domain sample data; The source domain sample data and target domain sample data are input into the time-mixed contrastive domain adaptation training module to calculate the contrastive loss, and the feature encoder is iteratively trained using the contrastive loss. The source domain sample data and target domain sample data are input into the fine-grained structural domain adaptation training module, and the domain discrimination loss and the fine-grained matching degree between the source domain sample data and the target domain sample data are calculated respectively. The feature encoder is then iteratively trained using the domain discrimination loss and the fine-grained matching degree respectively. The source domain sample data is input into the feature encoder to extract the source domain features. The source domain features are input into the regression predictor for prediction. The regression prediction loss is calculated and the regression predictor is iteratively trained using the regression prediction loss. The iteration stops when the preset number of iterations is reached, resulting in a trained remaining lifetime prediction neural network.

2. The method for predicting the remaining life of rolling bearings according to claim 1, characterized in that, The temporal mixing contrastive domain adaptation training module includes a feature encoder, a cross-domain temporal mixing generation module, a momentum encoder, and a projector. Its training process includes: inputting the source domain sample data and target domain sample data into the cross-domain temporal mixing generation module to generate target dominant domain sample data based on the cross-domain temporal mixing strategy; extracting target dominant domain features from the target dominant domain sample data using the momentum encoder; extracting target domain features from the target domain sample data using the feature encoder; embedding the target domain features using the projector to obtain embedded features; calculating the contrastive loss between the target dominant domain features and the embedded features; and iteratively training the projector, feature encoder, and momentum encoder sequentially based on backpropagation of the contrastive loss.

3. The method for predicting the remaining life of rolling bearings according to claim 2, characterized in that, The process of generating target dominant domain sample data includes: generating aggregated virtual sample data from multiple source domain samples through cyclic shifting operations; and generating target dominant domain sample data based on the virtual sample data and target domain sample data using a cross-domain temporal mixing strategy.

4. The method for predicting the remaining life of rolling bearings according to claim 1, characterized in that, The fine-grained structural domain adaptation module includes a feature encoder, a fine-grained computation module, and a domain discriminator. Its training process includes: extracting source domain features and target domain features from the source domain sample data and the target domain sample data using the feature encoder; inputting the source domain features and target domain features into the domain discriminator for adversarial training to confuse the source domain features and target domain features, calculating the domain classification loss, and iteratively training the domain discriminator and feature encoder based on the domain classification loss through backpropagation; calculating the Sinkhorn divergence between the source domain features and target domain features through the fine-grained computation module, where the result measured by the Sinkhorn divergence is the fine-grained matching degree between the source domain features and target domain features, and iteratively training the feature encoder based on the fine-grained matching degree through backpropagation.

5. The method for predicting the remaining life of rolling bearings according to claim 2, characterized in that, The network structure of the momentum encoder is the same as that of the feature extractor. The parameters of the momentum encoder are updated according to the parameters of the feature extractor and based on the momentum update principle.

6. A rolling bearing remaining life prediction system, characterized in that, include: Input interface, output interface, processor, computer-readable storage medium, and stored program instructions, wherein: Input interface: Used to obtain training sample data or sample data to be detected; The stored program instructions include: Program instructions programmed to acquire training sample data or sample data to be tested; The program instructions are programmed to train a remaining lifetime prediction neural network to obtain a trained remaining lifetime prediction neural network, which includes a feature extractor and a regression predictor. The program instructions are programmed to make predictions using a trained residual lifetime prediction neural network and output the prediction results. Processor: Retrieves program instructions to train the remaining lifetime prediction neural network, resulting in a trained remaining lifetime prediction neural network; The program instructions of the trained residual prediction neural network are retrieved, and the feature extractor is instructed to extract features from the rolling bearing vibration data to be detected. The extracted features are then input into the regression predictor for prediction processing to obtain the prediction result. Output interface: Used to output the predicted remaining lifetime value.

7. The rolling bearing remaining life prediction system according to claim 6, characterized in that, The stored program instructions also include: The program instructions are programmed into a feature encoder to extract source domain features and target domain features; The program instructions are programmed into a temporal hybrid contrastive domain adaptation training module for calculating the contrastive loss; The program instructions are programmed into a fine-grained structural domain adaptation training module, which is used to calculate the fine-grained matching degree between source domain features and target domain features, as well as the domain classification loss; The program instructions are programmed as a regression predictor for regression prediction and to calculate the regression prediction loss.

8. The rolling bearing remaining life prediction system according to claim 7, characterized in that, The processor fetches program instructions to train the remaining lifetime prediction neural network. The training process includes: Initialize system parameters and set the number of training iterations; Obtain training sample data through the input interface; The processor retrieves program instructions from the temporal mixed contrast domain adaptation training module, instructs the temporal mixed contrast domain adaptation module to calculate the contrast loss, and uses the contrast loss to iteratively train the feature encoder; The processor calls the program instructions of the fine-grained structural domain adaptation training module, instructing the fine-grained structural domain adaptation module to calculate the fine-grained matching degree and domain classification loss between the source domain sample data and the target domain sample data respectively, and iteratively trains the feature encoder using the fine-grained matching degree and domain classification loss respectively; The processor feature encoder's program instructions extract source domain features, call the regression predictor's program instructions, input the source domain features into the regression predictor for prediction, calculate the regression prediction loss, and iteratively train the regression predictor using the regression prediction loss. The iteration stops when the preset number of iterations is reached, resulting in a trained remaining lifetime prediction neural network.

9. The rolling bearing remaining life prediction system according to claim 6, characterized in that, Also includes: Vibration sensors for collecting rolling bearing vibration data and program instructions programmed to preprocess the collected rolling bearing vibration data. The preprocessing includes: dividing the collected rolling bearing vibration data into a source domain sample set with full life cycle data and a target domain sample set with incomplete life cycle data; and using the ratio of the remaining service life of the rolling bearing to the total service life as the label of the source domain sample.

Images