Remaining life prediction method and device

Abstract

The present invention provides a method and device for predicting remaining life, which belongs to the technical field of equipment life prediction. The method comprises: obtaining source domain samples and target domain samples, and extracting features from the source domain samples and target domain samples based on a feature extraction module to obtain source domain degradation features and target domain degradation features; pre-training an initial prediction model based on the source domain degradation features to obtain a transition prediction model, and training the transition prediction model based on the target fusion degradation features to obtain a target prediction model; predicting the remaining life of the equipment based on the target prediction model; wherein the feature extraction module comprises a parallel local decay feature extraction submodule, a degradation trend extraction submodule, and a feature fusion submodule. The present invention improves the accuracy of the remaining life prediction of mechanical equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of equipment life prediction, and specifically to a remaining life prediction method and device. Background Art

[0002] The remaining life prediction of mechanical equipment is of great significance to ensure the safe and reliable operation of the equipment. Due to the complexity and variability of the working environment of mechanical equipment, it is in different working environments with large differences for a long time, resulting in obvious data distribution differences in the collected multi-condition vibration monitoring signals. At present, most of the research on remaining life prediction methods based on deep learning mainly uses supervised learning to train the model, and assumes that the training samples and test samples of the remaining life prediction model meet the characteristics of independent and identical distribution of data, making this type of research only applicable to the remaining life prediction of a certain type of equipment under a certain specific working condition.

[0003] However, in actual complex engineering cases, significant changes in the working conditions of mechanical equipment will cause differences in the distribution of training data and test data, making it difficult to meet the idealized conditions of independent and identical distribution. Therefore, this limitation seriously restricts the application of existing research on remaining life prediction methods in engineering practice. Therefore, it is of great practical value to study cross-domain (cross-equipment, cross-working conditions) mechanical equipment remaining life prediction methods.

[0004] The most commonly used deep learning framework in the prior art is the convolutional neural network and its variants. However, the problems with this type of framework based on a simple feedforward neural network are that information can only be transmitted between adjacent layers and cannot be shared in the same layer, and time series information cannot be processed and memorized. It is also difficult to simultaneously capture local information, global information, and relevant knowledge of statistical characteristics in degradation data. The remaining life is a feature that is strongly related to time. Therefore, predicting the remaining life through the deep learning framework in the prior art will lead to technical problems such as inaccurate prediction results.

[0005] There is an urgent need to provide a remaining life prediction method and device to solve the above technical problems. Summary of the invention

[0006] In view of this, it is necessary to provide a remaining life prediction method and device to solve the technical problem of inaccurate remaining life prediction in the prior art.

[0007] On the one hand, in order to solve the above technical problems, the present invention provides a remaining life prediction method, comprising:

[0008] Acquire source domain samples and target domain samples, and perform feature extraction on the source domain samples and the target domain samples based on a feature extraction module to obtain source domain degradation features and target domain degradation features;

[0009] Pre-training an initial prediction model based on the source domain degradation feature to obtain a transition prediction model, and training the transition prediction model based on the target fusion degradation feature to obtain a target prediction model;

[0010] Predicting the remaining life of the equipment based on the target prediction model;

[0011] The feature extraction module includes a parallel local decay feature extraction submodule, a degradation trend extraction submodule and a feature fusion submodule.

[0012] In a possible implementation, the local decay feature extraction submodule includes a decay feature fuzzy positioning unit, a decay feature precise positioning unit, and a jump fusion unit; the feature extraction module is used to extract features from the source domain samples to obtain source domain degradation features, including extracting features from the source domain samples based on the local decay feature extraction submodule to obtain local decay features; the local decay feature extraction submodule is used to extract features from the source domain samples to obtain local decay features, including:

[0013] Extracting features from the source domain samples based on the decay feature fuzzy positioning unit to obtain fuzzy decay features;

[0014] Extracting features from the fuzzy decay features based on the decay feature precise positioning unit to obtain precise decay features;

[0015] The fuzzy decay feature and the precise decay feature are jump-fused based on the jump-fusion unit to obtain the local decay feature.

[0016] In a possible implementation, the decay feature fuzzy positioning unit includes a multi-core convolution layer, a fuzzy average pooling layer, a fuzzy activation function layer, a first multiplication operation layer, and a second multiplication operation layer; the feature extraction of the source domain sample based on the decay feature fuzzy positioning unit to obtain the fuzzy decay feature includes:

[0017] Performing a multi-core convolution operation on the source domain sample based on the multi-core convolution layer to obtain convolution features;

[0018] Performing average pooling processing on the convolutional features based on the fuzzy average pooling layer to obtain pooling features;

[0019] Normalizing the pooled features based on the fuzzy activation function layer, and obtaining a probability distribution of the pooled features;

[0020] Based on the first multiplication operation layer, sequentially multiplying and element-wise adding the convolution feature and the probability distribution to obtain a fuzzy positioning weight;

[0021] Based on the second multiplication operation layer, the fuzzy positioning weight and the convolution feature are element-wise multiplied to obtain the fuzzy decay feature.

[0022] In a possible implementation, the decay feature precise positioning unit includes a first convolution block, a second convolution block and a third convolution block; the first convolution block includes a first convolution layer, a first batch quantization normalization layer and a first activation function layer, the second convolution block includes a second convolution layer, a second batch normalization layer and a second activation function layer, and the third convolution block includes a third convolution layer, a third batch normalization layer and a third activation function layer; the feature extraction of the fuzzy decay feature based on the decay feature precise positioning unit to obtain the precise decay feature includes:

[0023] Extracting the fuzzy decay feature based on the first convolution block to obtain a first decay feature;

[0024] Performing feature extraction on the first decay feature based on the second convolution block to obtain a second decay feature;

[0025] Feature extraction is performed on the second decay feature based on the third convolution block to obtain the precise decay feature.

[0026] In a possible implementation, the jump fusion unit includes a first jump feature fusion layer, a second jump feature fusion layer, and a jump average pooling layer, the first jump feature fusion layer is arranged between the first convolution block and the second convolution block, and the second jump feature fusion layer is arranged between the third batch normalization layer and the third activation function layer; the jump fusion of the fuzzy decay feature and the precise decay feature based on the jump fusion unit to obtain the local decay feature includes:

[0027] Based on the first jump feature fusion layer, the first decay feature and the output of the second batch normalization layer are fused to obtain a first fused feature;

[0028] Based on the second jump feature fusion layer, the output of the first batch normalization layer, the output of the second batch normalization layer, and the output of the third batch normalization layer are fused to obtain a second fused feature;

[0029] The third activation function layer performs activation processing on the second fusion feature to obtain an activation feature;

[0030] The jump average pooling layer performs average pooling processing on the activation features to obtain the local decay features.

[0031] In a possible implementation, the degradation trend extraction submodule includes a first dense block, a second dense block, a third dense block, a third multiplication operation layer, and a fourth dense block; the feature extraction module is based on the source domain sample to extract features to obtain source domain degradation features, including extracting features from the source domain sample based on the degradation trend extraction submodule to obtain degradation trend features; the feature extraction from the source domain sample based on the degradation trend extraction submodule to obtain degradation trend features includes:

[0032] Extracting features from the source domain samples based on the first dense block to obtain first dense features;

[0033] Performing feature extraction on the first dense feature based on the second dense block to obtain a second dense feature;

[0034] Performing feature extraction on the second dense feature based on the third dense block to obtain a time series feature weight;

[0035] Based on the third multiplication operation layer, a same-position element multiplication operation is performed on the time series feature weight and the source domain sample to obtain a global time series feature vector;

[0036] Feature extraction is performed on the global time series feature vector based on the fourth dense block to obtain the degradation trend feature.

[0037] In a possible implementation, the first dense block, the second dense block, the third dense block and the fourth dense block each include a dense layer, a dense batch normalization layer and a dense activation function layer.

[0038] In a possible implementation manner, the first loss function when pre-training the initial prediction model based on the source domain degradation feature is:

[0039]

[0040] In the formula, is the first loss function; y S i is the true remaining life percentage value of the i-th source domain sample; Predict the remaining lifetime percentage value for the i-th source domain sample; M is the total number of source domain samples.

[0041] In a possible implementation manner, the second loss function when training the transition prediction model based on the target fusion degradation feature is:

[0042]

[0043]

[0044]

[0045]

[0046] In the formula, J total is the second loss function; μ is the first equilibrium parameter; is the cross-domain transfer learning loss function; γ is the second equilibrium function; Adaptation transfer loss function for adversarial domain; JS(F fusion_source ,F fusion_target ) is the source domain degradation feature F fusion_source and the target domain degradation feature F fusion_target JS divergence; N is the total number of samples in the target domain; is the pre-classification output result of the lth target domain sample; k l is the true classification output result of the lth target domain sample.

[0047] On the other hand, the present invention also provides a remaining life prediction device, comprising:

[0048] A degradation feature extraction unit, used to obtain source domain samples and target domain samples, and perform feature extraction on the source domain samples and the target domain samples based on a feature extraction module to obtain source domain degradation features and target domain degradation features;

[0049] A target prediction model determination unit, configured to pre-train an initial prediction model based on the source domain degradation feature to obtain a transition prediction model, and to train the transition prediction model based on the target fused degradation feature to obtain a target prediction model;

[0050] A remaining life prediction unit, used to predict the remaining life of the equipment based on the target prediction model;

[0051] The feature extraction module includes a parallel local decay feature extraction submodule, a degradation trend extraction submodule and a feature fusion submodule.

[0052] The beneficial effects of the present invention are as follows: the remaining life prediction method provided by the present invention can accurately extract the local decay features of the failure occurrence area from the non-stationary and nonlinear source domain samples and target domain samples under a strong noise background through the local decay feature extraction submodule in the feature extraction module, and the degradation trend features of the source domain samples and the target domain samples can be obtained by setting the degradation trend submodule, that is, compared with the problem of insufficient ability to extract time series features in traditional convolutional neural networks, the prediction accuracy of the remaining life can be improved by setting the degradation trend submodule to extract the time series feature of the degradation trend feature. Furthermore, the present invention can achieve the fusion of multi-channel data, adaptively mine the global consistency of degradation data, extract statistical characteristics and energy characteristics that effectively reflect the equipment degradation mechanism, and further improve the prediction accuracy of the remaining life by setting the feature fusion submodule to fuse the local decay features extracted by the local decay feature extraction submodule and the degradation trend features extracted by the degradation trend submodule. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0054] Figure 1 A schematic diagram of a flow chart of an embodiment of the remaining life prediction method provided by the present invention;

[0055] Figure 2 A schematic diagram of the structure of an embodiment of a feature extraction module provided by the present invention;

[0056] Figure 3 A schematic diagram of a flow chart of an embodiment of obtaining local decay characteristics provided by the present invention;

[0057] Figure 4 A schematic diagram of the structure of an embodiment of the decay feature fuzzy positioning unit provided by the present invention;

[0058] Figure 5 For the present invention Figure 3 A schematic flow chart of an embodiment of S301;

[0059] Figure 6 A schematic diagram of the structure of an embodiment of a recession characteristic precise positioning unit provided by the present invention;

[0060] Figure 7 For the present invention Figure 3 A schematic flow chart of an embodiment of S302;

[0061] Figure 8 A schematic diagram of the structure of an embodiment of the degradation trend extraction submodule provided by the present invention;

[0062] Fig. 9 A schematic diagram of a flow chart of an embodiment of obtaining degradation trend characteristics provided by the present invention;

[0063] Fig.10 A schematic diagram of the structure of an embodiment of the remaining life prediction device provided by the present invention;

[0064] Fig.11 A schematic structural diagram of an embodiment of an electronic device provided by the present invention. DETAILED DESCRIPTION

[0065] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0066] It should be understood that the schematic drawings are not drawn to scale. The flowchart used in the present invention shows the operations implemented according to some embodiments of the present invention. It should be understood that the operations of the flowchart can be implemented out of order, and the steps without logical context can be reversed in order or implemented simultaneously. In addition, those skilled in the art, under the guidance of the content of the present invention, can add one or more other operations to the flowchart, and can also remove one or more operations from the flowchart. Some of the block diagrams shown in the accompanying drawings are functional entities, which do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor systems and / or microcontroller systems.

[0067] Reference to an "embodiment" herein means that a particular feature, structure, or characteristic described in conjunction with the embodiment may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0068] The present invention provides a remaining life prediction method and device, which are described below respectively.

[0069] Figure 1 A schematic diagram of a flow chart of an embodiment of the remaining life prediction method provided by the present invention. Figure 2 A schematic diagram of the structure of an embodiment of the feature extraction module provided by the present invention is shown in FIG. Figure 1 and Figure 2 As shown, the remaining life prediction method includes:

[0070] S101, obtaining source domain samples and target domain samples, and performing feature extraction on the source domain samples and the target domain samples based on a feature extraction module to obtain source domain degradation features and target domain degradation features;

[0071] S102, pre-training the initial prediction model based on the source domain degradation feature to obtain a transition prediction model, and training the transition prediction model based on the target fusion degradation feature to obtain a target prediction model;

[0072] S103, predicting the remaining life of the equipment based on the target prediction model;

[0073] The feature extraction module 10 includes a parallel local decay feature extraction submodule 100 , a degradation trend extraction submodule 200 , and a feature fusion submodule 300 .

[0074] It should be noted that the source domain samples are samples of mechanical equipment under the first working condition, and the target domain samples are samples of mechanical equipment under the second working condition. The initial prediction model is pre-trained based on the source domain samples, and then the pre-trained transition prediction model is further trained based on the target domain samples to obtain the target prediction model, which can accurately predict the remaining life of the mechanical equipment under the second working condition. Specifically, the first working condition is a light speed and light load working condition, and the second working condition is a heavy speed and heavy load working condition.

[0075] In a specific embodiment of the present invention, both the source domain samples and the target domain samples are vibration data. It can be understood that the source domain samples and the target domain samples can also be other data besides vibration data, such as strain data, temperature data, etc.

[0076] It can be understood that the source domain degradation feature is a feature representing the degradation degree of the source domain sample, and the target domain degradation feature is a feature representing the degradation degree of the target domain sample.

[0077] It should be further explained that the local decay feature extraction submodule 100 is specifically used to extract features from source domain samples and target domain samples to obtain local decay features, the degradation trend extraction submodule 200 is specifically used to extract features from source domain samples and target domain samples to obtain degradation trend features, and the feature fusion submodule 300 is specifically used to fuse local decay features and degradation trend features to obtain source domain degradation features and target domain degradation features.

[0078] Compared with the prior art, the remaining life prediction method provided by the embodiment of the present invention can accurately extract the local decay features of the failure occurrence area from the non-stationary and nonlinear source domain samples and target domain samples under the strong noise background through the local decay feature extraction submodule 100 in the feature extraction module, and can obtain the degradation trend features of the source domain samples and the target domain samples by setting the degradation trend submodule 200, that is, compared with the problem of insufficient extraction ability of time series features in the traditional convolutional neural network, by setting the degradation trend submodule to extract the time series feature of the degradation trend feature, the prediction accuracy of the remaining life can be improved. Furthermore, the embodiment of the present invention can achieve the fusion of multi-channel data, adaptively mine the global consistency of degradation data, extract statistical characteristics and energy characteristics that effectively reflect the equipment degradation mechanism, and further improve the prediction accuracy of the remaining life by setting the feature fusion submodule 300 to fuse the local decay features extracted by the local decay feature extraction submodule and the degradation trend features extracted by the degradation trend submodule.

[0079] In some embodiments of the present invention, the specific method of obtaining source domain samples and target domain samples in step S101 is: constructing a source domain sample set and a target domain sample set based on sensors, extracting any sample from the source domain sample set as the source domain sample, and extracting any sample from the target domain sample set as the target domain sample.

[0080] It should be noted that the feature extraction module 10 performs the same steps for extracting features from source domain samples and target domain samples, and the following description will be made by taking feature extraction from source domain samples as an example.

[0081] In some embodiments of the present invention, Figure 2 As shown, the local decay feature extraction submodule 100 includes a decay feature fuzzy positioning unit 110, a decay feature precise positioning unit 120 and a jump fusion unit 130; then the feature extraction module 10 is used to extract features from the source domain samples to obtain source domain degradation features in step S101, including extracting features from the source domain samples based on the local decay feature extraction submodule 100 to obtain local decay features; as shown Figure 3 As shown, based on the local decay feature extraction submodule 100, feature extraction is performed on the source domain sample to obtain the local decay feature, including:

[0082] S301, extracting features of source domain samples based on the decay feature fuzzy positioning unit 110 to obtain fuzzy decay features;

[0083] S302, extracting the fuzzy decay feature based on the decay feature precise positioning unit 120 to obtain a precise decay feature;

[0084] S303 , performing jump fusion on the fuzzy decay feature and the precise decay feature based on the jump fusion unit 130 to obtain a local decay feature.

[0085] The embodiments of the present invention take into account the characteristics of strong nonlinearity, strong noise interference and instability in the wear and degradation process of mechanical equipment, and the technical problem that it is difficult to achieve ideal results by directly predicting the vibration acceleration signal collected by the sensor. In order to extract local features of nonlinear degradation data under strong noise background, in the process of noise reduction and compression of the original vibration signal of the whole life cycle, the decay feature fuzzy positioning unit 110 is used to automatically extract spatial local features; the decay feature precise positioning unit 120 is used to adaptively and accurately locate important decay features; the jump fusion unit 130 is used to realize jump adaptive data fusion based on multi-level local features, so that the extracted local decay features can more accurately reflect the decay type characteristics of the equipment, thereby improving the accuracy of the remaining life prediction.

[0086] In some embodiments of the present invention, Figure 4 As shown, the decay feature fuzzy positioning unit 110 includes a multi-core convolution layer, a fuzzy average pooling layer, a fuzzy activation function layer, a first multiplication operation layer and a second multiplication operation layer. Figure 5 As shown, step S301 includes:

[0087] S501, performing a multi-core convolution operation on the source domain sample based on the multi-core convolution layer to obtain convolution features;

[0088] S502, performing average pooling processing on the convolution features based on the fuzzy average pooling layer to obtain pooling features;

[0089] S503, normalizing the pooled features based on the fuzzy activation function layer, and obtaining the probability distribution of the pooled features;

[0090] S504, performing multiplication and element-wise addition operations on the convolution features and the probability distribution in sequence based on the first multiplication operation layer to obtain a fuzzy positioning weight;

[0091] S505 , performing an element-wise multiplication operation on the fuzzy positioning weight and the convolution feature based on the second multiplication operation layer to obtain a fuzzy decay feature.

[0092] In a specific embodiment of the present invention, the multi-core convolution layer is 16 32×1 convolution kernels, and the convolution feature can be expressed as:

[0093] c i j =∑x i ·K j ,j=1,...,16

[0094] In the formula, c ij is the convolution feature; K j is the jth convolution kernel; x i is the sample vector of i source domain samples; is the inner product operator symbol of the vector.

[0095] In a specific embodiment of the present invention, the activation function of the fuzzy activation function layer is a softmax function, and the probability distribution p of the pooling feature i j for:

[0096]

[0097] In the formula, a i j is the pooled feature of the i-th source domain sample after the j-th convolution kernel.

[0098] Among them, the fuzzy positioning weight Ms is:

[0099]

[0100] Among them, the fuzzy decay feature Fs is:

[0101] Fs=x i ·Ms

[0102] In some embodiments of the present invention, Figure 6 As shown, the decay feature precise positioning unit 120 includes a first convolution block, a second convolution block and a third convolution block; the first convolution block includes a first convolution layer, a first batch quantization normalization layer and a first activation function layer, the second convolution block includes a second convolution layer, a second batch normalization layer and a second activation function layer, and the third convolution block includes a third convolution layer, a third batch normalization layer and a third activation function layer; then step S302 is to extract the fuzzy decay feature based on the first convolution block, the second convolution block and the third convolution block in sequence to obtain the precise decay feature, specifically:

[0103] Extracting the fuzzy decay feature based on the first convolution block to obtain a first decay feature;

[0104] Extracting the first decay feature based on the second convolution block to obtain a second decay feature;

[0105] The second decay feature is extracted based on the third convolution block to obtain an accurate decay feature.

[0106] The embodiment of the present invention can extract local decay features with strong representation in vibration data by setting the decay feature precise positioning unit 120, and introduce a batch normalization layer to reduce the dimension, which can not only extract local spatial features but also effectively avoid linear overfitting.

[0107] Furthermore, considering that the traditional ReLU activation function converts all non-negative inputs to zero, it is easy to cause transition smoothness and lead to information loss. In a specific embodiment of the present invention, the activation function of the first activation function layer and the second activation function layer is a Leaky ReLU activation function, and its expression is:

[0108]

[0109] Among them, γ∈(0,1) is the non-zero slope assigned by the Leaky ReLU activation function to all negative values, which is usually set around 0.01. This activation function helps to retain valuable information in the data.

[0110] Among them, the activation function of the third activation function layer is a softmax activation function, which obtains a more accurate positioning representation of the decay feature with a more spatial local representation, that is, a precise decay feature.

[0111] Among them, the number of convolution kernels of the first convolution layer, the second convolution layer, and the third convolution layer are 32, 32, and 64 respectively, and the sizes of one-dimensional convolution kernels are 32×1, 3×1, and 3×1 respectively.

[0112] In some embodiments of the present invention, Figure 6 As shown, the jump fusion unit 130 includes a first jump feature fusion layer, a second jump feature fusion layer and a jump average pooling layer, the first jump feature fusion layer is arranged between the first convolution block and the second convolution block, and the second jump feature fusion layer is arranged between the third batch normalization layer and the third activation function layer; Figure 7 As shown, step S303 includes:

[0113] S701, fusing the first decay feature and the output of the second batch normalization layer based on the first jump feature fusion layer to obtain a first fused feature;

[0114] S702, based on the second jump feature fusion layer, the output of the first batch of quantized normalized layers, the output of the second batch of normalized layers, and the output of the third batch of normalized layers are fused to obtain a second fused feature;

[0115] S703, the third activation function layer performs activation processing on the second fusion feature to obtain an activation feature;

[0116] S704, the jump average pooling layer performs average pooling processing on the activated features to obtain local decay features.

[0117] The embodiment of the present invention obtains local decay features by performing three-level jump fusion on fuzzy decay features and precise decay features, which can effectively avoid the gradient vanishing phenomenon caused by a large number of network layers, and further improve the accuracy of remaining life prediction.

[0118] In some embodiments of the present invention, Figure 8 As shown, the degradation trend extraction submodule 200 includes a first dense block, a second dense block, a third dense block, a third multiplication operation layer and a fourth dense block; then the feature extraction module in step S101 extracts features from the source domain samples to obtain source domain degradation features, including extracting features from the source domain samples based on the degradation trend extraction submodule to obtain degradation trend features; then Fig. 9 As shown, based on the degradation trend extraction submodule 200, feature extraction is performed on the source domain samples to obtain degradation trend features, including:

[0119] S901, extracting features from source domain samples based on the first dense block to obtain first dense features;

[0120] S902, extracting features from the first dense features based on the second dense blocks to obtain second dense features;

[0121] S903, extracting features from the second dense features based on the third dense block to obtain a time series feature weight;

[0122] S904, performing a same-position element multiplication operation on the time series feature weight and the source domain sample based on the third multiplication operation layer to obtain a global time series feature vector;

[0123] S905 . Perform feature extraction on the global time series feature vector based on the fourth dense block to obtain degradation trend features.

[0124] The embodiment of the present invention takes into account that the convolution operation mainly works in local areas and cannot identify the long-term interactivity and statistical characteristics of the input data. The first dense block, the second dense block, the third dense block and the fourth dense block are designed to obtain the global time series feature vector of the source domain data, thereby improving the prediction accuracy of the remaining life.

[0125] Furthermore, the embodiment of the present invention introduces a fourth dense block after the third dense block to extract deep time series information of the global time series feature vector, which not only solves the problem of long-term memory dependence that is common in time series signals, but also can better mine the deep features FGT that can represent the degradation trend from the time dimension, thereby reflecting the changes in the failure process and remaining life, and further improving the prediction accuracy of the remaining life.

[0126] In a specific embodiment of the present invention, the first dense block, the second dense block, the third dense block and the fourth dense block each include a dense layer, a dense batch normalization layer and a dense activation function layer.

[0127] Among them, the activation functions of the dense activation function layers of the first dense block and the second dense block are Leaky ReLU functions, the activation function of the dense activation function layer of the third dense block is the softmax function, and the activation function of the dense activation function layer of the fourth dense block is the Tanh function.

[0128] Specifically, the number of nodes in the first dense block and the second dense block is 32, the number of nodes in the third dense block is set according to the size of the input source domain sample, and the number of nodes in the fourth dense block is 32.

[0129] During pre-training, the learning rate of the feature extraction module is set to 10 -3 , the batch size is set to 200 and the epoch is set to 50.

[0130] In some embodiments of the present invention, the initial prediction model includes a plurality of fully connected layers and a linear regression output layer. The first loss function when pre-training the initial prediction model based on the source domain degradation feature in step S102 is:

[0131]

[0132] In the formula, is the first loss function; y S i is the true remaining life percentage value of the i-th source domain sample; Predict the remaining lifetime percentage value for the i-th source domain sample; M is the total number of source domain samples.

[0133] The remaining life percentage value is the percentage of the remaining life time to the total life time.

[0134] In some embodiments of the present invention, the second loss function when training the transition prediction model based on the target fusion degradation feature in step S102 is:

[0135]

[0136]

[0137]

[0138]

[0139] In the formula, J total is the second loss function; μ is the first equilibrium parameter; is the cross-domain transfer learning loss function; γ is the second equilibrium function; Adaptation transfer loss function for adversarial domain; JS(F fusion_source ,F fusion_target) is the source domain degradation feature F fusion_source and the target domain degradation feature F fusion_target JS divergence; N is the total number of samples in the target domain; is the pre-classification output result of the lth target domain sample; k l is the true classification output result of the lth target domain sample.

[0140] The embodiment of the present invention sets the second loss function to include the first loss function, the cross-domain transfer learning loss function and the adversarial domain adaptation transfer loss function. The cross-domain transfer learning loss function is used to measure the correlation between the distribution of source domain degradation features and target domain degradation features, thereby evaluating the transfer learning performance and making the source domain data distribution as close as possible to the target domain data distribution, which can further improve the prediction accuracy of the trained target prediction model for the target domain data. Furthermore, by adversarial and adaptive transfer loss functions, the mined features have good domain invariance and degradation degree discrimination at the same time, achieving unsupervised domain adaptability, thereby further improving the prediction accuracy of the target prediction model for the target domain data, and using the model pre-trained with labeled samples in the source domain to efficiently migrate to a large number of unlabeled samples in the target domain, reducing the sample demand.

[0141] Taking into account the asymmetry of KL divergence, the embodiment of the present invention sets the cross-domain transfer learning loss function as JS divergence, eliminating the problem of gradient vanishing when the probability distribution distance between the two data sets is large or even completely non-overlapping. JS divergence is used to solve the asymmetric problem of KL divergence, that is, when the support sets of the two data distributions are very small, JS divergence can still find a constant. Among them, the smaller the JS divergence, the more similar the distributions between the two domain features are. The calculation formula of JS correlation metric is as follows:

[0142]

[0143]

[0144] Where q(x) is the probability density function of the source domain degradation feature; p(x) is the probability density function of the target domain degradation feature; KL() is the KL divergence calculation symbol.

[0145] Among them, the specific construction process and principle of the adversarial domain adaptation transfer loss function are as follows:

[0146] First, the Kernel K-Means Clustering algorithm is used to cluster the source domain degradation features and the target domain degradation features to obtain cluster centers representing different degradation degrees in the two domains, which are expressed as: and

[0147] Then, the cosine similarity between the cluster center and the target domain feature is used to pre-classify the domain affiliation of the feature, that is, to determine whether the feature is from the source domain or the target domain. The output of the adversarial domain adaptation transfer module for a target domain feature is:

[0148]

[0149] In the above formula, δ(·) is the cosine similarity function. The model output 0 represents that the feature vector comes from the source domain, and the output 1 represents that the feature vector comes from the target domain.

[0150] Based on the above output, the adversarial domain adaptation transfer loss function can be constructed.

[0151] It should be noted that the stochastic gradient descent algorithm is used to optimize the above overall loss function until convergence, complete domain adversarial training, and make the feature space of the source domain as close to the feature space of the target domain as possible, so that the distribution correlation of the fusion features of the source domain and the target domain is as large as possible, that is, based on the correlation between the source domain and the target domain data, the decay features of the two domains are migrated to a new feature space, thereby realizing the cross-domain feature migration of the source domain data and the target domain data. The optimization process of the loss function is described as follows:

[0152]

[0153]

[0154] In the above formula, t represents the tth iteration training process, η is the learning rate of the gradient descent algorithm, and its value is 0.001. Through multiple rounds of iterations, the parameter set corresponding to the saddle point in the model training process is found and That is, the optimal parameter set enables the adaptive degradation trend analysis model to extract deep fusion features with the smallest possible distribution difference, and at the same time, enables the prediction accuracy of the remaining life prediction model trained based on labeled source domain data to be as high as possible. It is worth mentioning that the training of the remaining life prediction model in the embodiment of the present invention does not rely on a large number of labeled samples in the target domain, which is more in line with the actual situation in the engineering field.

[0155] It should be noted that the process of training the transition prediction model is an adversarial domain adaptive migration technology based on unsupervised clustering comparison. The degradation trend extraction submodule, the target prediction model and the adversarial domain adaptive migration loss function are optimized simultaneously to realize the construction of a target prediction model that integrates adaptive deep degradation trend analysis and clustering comparison adversarial domain adaptive migration, thereby improving the accuracy of the remaining life prediction of mechanical equipment.

[0156] In order to better implement the remaining life prediction method in the embodiment of the present invention, based on the remaining life prediction method, the embodiment of the present invention also provides a remaining life prediction device, such as Fig.10 As shown, the remaining life prediction device 1000 includes:

[0157] The degradation feature extraction unit 1001 is used to obtain source domain samples and target domain samples, and perform feature extraction on the source domain samples and the target domain samples based on the feature extraction module to obtain source domain degradation features and target domain degradation features;

[0158] A target prediction model determination unit 1002 is used to pre-train the initial prediction model based on the source domain degradation feature to obtain a transition prediction model, and to train the transition prediction model based on the target fusion degradation feature to obtain a target prediction model;

[0159] The remaining life prediction unit 1003 is used to predict the remaining life of the equipment based on the target prediction model;

[0160] The feature extraction module includes a parallel local decay feature extraction submodule, a degradation trend extraction submodule and a feature fusion submodule.

[0161] The remaining life prediction device 1000 provided in the above embodiment can implement the technical solution described in the above remaining life prediction method embodiment. The specific implementation principles of the above modules or units can refer to the corresponding contents in the above remaining life prediction method embodiment, which will not be repeated here.

[0162] like Fig.11 As shown, the present invention also provides an electronic device 1100. The electronic device 1100 includes a processor 1101, a memory 1102 and a display 1103. Fig.11 Only some components of the electronic device 1100 are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0163] In some embodiments, the processor 1101 may be a central processing unit (CPU), a microprocessor or other data processing chip, used to run program codes or process data stored in the memory 1102, such as the remaining life prediction method of the present invention.

[0164] In some embodiments, the processor 1101 may be a single server or a server group. The server group may be centralized or distributed. In some embodiments, the processor 1101 may be local or remote. In some embodiments, the processor 1101 may be implemented in a cloud platform. In one embodiment, the cloud platform may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-cloud, etc., or any combination thereof.

[0165] In some embodiments, the memory 1102 may be an internal storage unit of the electronic device 1100, such as a hard disk or memory of the electronic device 1100. In other embodiments, the memory 1102 may also be an external storage device of the electronic device 1100, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the electronic device 1100.

[0166] Furthermore, the memory 1102 may include both an internal storage unit of the electronic device 1100 and an external storage device. The memory 1102 is used to store application software installed in the electronic device 1100 and various data.

[0167] In some embodiments, the display 1103 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, an OLED (Organic Light-Emitting Diode) touch device, etc. The display 1103 is used to display information of the electronic device 1100 and to display a visual user interface. The components 1101-1103 of the electronic device 1100 communicate with each other via a system bus.

[0168] In one embodiment, when the processor 1101 executes the remaining life prediction program in the memory 1102, the following steps may be implemented:

[0169] Obtain source domain samples and target domain samples, and extract features from the source domain samples and the target domain samples based on a feature extraction module to obtain source domain degradation features and target domain degradation features;

[0170] Pre-training the initial prediction model based on the source domain degradation features to obtain a transition prediction model, and training the transition prediction model based on the target fusion degradation features to obtain a target prediction model;

[0171] Predict the remaining life of the equipment based on the target prediction model;

[0172] The feature extraction module includes a parallel local decay feature extraction submodule, a degradation trend extraction submodule and a feature fusion submodule.

[0173] It should be understood that: when the processor 1101 executes the remaining life prediction program in the memory 1102, in addition to the above functions, other functions can also be implemented. For details, please refer to the description of the corresponding method embodiment above.

[0174] Furthermore, the embodiment of the present invention does not specifically limit the type of the electronic device 1100 mentioned, and the electronic device 1100 may be a portable electronic device such as a mobile phone, a tablet computer, a personal digital assistant (PDA), a wearable device, a laptop computer, etc. Exemplary embodiments of portable electronic devices include but are not limited to portable electronic devices equipped with IOS, Android, Microsoft or other operating systems. The above-mentioned portable electronic devices may also be other portable electronic devices, such as a laptop computer with a touch-sensitive surface (e.g., a touch panel). It should also be understood that in some other embodiments of the present invention, the electronic device 1100 may not be a portable electronic device, but a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0175] Accordingly, an embodiment of the present application also provides a computer-readable storage medium, which is used to store computer-readable programs or instructions. When the program or instructions are executed by a processor, the steps or functions in the remaining life prediction method provided by the above-mentioned method embodiments can be implemented.

[0176] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware (such as a processor, a controller, etc.) through a computer program, and the computer program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0177] The above is a detailed introduction to the remaining life prediction method and device provided by the present invention. Specific examples are used in this article to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea; at the same time, for technical personnel in this field, according to the idea of ​​the present invention, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Claims

1. A method for predicting remaining life, characterized in that: include: Acquire source domain samples and target domain samples, and perform feature extraction on the source domain samples and the target domain samples based on a feature extraction module to obtain source domain degradation features and target domain degradation features; Pre-training an initial prediction model based on the source domain degradation feature to obtain a transition prediction model, and training the transition prediction model based on the target domain degradation feature to obtain a target prediction model; Predicting the remaining life of the equipment based on the target prediction model; Wherein, the feature extraction module includes a parallel local decay feature extraction submodule, a degradation trend extraction submodule and a feature fusion submodule; The local decay feature extraction submodule includes a decay feature fuzzy positioning unit, a decay feature precise positioning unit and a jump fusion unit; the feature extraction module is used to extract features from the source domain samples to obtain source domain degradation features, including extracting features from the source domain samples based on the local decay feature extraction submodule to obtain local decay features; the local decay feature extraction submodule is used to extract features from the source domain samples to obtain local decay features, including: Extracting features from the source domain samples based on the decay feature fuzzy positioning unit to obtain fuzzy decay features; Extracting features from the fuzzy decay features based on the decay feature precise positioning unit to obtain precise decay features; Performing jump fusion on the fuzzy decay feature and the precise decay feature based on the jump fusion unit to obtain the local decay feature; The decay feature fuzzy positioning unit includes a multi-core convolution layer, a fuzzy average pooling layer, a fuzzy activation function layer, a first multiplication operation layer and a second multiplication operation layer; the feature extraction of the source domain sample based on the decay feature fuzzy positioning unit to obtain the fuzzy decay feature includes: Performing a multi-core convolution operation on the source domain sample based on the multi-core convolution layer to obtain convolution features; Performing average pooling processing on the convolutional features based on the fuzzy average pooling layer to obtain pooling features; Normalizing the pooled features based on the fuzzy activation function layer, and obtaining a probability distribution of the pooled features; Based on the first multiplication operation layer, sequentially multiplying and element-wise adding the convolution feature and the probability distribution to obtain a fuzzy positioning weight; Performing an element-wise multiplication operation on the fuzzy positioning weight and the convolution feature based on the second multiplication operation layer to obtain the fuzzy decay feature; The decay feature precise positioning unit includes a first convolution block, a second convolution block and a third convolution block; the first convolution block includes a first convolution layer, a first batch quantization normalization layer and a first activation function layer, the second convolution block includes a second convolution layer, a second batch normalization layer and a second activation function layer, and the third convolution block includes a third convolution layer, a third batch normalization layer and a third activation function layer; the feature extraction of the fuzzy decay feature based on the decay feature precise positioning unit to obtain the precise decay feature includes: Extracting the fuzzy decay feature based on the first convolution block to obtain a first decay feature; Performing feature extraction on the first decay feature based on the second convolution block to obtain a second decay feature; Performing feature extraction on the second decay feature based on the third convolution block to obtain the precise decay feature; The jump fusion unit includes a first jump feature fusion layer, a second jump feature fusion layer and a jump average pooling layer, the first jump feature fusion layer is arranged between the first convolution block and the second convolution block, and the second jump feature fusion layer is arranged between the third batch normalization layer and the third activation function layer; the jump fusion of the fuzzy decay feature and the precise decay feature based on the jump fusion unit to obtain the local decay feature includes: Based on the first jump feature fusion layer, the first decay feature and the output of the second batch normalization layer are fused to obtain a first fused feature; Based on the second jump feature fusion layer, the output of the first batch normalization layer, the output of the second batch normalization layer, and the output of the third batch normalization layer are fused to obtain a second fused feature; The third activation function layer performs activation processing on the second fusion feature to obtain an activation feature; The jump average pooling layer performs average pooling processing on the activation feature to obtain the local decay feature; The degradation trend extraction submodule includes a first dense block, a second dense block, a third dense block, a third multiplication operation layer and a fourth dense block; the feature extraction module is used to extract features from the source domain samples to obtain source domain degradation features, including extracting features from the source domain samples based on the degradation trend extraction submodule to obtain degradation trend features; the feature extraction from the source domain samples based on the degradation trend extraction submodule to obtain degradation trend features includes: Extracting features from the source domain samples based on the first dense block to obtain first dense features; Performing feature extraction on the first dense feature based on the second dense block to obtain a second dense feature; Performing feature extraction on the second dense feature based on the third dense block to obtain a time series feature weight; Based on the third multiplication operation layer, a same-position element multiplication operation is performed on the time series feature weight and the source domain sample to obtain a global time series feature vector; Performing feature extraction on the global time series feature vector based on the fourth dense block to obtain the degradation trend feature; The first dense block, the second dense block, the third dense block and the fourth dense block each include a dense layer, a dense batch normalization layer and a dense activation function layer.

2. The remaining life prediction method according to claim 1, characterized in that: The first loss function when pre-training the initial prediction model based on the source domain degradation feature is: In the formula, is the first loss function; For the i The true remaining life percentage value of source domain samples; For the i The remaining life percentage value predicted by source domain samples; is the total number of source domain samples.

3. The remaining life prediction method according to claim 2, characterized in that: The second loss function when training the transition prediction model based on the target fusion degradation feature is: In the formula, is the second loss function; is the first equalization parameter; Loss function for cross-domain transfer learning; is the second equilibrium function; Adapting the transfer loss function for adversarial domains; is the source domain degradation feature and the target domain degradation characteristics JS divergence; N is the total number of samples in the target domain; For the l The pre-classification output results of target domain samples; For the l The actual classification output results of target domain samples.

4. A remaining life prediction device, characterized in that: The method for predicting remaining life according to any one of claims 1 to 3, wherein the device comprises: A degradation feature extraction unit, used to obtain source domain samples and target domain samples, and perform feature extraction on the source domain samples and the target domain samples based on a feature extraction module to obtain source domain degradation features and target domain degradation features; a target prediction model determination unit, configured to pre-train an initial prediction model based on the source domain degradation feature to obtain a transition prediction model, and to train the transition prediction model based on the target domain degradation feature to obtain a target prediction model; A remaining life prediction unit, used to predict the remaining life of the equipment based on the target prediction model; The feature extraction module includes a parallel local decay feature extraction submodule, a degradation trend extraction submodule and a feature fusion submodule.

Images