Method and device for predicting residual service life of equipment

Abstract

The invention discloses a method and device for predicting the remaining service life of equipment, and belongs to the technical field of artificial intelligence. The method comprises the following steps: training an initial model based on a Transform model based on a control condition vector of sample equipment and time sequence data of operation parameters of the sample equipment to obtain a baseline model; based on the time sequence data of the historical operation parameters of the target equipment and the control condition vector of the target equipment, performing unsupervised transfer learning on the baseline model to obtain a life prediction model; and inputting the time sequence data of the operation parameters of the target equipment and the control condition vector of the target equipment into the life prediction model to obtain a prediction result of the residual service life of the target equipment. According to the method and the device for predicting the residual service life of the equipment provided by the invention, the accuracy of predicting the residual service life of the equipment can be improved by utilizing the Transform model with the control condition vector and the unsupervised transfer learning.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of artificial intelligence, and particularly relates to a device remaining useful life prediction method and device. BACKGROUND

[0002] At present, the remaining useful life (RUL) prediction technology of industrial equipment mainly relies on historical operation data and supervised learning methods. The above methods are usually trained and verified on open source or laboratory condition data sets. However, the acquisition cost of the above data sets is high, and it is often difficult to obtain abnormal operation data of the equipment in the actual industrial environment. Moreover, the existing RUL prediction methods face the problem of class imbalance, that is, due to the serious consequences of equipment failure and the potential long degradation process before failure, failure events are extremely rare in production systems.

[0003] Therefore, the existing device remaining useful life prediction method has the defect of low accuracy. SUMMARY

[0004] The present application aims to at least solve one of the technical problems existing in the prior art. To this end, the present application provides a device remaining useful life prediction method and device, which can improve the accuracy of the device remaining useful life prediction method.

[0005] In a first aspect, the present application provides a device remaining useful life prediction method, which comprises:

[0006] training an initial model based on a Transformer model based on a control condition vector of a sample device and time series data of an operating parameter of the sample device, to obtain a baseline model;

[0007] performing unsupervised transfer learning on the baseline model based on time series data of historical operating parameters of a target device and a control condition vector of the target device, to obtain a life prediction model;

[0008] inputting time series data of an operating parameter of the target device and a control condition vector of the target device into the life prediction model, to obtain a prediction result of the remaining useful life of the target device.

[0009] According to the equipment remaining useful life prediction method of the application, the characteristics of the equipment are effectively extracted from complex industrial data by using the Transformer model with the control condition vector, the accuracy and real-time performance of the equipment operation state monitoring can be ensured, good adaptability and stability are shown under different equipment and working conditions, the efficiency and accuracy of feature extraction can be improved, the performance of the model in the unsupervised remaining useful life prediction of the target equipment is improved through effective transfer learning under unsupervised conditions, and the accuracy of the equipment remaining useful life prediction is improved. Further, by accurately predicting the remaining life of the equipment and timely discovering faults, the maintenance cost of the equipment can be effectively reduced, the service life of the equipment can be prolonged, and the operation reliability of the industrial equipment can be improved. Further, a large amount of abnormal data collection is not required, and the normal operation parameters are directly used for prediction, which is very beneficial to industrial manufacturers, and can solve the problem that abnormal data is usually difficult to obtain and the collection cost is high, and only laboratories have funds to provide.

[0010] According to an embodiment of the application, the time series data of the operation parameters of the sample equipment is obtained from cross-domain multi-data sets.

[0011] According to an embodiment of the application, the unsupervised transfer learning of the baseline model based on the time series data of the historical operation parameters of the target equipment and the control condition vector of the target equipment is performed to obtain a life prediction model, including:

[0012] The unsupervised transfer learning of the baseline model based on the time series data of the historical operation parameters of the target equipment and the control condition vector of the target equipment is performed in combination with a domain adversarial network and a distance preserving regularization loss to obtain the life prediction model.

[0013] According to an embodiment of the application, the unsupervised transfer learning of the baseline model based on the time series data of the historical operation parameters of the target equipment and the control condition vector of the target equipment is performed in combination with a domain adversarial network and a distance preserving regularization loss to obtain the life prediction model, including:

[0014] The unsupervised transfer learning of the baseline model based on the time series data of the historical operation parameters of the target equipment and the control condition vector of the target equipment is performed in combination with a domain adversarial network and a distance preserving regularization loss to obtain a feature extraction layer in the life prediction model.

[0015] Based on multiple optimizers, the extreme learning machine in the life prediction model is trained according to the output of the feature extraction layer; the extreme learning machine is used to predict the remaining useful life of the equipment according to the output of the feature extraction layer.

[0016] According to an embodiment of the present application, the method for predicting the remaining service life of the equipment comprises the following steps:

[0017] inputting the time series data of the operating parameters of the target equipment and the control condition vector of the target equipment into the life prediction model to obtain a first result;

[0018] adjusting the first result based on the equipment health index difference coefficient to obtain the prediction result of the remaining service life of the target equipment; the equipment health index difference coefficient of the sample equipment is obtained by dividing the difference between the predicted health index of the sample equipment in the normal operating state and the predicted health index of the sample equipment in the case of zero expected life by the standard deviation of the predicted health index of the sample equipment in the normal operating state.

[0019] According to an embodiment of the present application, the method for predicting the remaining service life of the equipment further comprises:

[0020] in the case of failure of the target equipment, updating the predicted health index of the sample equipment in the case of zero expected life to the predicted health index of the target equipment in the case of failure to update the equipment health index difference coefficient.

[0021] In a second aspect, the present application provides a device for predicting the remaining service life of the equipment, which comprises:

[0022] a baseline training module for training an initial model based on the control condition vector of the sample equipment and the time series data of the operating parameters of the sample equipment to obtain a baseline model;

[0023] a transfer learning module for performing unsupervised transfer learning on the baseline model based on the time series data of the historical operating parameters of the target equipment and the control condition vector of the target equipment to obtain a life prediction model;

[0024] a life prediction module for inputting the time series data of the operating parameters of the target equipment and the control condition vector of the target equipment into the life prediction model to obtain a prediction result of the remaining service life of the target equipment.

[0025] According to the device remaining service life prediction device provided in the application, the characteristics of the device are effectively extracted from complex industrial data by using the Transformer model with a control condition vector, the accuracy and real-time performance of the device operation state monitoring can be ensured, good adaptability and stability are exhibited under different devices and working conditions, the efficiency and accuracy of feature extraction can be improved, the performance of the model in the unsupervised remaining service life prediction of the target device is improved through effective transfer learning under the unsupervised condition, and thus the accuracy of the device remaining service life prediction can be improved. Further, by accurately predicting the remaining life of the device and timely discovering faults, the maintenance cost of the device can be effectively reduced, the service life of the device can be prolonged, and the operation reliability of the industrial device can be improved. Further, a large amount of abnormal data collection is not required, and the normal operation parameters are directly used for prediction, which is very beneficial to industrial manufacturers, and can solve the problem that abnormal data is usually difficult to obtain and the collection cost is high, and only laboratories usually have funds to provide.

[0026] In a third aspect, the application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the device remaining service life prediction method according to the first aspect.

[0027] In a fourth aspect, the application provides a non-transitory computer-readable storage medium having a computer program stored thereon, and the computer program is executable by a processor to implement the device remaining service life prediction method according to the first aspect.

[0028] In a fifth aspect, the application provides a chip, including a processor and a communication interface, the communication interface and the processor are coupled, and the processor is used to run a program or an instruction to implement the device remaining service life prediction method according to the first aspect.

[0029] In a sixth aspect, the application provides a computer program product, including a computer program, and the computer program is executable by a processor to implement the device remaining service life prediction method according to the first aspect.

[0030] Additional aspects and advantages of the application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0031] The above and / or additional aspects and advantages of the application will become apparent and be readily appreciated from the following description, including the appended drawings, wherein:

[0032] Figure 1 is a flowchart of the device remaining service life prediction method provided by the embodiments of the application.

[0033] Figure 2 FIG. 1 is a structural schematic diagram of a baseline model in a device remaining useful life prediction method provided by an embodiment of the present application;

[0034] Figure 3 FIG. 2 is a schematic diagram of a source domain search process in the device remaining useful life prediction method provided by an embodiment of the present application;

[0035] Figure 4 FIG. 3 is a structural schematic diagram of a life prediction model of the device remaining useful life prediction method provided by an embodiment of the present application;

[0036] Figure 5 FIG. 4 is a schematic diagram of a device remaining useful life curve in the device remaining useful life prediction method provided by an embodiment of the present application;

[0037] Figure 6 FIG. 5 is a structural schematic diagram of a device remaining useful life prediction apparatus provided by an embodiment of the present application;

[0038] Figure 7 FIG. 6 is a structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0039] The technical solutions in the embodiments of the present application will be clearly described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all of them. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art belong to the scope of protection of the present application.

[0040] The terms "first", "second", and the like in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than that illustrated or described herein, and the objects distinguished by "first", "second", etc. are usually a category and do not limit the number of objects, for example, the first object can be one or more. In addition, "and / or" in the specification and claims means at least one of the connected objects, and the character " / " generally represents an "or" relationship between the front and rear associated objects.

[0041] In the related art, RUL prediction relies on data sets from specific sources, which are usually obtained by large enterprise laboratories in professional condition laboratories, and the cost is high. It is difficult for general enterprises to obtain training data, and the data acquisition is difficult and costly.

[0042] Due to the scarcity of failure events and the existence of class imbalance, traditional supervised learning methods are ineffective when dealing with unbalanced datasets. In recent years, generative adversarial networks (GANs) and transfer learning techniques have been used to address this problem, but their use has been limited to laboratory validation and lacks practical application in real-world industrial applications.

[0043] Due to their reliance on large amounts of labeled data and specific data distribution, traditional RUL prediction methods have limited applicability and flexibility in the context of new types of equipment or changing industrial environments.

[0044] Below, in conjunction with the accompanying drawings, the equipment remaining service life prediction method, equipment remaining service life prediction device, electronic device and readable storage medium provided in the embodiments of the present application are described in detail through specific embodiments and their application scenarios.

[0045] The method for predicting the remaining useful life of a device may be applied to a terminal, and may be specifically executed by hardware or software in the terminal.

[0046] The terminal includes, but is not limited to, a portable communication device such as a mobile phone or tablet computer having a touch-sensitive surface (e.g., a touch screen display and / or a touch pad). It should also be understood that, in some embodiments, the terminal may not be a portable communication device, but a desktop computer having a touch-sensitive surface (e.g., a touch screen display and / or a touch pad).

[0047] In the following embodiments, a terminal including a display and a touch-sensitive surface is described. However, it should be understood that the terminal may include one or more other physical user interface devices such as a physical keyboard, a mouse, and a joystick.

[0048] The embodiment of the present application provides a method for predicting the remaining useful life of a device. The execution subject of the method can be an electronic device or a functional module or functional entity in the electronic device that can implement the method for predicting the remaining useful life of the device. The electronic devices mentioned in the embodiment of the present application include but are not limited to mobile phones, tablet computers, computers, cameras and wearable devices, etc. The method for predicting the remaining useful life of a device provided in the embodiment of the present application is explained below using an electronic device as an example of the execution subject.

[0049] like Figure 1 As shown, the method for predicting the remaining useful life of equipment includes: step 110, step 120 and step 130.

[0050] Step 110 : Based on the control condition vector of the sample device and the time series data of the operating parameters of the sample device, an initial model based on the Transformer model is trained to obtain a baseline model.

[0051] In actual implementation, an initial model based on a Transformer model can be trained based on an open industrial dataset to construct a baseline model. The Transformer layer (composed of a Transformer network) in the baseline model can be used as a feature extractor for subsequent steps.

[0052] In the embodiments of the present application, a control condition vector is added to the input of the initial model based on the conventional Transformer model. The Transformer network with the control condition vector is used as the core architecture, and the PyTorch framework can be used for implementation for subsequent training and deployment.

[0053] As shown in Figure 2 , the baseline model can include an input layer, a Transformer layer, a concatenation layer, and an output layer.

[0054] The input layer is used to receive the input of the baseline model. The input of the baseline model can include a controllable condition vector of a sample device and a series of time series data of operating parameters of the sample device. n .

[0055] The control condition vector is an optimized encoding vector that can be updated during training. Different devices and different working conditions can correspond to different control condition vectors. The control condition vector is a vector used to represent the control condition of the sample device.

[0056] In some embodiments, the time series data of operating parameters can be sensor data or other serialized inputs. The K times are repeated to obtain richer feature representations.

[0057] The core part of the Transformer layer can be composed of multiple Transformer blocks Figure 2 , where N is a positive integer, and ×N indicates N Transformer blocks. Each Transformer block can include a multi-head attention unit (for implementing a multi-head attention mechanism) and a feedforward network. The multi-head attention mechanism enables the baseline model to simultaneously focus on the operating parameters of other time points when processing the operating parameters of each time point, thereby extracting cross-time point correlation features. The feedforward network is used to further convert the information processed by the attention mechanism. Each Transformer block can also include an addition and normalization layer to improve the stability and efficiency of model training.

[0058] The concatenation layer is used to concatenate the outputs of all Transformer blocks to form a composite feature representation that integrates information of all time points.

[0059] The output layer converts the concatenated features into the final output format. The predicted output y' is the final prediction made by the baseline model based on the input data.

[0060] In some embodiments, the time series data of the operating parameters of the sample device can adopt an open source dataset, that is, an industrial remaining useful life dataset that has been publicly disclosed.

[0061] In some embodiments, training the initial model based on the Transformer model based on the control condition vector of the sample device and the time series data of the operating parameters of the sample device can include sub-steps such as data screening, data normalization, normalized remaining useful life label, and data division. The order of execution of the above sub-steps is not specifically limited in the embodiments of the present application.

[0062] In some embodiments, the data screening step can include screening the operating parameter data output by the sensor and eliminating data sequences that are always stable under the same working condition. For example, in the C-MAPSS turbofan engine simulation data, one-third of the data sequences are always constant, which has nothing to do with the expected life and can be discarded.

[0063] In some embodiments, the data normalization step can include using the mean normalization method of PyTorch to uniformly normalize the operating parameter data of each dimension collected by the sensor under different working conditions to a data range centered on zero. The formula is: where σ is the standard deviation of the operating parameter data of each dimension; μ is the mean of the operating parameter data of each dimension.

[0064] In some embodiments, the normalized remaining useful life label step can include setting the normalized remaining useful life label at each time point. The remaining expected life in industrial applications is often represented as the unavailability of the device caused by the deviation of a certain target parameter from the normal value. For example, in the operating parameter data of the PCoE lithium battery, the periodic discharge capacity falling to a certain value represents the unavailability of the lithium battery. The remaining expected life within the normal value is normalized to 1, and the expected life in the degradation period is linearly normalized from 1 to 0, setting the normalized label for the training data.

[0065] In some embodiments, the data division step can include dividing the operating parameter data of different devices and different working conditions, and using independent optimizable control condition vectors ξ for different division data, and encoding the corresponding optimizable control condition vectors after self-labeling the id.

[0066] In some embodiments, in the training of the initial model based on the Transformer model, the mean square error (MSE) can be used to calculate the back gradient of the model as the evaluation function of the training loss. The calculation formula of the mean square error is as follows:

[0067]

[0068] Based on the above evaluation function, an optimization algorithm such as the ADAM optimizer algorithm can be used to update the parameters of the initial model, and the learning rate can be set according to the actual setting (for example, 10 -4 and the like).

[0069] On this basis, the cosine annealing learning rate scheduling can be used to periodically reduce the learning rate, so as to further improve the accuracy of the model.

[0070] In some embodiments, in the training of the initial model based on the Transformer model, the root mean square error (RMSE) and the relative error evaluation function can be used to evaluate the accuracy of the model prediction. The specific calculation formula is as follows:

[0071]

[0072] Wherein, The superscript of indicates the predicted value; the subscript gt indicates the true value; n indicates the number of test set samples; y indicates the predicted normalized life expectancy value; i indicates the period length of the threshold value σ low and the threshold value σ high , which is used to evaluate the window length prediction accuracy of the equipment degradation process.

[0073] In some embodiments, during the training process, when the evaluation function reaches a predetermined threshold, the network parameters of the baseline model, the IDs of the equipment and working condition labels, and the corresponding control condition vectors are serialized and stored, and the model training is serialized. These serialized data will be used for subsequent step 120 of transfer learning.

[0074] In some embodiments, the serialization process includes serialization of each network parameter in the baseline model, storage of the IDs of the equipment and working condition labels, and storage of the control condition vectors.

[0075] The serialization of the network parameters in the baseline model includes saving all network parameters of the Transformer model, including weights and biases, so that the model can be reproduced in subsequent steps. The storage of the IDs of the device and working condition labels includes recording the labeled IDs of each sample device and working condition to ensure that different data sets can be correctly identified and used during transfer learning. The storage of the control condition vector includes saving the controllable control condition vector corresponding to each device and working condition, so that these vectors can be correctly applied during transfer learning. The serialized file can be stored and used in step 120 for transfer learning of the baseline model to improve the adaptability and prediction accuracy of the model under different devices and working conditions.

[0076] Step 120, based on the time series data of the historical operating parameters of the target device and the control condition vector of the target device, unsupervised transfer learning is performed on the baseline model to obtain a life prediction model.

[0077] In actual implementation, after the baseline model is trained in step 110, the main structure (Transformer layer) thereof can be used as an extractor of device parameter features, and further optimized through a transfer learning method.

[0078] In some embodiments, after reading the baseline model and the serialized model parameters, the historical operating parameters of the normal operation of the target device can be used as input to obtain the main structure for predicting the remaining useful life of the target device.

[0079] The purpose of step 120 is to make the life prediction model applicable to RUL prediction with only unlabeled data in the case where fault data is difficult to collect or label in general industry, and only normal operating parameter data is collected, so as to improve the accuracy and practicality of the prediction of the remaining useful life of the device in the actual industrial environment. Even in the case of lack of fault data, the life prediction model can accurately predict the remaining useful life of the device through the features learned from the normal operating data.

[0080] For example, in the scenario of predicting the remaining useful life of a certain type of semiconductor wafer cutting machine, only time series parameter data such as cutting depth, cutting force and feed speed of the device in normal working state can be collected, and data of the device in failure state is lacking. Therefore, the traditional RUL prediction method cannot effectively train in this scenario. By using the transfer learning process in step 120, the data distribution of the milling machine processing parameters and the wear data set in step 110 can be automatically matched. With the similarity and consistency between the two data sets, the life prediction model can apply the features learned from the milling machine processing parameters and the wear data set to the semiconductor wafer cutting machine through transfer learning, thereby improving the accuracy of predicting the remaining useful life. In this way, even in the absence of failure data, the life prediction model can accurately predict the remaining useful life of the equipment based on the features extracted from the normal operation data, thereby improving the effectiveness and reliability of equipment maintenance and management in actual industrial environments.

[0081] Step 130, input the time series data of the operating parameters of the target device and the control condition vector of the target device into the life prediction model, and obtain the prediction result of the remaining useful life of the target device.

[0082] In actual implementation, the life prediction model obtained by transfer training in step 120 can be deployed to the actual production line to predict the remaining useful life of the target device on the actual production line.

[0083] The time series data of the operating parameters collected during the actual operation of the target device can be collected, and the control condition vector of the target device can be combined to predict the remaining useful life of the target device, and obtain the prediction result of the remaining useful life of the target device.

[0084] In some embodiments, the operating parameters can include, but are not limited to, various types of industrial parameter data such as voltage, current, temperature, vibration signal, strain, rotational speed, pressure, flow rate, and spindle speed.

[0085] In some embodiments, the time series data of the operating parameters of the target device will undergo the same preprocessing operations as the feature selection and data normalization in the processing of the open source data set in step 110, and then input into the network as the life prediction model.

[0086] According to the device remaining useful life prediction method provided in the embodiments of the present application, the features of the device are effectively extracted from complex industrial data by using the Transformer model with a control condition vector, the accuracy and real-time performance of the device operation state monitoring can be ensured, good adaptability and stability are exhibited under different devices and working conditions, the efficiency and accuracy of feature extraction can be improved, the performance of the model in unsupervised remaining useful life prediction of the target device is improved through effective transfer learning under unsupervised conditions, and thus the accuracy of device remaining useful life prediction can be improved. Further, by accurately predicting the remaining life of the device and timely discovering faults, the maintenance cost of the device can be effectively reduced, the service life of the device can be prolonged, and the operation reliability of the industrial device can be improved. Further, a large amount of abnormal data collection is not required, and the normal operation parameters are directly used for prediction, which is very beneficial to industrial manufacturers, and can solve the problem that abnormal data is usually difficult to obtain and the collection cost is high, and only laboratories usually have funds to provide.

[0087] In some embodiments, the time series data of the operation parameters of the sample device is obtained from the cross-domain multi-data set.

[0088] In actual implementation, cross-domain multi-data set joint training can be performed on multiple groups of open source data sets, that is, the time series data of the operation parameters of the sample device can be obtained from the cross-domain multi-data set.

[0089] In some embodiments, on the basis of the above training set division, the initial model can be trained simultaneously by using each data set divided. In the process of training the initial model by using different data sets, the sharing of the internal parameters of the Transformer layer can be maintained, different data sets can be retrieved by using the independent ID and trained by using the respective optimizable control condition vector.

[0090] In some embodiments, the cross-domain multi-data set can include, but is not limited to, an open source CFRP panel tensile fatigue experiment data set, a milling machine processing parameter and wear data set, a bearing experiment data set, a turbofan engine degradation simulation data set, and an IGBT accelerated aging data set.

[0091] According to the device remaining useful life prediction method provided in the embodiments of the present application, the time series data of the operation parameters of the sample device can be obtained from the cross-domain multi-data set, the baseline model can extract the common data distribution structure in the device remaining useful life problem of different fields, and thus the baseline model can capture the general characteristics of the RUL problem by using these industrial RUL data sets. The Transformer is trained on the existing open source data set, which helps the model to obtain the general characteristics of the industrial RUL problem, thereby avoiding the problem that it is difficult and expensive for general enterprises to obtain training data.

[0092] In some embodiments, based on the time series data of the historical operating parameters of the target device and the control condition vector of the target device, unsupervised transfer learning is performed on the baseline model to obtain a life prediction model, including: based on the time series data of the historical operating parameters of the target device and the control condition vector of the target device, combining a domain adversarial network and a distance preserving regularization loss, performing unsupervised transfer learning on the baseline model to obtain a life prediction model.

[0093] In actual implementation, unsupervised transfer learning can be performed in combination with a domain adversarial network and a distance preserving regularization loss.

[0094] The purpose of adopting the domain adversarial network is to perform adversarial domain transfer. In some embodiments, in the transfer learning step, adversarial domain transfer method with a domain classifier can be used for transfer learning. Specifically, the adversarial domain transfer refers to introducing an adversarial training mechanism in the transfer learning, and distinguishing the source domain and the target domain through a gradient reversal layer and a domain classifier, so as to realize feature extraction and transfer learning.

[0095] In some embodiments, the domain feature vector can be transmitted to the domain classifier through a gradient reversal layer after being converted by the Transformer. The domain classifier outputs a positive value from 0 to 1, which is used to judge whether the sample is from the source domain (the aforementioned training sample space) or the target domain (the training sample space of the corresponding target device). The classification ability of the domain classifier and the optimizable control condition vector corresponding to the target device are trained using an adversarial loss function, so as to perform feature space registration. The loss function is composed of two parts of domain discriminator loss and feature extractor loss, and the mathematical expression is as follows:

[0096] L C (x)=L D (x)+λL F (x)

[0097]

[0098] Wherein, subscript s represents the source domain; subscript t represents the target domain; x represents the input sample data; P represents the domain to which the sample data belongs; F represents the encoder of the sample data, which corresponds to the Transformer layer trained in step 110; D represents the domain classifier trained in the adversarial domain transfer; L D represents the loss function for trying to classify the data from the source domain as the source domain and the data from the target domain as the target domain; L F represents the loss function for trying to deceive the domain discriminator so that it cannot distinguish the source domain and the target domain; λ is a pre-set hyperparameter.

[0099] The adversarial domain transfer method using the domain adversarial network can make the data distribution of the source domain and the target domain in the feature space as close as possible through feature space registration, thereby improving the performance of the life prediction model on the target domain.

[0100] The distance preserving regular loss is used to preserve the data distribution in the domain. In the transfer learning training process, the distance preserving regular loss term is also added to constrain the network parameters.

[0101] In some embodiments, the distance preserving regular loss term is a multidimensional scaling regular loss to ensure the consistency of the data distribution in the domain. The specific expression is as follows:

[0102]

[0103] where S is the data sample set of the target domain or the source domain; for the source domain, η = 1; and in the target domain, η is a hyperparameter participating in optimization.

[0104] The distance preserving regular loss can maintain the consistency of the data distribution structure in the domain after feature extraction of the target domain and the source domain with the original data. By using the distance preserving regular loss, the distance relationship of the samples in the feature space can be matched with the distance relationship in the input space as much as possible in the transfer learning process, thereby improving the generalization ability and prediction accuracy of the life prediction model on the target domain.

[0105] In some embodiments, a source domain search method is also used in the transfer learning process to match the closest source domain task for the target domain.

[0106] In actual implementation, the source domain samples are randomly sampled at the beginning of training. For this purpose, the domain classifier D needs to add an additional control condition vector ξ, that is, D(F(x t )ξ) is optimized for a certain number of rounds, and the source domain that minimizes the adversarial transfer loss L C (xξ) is calculated, and its samples are used as the source domain set to further optimize the adversarial domain transfer loss and the distance preserving regular loss.

[0107] For example, in the process of transfer learning of engine parameter samples, a large amount of random open source data set is used as source domain data to randomly sample the training field classifier. After 500 rounds of training, it is found that the C-MAPSS turbofan engine simulation data corresponds to the smallest adversarial transfer loss, so the C-MAPSS turbofan engine simulation data is selected as the source domain data, and the adversarial field transfer loss and distance preserving regularization loss are further optimized for transfer learning training. Through this source domain search method, the most matched source domain data can be effectively selected to ensure that the feature distribution difference between the source domain and the target domain is minimized during the transfer learning process, thereby improving the generalization ability and prediction accuracy of the model in the target domain.

[0108] The process of source domain search can be as shown in Figure 3 For example, for the same target domain, it is found that the adversarial transfer loss corresponding to the source domain 2 is the smallest among the adversarial transfer losses corresponding to the n source domains (source domain 1 to source domain n, n is a positive integer), which is only 0.08. Therefore, the source domain 2 is used as the source domain for subsequent optimization of the target device, i.e., as the matching domain. A typical case is that when predicting the remaining useful life of a certain type of semiconductor wafer cutting machine, the source domain search finds that the adversarial field transfer loss corresponding to the data distribution of the milling machine processing data is the smallest, so the milling machine processing data is used as the source domain for subsequent learning and optimization of the certain type of semiconductor wafer cutting machine.

[0109] During the training process, the cross-field multi-data set joint training and the source domain search method can extract common features from data in different fields, thereby improving the adaptability and generalization ability of the model in unknown environments.

[0110] In some embodiments, based on the time series data of the historical running parameters of the target device and the control condition vector of the target device, the baseline model is subjected to unsupervised transfer learning in combination with the field adversarial network and the distance preserving regularization loss to obtain a life prediction model, including: based on the time series data of the historical running parameters of the target device and the control condition vector of the target device, the baseline model is subjected to unsupervised transfer learning in combination with the field adversarial network and the distance preserving regularization loss to obtain a feature extraction layer in the life prediction model.

[0111] In actual implementation, the overall structure of the life prediction model can be as shown in Figure 4 During the transfer learning process, the Transformer layer serves as a device parameter feature extractor, and its main parameters remain frozen and do not participate in gradient optimization. Only the additional control condition vector ξ targetThe parameters are not fixed to participate in the optimization process. This strategy aims to utilize the remaining useful life dataset distribution characteristics that the Transformer layer has learned in advance and adapt to the specific needs of the target domain by optimizing the control condition vector to achieve accurate prediction of the remaining useful life of the target device. The contrastive domain transfer loss includes comparing the corresponding adversarial domain transfer losses of each source domain to determine the minimum value.

[0112] In some embodiments, optimizing the Transformer layer includes setting the control condition vector as an optimizable embedding and adding it to the neural network framework to optimize it using the unified gradient backpropagation mode and optimizer used by the aforementioned field classifier in adversarial domain transfer. Using the control condition vector can guide the encoder to generate an encoding mode as close as possible to the operating parameter data distribution of the target device, and avoid negatively affecting the learned domain characteristics, thereby maintaining the accuracy of the prediction. By optimizing the control condition vector, the Transformer layer can adapt to the specific needs of the target domain while utilizing the distribution characteristics of the remaining useful life dataset learned by the pre-trained model to achieve accurate prediction of the remaining useful life of the target device.

[0113] Based on the output of the feature extraction layer, the extreme learning machine in the life prediction model is trained using various optimizers; the extreme learning machine is used to predict the remaining useful life of the device based on the output of the feature extraction layer.

[0114] In actual implementation, after the transfer learning is completed, the feature vectors of the source domain and the target domain are mapped to a unified distribution space. After that, an extreme learning machine (ELM) can be trained using samples from the source domain and the target domain to predict the remaining useful life of the device.

[0115] The source domain samples have diverse life labels, while the life labels of the target domain samples can be unified to 1. During training, the pseudo-inverse matrix H + =(H T H) -1 H T is first calculated, where H is the hidden layer output matrix. Then, the difference between the pseudo-inverse matrix and the extreme learning machine parameters is input to the optimizer as the gradient of backpropagation.

[0116] In some embodiments, various optimizers such as SGD or ADAM can be used to train the extreme learning machine. By using the optimizer to train the extreme learning machine, the simple structure, good generalization, and efficient training of the extreme learning machine can be utilized to maintain the distribution structure of the feature vector as much as possible while learning the remaining useful life based on labeled data.

[0117] Through the above process, the labeled remaining useful life data samples of the source domain can be effectively utilized to migrate and learn the remaining useful life (i.e., the default normal operation state sample with a remaining useful life of 1) of the unlabeled data samples of the target domain. This strategy not only improves the prediction accuracy of the model on the target domain, but also fully utilizes the data characteristics of the source domain and the target domain, achieving efficient transfer learning.

[0118] In some embodiments, the time series data of the operating parameters of the target device and the control condition vector of the target device are input into the life prediction model to obtain a first result.

[0119] In actual implementation, the first result can be the output of the life prediction model.

[0120] The first result is adjusted based on a device health index difference coefficient of a sample device to obtain the prediction result of the remaining useful life of the target device. The device health index difference coefficient of the sample device is obtained by dividing the difference between the predicted health index of the sample device in the normal operating state and the predicted health index of the sample device in the case of an expected life of zero by the standard deviation of the predicted health index of the sample device in the normal operating state.

[0121] In actual implementation, the first result can also be adjusted according to further operating parameters collected during actual operation of the target device.

[0122] It should be noted that the remaining useful life predicted by the life prediction model is actually a normalized value, and the normalization range depends on the parameter range predefined by the user. Moreover, the relative life predicted by the life prediction model is defined based on the source domain data. Therefore, the purpose of runtime adjustment is to match the prediction result of the life prediction model with the operating state of the target device, so as to continuously optimize and improve the accuracy of the prediction.

[0123] In some embodiments, adaptive adjustment of the system threshold based on the distribution of the feature vector can be performed. A device health index difference coefficient can be defined, and the health index difference coefficient corresponding to the source device on the open source data set can be calculated. The standard deviation of the predicted health index of the target device during actual operation is calculated using the difference coefficient, so as to adaptively adjust the expected life range corresponding to the target device.

[0124] The health index difference coefficient is defined as the difference between the predicted health index of the device in the normal operating state and the predicted health index of the device when the expected life is zero, and the health index difference coefficient is divided by the standard deviation of the health index of the device in the normal operating state. The mathematical expression is: wherein, F HThe network represents the remaining service life of the target device in step 120, which can also be regarded as the health index of the target device; x normal The time series of the operating parameters of the target device in the normal operating state; x failure The time series of the operating parameters of the target device when the expected service life is zero; σ is the standard deviation symbol. The health index difference coefficient is used to measure the fluctuation of the health index of the device in the normal operating state and the difference value after the device fails.

[0125] The reference device is defined as the difference value of 1, so the health index difference coefficient of the device is the standard deviation of the health index in the normal operating state. In the embodiments of the present application, it is assumed that the health index difference coefficients of different devices are consistent, and the following expression is used to calculate the health index of the device in the adaptive adjustment

[0126]

[0127] Where D Hsource The device health index difference coefficient of the sample device, source represents the sample device; H predict The first result of the output; H adjusted The first result corrected based on the device health index difference coefficient of the sample device.

[0128] When predicting the normalized expected service life of the deployed device, the health index is normalized again for linear mapping, and the prediction result of the remaining service life of the target device is obtained:

[0129] In some embodiments, during the deployment of the target device, the operating parameters of the target device are continuously collected, and the records cached in a time window are stored in the local device. When the target device has obvious abnormal data or is manually updated with the fault time, the cached record sequence in the time window is serialized and saved in the target device. After a certain service time or a specified number of faults of the target device, the key fault sequence data of the continuous integration record is used to further fine-tune the life prediction model based on the Transformer. Through such continuous data integration and training, the prediction performance of the model can be continuously improved to adapt to changes in devices and working conditions, ensuring the reliability and practicality of the prediction results.

[0130] In some embodiments, when the device expected life is less than the remaining life threshold, a reminder can be pushed to the maintenance personnel, and a real-time expected life and health index monitoring interface can be provided. The real-time reminder and monitoring interface can help the maintenance personnel to understand the device status in a timely manner, take necessary maintenance measures, and avoid losses caused by device failure. Through the above steps, efficient and accurate device life prediction can be achieved in actual applications, and the reliability and practicality of the prediction results can be ensured through real-time adjustment and continuous improvement.

[0131] The specific value of the remaining life threshold is not limited in the embodiments of the present application. For example, the remaining life threshold can be 0.5 or 0.3, etc. A typical URL change curve can be as shown in Figure 5 .

[0132] According to the device remaining use life prediction method provided by the embodiments of the present application, the prediction result of the device remaining use life is adjusted in real time according to the operating parameters collected in the actual operation of the target device, and the threshold of the system is adjusted adaptively and the threshold of the failure time is set manually, so that the accuracy and reliability of the prediction result can be ensured.

[0133] In some embodiments, the device remaining use life prediction method further comprises: in the case that the target device fails, updating the predicted health index of the sample device in the case that the expected life is zero to the predicted health index of the target device in the case that the target device fails, to update the device health index difference coefficient.

[0134] In actual execution, the health index difference coefficient can also be updated based on the threshold adjustment mode of manually setting the failure node by the human. When the maintenance personnel reports the device failure, the failure time can be manually set. After receiving the failure time set by the maintenance personnel, the health index at this time is taken as a new threshold zero point, and in subsequent prediction of the normalized expected life of the device, the health index is normalized and linearly mapped again: wherein H failure represents the predicted health index of the target device in the case that the target device fails, H failure = 1-H adjusted .

[0135] According to the device remaining use life prediction method provided by the embodiments of the present application, the prediction result of the device remaining use life is adjusted in real time according to the operating parameters collected in the actual operation of the target device, and the threshold of the system is adjusted adaptively and the threshold of the failure time is set manually, so that the accuracy and reliability of the prediction result can be ensured.

[0136] The remaining service life prediction method for equipment provided in the embodiment of the present application can be executed by a device for predicting the remaining service life of equipment. In the embodiment of the present application, the remaining service life prediction method for equipment provided in the embodiment of the present application is described by taking the device for predicting the remaining service life of equipment as an example.

[0137] An embodiment of the present application also provides a device for predicting the remaining useful life of equipment.

[0138] like Figure 6 As shown, the device for predicting the remaining useful life of the equipment includes: a baseline training module 610, a transfer learning module 620 and a life prediction module 630.

[0139] A baseline training module 610 is configured to train an initial model based on a Transformer model based on the control condition vector of the sample device and the time series data of the operating parameters of the sample device to obtain a baseline model;

[0140] A transfer learning module 620 is configured to perform unsupervised transfer learning on the baseline model based on the time series data of the historical operating parameters of the target device and the control condition vector of the target device to obtain a life prediction model;

[0141] The life prediction module 630 is used to input the time series data of the operating parameters of the target device and the control condition vector of the target device into the life prediction model to obtain the prediction result of the remaining service life of the target device.

[0142] According to the device for predicting the remaining useful life of equipment provided by the embodiment of the present application, by utilizing a Transformer model with a control condition vector, the features of the equipment can be effectively extracted from complex industrial data, which can ensure the accuracy and real-time performance of equipment operation status monitoring, and exhibit good adaptability and stability under different equipment and working conditions, which can improve the efficiency and accuracy of feature extraction. By effectively performing transfer learning under unsupervised conditions, the performance of the model in unsupervised remaining useful life prediction for target equipment is improved, thereby improving the accuracy of the remaining useful life prediction of the equipment. Furthermore, by accurately predicting the remaining life of the equipment and timely discovering faults, the maintenance cost of the equipment can be effectively reduced, the service life of the equipment can be extended, and the operational reliability of the industrial equipment can be improved. Furthermore, there is no need to collect a large amount of abnormal data, and normal operating parameters can be directly used for prediction, which is very beneficial to industrial manufacturers and can solve the problem that abnormal data is usually difficult to obtain, the collection cost is high, and usually only laboratories have the funds to provide it.

[0143] In some embodiments, the time series data of the operating parameters of the sample devices are obtained from multiple cross-domain datasets.

[0144] In some embodiments, the transfer learning module 620 can be specifically configured to perform unsupervised transfer learning on the baseline model based on the time series data of the historical operating parameters of the target device and the control condition vector of the target device, in combination with the domain adversarial network and the distance preserving regular loss, to obtain the life prediction model.

[0145] In some embodiments, the transfer learning module 620 can include:

[0146] a learning unit configured to perform unsupervised transfer learning on the baseline model based on the time series data of the historical operating parameters of the target device and the control condition vector of the target device, in combination with the domain adversarial network and the distance preserving regular loss, to obtain a feature extraction layer in the life prediction model;

[0147] a training unit configured to train an extreme learning machine in the life prediction model based on the output of the feature extraction layer and a plurality of optimizers; and the extreme learning machine configured to predict the remaining useful life of the device based on the output of the feature extraction layer.

[0148] In some embodiments, the life prediction module 630 can include:

[0149] a prediction unit configured to input the time series data of the operating parameters of the target device and the control condition vector of the target device into the life prediction model to obtain a first result;

[0150] an adjustment unit configured to adjust the first result based on a device health index difference coefficient of a sample device to obtain a prediction result of the remaining useful life of the target device; and the device health index difference coefficient of the sample device is obtained by dividing the difference between the predicted health index of the sample device in a normal operating state and the predicted health index of the sample device in a case where the expected life is zero by the standard deviation of the predicted health index of the sample device in the normal operating state.

[0151] In some embodiments, the life prediction module 630 can further include:

[0152] an updating unit configured to update the predicted health index of the sample device in the case where the expected life is zero to the predicted health index of the target device in the case where the target device fails, to update the device health index difference coefficient, in the case where the target device fails.

[0153] The device residual service life prediction apparatus in the embodiments of the present application can be an electronic device, or a component in an electronic device, such as an integrated circuit or a chip. The electronic device can be a terminal, or other devices other than a terminal. For example, the electronic device can be a mobile phone, a tablet computer, a notebook computer, a palm computer, a vehicle-mounted electronic device, a Mobile Internet Device (MID), an augmented reality (AR) / virtual reality (VR) device, a robot, a wearable device, an ultra-mobile personal computer (UMPC), a netbook, or a personal digital assistant (PDA), and can also be a server, a Network Attached Storage (NAS), a personal computer (PC), a television (TV), a teller machine, or a self-service machine, and the like, and the embodiments of the present application are not limited in this regard.

[0154] The device residual service life prediction apparatus in the embodiments of the present application can be a device with an operating system. The operating system can be an Android operating system, an iOS operating system, or other possible operating systems, and the embodiments of the present application are not limited in this regard.

[0155] The device residual service life prediction apparatus provided in the embodiments of the present application can implement the method embodiments Figures 1 to 5 The method embodiments implement various processes, and to avoid repetition, the details are not described herein.

[0156] In some embodiments, as shown in Figure 7 The embodiments of the present application also provide an electronic device 700, which includes a processor 710, a memory 720, and a computer program stored in the memory 720 and executable on the processor 710. When the computer program is executed by the processor 710, various processes of the device residual service life prediction method embodiments described above are implemented, and the same technical effects are achieved. To avoid repetition, the details are not described herein.

[0157] It should be noted that the electronic device in the embodiments of the present application includes the mobile electronic device and the non-mobile electronic device described above.

[0158] The embodiment of the present application further provides a non-transitory computer readable storage medium, which stores a computer program. The computer program is executed by a processor to implement each process of the device residual service life prediction method embodiment and achieve the same technical effects. To avoid repetition, details are not described herein.

[0159] The processor is the processor in the electronic device in the above embodiments. The readable storage medium includes a computer readable storage medium, such as a computer readable only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0160] The embodiment of the present application further provides a computer program product, which includes a computer program. The computer program is executed by a processor to implement the device residual service life prediction method.

[0161] The processor is the processor in the electronic device in the above embodiments. The readable storage medium includes a computer readable storage medium, such as a computer readable only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0162] The embodiment of the present application further provides a chip, which includes a processor and a communication interface. The communication interface is coupled with the processor. The processor is configured to run a program or an instruction to implement each process of the device residual service life prediction method and achieve the same technical effects. To avoid repetition, details are not described herein.

[0163] It should be understood that the chip mentioned in the embodiment of the present application can also be referred to as a system level chip, a system chip, a chip system or a system on chip, etc.

[0164] It should be noted that, in this document, the term "comprising" or "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or apparatus including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or apparatus. Without more limitations, the element defined by the statement "including a" does not exclude the presence of additional identical elements in the process, method, article or apparatus including the element. In addition, it should be pointed out that the scope of the method and apparatus in the present application is not limited to the order of performing the functions shown or discussed, but can also include performing the functions in a substantially simultaneous manner or in a reverse order, for example, the described method can be performed in an order different from that described, and various steps can be added, omitted or combined. In addition, the features described with reference to some examples can be combined in other examples.

[0165] Through the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned example methods can be realized by means of software and necessary general hardware platforms, and of course, can also be realized by hardware, but in many cases, the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a computer software product in essence or in the form of a part of the prior art that makes a contribution. The computer software product is stored in a storage medium (such as a ROM / RAM, a magnetic disc, an optical disc), and includes a plurality of instructions for causing a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in various embodiments of the present application.

[0166] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the above-mentioned specific embodiments, and the above-mentioned specific embodiments are only illustrative and not restrictive. Those skilled in the art can make many forms under the inspiration of the present application without departing from the purpose of the present application and the scope protected by the claims, and all of them belong to the protection of the present application.

[0167] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an illustrative embodiment", "an example", "a specific example", or "some examples" means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily mean the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0168] Although the embodiments of the present application have been shown and described, those skilled in the art can understand that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and purposes of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A method for predicting the remaining useful life of equipment, characterized in that: include: Training an initial model based on a Transformer model based on a control condition vector of a sample device and time series data of an operating parameter of the sample device to obtain a baseline model; Based on the time series data of historical operating parameters of the target device and the control condition vector of the target device, performing unsupervised transfer learning on the baseline model to obtain a life prediction model; The time series data of the operating parameters of the target device and the control condition vector of the target device are input into the life prediction model to obtain a prediction result of the remaining service life of the target device.

2. The method for predicting the remaining useful life of equipment according to claim 1, characterized in that: The time series data of the operating parameters of the sample equipment are obtained from multiple cross-domain data sets.

3. The method for predicting the remaining useful life of equipment according to claim 1, characterized in that: The method of performing unsupervised transfer learning on the baseline model based on the time series data of historical operating parameters of the target device and the control condition vector of the target device to obtain a life prediction model includes: Based on the time series data of the historical operating parameters of the target device and the control condition vector of the target device, the baseline model is subjected to unsupervised transfer learning in combination with a domain adversarial network and a distance-preserving regularized loss to obtain the life prediction model.

4. The method for predicting the remaining useful life of equipment according to claim 3, characterized in that: The method of performing unsupervised transfer learning on the baseline model based on the time series data of the historical operating parameters of the target device and the control condition vector of the target device in combination with a domain adversarial network and a distance-preserving regularized loss to obtain the life prediction model includes: Based on the time series data of the historical operating parameters of the target device and the control condition vector of the target device, unsupervised transfer learning is performed on the baseline model in combination with a domain adversarial network and a distance-preserving regularized loss to obtain a feature extraction layer in the life prediction model; Based on multiple optimizers, the extreme learning machine in the life prediction model is trained according to the output of the feature extraction layer; the extreme learning machine is used to predict the remaining service life of the equipment according to the output of the feature extraction layer.

5. The method for predicting the remaining useful life of equipment according to any one of claims 1 to 4, characterized in that: Inputting the time series data of the operating parameters of the target device and the control condition vector of the target device into the life prediction model to obtain the prediction result of the remaining service life of the target device includes: Inputting the time series data of the operating parameters of the target device and the control condition vector of the target device into the life prediction model to obtain a first result; Based on the equipment health index difference coefficient, the first result is adjusted to obtain the predicted result of the remaining service life of the target equipment; the equipment health index difference coefficient of the sample equipment is obtained by dividing the difference between the predicted health index of the sample equipment in normal operation and the predicted health index of the sample equipment when the expected life is zero by the standard deviation of the predicted health index of the sample equipment in normal operation.

6. The method for predicting the remaining useful life of equipment according to claim 5, characterized in that: Also includes: When the target device fails, the predicted health index of the sample device when the expected life is zero is updated to the predicted health index when the target device fails, so as to update the device health index difference coefficient.

7. A device for predicting the remaining useful life of equipment, characterized in that: include: A baseline training module, configured to train an initial model based on a Transformer model based on a control condition vector of a sample device and time series data of an operating parameter of the sample device to obtain a baseline model; a transfer learning module, configured to perform unsupervised transfer learning on the baseline model based on time series data of historical operating parameters of the target device and a control condition vector of the target device to obtain a life prediction model; The life prediction module is used to input the time series data of the operating parameters of the target device and the control condition vector of the target device into the life prediction model to obtain the prediction result of the remaining service life of the target device.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method for predicting the remaining useful life of equipment according to any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for predicting the remaining useful life of equipment as described in any one of claims 1 to 6 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the method for predicting the remaining useful life of equipment according to any one of claims 1 to 6 is implemented.

Images