Device health management method and device based on small sample data and adaptive migration

Abstract

The application discloses a device health management method and device based on small sample data and adaptive migration, relates to the technical field of device health management, and health indexes are constructed for sensor data through a BLCAE model and a Pearson correlation coefficient; secondly, the initial degradation start time of the device is determined through a degradation stage detection method, and degradation stage data is extracted; finally, migration is performed between the same degradation categories, so that the degradation information can be effectively shared between similar fields or tasks to improve the accuracy and stability of the model. The method comprises the following steps: constructing health indexes of a target device based on sensor signal data, detecting whether the target device has degraded according to the health indexes; if it is determined that the target device has degraded according to the health indexes, the degradation type of the target device is determined, and degradation stage data is extracted; and the degradation stage data is input into a health state prediction model corresponding to the degradation type, so as to obtain a health state prediction result output by the health state prediction model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment health management, and in particular to an equipment health management method and device based on small sample data and adaptive transfer. BACKGROUND

[0002] With the development of equipment management mode, deep learning has been widely concerned and researched in the field of equipment health management. Since health management is the key to ensure the future efficient and sustainable service of equipment, by predicting trends, behavior patterns and correlations through statistical or machine learning models, potential failures can be predicted in advance, so as to improve the decision-making process of maintenance activities, avoid downtime, and reduce personnel injuries and property losses caused by sudden failures. Therefore, real-time health management of equipment, especially in small sample cases, has positive practical significance. However, the small sample problem is one of the main reasons hindering the wide application of deep learning in this field.

[0003] In related technologies, some scholars apply transfer learning to equipment health management under small samples, share and transfer information between similar fields or tasks, and a feature-based deep transfer learning method reduces the distribution deviation between source and target domains by minimizing the maximum average difference between them to solve the small sample problem, data distribution difference problem and individualization problem.

[0004] In the process of implementing the present application, the applicant found that the related technologies at least have the following problems:

[0005] Affected by working environment, working conditions and individual differences, the degradation process and its probability distribution of the same equipment or parts may be significantly different. In the case of excessive field distribution difference and poor data quality, forced transmission may lead to negative transfer, and the model performance will be worse. Especially in small sample cases, the model is difficult to avoid the interference of extreme data on the overall data. Moreover, the degradation stage judgment method using a single identification strategy and fixed alarm threshold cannot adapt to individual differences, is difficult to cope with the interference of random noise in data, cannot accurately extract degradation stage data, and not only wastes computing resources and reduces model training efficiency, but also affects the accuracy of equipment health management, leading to misjudgment. SUMMARY

[0006] Therefore, the application provides a device health management method and device based on small sample data and adaptive migration, mainly aiming to solve the problem that the degradation process and probability distribution of the same device or part may be significantly different due to the influence of working environment, working condition and individual differences, especially in the case of small sample, the model is difficult to avoid the interference of extreme data on the overall data. And the degradation stage judgment method using a single identification strategy and fixed alarm threshold cannot adapt to individual differences, is difficult to cope with the interference of random noise in the data, cannot accurately extract the degradation stage data, not only wastes computing resources and reduces the model training efficiency, but also affects the accuracy of device health management, leading to misjudgment.

[0007] According to the first aspect of the application, a device health management method based on small sample data and adaptive migration is provided, which comprises:

[0008] Obtaining sensor signal data of a target device, and constructing a health index of the target device based on the sensor signal data;

[0009] Detecting whether the target device has degraded according to the health index;

[0010] If it is determined that the target device has degraded according to the health index, determining the degradation type of the target device, and extracting degradation stage data in the sensor signal data based on a degradation stage detection method and the health index;

[0011] Obtaining a health state prediction model corresponding to the degradation type, inputting the degradation stage data into the health state prediction model, and obtaining a health state prediction result output by the health state prediction model.

[0012] According to the second aspect of the application, a device health management device based on small sample data and adaptive migration is provided, which comprises:

[0013] The construction module is configured to obtain sensor signal data of a target device, and construct a health index of the target device based on the sensor signal data;

[0014] The detection module is configured to detect whether the target device has degraded according to the health index;

[0015] The extraction module is configured to determine the degradation type of the target device if it is determined that the target device has degraded according to the health index, and extract degradation stage data in the sensor signal data based on a degradation stage detection method and the health index;

[0016] The prediction module is configured to acquire a health state prediction model corresponding to the degradation type, input the degradation stage data into the health state prediction model, and obtain a health state prediction result output by the health state prediction model.

[0017] Through the above ending scheme, the device health management method and device based on small sample data and adaptive migration are provided. The sensor signal data of a target device is acquired, a health index of the target device is constructed based on the sensor signal data, and whether the target device has degradation is detected according to the health index. If it is determined that the target device has degradation according to the health index, the degradation type of the target device is determined, degradation stage data is extracted from the sensor signal data based on a degradation stage detection method and the health index, a health state prediction model corresponding to the degradation type is acquired, the degradation stage data is input into the health state prediction model, and a health state prediction result output by the health state prediction model is obtained. First, a reasonable health index is constructed for sensor data by using a BLCAE (Bidirectional Long Short-Term Memory Convolutional Neural Network Model) and a Pearson correlation coefficient. Second, the initial degradation start time of the device is determined by using an OSW-PI-ACED (Overlapping Sliding Window-Pettitt Mutation Point, Inflection Point Extraction-Adaptive Continuous Exceeding Detection Verification) degradation stage detection method, and degradation stage data of each data is extracted. Finally, migration is performed between the same degradation categories, so that the degradation information can be effectively shared between similar fields or tasks, the occurrence of negative migration is avoided, and the reusability of historical knowledge and information is improved. By performing necessary processing on the limited data collected by the sensor, the data is classified to reduce the feature distribution difference between the source domain and the target domain, and domain adaptive migration is performed between the same categories, so that the degradation information can be shared between similar fields or tasks, the occurrence of negative migration is avoided, and the degradation stage data of the device can be accurately extracted, reducing the interference of irrelevant information on the prediction of the health state of the device. Even in the case of small samples, the degradation information can be captured from a small amount of data, the data cross-domain deviation is reduced, the precision and stability of the model are improved, and the health management of the device is realized.

[0018] The above description is only a summary of the technical solutions of the present application. In order to more clearly understand the technical means of the present application, the present application can be implemented according to the content of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific embodiments of the present application are described as follows. BRIEF DESCRIPTION OF DRAWINGS

[0019] Various other advantages and benefits will become apparent to those of ordinary skill in the art, upon reading the following detailed description of the preferred embodiment. The accompanying drawings are included to provide a description of preferred embodiments, and are not intended to limit the scope of the present application. Furthermore, the drawings are included to illustrate embodiments of the application, and, to provide a better understanding of the application, the embodiments of which will be described and illustrated in connection with the appended drawings. In the drawings:

[0020] Figure 1 A method flow diagram of device health management based on small sample data and adaptive migration provided by an embodiment of the application is shown;

[0021] Figure 2A Another method flow diagram of device health management based on small sample data and adaptive migration provided by an embodiment of the application is shown;

[0022] Figure 2B A health state prediction flow diagram provided by an embodiment of the application is shown;

[0023] Figure 3A A structure diagram of device health management based on small sample data and adaptive migration provided by an embodiment of the application is shown;

[0024] Figure 3B Another structure diagram of device health management based on small sample data and adaptive migration provided by an embodiment of the application is shown;

[0025] Figure 4 An apparatus structure diagram of a computer device provided by an embodiment of the application is shown. DETAILED DESCRIPTION

[0026] Exemplary embodiments of the present application will be described herein below with reference to the accompanying drawings. Although exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided so that the present application can be more thoroughly and completely understood, and so that the scope of the present application can be accurately conveyed to those skilled in the art.

[0027] An embodiment of the present application provides a device health management method based on small sample data and adaptive migration, as shown in the figure, the method comprises: Figure 1

[0028] 101, obtaining sensor signal data of a target device, and constructing a health index of the target device based on the sensor signal data.

[0029] ​The health stage data provides little information for equipment health state prediction, and the degradation stage judgment method currently adopts a single identification strategy and fixed alarm threshold, which cannot adapt to individual differences and is difficult to cope with random noise interference in the data. If the degradation stage data cannot be accurately extracted, not only will it waste computing resources and reduce model training efficiency, but it will also affect the accuracy of equipment health management, leading to misjudgment. The feature-based deep transfer learning method reduces the distribution deviation between the source domain and the target domain by minimizing the maximum average deviation between them. However, if forced transmission is performed in the case of large domain distribution difference and poor data quality, it may lead to negative transfer, and the model performance will be worse. Especially in the case of small sample, the model is difficult to cope with the interference of extreme data on the overall data.

[0030] To solve this problem, the present application proposes a device health management method based on small sample data and adaptive transfer. First, a reasonable health index is constructed for sensor data through BLCAE (Bidirectional Long Short-Term Memory Convolutional Neural Network Model) and Pearson correlation coefficient (Pearson correlation coefficient); second, the initial degradation start time of the equipment is determined by the degradation stage detection method of OSW-PI-ACED (based on overlapping sliding window-Pettitt mutation point, inflection point extraction-adaptive continuous exceedance detection verification), and the degradation stage data of each data is extracted; finally, transfer is performed between the same degradation category, so that the degradation information can be effectively shared between similar fields or tasks, while avoiding the occurrence of negative transfer and improving the reusability of historical knowledge and information. The execution subject of the present application can be a device health management system. The device health management system relies on the computing power of the server to provide services for users. The server can be a standalone server, or it can provide cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. The server of the basic cloud computing is used to quickly, accurately and stably realize the device health management of the device health management system.

[0031] In the embodiments of the present application, the device health management system obtains sensor signal data of the target device, and constructs health indicators of the target device based on the sensor signal data. The sensor signal data in actual application can be any sensor signal data containing equipment degradation information, such as vibration data, temperature data, humidity data, etc., which can be obtained from the sensors installed on the equipment. It can also be sensor signal data of the equipment under different working conditions. At this time, the sensor signal data refers to the data containing equipment degradation information from the first time the equipment starts running to failure, such as full-life vibration signal data, full-life temperature signal data, etc.

[0032] 102. Determine whether the target device has degraded according to the health index.

[0033] In the embodiment of the present application, the device health management system determines whether the target device has degraded according to the health index. Taking the change of vibration amplitude in the bearing degradation process as an example, in its full life vibration data, the amplitude of vibration is in a fluctuating state, not in a linear relationship with time, and the change of vibration amplitude is a process from slight to significant increase. When the amplitude change shows an upward trend, it means that the bearing has begun to degrade. The value range of the health index is from 1 to 0, which can adapt to sensor data of different components and different working conditions, and further more clearly determine whether the target device has degraded, wherein, in the health stage, the health index is basically between [0.95, 1], and in the degradation stage, the value of the health index gradually decreases to 0. In short, for example, [2, 2.1, 1.99, 1.98, 2…2.5, 2.7, 2.6, 2.8…3…20] is regarded as full life vibration data, and [1, 0.99, 0.99, 1, 0.97…0.99, 0.95, 0.92, …0.2…0] is regarded as the health index corresponding to the full life vibration data. If the health index constructed at the time of detection is [0.99, 1, 0.99, 0.98…0.99], it means that the current bearing is in a healthy state and does not need health state prediction; if the health index constructed at the time of detection is [0.98, 0.99, 1, 0.97, 0.8…0.2…0.01], that is, the health index gradually decreases as a whole with the increase of time, it means that the bearing has begun to degrade at this time and needs health state prediction.

[0034] 103. If it is determined that the target device has degraded according to the health index, determine the degradation type of the target device, and extract the degradation stage data from the sensor signal data based on the degradation stage detection method and the health index.

[0035] In the embodiment of the present application, if the device health management system determines that the target device has degraded according to the health index, the degradation type of the target device is determined, and the degradation stage data is extracted from the sensor signal data based on the degradation stage detection method and the health index, wherein the degradation type includes rapid degradation type and slow degradation type. When it is determined that the bearing has begun to degrade, the degradation type of the test bearing needs to be determined, and then the degradation stage data is extracted and put into the model corresponding to the degradation type for health state prediction.

[0036] 104. Obtain a health state prediction model corresponding to the degradation type, input the degradation stage data into the health state prediction model, and obtain a health state prediction result output by the health state prediction model.

[0037] In the embodiment of the present application, the device health management system obtains a health state prediction model corresponding to the degradation type, inputs the degradation stage data into the health state prediction model, and obtains the health state prediction result output by the health state prediction model. The present application trains two models using two types of degradation data, constructs a health index for the device through the sensor signal, judges the current device health state according to the health index, extracts the data of the device in the degradation stage, simultaneously completes data classification, and migrates between the same degradation categories, reduces the data distribution difference, avoids the occurrence of negative migration, thereby realizing device health management under small sample.

[0038] The method provided by the embodiment of the present application acquires sensor signal data of a target device, constructs a health index of the target device based on the sensor signal data, and detects whether the target device has degraded according to the health index; if it is determined that the target device has degraded according to the health index, the degradation type of the target device is determined, degradation stage data is extracted from the sensor signal data based on a degradation stage detection method and the health index; a health state prediction model corresponding to the degradation type is acquired, the degradation stage data is input into the health state prediction model, and a health state prediction result output by the health state prediction model is obtained. First, a reasonable health index is constructed for the sensor data through BLCAE (Bidirectional Long Short-Term Memory Convolutional Neural Network Model) and Pearson correlation coefficient (Pearson correlation coefficient); second, the initial degradation start time of the device is determined by the degradation stage detection method of OSW-PI-ACED (based on overlapping sliding window-Pettitt mutation point, inflection point extraction-adaptive continuous exceeding detection verification), and the degradation stage data of each data is extracted; finally, migration is performed between the same degradation categories, so that the degradation information can be effectively shared between similar fields or tasks, while avoiding the occurrence of negative migration and improving the reusability of historical knowledge and information. By performing necessary processing on the limited data collected by the sensor, the data is classified to reduce the feature distribution difference between the source domain and the target domain, and domain adaptive migration is performed between the same categories, so that the degradation information can be shared between similar fields or tasks, the occurrence of negative migration is avoided, and the degradation stage data of the device can be accurately extracted, reducing the interference of irrelevant information on the device health state prediction. Even in the case of small sample, the degradation information can be captured from a small amount of data, the data cross-domain deviation is reduced, the precision and stability of the model are improved, and the health management of the device is realized.

[0039] Further, as a refinement and expansion of the specific implementation of the above embodiment, in order to completely describe the specific implementation process of the embodiment, the embodiment of the present application provides another device health management method based on small sample data and adaptive migration, as shown in Figure 2A The method comprises the following steps.

[0040] 201、obtain sensor signal sample data, obtain sample data containing labels as source domain data in the sensor signal sample data, and obtain sample data not containing labels as target domain data in the sensor signal sample data.

[0041] In the embodiment of the present application, the equipment health management system first obtains sensor signal sample data, wherein the sensor signal sample data is equipment time series degradation data under different working conditions. Since the vibration of the equipment changes during the degradation process, the equipment time series degradation data in the embodiment of the present application mainly refers to vibration data. For example, the sampling time is set to 2s, the sampling interval time is set to 1min, the vibration amplitude, acceleration and other signals of the equipment are collected through the sensor on the equipment, and the equipment time series degradation data is obtained. The equipment time series degradation data includes vibration signal data of the equipment from health to failure. The data collected in the embodiment of the present application is vibration data in the bearing degradation process, including PHM 2012 Dataset and XJTU-SY Dataset. Taking XJTU-SY as an example, the dataset contains three working conditions, and each working condition has five full-life vibration data of bearings from health to failure, and PHM 2012 is similar. Then, the equipment health management system obtains sample data containing labels as source domain data in the sensor signal sample data, and obtains sample data not containing labels as target domain data in the sensor signal sample data. Through necessary processing of the limited data collected by the sensor, the data is classified to reduce the feature distribution difference between the source domain and the target domain, and domain adaptive migration is performed between the same categories, so that the degradation information can be shared between similar fields or tasks, and the occurrence of negative migration is avoided.

[0042] It should be noted that in actual application, the sensor signal sample data includes sample data containing labels and sample data not containing labels. However, if the data in the two data sets used for training and verification does not have labels, several groups of data need to be extracted from all the data and labeled as source domains, where the data refers to the sensor signal sample data. When training and verifying the model using the two unlabeled data sets, in order to simplify the process, part of the data is extracted from the two categories after the degradation process classification is completed, and then labeled as the source domain.

[0043] 202、divide the target domain data into a training data set and a test data set according to a preset ratio, and use the source domain data and the training data set as a target training data set.

[0044] In the embodiment of the present application, the equipment health management system divides the target domain data into a training data set and a test data set according to a preset ratio, and uses the source domain data and the training data set as a target training data set, wherein the preset ratio can be set to 80% for the training set and 20% for the test set.

[0045] 203、obtain a pre-constructed bidirectional long short-term memory convolutional neural network model and a Pearson correlation coefficient, process the target training data set by using the bidirectional long short-term memory convolutional neural network model and the Pearson correlation coefficient, and obtain a first health index.

[0046] In the embodiment of the present application, the device health management system obtains a pre-constructed bidirectional long short-term memory convolutional neural network model and a Pearson correlation coefficient, processes the target training data set by using the bidirectional long short-term memory convolutional neural network model and the Pearson correlation coefficient, and obtains a first health index. The BLCAE model is a Bi-LSTM-CNN automatic encoder, which is used to extract features containing degradation information in device time-series degradation data. The features output by the BLCAE model are a three-dimensional matrix, which can not only effectively filter out random noise in the original data, but also make the constructed health index more representative. By extracting features through the BLCAE model and using the Pearson correlation coefficient to calculate the correlation between the features, the health index can be obtained, which can adapt to sensor data of different components and different working conditions, thereby improving the stability and universality of the health state prediction model.

[0047] 204、divide the target training data set into an abrupt degradation data set and a slow degradation data set according to the first health index, and extract abrupt degradation stage data in the abrupt degradation data set and slow degradation stage data in the slow degradation data set by using a degradation stage detection method.

[0048] The health index can describe the degradation process of the device. Some devices need a long time from the beginning of degradation to failure, which can be called slow degradation. Some devices only need a short time, which can be called abrupt degradation. According to the constructed health index, data of the same degradation type can be divided into a category, which can reduce the difference in data distribution and avoid negative transfer. In the embodiment of the present application, the device health management system divides the first health index into a health index corresponding to the abrupt degradation type and a health index corresponding to the slow degradation type according to the degradation type, and then divides the target training data set into an abrupt degradation data set and a slow degradation data set based on the health index corresponding to the abrupt degradation type and the health index corresponding to the slow degradation type. The degradation type can be simply understood as the ratio of the time t from the beginning of degradation to failure to the time T from the beginning of use to failure, i.e. t / T. According to the ratio and the size of the data amount, it can be divided into two categories or n categories. In the embodiment of the present application, the degradation type is divided into abrupt degradation (t / T is relatively small) and slow degradation (t / T is relatively large).

[0049] Then, the device health management system extracts rapid degradation phase data in the rapid degradation data set and slow degradation phase data in the slow degradation data set by using a degradation phase detection method. The degradation phase detection method is a degradation phase detection method based on an overlapping sliding window-Pettitt mutation point, inflection point extraction-adaptive continuous exceeding detection verification, which is used to determine the initial degradation start time of the device and extract the degradation phase data.

[0050] Specifically, the device health management system extracts a target health indicator in the health indicators corresponding to the rapid degradation type by using the 2sigma principle, and determines a detection threshold value by using the target health indicator. Then, the device health management system slides an overlapping sliding window above the target health indicator, and extracts mutation points of the target health indicator in a plurality of data points included in the overlapping sliding window based on a mutation point detection algorithm during the sliding process, to obtain a plurality of mutation points corresponding to each overlapping sliding window.

[0051] For each overlapping sliding window, the device health management system determines a plurality of inflection points by using a plurality of data points included in the overlapping sliding window. Specifically, for each data point, a target data point is determined in the plurality of data points according to the data point, where the target data point is the third data point behind the data point. Then, a slope value of the data point and the target data point is calculated; if the slope value is less than a preset threshold value, the data point is taken as an inflection point, where the preset threshold value is -0.002. Then, the device health management system sorts the plurality of mutation points and the plurality of inflection points in time sequence to obtain a time sequence. The device health management system extracts a plurality of target time points in the time sequence according to a continuous exceeding detection verification principle. Specifically, the first time point of the time sequence is detected according to the continuous exceeding detection verification principle; if the i-th time point in the time sequence is greater than the detection threshold value, the i+1-th time point is continuously detected until the i+j-th time point is detected, and the calculation formula of the parameter j is as follows:

[0052] Formula 1:

[0053] Where n is the number of time points in the time sequence. When it is detected that there are j-1 time points greater than the detection threshold value, the i-th time point is taken as a target time point. The last time point of the time sequence is detected until a plurality of target time points are obtained.

[0054] Subsequently, the device health management system sorts the plurality of target time points in time sequence to obtain a sorting result, and reads the target time point ranked first in the sorting result as an initial degradation time. Finally, the device health management system extracts rapid degradation phase data in the rapid degradation data set according to the initial degradation time.

[0055] With a set of full life data as an example, first, the health indicators are extracted using overlapping sliding windows to avoid misjudgment of data at the end of the window, wherein the window size and the overlapping size are set to 100 and 15 respectively. Then, the Pettitt (mutation point detection) method is used to extract the mutation points in the window, and the slope between each point and the third point after it in each window is calculated. If the slope is less than the set threshold, i.e. -0.002, the point is considered as an inflection point and is likely to be a degradation point. Subsequently, the detected mutation points and inflection points are placed in the same list in chronological order. Finally, according to the 2sigma principle, the first 10% of the health indicators are used to construct the detection threshold, and according to the continuous exceeding detection verification principle, it is first judged whether the ith point in the list exceeds the detection threshold. If it exceeds the detection threshold, it is judged whether j-1 points among i+1, i+2,..., i+j points exceed the threshold, wherein j is determined by the total number n of mutation points in the list. If there are j-1 points exceeding the threshold, it means that the ith point meets the requirements, and then all the points meeting the requirements are placed in a new list in chronological order, and the time corresponding to the first value in the new list is taken as the initial degradation time. The original data after the initial degradation time point is extracted as the degradation stage data.

[0056] The health indicators are used to represent the degradation process of the device based on the sensor signal data, and then the OSW-PI-ACED method is used to determine the initial degradation time based on the constructed health indicators. Compared with the method of directly determining the initial degradation time based on the sensor signal data using the OSW-PI-ACED method, the health indicators constructed in the embodiment can filter out invalid data, simply and clearly represent the initial degradation time and degradation process of the device, and also reduce the interference caused by random noise in the process of determining the initial degradation time, thereby improving the accuracy of the health state prediction model and the efficiency of the device health management.

[0057] 205, the rapid degradation stage data and the slow degradation stage data are used for model construction respectively, and the health state prediction model corresponding to the rapid degradation type and the health state prediction model corresponding to the slow degradation type are obtained.

[0058] In the embodiment of the present application, the device health management system obtains the MBR migration model corresponding to the rapid degradation type, the LP migration model corresponding to the rapid degradation type, the MBR migration model corresponding to the slow degradation type, and the LP migration model corresponding to the slow degradation type. It should be noted that the parameters of the MBR and LP migration models are set at the same time of obtaining the migration models, including learning rate, batch, initialization bias and weight, etc. Then, the device health management system inputs the rapid degradation stage data into the MBR model corresponding to the rapid degradation type to obtain the rapid degradation feature, and inputs the slow degradation stage data into the MBR model corresponding to the slow degradation type to obtain the slow degradation feature. Subsequently, the device health management system inputs the rapid degradation feature into the LP model corresponding to the rapid degradation type to obtain the health status label of the rapid degradation type, and inputs the slow degradation feature into the LP model corresponding to the slow degradation type to obtain the health status label of the slow degradation type.

[0059] Then, the device health management system calculates the first loss function using the health status label of the rapid degradation type, and calculates the second loss function using the health status label of the slow degradation type. Specifically, the device health management system calculates the health status label of the rapid degradation type using the loss function calculation formula to obtain the first loss function, and the loss function calculation formula is as follows formula 2:

[0060] Formula 2: L total = λL RUL + μL MMD

[0061]

[0062] Wherein, y i is the real health status label, is the health status label of the rapid degradation type, λ is the first weight parameter, μ is the second weight parameter, n is the sampling number at the degradation stage, D S is the source domain feature, D T is the target domain feature, n S is the sample size of the source domain d S , n T is the sample size of the target domain d T , ‖·‖ H is the reproducing kernel Hilbert space, is a calculation function for mapping data to the reproducing kernel Hilbert space. It should be noted that the calculation process of the second loss function is the same as that of the first loss function. Wherein, L RUL is the mean square difference between the actual health status label and the predicted health status label obtained by the model, that is, part of the loss function.

[0063] Then, the device health management system updates the network parameters of the MBR migration model corresponding to the rapid degradation type and the LP migration model corresponding to the rapid degradation type based on the back propagation algorithm using the first loss function, and updates the network parameters of the MBR migration model corresponding to the slow degradation type and the LP migration model corresponding to the slow degradation type using the second loss function. The training of the health state prediction model is completed by continuously reducing the gap between the actual health state label and the predicted health state label, that is, continuously reducing the loss function value.

[0064] Finally, the device health management system updates the network parameters of the MBR migration model corresponding to the rapid degradation type, the LP migration model corresponding to the rapid degradation type, the MBR migration model corresponding to the slow degradation type, and the LP migration model corresponding to the slow degradation type according to the preset iteration number, and obtains the health state prediction model corresponding to the rapid degradation type and the health state prediction model corresponding to the slow degradation type.

[0065] 206, the test data set is processed using the bidirectional long short-term memory convolutional neural network model and the Pearson correlation coefficient to obtain a second health indicator, the specified degradation phase data is extracted from the test data set based on the degradation phase detection method and the second health indicator, and the health state prediction model is tested using the specified degradation phase data.

[0066] In the embodiments of the present application, the device health management system processes the test data set using the bidirectional long short-term memory convolutional neural network model and the Pearson correlation coefficient to obtain a second health indicator. It should be noted that the process of constructing the second health indicator is the same as the process of constructing the first health indicator described above.

[0067] If it is determined that the device has degraded according to the second health indicator detection, the degradation type of the device is determined, and the specified degradation phase data is extracted from the test data set based on the degradation phase detection method. It should be noted that the process of extracting the degradation phase data based on the degradation phase detection method is the same as the process of extracting the rapid degradation phase data from the rapid degradation data set and the slow degradation phase data from the slow degradation data set using the degradation phase detection method. The health state prediction model corresponding to the degradation type is obtained, the specified degradation phase data is input into the health state prediction model, and the health state prediction result output by the health state prediction model is obtained. Then, the device health management system obtains the real health state result of the device in the sensor signal sample data, calculates the model accuracy using the health state prediction result and the health state real result, and if the model accuracy is greater than or equal to the preset accuracy threshold, it indicates that the accuracy of the trained health state prediction model meets the requirements, and the health state prediction model is stored, which can ensure the accuracy and accuracy of the model, thereby realizing efficient management of the device health.

[0068] Based on the above process, the health state prediction model of the two degradation types can be trained, the data of the whole process from health to failure of the sensor acquisition device is collected, and after the degradation process classification and degradation stage data extraction, the data is input into the MBR-LP migration model to obtain the device health state prediction result. In the case of small sample, the cross-condition health state prediction accuracy of 12 groups of data can reach more than 75%, which is much higher than the national standard requirement of 60%, which shows the advancement and superiority of the device health management method based on small sample data and adaptive migration proposed in the application.

[0069] 207、Obtaining sensor signal data of the target device, and constructing a health index of the target device based on the sensor signal data.

[0070] In the embodiment of the present application, the device health management system acquires a preset sampling interval time, acquires sensor signal data by performing multiple data acquisition on the target device through the sensor according to the preset sampling interval. For example, the sampling time is set to 2S, the sampling interval time is set to 1min, the entire degradation process of the target device needs to be sampled N times, assuming that the data collected within 2S is a group of data, then after N times of sampling, N groups of data are obtained, which are the sensor signal data. Then, the device health management system acquires the pre-constructed bidirectional long short-term memory convolutional neural network model, inputs the sensor signal data into the bidirectional long short-term memory convolutional neural network model, and acquires the feature data set output by the bidirectional long short-term memory convolutional neural network model. For example, inputting the above obtained N groups of data into the BLCAE model will obtain N feature data, and the N feature data is the feature data set A, wherein A=(n1, n2,... nm), m is the sampling number of the target device from the beginning to the failure. Subsequently, the device health management system acquires a preset range, selects multiple feature data in the feature data set according to the preset range, and takes the average value of the multiple feature data as the reference data. Since the sensor signal data is collected in time sequence, after extracting the features through the BLCAE model, the features in the feature data set are also in time sequence, therefore, multiple feature data in the feature data set is directly selected according to the preset range. The preset range is the first 10%-20% of the data feature set, because the new parts have not been run-in when they are used for the first time, the initial vibration may be large, therefore, selecting features from the first data will result in low accuracy of the constructed health index. Finally, the device health management system acquires the Pearson correlation coefficient, calculates the correlation between each feature data and the reference data in time sequence using the Pearson correlation coefficient, and obtains the health index. It should be noted that the construction process of the health index is the same as the construction process of the first health index and the second health index, which will not be described in detail here. By extracting features through the bidirectional long short-term memory convolutional neural network model, and then calculating the correlation between the features using the Pearson correlation coefficient, the health index is obtained, which can adapt to sensor data of different components and different working conditions, not only can accurately judge the health state of the device, but also can improve the stability, universality and accuracy of model training, thereby realizing device health management under small sample.

[0071] 208、According to the health index, it is detected whether the target device has degraded, if it is determined that the target device has degraded according to the health index, the degradation type of the target device is determined, and the degradation stage data is extracted from the sensor signal data based on the degradation stage detection method and the health index.

[0072] In the embodiment of the present application, the device health management system detects whether the target device has degraded according to the health indicators. If it is determined that the target device has degraded according to the health indicators, the type of degradation of the target device is determined. Then, the device health management system extracts the degradation stage data in the sensor signal data by using the degradation stage detection method and the health indicators. It should be noted that the process of extracting the degradation stage data in the sensor signal data based on the degradation stage detection method is the same as the process of extracting the rapid degradation stage data in the rapid degradation data set and the slow degradation stage data in the slow degradation data set by using the degradation stage detection method.

[0073] Specifically, the device health management system extracts a target health indicator in the health indicators by using the 2sigma principle, and determines a detection threshold value by using the target health indicator. Then, the device health management system slides an overlapping sliding window above the target health indicator, and extracts a mutation point of the target health indicator in a plurality of data points included in the overlapping sliding window based on the mutation point detection algorithm during the sliding process, to obtain a plurality of mutation points corresponding to each overlapping sliding window. For each overlapping sliding window, the device health management system determines a plurality of inflection points by using a plurality of data points included in the overlapping sliding window. Specifically, for each data point, a target data point is determined in the plurality of data points according to the data point, where the target data point is the third data point behind the data point. Then, the device health management system calculates a slope value of the data point and the target data point; if the slope value is less than a preset threshold value, it means that the data point may be a degradation point, and the data point is taken as an inflection point, where the preset threshold value is -0.002. Then, the device health management system sorts the plurality of mutation points and the plurality of inflection points in time sequence to obtain a time sequence, and extracts a plurality of target time points in the time sequence according to the continuous exceeding detection verification principle. Specifically, the device health management system starts detecting from the first time point of the time sequence according to the continuous exceeding detection verification principle; if the i th time point in the time sequence is greater than the detection threshold value, the i+1 th time point is continuously detected until the i+j th time point is detected, where the calculation formula of the parameter j is formula 1.

[0074] Subsequently, the device health management system sorts the plurality of target time points in time sequence to obtain a sorting result, and reads the target time point ranked first in the sorting result as an initial degradation time. Finally, the device health management system extracts the degradation stage data in the sensor signal data according to the initial degradation time. Based on the degradation stage detection method, the device health management system can quickly and accurately extract the degradation stage data of the target device, so as to realize health state detection of different components and different working conditions of the same component, and realize health management of the device.

[0075] 209、obtaining a health state prediction model corresponding to the degradation type, inputting the degradation stage data into the health state prediction model, and obtaining a health state prediction result output by the health state prediction model.

[0076] In the embodiments of the present application, the device health management system obtains a health state prediction model corresponding to the degradation type, inputs the degradation stage data into the health state prediction model, and obtains a health state prediction result output by the health state prediction model. As described above, the device health management system constructs reasonable health indicators for sensor data through BLCAE and Pearson correlation coefficient. Secondly, the initial degradation start time of the device is determined through the degradation stage detection method of OSW-PI-ACED, and the degradation stage data of each data is extracted. Finally, migration is performed between the same degradation categories, so that the degradation information can be effectively shared between similar fields or tasks, while avoiding negative migration and improving the reusability of historical knowledge and information. Compared with other methods, the present application can achieve high precision under small sample conditions, avoid high cost of data preparation in some specific devices, and perform more excellent and stable cross-domain performance.

[0077] From the above process, a health state prediction schematic diagram proposed by the embodiments of the present application is as follows:

[0078] As Figure 2BAs shown, the original sensor signal data, i.e. vibration data in the bearing degradation process, is obtained, the original sensor signal data is divided into labeled source domain data and unlabeled target domain data, the source domain data is all used as a training set, and the target domain data is randomly divided into two parts, of which 80% is a training set and 20% is a test set. The training set is used to construct a health index, and then classification is performed. The health index can describe the degradation process of the equipment. Some equipment needs a long time from the beginning of degradation to failure, which can be called slow degradation. While some only need a short time, which can be called rapid degradation. The classification process is to divide the original data into rapid degradation or slow degradation categories according to the degradation process described by the health index. Then, the initial degradation start time of the equipment is judged by a degradation stage detection method based on an overlapping sliding window-Pettitt mutation point, inflection point extraction-adaptive continuous exceeding detection verification, and the degradation stage data of each data in the data set is extracted. Subsequently, in order to improve the performance of the model in predicting the health status of the equipment and health management under small samples, the model is trained in the following way: features are extracted using the MBR model; the features are input into the LP model to obtain the health status label; the health status label is used to calculate the loss function, and the network parameters of each layer of the MBR and LP models are updated through the back propagation algorithm until the set iteration number is reached, and the training is stopped, and a health status prediction model is obtained. The device health management method for small sample data classification degradation and domain adaptive migration proposed in the present application reduces the interference of random noise in the device degradation stage searching process by constructing a health index, and sets adaptive threshold and adaptive exceeding detection number for different data, so that the model has better generalization. After data classification, migration is performed between the same degradation categories to reduce the data distribution difference and avoid the occurrence of negative migration. At the same time, the health status prediction and health management of the equipment under small samples are completed.

[0079] The method provided by the embodiment of the application obtains sensor signal data of a target device, constructs a health index of the target device based on the sensor signal data, and detects whether the target device has degraded according to the health index; if it is determined that the target device has degraded according to the health index, a degradation type of the target device is determined, degradation stage data is extracted from the sensor signal data based on a degradation stage detection method and the health index, a health state prediction model corresponding to the degradation type is obtained, the degradation stage data is input into the health state prediction model, and a health state prediction result output by the health state prediction model is obtained. First, a reasonable health index is constructed for sensor data by using a BLCAE (bidirectional long short-term memory convolutional neural network model) and a Pearson correlation coefficient (Pearson correlation coefficient); second, an initial degradation start time of the device is determined by using an OSW-PI-ACED (overlapping sliding window-based Pettitt mutation point, inflection point extraction, and adaptive continuous exceeding detection and verification) degradation stage detection method, and degradation stage data of each data is extracted; and finally, domain adaptation migration is performed between the same degradation categories, so that the degradation information can be effectively shared between similar fields or tasks, the occurrence of negative migration is avoided, and the reusability of historical knowledge and information is improved. By performing necessary processing on limited data collected by the sensor, the data is classified to reduce the feature distribution difference between the source domain and the target domain, and domain adaptation migration is performed between the same categories, so that the degradation information can be shared between similar fields or tasks, the occurrence of negative migration is avoided, and the degradation stage data of the device can be accurately extracted, reducing the interference of irrelevant information on the prediction of the health state of the device. Even in the case of a small sample, the degradation information can be captured from a small amount of data, the data cross-domain deviation is reduced, the precision and stability of the model are improved, and the health management of the device is realized.

[0080] Further, as Figure 1 A specific implementation of the method, the embodiment of the application provides a device health management device based on small sample data and adaptive migration, as Figure 3A The device includes a construction module 301, a detection module 302, a determination module 303, and a prediction module 304.

[0081] The construction module 301 is configured to obtain sensor signal data of a target device, and construct a health index of the target device based on the sensor signal data.

[0082] The detection module 302 is configured to detect whether the target device has degraded according to the health index.

[0083] The extraction module 303 is configured to determine a degradation type of the target device if it is determined that the target device has degraded according to the health index, extract degradation stage data from the sensor signal data based on a degradation stage detection method and the health index.

[0084] The prediction module 304 is configured to acquire a health state prediction model corresponding to the degradation type, input the degradation stage data into the health state prediction model, and obtain a health state prediction result output by the health state prediction model.

[0085] In a specific application scenario, the construction module 301 is configured to acquire a preset sampling interval time, perform multiple data acquisitions on the target device through a sensor according to the preset sampling interval time, and obtain sensor signal data; acquire a pre-constructed bidirectional long short-term memory convolutional neural network model, input the sensor signal data into the bidirectional long short-term memory convolutional neural network model, acquire feature data sets output by the bidirectional long short-term memory convolutional neural network model; acquire a preset range, select multiple feature data in the feature data sets according to the preset range, and take an average value of the multiple feature data as reference data; acquire a Pearson correlation coefficient, and calculate a correlation between each feature data and the reference data in a time sequence according to the Pearson correlation coefficient to obtain the health index.

[0086] In a specific application scenario, the extraction module 303 is configured to acquire a 2sigma principle, extract a target health index in the health index by using the 2sigma principle, and determine a detection threshold value by using the target health index; acquire an overlapping sliding window and a mutation point detection algorithm of the degradation stage detection method, slide the overlapping sliding window on the target health index, extract a mutation point of the target health index in multiple data points included in the overlapping sliding window based on the mutation point detection algorithm in a sliding process, and obtain multiple mutation points corresponding to each overlapping sliding window; for each overlapping sliding window, determine multiple inflection points by using multiple data points included in the overlapping sliding window, sort the multiple mutation points and the multiple inflection points in a time sequence to obtain a time sequence; acquire a continuous exceeding detection verification principle, extract multiple target time points in the time sequence according to the continuous exceeding detection verification principle, sort the multiple target time points in a time sequence to obtain a sorting result; read a target time point ranked first in the sorting result as an initial degradation time of the target device, and extract the degradation stage data in the sensor signal data according to the initial degradation time.

[0087] In a specific application scenario, the extraction module 303 is configured to, for each data point, determine a target data point according to the data point in the plurality of data points, the target data point being the third data point behind the data point; obtain a preset threshold, calculate a slope value of the data point and the target data point; if the slope value is less than the preset threshold, take the data point as an inflection point; and perform calculation processing on each data point of the overlapping sliding window to obtain the plurality of inflection points.

[0088] In a specific application scenario, the extraction module 303 is configured to start detection from a first time point of the time sequence according to the continuous exceeding detection verification principle; if an i th time point in the time sequence is greater than the detection threshold, continue to detect an i+1 th time point until an i+j th time point is detected, wherein,

[0089]

[0090] wherein n is the number of time points in the time sequence; when it is detected that there are j-1 time points greater than the detection threshold, take the i th time point as the target time point; and continue to detect according to the continuous exceeding detection verification principle until a last time point of the time sequence, to obtain the plurality of target time points.

[0091] In a specific application scenario, as shown in Figure 3B the device further includes a training module 305.

[0092] The training module 305 is configured to acquire sensor signal sample data, acquire sample data containing labels as source domain data in the sensor signal sample data, acquire sample data not containing labels as target domain data in the sensor signal sample data, and acquire equipment time sequence degradation data under different working conditions as the sensor signal sample data; acquire a preset proportion, divide the target domain data into a training data set and a test data set according to the preset proportion, and take the source domain data and the training data set as a target training data set; acquire a pre-constructed bidirectional long short-term memory convolutional neural network model and a Pearson correlation coefficient, process the target training data set by using the bidirectional long short-term memory convolutional neural network model and the Pearson correlation coefficient, and obtain a first health index; divide the first health index into a health index corresponding to an abrupt degradation type and a health index corresponding to a slow degradation type according to a degradation type, the degradation type includes the abrupt degradation type and the slow degradation type, divide the target training data set into an abrupt degradation data set and a slow degradation data set based on the health index corresponding to the abrupt degradation type and the health index corresponding to the slow degradation type; acquire a degradation stage detection method, extract abrupt degradation stage data in the abrupt degradation data set by using the degradation stage detection method, and extract slow degradation stage data in the slow degradation data set by using the degradation stage detection method; and respectively use the abrupt degradation stage data and the slow degradation stage data to construct a model, and obtain a health state prediction model corresponding to the abrupt degradation type and a health state prediction model corresponding to the slow degradation type.

[0093] In a specific application scenario, the training module 305 is configured to obtain an MBR migration model corresponding to the rapid degradation type, an LP migration model corresponding to the rapid degradation type, an MBR migration model corresponding to the slow degradation type, and an LP migration model corresponding to the slow degradation type; input the rapid degradation stage data into the MBR model corresponding to the rapid degradation type to obtain rapid degradation features, and input the slow degradation stage data into the MBR model corresponding to the slow degradation type to obtain slow degradation features; input the rapid degradation features into the LP model corresponding to the rapid degradation type to obtain a health status label of the rapid degradation type, and input the slow degradation features into the LP model corresponding to the slow degradation type to obtain a health status label of the slow degradation type; calculate a first loss function by using the health status label of the rapid degradation type, and calculate a second loss function by using the health status label of the slow degradation type; obtain a back propagation algorithm, update network parameters of the MBR migration model corresponding to the rapid degradation type and the LP migration model corresponding to the rapid degradation type by using the first loss function based on the back propagation algorithm; update network parameters of the MBR migration model corresponding to the slow degradation type and the LP migration model corresponding to the slow degradation type by using the second loss function based on the back propagation algorithm; obtain a preset number of iterations, and update network parameters of the MBR migration model corresponding to the rapid degradation type, the LP migration model corresponding to the rapid degradation type, the MBR migration model corresponding to the slow degradation type, and the LP migration model corresponding to the slow degradation type according to the preset number of iterations to obtain a health status prediction model corresponding to the rapid degradation type and a health status prediction model corresponding to the slow degradation type.

[0094] In a specific application scenario, the training module 305 is configured to obtain a loss function calculation formula, calculate the health status label of the rapid degradation type by using the loss function calculation formula to obtain the first loss function, wherein,

[0095] L total =λL RUL +μL MMD

[0096]

[0097] wherein y i is a real health status label, is the health status label of the rapid degradation type, λ is a first weight parameter, μ is a second weight parameter, n is a sampling number at a degradation stage, D S is a source domain feature, D T is a target domain feature, n Sn is the sample size of the source domain d S n is the sample size of the source domain d T n is the sample size of the source domain d T n is the sample size of the source domain d H n is the sample size of the source domain d n is the sample size of the source domain d

[0098] In a specific application scenario, the training module 305 is configured to process the test data set by using the bidirectional long short-term memory convolutional neural network model and the Pearson correlation coefficient to obtain a second health indicator; if it is determined that the device has degraded according to the second health indicator, the type of degradation of the device is determined, a health state prediction model corresponding to the type of degradation is obtained, specified degradation stage data is extracted from the test data set based on the degradation stage detection method and the second health indicator; the specified degradation stage data is input into the health state prediction model to obtain a health state prediction result output by the health state prediction model; a real health state result of the device is obtained from the sensor signal sample data, and a model accuracy is calculated by using the health state prediction result and the health state real result calculation model; a preset accuracy threshold is obtained, and if the model accuracy is greater than or equal to the preset accuracy threshold, the health state prediction model is stored.

[0099] The device provided by the embodiment of the application obtains sensor signal data of a target device, constructs a health index of the target device based on the sensor signal data, and detects whether the target device has degraded according to the health index; if it is determined that the target device has degraded according to the health index, a degradation type of the target device is determined, degradation stage data is extracted based on a degradation stage detection method and the health index in the sensor signal data; a health state prediction model corresponding to the degradation type is obtained, the degradation stage data is input into the health state prediction model, and a health state prediction result output by the health state prediction model is obtained. First, a reasonable health index is constructed for sensor data by using a BLCAE (bidirectional long short-term memory convolutional neural network model) and a Pearson correlation coefficient (Pearson correlation coefficient); second, an initial degradation start time of the device is determined by using an OSW-PI-ACED (overlapping sliding window-based Pettitt mutation point, inflection point extraction, and adaptive continuous exceeding detection and verification) degradation stage detection method, and degradation stage data of each data is extracted; and finally, migration is performed between the same degradation categories, so that the degradation information can be effectively shared between similar fields or tasks, the occurrence of negative migration is avoided, and the reusability of historical knowledge and information is improved. The limited data collected by the sensor is processed as necessary, the data is classified to reduce the feature distribution difference between the source domain and the target domain, and domain adaptive migration is performed between the same categories, so that the degradation information can be shared between similar fields or tasks, the occurrence of negative migration is avoided, and the degradation stage data of the device can be accurately extracted, reducing the interference of irrelevant information on the prediction of the health state of the device. Even in the case of a small sample, the degradation information can be captured from a small amount of data, the data cross-domain deviation is reduced, the precision and stability of the model are improved, and the health management of the device is realized.

[0100] It should be noted that other corresponding descriptions of the various functional units involved in the device health management device based on small sample data and adaptive migration provided by the embodiments of the present application can be referred to Figure 1 and Figures 2A-2B , and will not be described here.

[0101] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties.

[0102] The technical features of the above embodiments can be combined in any way. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the present application.

[0103] The above-described embodiments are merely illustrative for the present application and are described in more detail and specifically, but should not be construed as limiting the scope of the present application. It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all fall within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

[0104] In the exemplary embodiments, referring to Figure 4 A computer device is also provided, which comprises a bus, a processor, a memory and a communication interface, and can further comprise an input / output interface and a display device, wherein the communication between various functional units can be completed through the bus. The memory stores a computer program, and the processor is configured to execute the program stored in the memory to execute the device health management method based on small sample data and adaptive migration in the above-described embodiments.

[0105] A computer readable storage medium, which stores a computer program, the computer program is executed by a processor to implement the steps of the device health management method based on small sample data and adaptive migration.

[0106] Through the above description of the embodiments, those skilled in the art can clearly understand that the present application can be implemented by hardware, or by means of software and necessary general hardware platform. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a U disk, a mobile hard disk, etc.), and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in various embodiments of the present application.

[0107] Those skilled in the art can understand that the accompanying drawings are only a schematic diagram of a preferred embodiment, and the modules or processes in the drawings are not necessarily required for implementing the present application.

[0108] Those skilled in the art can understand that the modules in the device in the embodiments can be distributed in the device in the embodiments as described in the embodiments, or can be changed and located in one or more devices different from the embodiments. The modules in the above-described embodiments can be combined into one module, or can be further split into a plurality of sub-modules.

[0109] The above-mentioned serial numbers of the present application are only for description, and do not represent the advantages or disadvantages of the embodiments.

[0110] The above disclosure is only for several specific embodiments of the present application, but the present application is not limited thereto, and any changes that can be thought of by those skilled in the art should fall within the protection scope of the present application.

Claims

1. A device health management method based on small sample data and adaptive migration, characterized in that, The method comprises the following steps: obtaining sensor signal data of a target device, and constructing a health index of the target device based on the sensor signal data; detecting whether the target device has degraded according to the health index; if it is determined that the target device has degraded according to the health index, determining a degradation type of the target device, and extracting degradation stage data in the sensor signal data based on a degradation stage detection method and the health index; obtaining sensor signal sample data, obtaining sample data containing labels as source domain data in the sensor signal sample data, and obtaining sample data not containing labels as target domain data in the sensor signal sample data, the sensor signal sample data being device time series degradation data under different working conditions; obtaining a preset proportion, dividing the target domain data into a training data set and a test data set according to the preset proportion, and taking the source domain data and the training data set as a target training data set; obtaining a pre-constructed bidirectional long short-term memory convolutional neural network model and a Pearson correlation coefficient, processing the target training data set by using the bidirectional long short-term memory convolutional neural network model and the Pearson correlation coefficient, and obtaining a first health index; dividing the first health index into a health index corresponding to an abrupt degradation type and a health index corresponding to a slow degradation type according to the degradation type, the degradation type including the abrupt degradation type and the slow degradation type; dividing the target training data set into an abrupt degradation data set and a slow degradation data set based on the health index corresponding to the abrupt degradation type and the health index corresponding to the slow degradation type; obtaining a degradation stage detection method, extracting abrupt degradation stage data in the abrupt degradation data set by using the degradation stage detection method, and extracting slow degradation stage data in the slow degradation data set by using the degradation stage detection method; respectively constructing a model by using the abrupt degradation stage data and the slow degradation stage data, obtaining a health state prediction model corresponding to the abrupt degradation type and a health state prediction model corresponding to the slow degradation type; obtaining the health state prediction model corresponding to the degradation type, inputting the degradation stage data into the health state prediction model, and obtaining a health state prediction result output by the health state prediction model.

2. The method of claim 1, wherein, The method comprises the following steps: obtaining a preset sampling interval time, performing multiple data acquisitions on the target device by a sensor according to the preset sampling interval time, and obtaining the sensor signal data; obtaining a pre-constructed bidirectional long short-term memory convolutional neural network model, inputting the sensor signal data into the bidirectional long short-term memory convolutional neural network model, and obtaining a feature data set output by the bidirectional long short-term memory convolutional neural network model; obtaining a preset range, selecting multiple feature data in the feature data set according to the preset range, and taking an average value of the multiple feature data as reference data; Obtaining a Pearson correlation coefficient, and sequentially calculating the correlation between each feature data and the reference data according to time sequence by using the Pearson correlation coefficient to obtain the health index.

3. The method of claim 1, wherein, The degradation stage detection method and the health index extract degradation stage data from the sensor signal data, including: Obtaining a 2sigma principle, extracting a target health index in the health index by using the 2sigma principle, and determining a detection threshold by using the target health index; Obtaining an overlapping sliding window of the degradation stage detection method and a mutation point detection algorithm, sliding the overlapping sliding window above the target health index, extracting a mutation point of the target health index in a plurality of data points included in the overlapping sliding window based on the mutation point detection algorithm during the sliding process, and obtaining a plurality of mutation points corresponding to each overlapping sliding window; For each overlapping sliding window, determining a plurality of inflection points by using a plurality of data points included in the overlapping sliding window, sorting the plurality of mutation points and the plurality of inflection points according to time sequence to obtain a time sequence; Obtaining a continuous exceeding detection verification principle, extracting a plurality of target time points in the time sequence according to the continuous exceeding detection verification principle, sorting the plurality of target time points according to time sequence to obtain a sorting result; Reading a target time point ranked first in the sorting result as an initial degradation time of the target device, and extracting the degradation stage data from the sensor signal data according to the initial degradation time.

4. The method of claim 3, wherein, The method for determining a plurality of inflection points by using a plurality of data points included in the overlapping sliding window includes: For each data point, determining a target data point in the plurality of data points according to the data point, the target data point being a third data point behind the data point; Obtaining a preset threshold, calculating a slope value of the data point and the target data point; If the slope value is less than the preset threshold, the data point is taken as an inflection point; The method for calculating and processing each data point of the overlapping sliding window to obtain the plurality of inflection points.

5. The method of claim 1, wherein, The method for respectively constructing a model by using the rapid degradation stage data and the slow degradation stage data to obtain a health state prediction model corresponding to the rapid degradation type and a health state prediction model corresponding to the slow degradation type includes: Obtaining an MBR migration model corresponding to the rapid degradation type, an LP migration model corresponding to the rapid degradation type, an MBR migration model corresponding to the slow degradation type, and an LP migration model corresponding to the slow degradation type; Inputting the rapid degradation stage data into the MBR model corresponding to the rapid degradation type to obtain rapid degradation features, and inputting the slow degradation stage data into the MBR model corresponding to the slow degradation type to obtain slow degradation features; Inputting the rapid degradation features into the LP model corresponding to the rapid degradation type to obtain a health state label of the rapid degradation type, and inputting the slow degradation features into the LP model corresponding to the slow degradation type to obtain a health state label of the slow degradation type; and A first loss function is calculated using the health state label of the rapid degradation type, and a second loss function is calculated using the health state label of the slow degradation type; An error back propagation algorithm is obtained, and network parameters of the MBR migration model corresponding to the rapid degradation type and the LP migration model corresponding to the rapid degradation type are updated using the first loss function based on the error back propagation algorithm; Network parameters of the MBR migration model corresponding to the slow degradation type and the LP migration model corresponding to the slow degradation type are updated using the second loss function based on the error back propagation algorithm; A preset number of iterations is obtained, and network parameters of the MBR migration model corresponding to the rapid degradation type, the LP migration model corresponding to the rapid degradation type, the MBR migration model corresponding to the slow degradation type, and the LP migration model corresponding to the slow degradation type are updated according to the preset number of iterations to obtain a health state prediction model corresponding to the rapid degradation type and a health state prediction model corresponding to the slow degradation type.

6. The method of claim 5, wherein, The first loss function is calculated using the health state label of the rapid degradation type, including: An error back propagation algorithm is obtained, and network parameters of the MBR migration model corresponding to the rapid degradation type and the LP migration model corresponding to the rapid degradation type are updated using the first loss function based on the error back propagation algorithm; wherein, is a true health state label, is a health state label of the rapid degradation type, is a first weight parameter, is a second weight parameter, n is a number of samples at a degradation stage, is a source domain feature, is a target domain feature, is a source domain sample size, is a target domain sample size, is a reproducing kernel Hilbert space, is a computational function mapping data to the reproducing kernel Hilbert space.

7. The method of claim 1, wherein, The method further includes: The test data set is processed using the bidirectional long short-term memory convolutional neural network model and the Pearson correlation coefficient to obtain a second health indicator; If it is determined that the device has degraded according to the second health indicator, the degradation type of the device is determined, and specified degradation stage data is extracted from the test data set based on the degradation stage detection method and the second health indicator; The health state prediction model corresponding to the degradation type is obtained, and the specified degradation stage data is input into the health state prediction model to obtain a health state prediction result output by the health state prediction model; The health state real result of the device is obtained in the sensor signal sample data, and the model accuracy is calculated using the health state prediction result and the health state real result; A preset accuracy threshold is obtained, and if the model accuracy is greater than or equal to the preset accuracy threshold, the health state prediction model is stored.

8. A device health management apparatus based on small sample data and adaptive migration, characterized by, It includes: A construction module is configured to obtain sensor signal data of a target device and construct a health indicator of the target device based on the sensor signal data; A detection module is configured to detect whether the target device has degraded according to the health indicator; An extraction module is configured to determine the degradation type of the target device if it is determined that the target device has degraded according to the health indicator, and extract degradation stage data from the sensor signal data based on a degradation stage detection method and the health indicator. The training module is configured to acquire sensor signal sample data, acquire sample data containing labels as source domain data in the sensor signal sample data, acquire sample data not containing labels as target domain data in the sensor signal sample data, and the sensor signal sample data is equipment time sequence degradation data under different working conditions; acquire a preset proportion, divide the target domain data into a training data set and a test data set according to the preset proportion, and take the source domain data and the training data set as a target training data set; acquire a pre-constructed bidirectional long short-term memory convolutional neural network model and a Pearson correlation coefficient, process the target training data set by using the bidirectional long short-term memory convolutional neural network model and the Pearson correlation coefficient, and obtain a first health index; divide the first health index into a health index corresponding to an abrupt degradation type and a health index corresponding to a slow degradation type according to degradation types, the degradation types include the abrupt degradation type and the slow degradation type, divide the target training data set into an abrupt degradation data set and a slow degradation data set based on the health index corresponding to the abrupt degradation type and the health index corresponding to the slow degradation type, acquire a degradation stage detection method, extract abrupt degradation stage data in the abrupt degradation data set by using the degradation stage detection method, and extract slow degradation stage data in the slow degradation data set by using the degradation stage detection method; respectively use the abrupt degradation stage data and the slow degradation stage data to construct a model, and obtain a health state prediction model corresponding to the abrupt degradation type and a health state prediction model corresponding to the slow degradation type; The prediction module is configured to acquire the health state prediction model corresponding to the degradation type, input the degradation stage data into the health state prediction model, and obtain a health state prediction result output by the health state prediction model.

Images