WO2023227071A1 - 一种多模型融合的航空电子产品健康评估方法 - Google Patents

一种多模型融合的航空电子产品健康评估方法 Download PDF

Info

Publication number
WO2023227071A1
WO2023227071A1 PCT/CN2023/096340 CN2023096340W WO2023227071A1 WO 2023227071 A1 WO2023227071 A1 WO 2023227071A1 CN 2023096340 W CN2023096340 W CN 2023096340W WO 2023227071 A1 WO2023227071 A1 WO 2023227071A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
model
module
test
subsystem
Prior art date
Application number
PCT/CN2023/096340
Other languages
English (en)
French (fr)
Inventor
文佳
赵晓虎
阎德劲
钱东
梁天辰
周榜兰
Original Assignee
中国电子科技集团公司第十研究所
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from CN202210922051.4A external-priority patent/CN114969990B/zh
Application filed by 中国电子科技集团公司第十研究所 filed Critical 中国电子科技集团公司第十研究所
Publication of WO2023227071A1 publication Critical patent/WO2023227071A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/10—Geometric CAD
    • G06F30/15—Vehicle, aircraft or watercraft design
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00—Computer-aided design [CAD]
    • G06F30/20—Design optimisation, verification or simulation
    • G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G06N20/20—Ensemble learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2119/00—Details relating to the type or aim of the analysis or the optimisation
    • G06F2119/02—Reliability analysis or reliability optimisation; Failure analysis, e.g. worst case scenario performance, failure mode and effects analysis [FMEA]

Definitions

  • the invention relates to the technical field of health assessment, in particular to a multi-model fusion avionics product health assessment method, which is used for the development of intelligent diagnostic system software for modern highly integrated modular avionics products, and provides technical support for system condition-based maintenance.
  • avionics systems play an increasingly important role in aircraft platforms.
  • the system functions are more complex and the hardware scale is larger.
  • a single product failure has a huge impact on the aircraft platform.
  • the potential impact on flight safety and mission reliability is even greater.
  • the highly integrated design of the avionics system has increased the complexity of the cross-linking of digital and radio frequency signals within the system.
  • the types of digital buses have expanded from the original single 1553B bus to RapidIO, CAN bus, 100/1000/10G Ethernet, etc.
  • the type and frequency range of radio frequency signals have more than doubled.
  • System mode switching and functional reconstruction have resulted in complex routing switching of digital and radio frequency signals, which has increased the uncertainty of fault propagation and superimposed the randomness and intermittent characteristics of electronic product faults. This makes system fault identification and health assessment more difficult.
  • the aircraft health management domain is structurally divided according to functions such as condition monitoring, fault diagnosis, trend analysis, fault prediction, display/recording, maintenance guidance, etc.
  • Highly integrated airborne systems generally adopt a multi-module integrated research and development approach.
  • the functions of a health management domain may be mapped to multiple different module contractors, and there is a many-to-many relationship between the functions and the contractors, making it more difficult to divide the contractor's work interface and process collaboration. .
  • Health assessment and prediction is to collect various data information of equipment by using advanced sensor means, with the help of various algorithms (such as Fourier transform, Kalman filter, etc.) and intelligent models (such as physical models, neural networks, expert systems), etc.
  • algorithms such as Fourier transform, Kalman filter, etc.
  • intelligent models such as physical models, neural networks, expert systems
  • the prediction method based on the failure physical model is to understand the fault mechanism of the product and the accurate parameters of product degradation by conducting failure physical experiments or simulations, and on this basis, build a failure physical model of the product. type to meet the needs of fault assessment and prediction.
  • This method has achieved certain results in laboratory verification, there are the following problems when applied to avionics products: (1) The failure physical model requires a sufficient and in-depth understanding of the product's failure mechanism, and the modeling process is often Independently personalized and difficult to inherit. At present, there is no particularly good universal model for avionics products, and the cost of customizing failure physical models is high and difficult to be widely used in actual projects. (2) During the operation of avionics products, they are affected by various factors such as humidity, heat, vibration, and complex electromagnetic environment. The failure mechanisms are complex, the boundaries of the failure physical model are difficult to determine, and the prediction results in engineering practice are inaccurate.
  • the advantage of the data-driven health assessment method is that it does not require in-depth study of fault mechanisms and the construction of accurate failure models. Especially for complex systems such as aerospace and aerospace, it is difficult to construct failure physical models that characterize the performance degradation and remaining life of electronic products, and these products However, it has a large amount of available condition monitoring information and test data. Therefore, the data-driven method has received widespread attention from NASA, many research institutions and enterprises.
  • the avionics system adopts a multi-level architecture, and models at different levels such as the module level, functional level, and subsystem level face model differences caused by differences in complex electromagnetic, mechanical, and environmental stresses and usage patterns of different aircraft platforms; 2)
  • the avionics products of the same aircraft platform are subject to model differences caused by differentiated use environments in different regions, such as freezing and extremely cold environments on plateaus, large temperature differences between day and night in deserts, and high-humidity and salt-spray environments on ocean islands and reefs.
  • a single model is difficult to meet the requirements when facing the above-mentioned multi-level and multi-scenario complex application conditions. It is necessary to propose a multi-model fusion avionics product health assessment method to improve algorithm stability and diagnostic prediction accuracy.
  • the present invention proposes a method with good stability and generalization. It is an avionics product health assessment method based on multi-model fusion that can be applied to widely different scenarios, has high trend fitting and accuracy, and can meet the needs of situation-based maintenance.
  • the relevant data are parameters that characterize the health status of the avionics product; the parameters include at least one of the following: operating voltage value, current value, temperature value, loading status, clock lock flag, and signal amplitude of the avionics product;
  • Data preprocessing includes at least one of the following: data statistics and data optimization; the statistical values of data statistics include at least one of the following: average, median, and frequent values; the methods used in data optimization include at least one of the following: singular value elimination , missing value filling, data smoothing, data dimensionality reduction.
  • the plurality of base models include at least a machine learning model, a tasteless particle filter model, and a stochastic process model; wherein the machine learning model includes at least one of the following: support vector machine, long short-term memory neural network, and deep belief network.
  • the training rules of the support vector machine include: the training set is composed of training data and labels.
  • the training data matrix is [m, n], m represents the number of samples, and n represents the characteristics of each sample;
  • the label matrix is [m, 1], indicating the classification label value corresponding to m samples;
  • the training sample size is 70% of the total sample size;
  • the test set composition the data structure is consistent with the training set, and the test sample size is 30% of the total sample size ;
  • Model hyperparameters include at least one of the following: kernel function selection, penalty factor, and kernel function kernel coefficient.
  • the training rules of the deep belief network include: training set: composed of training samples and prediction labels.
  • the number of features of the training samples is determined by the number of channels for collecting data and the number of samples required for each channel; training The data matrix is [m, n] and the label matrix is [m, 1], which represents the predicted label value of m samples; the training sample size is 70% of the total sample size; test set: the test data matrix is [o, 1] , represents o data after the training sample; the test sample size is 30% of the total sample size;
  • the model structure parameters include at least one of the following: the number of network layers, attributes of each layer, and the number of neurons in each layer; the model hyperparameters include at least the following One: training times, batch training size, learning rate and activation function.
  • the training of the tasteless particle filter model and the stochastic process model is to determine the prediction model and prior distribution from historical data, continuously update the weights based on the samples, and obtain the posterior distribution of the state, thereby completing the update of the prediction model.
  • multiple base models are trained, including:
  • the model uses the first data as a training sample to obtain the model's output value for the training sample;
  • the input training samples are looped until the error converges or the maximum number of loops is reached.
  • quantitative measurements are divided into two categories: health assessment and trend prediction; wherein the quantitative measurement parameters of health assessment include at least one of the following: accuracy rate and precision rate; wherein accuracy rate accounts for the total of all correct classification results. Ratio; precision rate is the proportion of correctly identified as a certain category to all identified as this category;
  • the quantitative measurement parameters of trend prediction use at least one of the following: root mean square error, mean absolute error and correlation coefficient; where the root mean square error is the square root of the ratio of the sum of squares of the deviations between the observed value and the true value and the number of observations; the mean absolute error Calculate the average value of the absolute value of the deviation between the observed value and the true value; the calculation formula of the correlation coefficient is:
  • X is the observed value
  • Y is the true value
  • Cov(X, Y) is the covariance of X and Y
  • the Adaboosting algorithm is used to optimize the combination strategy, and a linear error function is used to iterate each base model to obtain a series of weight coefficients of weak models;
  • the number of base models is T; where xi is the training data of the i-th sample, yi is the label of the i-th sample, and N is the number of training samples.
  • the acquisition process of the integrated model is:
  • w 1, i is the error weight of the i-th sample on the final error
  • w 1, N is the error weight of the N-th sample on the final error
  • h t ( xi ) is the calculation result of the t-th basic model on the i-th sample
  • w t, i is the influence weight of the i-th sample of the t-th basic model on the final error
  • w t+1, N is the new error weight of the Nth sample
  • K represents the median of all base model outputs.
  • collecting relevant data of avionics products includes: arranging the data processing and distribution layer above the data collection and preprocessing layer; wherein, the data processing and distribution layer is connected with the collection and preprocessing layer, the data transmission layer,
  • the display control and storage layers together constitute the aircraft comprehensive status monitoring and diagnosis system, and the data processing and distribution layer is the airborne health management domain;
  • the data acquisition and preprocessing layer collects, preprocesses and packages the status monitoring data of the airborne system. Form a status monitoring data package and upload the status monitoring data package to the airborne health management domain through the data transmission layer.
  • the method further includes: dividing the status of the airborne health management domain into four types: power-on self-check status, periodic self-check status, maintenance self-check status and fault status; if the power-on initialization is successful, then Transition from power-on self-test state to periodic self-test state; if power-on initialization fails, transition from power-on self-test state to fault state; in periodic self-test state, if a fatal fault occurs, transition to fault state, if After receiving the maintenance self-test instruction, it enters the maintenance self-test state; in the fault state, if it receives the maintenance self-test instruction, it enters the maintenance self-test state; in the maintenance self-test state, if the maintenance self-test fails, then Transition to fault state, if an instruction to exit maintenance mode is received, then Transition to periodic self-test state.
  • the method further includes: dividing the external interface messages of the airborne health management domain into interface messages between the data collection and preprocessing layer and the airborne health management domain, and the interface messages between the airborne health management domain and the display control layer. and interface messages between storage layers.
  • the method further includes: dividing the airborne health management domain into four levels: module level, functional thread level, subsystem level, and system level, and setting tasks, input and output information for each level.
  • the module-level tasks are test coordination management and control between multi-channel circuit units within the module, failure time stress analysis and module health assessment; module-level inputs include at least one of the following: module model update instructions and module startup Self-test instructions; module-level output includes at least one of the following: module work and environmental stress monitoring parameters, module health assessment and diagnosis results; functional thread-level tasks are multi-module test collaborative management and control, inter-module fault correlation analysis and functional health status Assessment; inputs at the functional thread level include at least one of the following: module work and environmental stress monitoring parameters output at the module level, module health assessment and diagnosis results, and functional thread model update instructions and function startup self-test instructions input at the subsystem level; The output at the functional thread level includes at least one of the following: functional thread status monitoring parameters, functional health assessment and diagnosis results; the subsystem level tasks are multi-thread test collaborative management and control, multi-thread fault correlation analysis and subsystem remaining capacity assessment; subsystem level
  • the input of the level includes at least one of the following: functional thread status monitoring parameters, functional health assessment and diagnosis
  • the method further includes: establishing a health assessment and diagnosis model of the airborne health management domain based on the constructed data management unit, diagnosis model, health assessment unit, enhanced diagnosis unit, fault prediction unit and diagnosis process management unit. .
  • the data management unit responds to the key information, updates the local cache, maps the external input information to the diagnostic model, and completes the conversion between the external input data and the diagnostic model;
  • the key information includes externally input fault reports, tests Data packets, configuration messages, consumables, and status parameters;
  • the diagnostic model manages prior knowledge related to system diagnostic status;
  • the health assessment unit performs functional thread and module anomaly detection and system-level remaining capacity assessment;
  • the enhanced diagnostic unit adopts the diagnostic model
  • a relatively independent universal diagnostic inference engine performs fault traceability, fault confirmation and fault correlation analysis;
  • the fault prediction unit adopts a prediction method based on feature trends to collect data and parameter degradation for products or components with obvious degradation characteristics and findable fault patterns.
  • the diagnosis process management unit collaboratively manages the model input data set and the health assessment, enhanced diagnosis and fault prediction processes, and transmits the fault or return status of the functional threads and modules output by the health assessment unit to the enhanced diagnosis unit or fault prediction unit to eliminate associated faults and match the mode to predict the time of fault occurrence; the diagnosis process management unit will feed back the output results of the enhanced diagnosis unit and fault prediction unit to the health assessment unit to provide input for the assessment of the remaining capacity of the airborne system.
  • the method further includes: dividing the database table of the airborne health management domain into a system table, a cross-subsystem diagnosis result table, a subsystem table, a software fault report table, a network node status table, a function table, and a function table.
  • the system table includes at least one of the following: system identification, cross-subsystem diagnosis results and subsystem health status summary information
  • the cross-subsystem diagnosis result table includes at least one of the following: diagnosis result identification, diagnosis time, fault isolation result and functional evaluation result information
  • the subsystem table includes at least one of the following: subsystem identification, subsystem health status details, and associated functions.
  • the software failure report table includes at least one of the following: software identification, failure time, failure type, class identification, processor node identification;
  • the network node status table includes at least one of the following: network identification , collection time, number of nodes, node identification, node status;
  • the function table includes at least one of the following: function identification, function health status, belonging module identification, function self-test results, function operating parameter information;
  • the function BIT result table includes at least the following One: function identification, collection time, number of test points, test point identification or ID, test point status;
  • the function operation parameter table includes at least one of the following: function identification, collection time, number of parameters, parameter identification or ID, parameter value ;
  • the module table includes at least one of the following: module identification, module health status, module BIT results and module working parameters;
  • the module BIT result table includes at least one of the following: module identification, collection time, number of test points, test point identification, test points Status;
  • the module working parameter table includes at least one of the following: module identification, collection
  • the method further includes: the system table is associated with the cross-subsystem diagnosis result table through the diagnosis result identification, and is associated with the subsystem table through the belonging subsystem identification; the subsystem table is associated with the software fault report table through the software identification, It is associated with the network node status table through the network identifier, and is associated with the function table through the function identifier it belongs to; the function table is associated with the function BIT result table and function operation parameter table through the function identifier, and is associated with the module table through the module identifier it belongs to; the module table is associated with the module identifier It is associated with the module BIT result table and module operating parameter table, and is associated with the static BIT configuration table through the test point identification.
  • the present invention has the following advantages:
  • the multi-model fusion avionics product health assessment method designed by the present invention uses the Adaboosting algorithm to integrate multiple base models, improves the stability and accuracy of prediction, and effectively solves the problem of small sample conditions in several life cycles.
  • This paper combines the needs of modern avionics system intelligent scheduling management and autonomous maintenance guarantee, and proposes a complete prediction model framework of the multi-model fusion Adaboosting algorithm for the problem of avionics product life prediction, which solves the problem of temporal information memory for fault prediction. Problems with model overfitting under small samples. Through simulation experiments, it is proved that compared with the current shallow learning model and the classic LSTM model, this method has better stability, trend fitting degree and prediction accuracy, and can be used for data-based The application of driven methods in avionics product health assessment, prediction and management provides reference.
  • the method of the present invention adopts a loosely coupled, hierarchical, and modular composition architecture, which can ensure the independence between different levels and different object models, facilitate design changes and update localization, support independent insertion of new technologies, and reduce The impact of technological updates or degradation on subsystems.
  • the object-oriented health management database design has good scalability and comprehensively records the status monitoring and diagnosis data of various types of objects in the airborne system, which can be used for further intermittent fault analysis and complex fault diagnosis off-board.
  • the method provided by the present invention can be used in the design and development of aircraft comprehensive condition monitoring and diagnosis systems, and has good economic benefits.
  • Figure 1 is a schematic flowchart of a multi-model fusion avionics product health assessment process according to an embodiment of the present invention
  • Figure 2 is a schematic diagram of an optional multi-model fusion avionics product health assessment process according to an embodiment of the present invention
  • Figure 4 is a schematic diagram comparing a prediction result and the real degradation process according to an embodiment of the present invention.
  • Figure 5 is a schematic diagram comparing the prediction results of an LSTM method and the real degradation process according to the embodiment of the present invention
  • Figure 6 is a schematic diagram comparing the prediction results of a BP neural network method and the real degradation process according to the embodiment of the present invention.
  • Figure 8 is a schematic diagram of state transition of the airborne health management domain according to the embodiment of the present invention.
  • Figure 9 is a schematic diagram of the airborne health management domain interface messages according to the embodiment of the present invention.
  • Figure 10 is a schematic diagram of the hierarchical division of airborne health management domains according to an embodiment of the present invention.
  • Figure 11 is a schematic structural diagram of a general diagnostic model in the airborne health management domain according to an embodiment of the present invention.
  • Figure 12 is a schematic diagram of the relationship between the airborne health management domain data tables according to the embodiment of the present invention.
  • the multi-model fusion avionics product health assessment process includes:
  • Step 01 Collect relevant data of avionics products
  • Step 02 Perform data preprocessing on relevant data, where the preprocessed data includes first data and second data;
  • Step 03 Based on the first data, train multiple base models
  • Step 04 Quantitatively measure and fuse multiple base models to obtain an integrated model
  • Step 05 Use the second data as a test sample and input it into the integrated model to obtain the health assessment results of avionics products.
  • the multi-model fusion avionics product health assessment process is divided into 5 steps: 1) data preprocessing, including data statistics and data optimization; 2) training rule management, including Deep learning model and shallow learning model training rule management; 3) model training; 4) model quantitative measurement, including health assessment and trend prediction quantitative measurement parameters; 5) model group fusion, separately for health assessment model group and trend prediction model group Perform fusion.
  • Data preprocessing Due to the complex external stresses in the airborne environment, there are strong noise interference, abnormal data values, data loss, etc. in the data collected by the sensors, which seriously affects the accuracy of aircraft health assessment and prediction. Therefore, it is very necessary to carry out massive data analysis. preprocessing.
  • Data preprocessing includes two parts: data statistics and data optimization: 1) Data statistics is to analyze the statistical rules of the data and manually extract static features to characterize the status of the monitored object.
  • Data optimization is mainly aimed at the collection anomalies in actual engineering, such as the existence of missing values and singular values, too high data dimensions, excessive data fluctuations, etc., and uses methods such as singular value elimination, missing value filling, data smoothing, and data dimensionality reduction to optimize the data. Process to eliminate the impact of data problems on model training.
  • Training rules management Users of intelligent diagnostic systems are usually ground maintenance personnel and aircraft operators who do not have relevant professional knowledge in intelligent diagnostic model rule management. Therefore, it is necessary to black-box the intelligent diagnostic model, summarize the training rules, and provide detailed descriptions of all interfaces.
  • the training rules are a summary of the structure definition of the training model data set, the intrinsic mechanism characteristics of the model, hyperparameters and structural parameter settings.
  • the system provides file import and manual input interfaces.
  • the shallow learning model takes support vector machine as an example. Its training rules include: 1) Training set composition: It consists of training data and labels. The training data matrix is [m, n], m represents the number of samples, and n represents the number of each sample. feature. The label matrix is [m, 1], which represents the classification label values corresponding to m samples. The training sample size usually defaults to 70% of the total sample size. 2) Test set composition: The data structure is consistent with the training set, and the test sample size defaults to 30% of the total sample size. 3) Model hyperparameters include: kernel function selection, penalty factor, kernel function kernel coefficient and other hyperparameters.
  • the deep learning model takes the deep belief network as an example.
  • Its training rules include: 1) Training set: It consists of training samples and prediction labels. The number of features of the training sample is the number of channels for collecting data and the number of samples required for each channel. decided.
  • the training data matrix is [m, n], and the label matrix is [m, 1], which represents the predicted label value of m samples.
  • the training sample size usually defaults to 70% of the total sample size.
  • Test set The test data matrix is [o, 1], which represents o data after the training sample. The test sample size defaults to 30% of the total sample size.
  • Model structure parameters include: number of network layers, attributes of each layer, and number of neurons in each layer.
  • Model hyperparameters include: training times, batch training size, learning rate, activation function, etc.
  • Model training The machine learning model (SVM, LSTM and DBN) model training process is as follows:
  • Step 1 Complete the model structure initialization according to the model structure parameters and hyperparameters in the training rules.
  • Step 3 Reversely correct the error between the label and output value of the model sample according to the gradient descent algorithm, and update the internal weights and biases of the model.
  • Step 4 Cyclically input training samples until the error converges or the maximum number of cycles is reached.
  • Unscented particle filtering and stochastic process model training determine the prediction model and prior distribution from historical data, continuously update the weights based on samples, and obtain the posterior distribution of the state, thereby completing the update of the prediction model.
  • Model quantification measures Model quantitative measurement is performed after model training to measure and evaluate the training effect of the model to determine whether the model meets the on-board requirements. Model metrics are divided into two categories: health assessment and trend prediction.
  • the main quantitative measurement parameters of health assessment are: (1) Accuracy rate: the total proportion of all correct classification results. (2) Precision rate: the proportion of correct identifications of a certain category among all classifications of that category.
  • Root mean square error the sum of squares of the deviation between the observed value and the true value and the square root of the ratio of the number of observations. It is very sensitive to particularly large or small errors in measurement and can better detect Reflects the precision of the measurement.
  • Model group fusion The essence of model group fusion is the process of integrating some weak learners into a strong learner. This process aims to maintain the diversity of features of each model while improving the accuracy and generalization ability of the model output results.
  • this article chooses the Adaboosting algorithm to combine strategies For optimization, a linear error function is used to iterate each base model to obtain a series of weight coefficients of weak models. If the integrated sample set The number of base models is T; where xi is the training data of the i-th sample, yi is the label of the i-th sample, and N is the number of training samples.
  • the integrated algorithm process is:
  • Step 1 Initialization. Initialize the error weight W 1 of each sample in the target scene integrated sample set. This weight is the impact of each sample on the final error synthesis.
  • the initialization formula is as follows Down:
  • Step 2 Calculate the base model error. According to the following linear loss function formula, calculate the base model output h t , the sample maximum error E t , the relative error e ti of the i-th sample, and the prediction error rate ⁇ t of the base learner in sequence;
  • h t ( xi ) is the calculation result of the t-th basic model on the i-th sample
  • w t, i is the influence weight of the i-th sample of the t-th basic model on the final error
  • W t+1 (w t+1, 1 , w t+1, 2 ...w t+1, N ) (9)
  • w t+1, N is the new error weight of the Nth sample
  • Step 5 Loop through steps 2 to 4 until the prediction error rate is 0 or the number of base models reaches T;
  • Step 6 Calculate the integrated output of the base model.
  • the calculation formula is as follows:
  • K represents the median of all base model outputs.
  • aviation batteries are the core component of aircraft power systems. They mainly provide power for key system equipment in the aircraft when the main power supply is abnormal. They have been widely used at home and abroad as emergency power supplies and auxiliary power supplies. Various aircraft platforms. When an aviation battery fails during flight, the avionics system will face the risk of power outage, which may even cause the entire aircraft to lose control with catastrophic consequences.
  • Aviation batteries mainly include nickel-cadmium batteries, lead-acid batteries, and lithium-ion batteries. As lithium-ion battery technology gradually matures, its lifespan, quality and capacity advantages gradually become apparent. It has gradually replaced nickel-cadmium batteries and lead-acid batteries as the main battery equipment. Therefore, this article uses lithium-ion aviation batteries as the research object to verify the proposed method.
  • the experimental data uses four groups of lithium-ion battery data from NASA. The experimental data are shown in Figure 3(a) to Figure 3(d).
  • the verification experiment first preprocessed 4 sets of lithium-ion battery data, obtained 4 preprocessed data sets through data statistics and data optimization, then selected the 1st to 3rd preprocessed data sets to construct a training sample set, and used the 4th preprocessed data set to construct a training sample set.
  • the data set constructs a verification sample set and a test sample set.
  • the two comparison methods set up in the verification experiment are LSTM and BP neural network.
  • the specific parameter settings are as follows:
  • Adaboosting algorithm is used to integrate 5 types of base models, including: support vector machine, long short-term memory neural network, deep belief network, tasteless particle filter model and random process model.
  • the support vector machine selects the RBF kernel function, the regularization parameter C is set to 0.5, and the Gamma parameter is set to 1;
  • the structure settings of the LSTM model are [41001], the learning rate is 0.1, and the gradient descent algorithm with momentum is used for training, and the momentum is set to 0.8 , the error convergence threshold is 0.01, the maximum number of iterations is 300;
  • the structure settings of the DBN model are [45005001], the learning rate is 0.1, the training uses the gradient descent algorithm with momentum, the momentum is set to 0.8, the error convergence threshold is 0.01, the maximum iteration
  • the number of times is 300;
  • the number of ions for tasteless example filtering is set to 10;
  • the random process uses the expectation maximization EM algorithm to estimate the Gamma process model.
  • BP Back propagation network
  • step 01, collecting relevant data of avionics products includes:
  • the data processing and distribution layer is set above the data acquisition and preprocessing layer; among them, the data processing and distribution layer, the acquisition and preprocessing layer, the data transmission layer, the display control and the storage layer together constitute the aircraft comprehensive status monitoring and diagnosis system.
  • the data processing and distribution layer is the airborne health management domain; the data collection and preprocessing layer collects, preprocesses and encapsulates the status monitoring data of the airborne system to form status monitoring data packets, and transmits the status monitoring data packets through data layer is uploaded to the airborne health management domain.
  • the airborne health management domain is in the aircraft integrated condition monitoring and diagnosis system.
  • the biggest difference between the aircraft comprehensive condition monitoring and diagnosis system and the traditional BIT system is that from a professional focus, a data processing and distribution layer is designed on top of the member-level health data collection and pre-processing layer, that is, the airborne health management domain.
  • the airborne health management domain integrates various intelligent reasoning algorithms and various health assessment and diagnosis models, maximizes the use of various status monitoring data collected at the bottom layer, improves system health perception and diagnosis capabilities, and supports sensor intelligent dispatch management and autonomous guarantee.
  • the data transmission part includes two types: one is member-level internal health data transmission, and the transmission methods include CAN bus, RS485 bus, 100M network, Gigabit network, etc.
  • CAN bus is used for health data transmission of modules in the rack
  • RS485 bus and 100M network are used for health data transmission of independent devices, antenna interface units or antennas
  • Gigabit network is used for health data transmission of display and control computer modules.
  • the other is used for large-scale health data transmission between each member subsystem and the system health management domain, health display and control, and storage servers. It mainly uses optical fiber 10G network.
  • the airborne health management domain has an upward interface relationship with the display, control and storage parts through data transmission.
  • the display and control unit provides corresponding health information display according to different use objects such as air crew, ground crew and maintenance. and control interface to complete function, module, bus network health status information display and maintenance self-check process control functions.
  • the storage unit uses a database to store health status information at each level.
  • the multi-model fusion avionics product health assessment method also includes:
  • the power-on self-check state is a self-check state that the airborne health management domain automatically enters after the aircraft is powered on. In this state, the airborne system performs module power-on self-check, bus network node status detection and functional thread Link resource self-test.
  • the periodic self-check state is a periodic self-check state that the airborne health management domain automatically enters after successful initialization. In this state, the airborne system periodically checks the machine without interrupting the normal operation of the system itself.
  • the maintenance self-check state is an in-depth self-check state performed in maintenance mode.
  • the airborne health management domain has all resource control permissions and has complete fault detection and isolation capabilities. It can target individual or Multiple functions and modules start the maintenance self-test process, set self-test parameters, and query and set fault thresholds.
  • the system fault state is a state entered after the airborne system fails to initialize or a fatal failure occurs. In this state, it can be converted to the maintenance self-test state through maintenance self-test. If the maintenance self-test fails, it will return to the system fault state.
  • the power-on initialization If the power-on initialization is successful, it will transition from the power-on self-test state to the periodic self-test state; if the power-on initialization fails, it will transition from the power-on self-test state to the fault state;
  • the multi-model fusion avionics product health assessment method also includes:
  • the external interface messages of the airborne health management domain are divided into interface messages between the data collection and preprocessing layer and the airborne health management domain, and the interfaces between the airborne health management domain and the display control and storage layers. information.
  • the interface messages between the data collection and preprocessing layer and the airborne health management domain are subdivided into subsystem health management requests, subsystem power-on self-test/startup self-test result summary information, subsystem power-on self-test results
  • subsystem health management requests There are eight message types: detailed information on inspection/startup self-test results, summary information on subsystem periodic self-test results, detailed information on subsystem periodic results, subsystem configuration data report, subsystem static BIT data report and subsystem software fault report.
  • the multi-model fusion avionics product health assessment method also includes:
  • the airborne health management domain adopts a loosely coupled, hierarchical, and modular open model architecture, and adopts the following design principles:
  • the hierarchical structure division of the airborne health management domain should be consistent with the physical or logical structure division of the system itself, divided into four levels: module level, functional thread level, subsystem level, and system level, and set the Task, input and output information.
  • the module-level tasks are test coordination management and control between multi-channel circuit units within the module, failure time stress analysis and module health assessment;
  • Module-level input includes module model update instructions and module startup self-test instructions
  • Module-level output includes module work and environmental stress monitoring parameters, module health assessment and diagnosis results
  • Functional thread-level tasks include multi-module test collaborative management and control, inter-module fault correlation analysis, and functional health status assessment;
  • Inputs at the functional thread level include module work and environmental stress monitoring parameters, module health assessment and diagnosis results output at the module level, and functional thread model update instructions and function startup self-test instructions input at the subsystem level;
  • Functional thread-level output includes functional thread status monitoring parameters, functional health assessment and diagnosis results
  • the tasks at the subsystem level are multi-threaded test collaborative management and control, multi-thread fault correlation analysis and subsystem remaining capacity assessment;
  • Inputs at the subsystem level include functional thread status monitoring parameters, functional health assessment and diagnosis result information output by the functional thread level, as well as subsystem model update instructions and subsystem startup self-test instructions input by the system level;
  • Subsystem-level output includes subsystem status monitoring parameters, software fault reports, and subsystem health assessment and diagnosis results
  • System-level output includes system health status summary information and system health status detailed information.
  • the multi-model fusion avionics product health assessment method also includes:
  • the data management unit is a converter between external input data and diagnostic models. It responds to input fault reports, test data packages, configuration messages, consumables, and status parameter messages, updates the local cache based on these messages, and maps the messages to The diagnostic model reacts to parameter changes and fault indications by triggering updates to the diagnostic model.
  • the data management unit also manages test requests for subsystems and establishes a mapping between data requests and test requests.
  • the diagnostic model is defined as a database structure object or a node containing related programs, including a priori knowledge related to diagnosis, such as parameter status, fault duration, prerequisites and node associations of the model, etc.
  • the diagnostic model should be able to reflect the randomness and intermittent characteristics of faults, as well as the uncertain behavior of faults propagating horizontally and vertically in complex system hierarchical models.
  • the health assessment unit implements anomaly detection and status assessment of functional threads and modules, as well as system-level residual capacity assessment. For the functional threads and modules of the airborne system, the health assessment unit identifies the failure mode of the functional thread or module through multi-parameter comprehensive anomaly detection and failure time stress analysis, and determines whether it is a "hard fault” or some kind of work and environmental stress. Intermittent failure under conditions.
  • the system remaining capacity assessment mainly integrates the current or future health status (normal, failure or degradation) of all functional threads and modules, combined with the current working mode and resource configuration of the system or the system working mode and resource configuration required for aircraft mission requirements in the future. Evaluate the remaining capacity of the system at the current or future time.
  • the enhanced diagnosis unit adopts a universal diagnostic inference engine that is relatively independent from the diagnosis model development to realize functions such as fault source tracing, fault confirmation, and fault correlation analysis (correlated fault elimination, false alarm and missed detection identification).
  • diagnostic inference engines needs to overcome engineering application problems such as unreliable test evidence, uncertain fault propagation, and coexistence of multiple faults.
  • the diagnostic process may require additional test data, thus generating test requests for the subsystem.
  • the diagnostic process management unit implements model input data collection and health assessment, enhanced diagnosis and Collaborative management of fault prediction processes.
  • the health assessment unit outputs the health status of functional threads and modules: normal, faulty, or degraded. For the fault state, start the enhanced diagnosis process, eliminate associated faults or false alarms, and locate the cause of the fault; for the degraded state, start the fault prediction process, match the degradation mode, analyze the degradation trend, and predict the time of fault occurrence. The results of enhanced diagnosis and fault prediction are then fed back to the health assessment unit to provide input for the system's remaining capacity assessment.
  • the multi-model fusion avionics product health assessment method also includes:
  • the airborne health management domain database adopts an object-oriented design concept and uses systems, subsystems, functional threads and modules as objects to design a highly cohesive and low-coupled data structure to form a health data record table for objects at each level. , to meet the multi-level architectural characteristics of airborne systems.
  • the database table of the airborne health management domain is divided into a system table, a cross-subsystem diagnosis result table, a subsystem table, a software fault report table, a network node status table, a function table, a function BIT result table, a function operation parameter table, and a module table. , module BIT result table, module working parameter table and static BIT configuration table.
  • the system table includes system identification, cross-subsystem diagnosis results and subsystem health status summary information
  • the cross-subsystem diagnosis results table includes diagnosis result identification, diagnosis time, fault isolation results and functional evaluation result information
  • the subsystem table includes subsystem identification, subsystem health status details, functional identification, software fault reports, and network node status information;
  • the network node status table includes network identification, collection time, number of nodes, node identification, and node status;
  • the function BIT result table includes function identification, collection time, number of test points, test point identification or ID, and test point status;
  • the function operation parameter table includes function identification, collection time, number of parameters, parameter identification or ID, and parameter value;
  • the module table includes module identification, module health status, module BIT results and module working parameters
  • the module BIT result table includes module identification, collection time, number of test points, test point identification, and test point status;
  • the module working parameter table includes module identification, collection time, parameter identification, and parameter values
  • the multi-model fusion avionics product health assessment method also includes:
  • the system table is associated with the cross-subsystem diagnosis result table through the diagnosis result identifier, and is associated with the subsystem table through the subsystem identifier to which it belongs;
  • the subsystem table is associated with the software fault report table through the software identification, the network node status table through the network identification, and the function table through the associated function identification;
  • the function table is associated with the function BIT result table and the function operation parameter table through the function identifier, and is associated with the module table through the module identifier to which it belongs;
  • the module table is associated with the module BIT result table and module working parameter table through the module identification, and is associated with the static BIT configuration table through the test point identification.
  • the solution provided by the embodiment of the present invention can be applied to the technical field of health assessment.
  • relevant data of avionics products is collected; data preprocessing is performed on the relevant data, where the preprocessed data includes first data and second data; based on the first data, multiple base models are trained; Quantitative measurement and fusion of multiple base models are performed to obtain an integrated model; the second data is used as a test sample and input into the integrated model to obtain the health assessment results of avionics products.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Geometry (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Computer Hardware Design (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Optimization (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Pure & Applied Mathematics (AREA)
  • Aviation & Aerospace Engineering (AREA)
  • Automation & Control Theory (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

本发明公开了一种多模型融合的航空电子产品健康评估方法，该方法包括以下步骤：采集航空电子产品的相关数据；对相关数据进行数据预处理，其中，预处理后的数据包括第一数据和第二数据；基于第一数据，训练多个基模型；对多个基模型进行量化度量和融合，得到集成模型；将第二数据作为测试样本，输入集成模型中，得到航空电子产品的健康评估结果。本发明采用Adaboosting算法对多种基模型进行集成，提高了预测的稳定性和精度，有效解决了在几个全寿命周期这一小样本条件下模型训练的过拟合问题；具有更好的稳定性、趋势拟合度和预测精度，可以为基于数据驱动的方法在航空电子产品健康评估、预测与管理上的应用提供参考。

Description

一种多模型融合的航空电子产品健康评估方法
相关申请的交叉引用
本申请要求于2022年8月2日提交中国国家知识产权局、申请号为202210922051.4、申请名称“一种多模型融合的航空电子产品健康评估方法”的中国专利申请的优先权，以及于2022年5月25日提交中国国家知识产权局、申请号为202210572890.8、申请名称“一种多层次模型融合的机载健康管理域设计方法”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本发明涉及健康评估技术领域，特别是一种多模型融合的航空电子产品健康评估方法，用于现代高度综合模块化航空电子产品的智能诊断系统软件开发，为系统视情维修提供技术支撑。
背景技术
随着现代电子信息技术的高速发展与大规模集成电路和芯片的广泛应用，航空电子系统在飞机平台中扮演着越来越重要的角色，系统功能更加复杂、硬件规模更加庞大，单个产品故障对飞行安全与任务可靠性的潜在影响更大。一方面，航空电子系统高度综合化的设计增加了系统内部数字、射频信号交联的复杂性，数字总线种类由原来单一1553B总线扩展到RapidIO、CAN总线、百/千/万兆以太网等，射频信号种类和频段范围扩展了1倍以上，系统模式切换与功能重构导致数字、射频信号路由切换复杂，增加了故障传播的不确定性，叠加了电子产品故障随机性与间歇性的特点，使得系统故障识别与健康评估的难度加大。另一方面，飞机健康管理域按照状态监测、故障诊断、趋势分析、故障预测、显示/记录、维修指南等功能进行结构划分，而高度综合化机载系统一般采用多模块集成的研发方式，在这种情况下，一个健康管理域的功能可能映射到多个不同的模块承制商，功能与承制商之间呈现多对多的关系，承制商工作界面划分与过程协同的难度加大。此外，针对各层次的模型设计缺乏通用的诊断模型结构定义，不同能力水平的承制厂商开发出来的诊断模型质量参差不齐。
健康评估与预测是通过运用先进的传感器手段采集设备的各类数据信息，借助各种算法(如傅里叶变换、卡尔曼滤波等)和智能模型(如物理模型、神经网络、专家系统)等来监测、评估与预测产品自身的健康状态，可有效解决现代高度综合化航空电子系统BIT故障检测隔离能力受限，缺乏精准化、定量化故障退化评估与预测手段的问题，是实现航空电子系统故障重构与视情维修的关键技术。
目前，健康评估与预测方法主要分为基于故障模型和基于数据驱动的2个方向。基于失效物理模型的预测方法是通过开展失效物理实验或仿真认知产品的故障机理与产品退化的准确参数，并在基础上构建产品的失效物理模 型以满足故障评估和预测的需求。该方法虽然在实验室验证中取得了一定的效果，但应用到航空电子产品上存在以下问题：(1)失效物理模型需要对产品的故障机理有着充分深入的认知，且建模过程往往是独立个性化且难以继承。目前针对航空电子产品还没有特别好的通用模型公布，而定制失效物理模型成本较高难以广泛应用在实际工程中。(2)航空电子产品运行过程中受湿热、振动、复杂电磁环境等多种因素影响，故障机理复杂，失效物理模型构建的边界难以确定，工程实践中预测结果准确性差。
基于数据驱动的健康评估方法优势在于不需要深入研究故障机理以及构建准确的失效模型，特别是对于航空航天等复杂系统，由于表征电子产品性能退化与剩余寿命的失效物理模型难以构建，而这些产品却具备大量可用的状态监测信息和测试数据，因此，数据驱动的方法得到了美国航空航天局、众多研究机构及企业的广泛重视。
目前，数据驱动的方法包括支持向量机、长短期记忆神经网络模型、深度置信网络模型、无味粒子滤波方法、随机过程模型等方法，这些方法在应用到复杂航空电子产品时主要面临两方面的技术难题：1)航空电子系统采用多层次架构，模块级、功能级、子系统级等不同层级模型面临不同飞机平台复杂电磁、机械和环境应力差异与使用模式差别带来的模型差异性；2)相同飞机平台航空电子产品在面向不同区域差异化使用环境带来的模型差异性，例如高原冰冻极寒环境、沙漠昼夜大温差环境、海洋岛礁高湿盐雾环境等。单一模型在面向上述多层次、多场景复杂应用条件时难以满足要求，需要提出一种多模型融合的航空电子产品健康评估方法，以提高算法稳定性和诊断预测精度。
发明内容
针对多区域部署恶劣环境条件和多平台复杂工况下，航空电子产品现有健康评估方法存在的泛化性不足、评估精度差等问题，本发明提出了一种稳定性和泛化性好，能适用大差异场景，趋势拟合度和精度高，能满足视情维修需求的基于多模型融合的航空电子产品健康评估方法。
本发明公开了一种多模型融合的航空电子产品健康评估方法，包括以下步骤：
采集航空电子产品的相关数据；
对相关数据进行数据预处理，其中，预处理后的数据包括第一数据和第二数据；
基于第一数据，训练多个基模型；
对多个基模型进行量化度量和融合，得到集成模型；
将第二数据作为测试样本，输入集成模型中，得到航空电子产品的健康评估结果。
在一个实施例中，相关数据为表征航空电子产品健康状态的参数；参数至少包括以下之一：航空电子产品的工作电压值、电流值、温度值、加载状态、时钟锁定标志、信号幅值；
数据预处理至少包括以下之一：数据统计和数据优化；数据统计的统计值至少包括以下之一：平均数、中位数、频繁值；数据优化采用的方法至少包括以下之一：奇异值剔除、缺失值填充、数据平滑、数据降维。
在一个实施例中，多个基模型至少包括机器学习模型、无味粒子滤波模型和随机过程模型；其中，机器学习模型至少包括以下之一：支持向量机、长短期记忆神经网络、深度置信网络。
在一个实施例中，支持向量机的训练规则包括：训练集构成：由训练数据和标签组成，训练数据矩阵为[m，n]，m表示样本数目，n表示每个样本的特征；标签矩阵为[m，1]，表示m个样本所对应的分类标签值；训练样本量为总样本量的70％；测试集构成：数据结构与训练集一致，测试样本量为总样本量的30％；模型超参数至少包括以下之一：核函数选择、惩罚因子、核函数内核系数。
在一个实施例中，深度置信网络的训练规则包括：训练集：由训练样本及预测标签组成，训练样本的特征数是由采集数据的通道数目以及每个通道所需的样本数量决定的；训练数据矩阵为[m，n]，标签矩阵为[m，1]，表示m个样本的预测标签值；训练样本量为总样本量的70％；测试集：测试数据矩阵为[o，1]，表示训练样本之后的o个数据；测试样本量为总样本量的30％；模型结构参数至少包括以下之一：网络层数、各层属性和各层神经元数；模型超参数至少包括以下之一：训练次数、批训练大小、学习率和激活函数。
在一个实施例中，无味粒子滤波模型和随机过程模型训练是从历史数据中确定预测模型和先验分布，根据样本不断更新权重，获取状态的后验分布，从而完成更新预测模型。
在一个实施例中，基于第一数据，训练多个基模型，包括：
根据训练规则中的模型结构参数和超参数完成模型结构初始化；
将第一数据作为训练样本，模型进行前馈运算，得到模型对训练样本的输出值；
根据梯度下降算法对训练样本的标签与输出值的误差进行反向修正，更新模型内部权重和偏置；
循环输入训练样本直到误差收敛或达到最大循环次数停止。
在一个实施例中，量化度量分为健康评估和趋势预测两类；其中，健康评估的量化度量参数至少包括以下之一：准确率和精确率；其中，准确率为所有正确分类结果的占总比；精确率为正确判别为某类占全部判别为该类的比例；
趋势预测的量化度量参数至少采用以下之一：均方根误差、平均绝对误差和相关系数；其中，均方根误差为观测值与真值偏差的平方和与观测次数比值的平方根；平均绝对误差为观测值与真值偏差绝对值求平均值；相关系数的计算公式为：
其中，X为观测值，Y为真实值，Cov(X，Y)为X与Y的协方差，Var|X|为X的方差，Var|Y|为Y的方差。
在一个实施例中，在对多个基模型进行量化度量和融合，得到集成模型中：
采用Adaboosting算法对组合策略进行优化，采用线性误差函数对每个基模型进行循环迭代得到一系列弱模型的权重系数；
若集成样本集基模型数量为T；其中，xi为第i个样本的训练数据，yi为第i个样本的标签，N为训练样本数量。
在一个实施例中，集成模型的获取过程为：
初始化，对目标场景的集成样本集的误差权重W1进行初始化，初始化公式如下：
其中，w1，i为第i个样本对最终误差影响的误差权重，w1，N为第N个样本对最终误差影响的误差权重；
计算基模型误差，根据如下线性损失函数公式，依次计算基模型输出ht、样本最大误差Et、第i个样本的相对误差eti、基学习器的预测误差率εt；
Et＝max|yi-ht(xi)|，i＝1，2...N   (3)

其中，ht(xi)为第t个基模型对第i个样本的计算结果，wt，i为第t个基模型的第i个样本对最终误差影响的影响权重；
计算基模型权重系数αt，计算公式如下：
更新样本集的误差权重，依次计算泛化因子Zt，样本的新误差权重wt+1，i和样本集的新误差权重Wt+1，计算公式如下：

Wt+1＝(wt+1，1，wt+1，2…wt+1，N)    (9)
其中，wt+1，N为第N个样本的新误差权重；
循环执行步骤42至步骤44，直到预测误差率为0或基模型数量达到T；
计算基模型的集成输出，计算公式如下：
其中，K表示所有基模型输出的中位数。
在一个实施例中，采集航空电子产品的相关数据包括：将数据处理及分发层设置在数据采集及预处理层之上；其中，数据处理及分发层与采集及预处理层、数据传输层、显控及存储层共同构成飞机综合状态监测与诊断系统，且数据处理及分发层为机载健康管理域；数据采集及预处理层对机载系统的状态监测数据进行采集、预处理和封装，形成状态监测数据包，并将状态监测数据包通过数据传输层上传至机载健康管理域。
在一个实施例中，该方法还包括：将机载健康管理域的状态划分为四种：上电自检状态、周期自检状态、维护自检状态和故障状态；若上电初始化成功，则由上电自检状态转换到周期自检状态；若上电初始化失败，则由上电自检状态转换到故障状态；在周期自检状态下，若发生致命故障，则转换到故障状态，若接收到维护自检指令，则进入到维护自检状态；在故障状态下，若接收到维护自检指令，则进入到维护自检状态；在维护自检状态下，若维护自检失败，则转换到故障状态，若接收到退出维护模式指令，则 转换到周期自检状态。
在一个实施例中，该方法还包括：将机载健康管理域的外部接口消息划分为数据采集及预处理层与机载健康管理域之间的接口消息，以及机载健康管理域与显控及存储层之间的接口消息。
在一个实施例中，该方法还包括：将机载健康管理域划分为模块级、功能线程级、子系统级、系统级共四个层级，并设置每个层级的任务、输入和输出信息。
在一个实施例中，模块级的任务为模块内部多通道电路单元之间的测试协同管控、故障时间应力分析与模块健康评估；模块级的输入至少包括以下之一：模块模型更新指令和模块启动自检指令；模块级的输出至少包括以下之一：模块工作与环境应力监测参数、模块健康评估与诊断结果；功能线程级的任务为多模块测试协同管控、模块间故障关联分析与功能健康状态评估；功能线程级的输入至少包括以下之一：模块级输出的模块工作与环境应力监测参数、模块健康评估与诊断结果，以及子系统级输入的功能线程模型更新指令、功能启动自检指令；功能线程级的输出至少包括以下之一：功能线程状态监测参数、功能健康评估与诊断结果；子系统级的任务为多线程测试协同管控、多线程故障关联分析与子系统剩余能力评估；子系统级的输入至少包括以下之一：功能线程级输出的功能线程状态监测参数、功能健康评估与诊断结果信息，以及系统级输入的子系统模型更新指令、子系统启动自检指令；子系统级的输出至少包括以下之一：子系统状态监测参数、软件故障报告和子系统健康评估与诊断结果；系统级的任务为子系统间测试协同管控、跨子系统故障诊断与系统剩余能力评估；系统级的输入至少包括以下之一：子系统输出的子系统状态监测参数、软件故障报告和子系统健康评估与诊断结果；系统级的输出至少包括以下之一：系统健康状态概要信息和系统健康状态详细信息。
在一个实施例中，该方法还包括：基于构建的数据管理单元、诊断模型、健康评估单元、增强诊断单元、故障预测单元和诊断过程管理单元，建立机载健康管理域的健康评估与诊断模型。
在一个实施例中，数据管理单元对关键信息进行响应，更新本地缓存，将外部输入信息映射到诊断模型，完成外部输入数据与诊断模型之间的转换；关键信息包括外部输入的故障报告、测试数据包、配置消息、消耗品、状态参数；诊断模型对系统诊断状态相关的先验知识进行管理；健康评估单元进行功能线程和模块异常检测、系统级剩余能力评估；增强诊断单元采用与诊断模型相对独立的通用诊断推理机，进行故障溯源、故障确认以及故障关联分析；故障预测单元采用基于特征趋势的预测方法，针对退化特征明显、故障规律可寻的产品或部件，进行数据收集、参数退化趋势跟踪与预测特征提取；诊断过程管理单元对模型输入数据集合以及健康评估、增强诊断和故障预测过程进行协同管理，将健康评估单元输出的功能线程和模块的故障或退货状态传输至增强诊断单元或故障预测单元，消除关联故障，匹配退 化模式，预测故障发生时间；诊断过程管理单元将增强诊断单元和故障预测单元输出的结果再反馈至健康评估单元，为机载系统剩余能力评估提供输入。
在一个实施例中，该方法还包括：将机载健康管理域的数据库表划分为系统表、跨子系统诊断结果表、子系统表、软件故障报告表、网络节点状态表、功能表、功能BIT结果表、功能运行参数表、模块表、模块BIT结果表、模块工作参数表和静态BIT配置表；系统表至少包括以下之一：系统标识、跨子系统诊断结果和子系统健康状态概要信息；跨子系统诊断结果表至少包括以下之一：诊断结果标识、诊断时间、故障隔离结果和功能评估结果信息；子系统表至少包括以下之一：子系统标识、子系统健康状态详细信息、所属功能标识、软件故障报告以及网络节点状态信息；软件故障报告表至少包括以下之一：软件标识、故障时间、故障类型、类标识、处理器节点标识；网络节点状态表至少包括以下之一：网络标识、采集时间、节点个数、节点标识、节点状态；功能表至少包括以下之一：功能标识、功能健康状态、所属模块标识、功能自检结果、功能运行参数信息；功能BIT结果表至少包括以下之一：功能标识、采集时间、测试点数量、测试点标识或ID、测试点状态；功能运行参数表至少包括以下之一：功能标识、采集时间、参数个数、参数标识或ID、参数值；模块表至少包括以下之一：模块标识、模块健康状态、模块BIT结果和模块工作参数；模块BIT结果表至少包括以下之一：模块标识、采集时间、测试点数量、测试点标识、测试点状态；模块工作参数表至少包括以下之一：模块标识、采集时间、参数标识、参数值；静态BIT配置表至少包括以下之一：测试点数量、测试点标识、滤波类型、阈值、测试参数类型。
在一个实施例中，该方法还包括：系统表通过诊断结果标识与跨子系统诊断结果表关联，通过所属子系统标识与子系统表关联；子系统表通过软件标识与软件故障报告表关联，通过网络标识与网络节点状态表关联，通过所属功能标识与功能表关联；功能表通过功能标识与功能BIT结果表、功能运行参数表关联，通过所属模块标识与模块表关联；模块表通过模块标识与模块BIT结果表、模块工作参数表关联，通过测试点标识与静态BIT配置表关联。
由于采用了上述技术方案，本发明具有如下的优点：
(1)本发明设计的多模型融合航空电子产品健康评估方法采用Adaboosting算法对多种基模型进行集成，提高了预测的稳定性和精度，有效解决了在几个全寿命周期这一小样本条件下模型训练的过拟合问题。
(2)本文结合现代航空电子系统智能调度管理与自主维护保障的需求，针对航空电子产品寿命预测问题，提出了多模型融合的Adaboosting算法的完整预测模型框架，解决了故障预测的时序信息记忆问题与小样本下的模型过拟合问题。通过仿真实验证明相较于目前浅层的学习模型和经典LSTM模型，本文方法有着更好的稳定性、趋势拟合度和预测精度，可以为基于数据 驱动的方法在航空电子产品健康评估、预测与管理上的应用提供参考。
(3)本发明的方法采用了松耦合、层次化、模块化的组成架构，可保证不同层次、不同对象模型之间的独立性，便于设计变更和更新本地化，支持独立插入新技术，减少技术更新或退化对子系统的影响。
(4)机载健康管理域通用诊断模型结构的建立可促进模型软件开发过程的规范化和标准化，提高模型软件的代码质量和稳定性。
(5)面向对象的健康管理数据库设计具备良好的可扩展性，全面记录了机载系统各类型对象的状态监测与诊断数据，可用于机下进一步的间歇故障分析和复杂故障诊断。
(6)本发明提供的方法可用于飞机综合状态监测与诊断系统设计开发，具备良好的经济效益。
附图说明
为了更清楚地说明本发明实施例中的技术方案，下面将对实施例描述中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本发明实施例中记载的一些实施例，对于本领域普通技术人员来讲，还可以根据这些附图获得其他的附图。
图1为本发明实施例的一种多模型融合的航空电子产品健康评估流程示意图；
图2为本发明实施例的可选的一种多模型融合的航空电子产品健康评估流程示意图；
图3(a)至图3(d)分别为本发明实施例的锂离子蓄电池全寿命退化数据示意图；
图4为本发明实施例的一种预测结果与真实退化过程对比示意图；
图5为本发明实施例的一种LSTM方法预测结果与真实退化过程对比示意图；
图6为本发明实施例的一种BP神经网络方法预测结果与真实退化过程对比示意图。
图7为本发明实施例的机载健康管理域外部接口关系示意图；
图8为本发明实施例的机载健康管理域状态转换示意图；
图9为本发明实施例的机载健康管理域接口消息示意图；
图10为本发明实施例的机载健康管理域层次划分示意图；
图11为本发明实施例的机载健康管理域通用诊断模型结构示意图；
图12为本发明实施例的机载健康管理域数据表关系示意图。
具体实施方式
结合附图和实施例对本发明作进一步说明，显然，所描述的实施例仅是本发明实施例一部分实施例，而不是全部的实施例。本领域普通技术人员所获得的所有其他实施例，都应当属于本发明实施例保护的范围。
参阅图1，多模型融合的航空电子产品健康评估流程包括：
步骤01：采集航空电子产品的相关数据；
步骤02：对相关数据进行数据预处理，其中，预处理后的数据包括第一数据和第二数据；
步骤03：基于第一数据，训练多个基模型；
步骤04：对多个基模型进行量化度量和融合，得到集成模型；
步骤05：将第二数据作为测试样本，输入集成模型中，得到航空电子产品的健康评估结果。
在一种可选的实施例中，参阅图2，多模型融合的航空电子产品健康评估流程分为5个步骤：1)数据预处理，包括数据统计和数据优化；2)训练规则管理，包括深度学习模型、浅层学习模型训练规则管理；3)模型训练；4)模型量化度量，包括健康评估、趋势预测量化度量参数；5)模型组融合，分别对健康评估模型组和趋势预测模型组进行融合。
数据预处理。由于机载环境中存在复杂的外部应力，传感器采集到的数据中存在强噪声干扰、数据值异常、数据丢失等情况，严重影响了飞机健康评估与预测的精准度，因此非常有必要进行海量数据的预处理。数据预处理包括数据统计和数据优化两部分：1)数据统计是通过分析数据的统计学规律，手动提取静态特征来表征监测对象状态，统计值包括平均数、中位数、频繁值等；2)数据优化主要针对工程实际中采集异常，比如存在缺失值和奇异值、数据维度过高、数据波动过大等问题，采用奇异值剔除、缺失值填充、数据平滑、数据降维等方法对数据进行处理，消除数据问题对模型训练造成的影响。
训练规则管理。智能诊断系统的使用者通常为地面维护人员和机上操作人员，不具备智能诊断模型规则管理的相关专业知识，因此需要将智能诊断模型黑盒处理，总结归纳训练规则并对所有接口进行详细说明。训练规则是对训练模型数据集结构定义、模型内在机理特性、超参数和结构参数设置总结，系统内提供文件导入和人工输入接口。
浅层学习模型以支持向量机为例，其训练规则包括：1)训练集构成：由训练数据和标签组成，训练数据矩阵为[m，n]，m表示样本数目，n表示每个样本的特征。标签矩阵为[m，1]，表示m个样本所对应的分类标签值。训练样本量通常默认为总样本量的70％。2)测试集构成：数据结构与训练集一致，测试样本量默认为总样本量的30％。3)模型超参数包括：核函数选择、惩罚因子、核函数内核系数等超参数。
深度学习模型以深度置信网络为例，其训练规则包括了：1)训练集：由训练样本及预测标签组成，训练样本的特征数是由采集数据的通道数目以及每个通道所需的样本数量决定的。训练数据矩阵为[m，n]，标签矩阵为[m，1]，表示m个样本的预测标签值。训练样本量通常默认为总样本量的70％。2)测试集：测试数据矩阵为[o，1]，表示训练样本之后的o个数据。测试样本量默认为总样本量的30％。3)模型结构参数包括：网络层数、各层属性和各层神经元数等。4)模型超参数包括：训练次数、批训练大小、学习率和激活函数等。
模型训练。机器学习模型(SVM、LSTM和DBN)模型训练流程如下：
Step 1：根据训练规则中的模型结构参数和超参数完成模型结构初始化。
Step 2：输入一批训练模型样本，模型进行前馈运算，得到模型对该样本的输出值。
Step 3：根据梯度下降算法对该模型样本的标签与输出值的误差进行反向修正，更新模型内部权重和偏置。
Step 4：循环输入训练样本直到误差收敛或达到最大循环次数停止。
无味粒子滤波和随机过程模型训练是从历史数据中确定预测模型和先验分布，根据样本不断更新权重，获取状态的后验分布，从而完成更新预测模型。
模型量化度量。模型量化度量是在模型训练之后执行，对模型的训练效果进行度量评估，判断模型是否满足机上需求。模型度量分为健康评估和趋势预测两类。
1)健康评估量化度量参数主要有：(1)准确率：所有正确分类结果的占总比。(2)精确率：正确判别为某类占全部判别为该类的比例。
2)趋势预测量化度量参数主要采用：(1)均方根误差：观测值与真值偏差的平方和与观测次数比值的平方根，对测量中特别大或特别小误差非常敏感，能较好地反映出测量的精密度。
(2)平均绝对误差：观测值与真值偏差绝对值求平均值，该值可以避免误差相互抵消，准确反映实际误差大小。(3)相关系数：计算公式如下，用以衡量观测值与真实值问相关强度。
其中，X为观测值，Y为真实值，Cov(X，Y)为X与Y的协方差，Var|X|为X的方差，Var|Y|为Y的方差。
模型组融合。模型组融合的本质是将一些弱学习器集成为一个强学习器的过程，该过程旨在保持各模型特征多样性的同时，提高模型输出结果精度和泛化能力。针对目前常用融合算法，如加权投票法、平均法、相对投票法等，只考虑单模型性能差异来设置投票权重，无法解决不同环境下不同基模型的性能差异问题，本文选择Adaboosting算法对组合策略进行优化，采用线性误差函数对每个基模型进行循环迭代得到一系列弱模型的权重系数。若集成样本集基模型数量为T；其中，xi为第i个样本的训练数据，yi为第i个样本的标签，N为训练样本数量。集成算法流程为：
步骤1：初始化，对目标场景集成样本集中每个样本的误差权重W1进行初始化，该权重是每个样本对最终误差综合的影响大小。初始化公式如 下：
其中，w1，i为第i个样本对最终误差影响的误差权重，w1，N为第N个样本对最终误差影响的误差权重；
步骤2：计算基模型误差，根据如下线性损失函数公式，依次计算基模型输出ht、样本最大误差Et、第i个样本的相对误差eti、基学习器的预测误差率εt；
Et＝max|yi-ht(xi)|，i＝1，2...N    (3)

其中，ht(xi)为第t个基模型对第i个样本的计算结果，wt，i为第t个基模型的第i个样本对最终误差影响的影响权重；
步骤3：计算基模型权重系数αt，计算公式如下：
步骤4：更新样本集的误差权重，依次计算泛化因子Zt，样本的新误差权重wt+1，i和样本集的新误差权重Wt+1，计算公式如下：

Wt+1＝(wt+1，1，wt+1，2…wt+1，N)    (9)
其中，wt+1，N为第N个样本的新误差权重；
步骤5：循环执行步骤2至步骤4，直到预测误差率为0或基模型数量达到T；
步骤6：计算基模型的集成输出，计算公式如下：
其中，K表示所有基模型输出的中位数。
参阅图3(a)至图3(d)，航空蓄电池是飞机电源系统的核心组成，主要在主电源异常时为机内关键系统设备供电，目前已经作为应急电源和辅助电源广泛应用于国内外各类飞机平台。当航空蓄电池在飞行中发生故障时，航空电子系统将面临断电危险，甚至可能导致整个飞机失控造成灾难性后果。航空蓄电池主要包括镍镉蓄电池、铅酸蓄电池、锂离子蓄电池三类。随着锂离子蓄电池工艺技术逐渐成熟，其寿命、质量和容量优势逐渐明显，目前已经逐渐取代镍镉蓄电池、铅酸蓄电池成为蓄电池的主要器材。因此本文以锂离子航空蓄电池作为研究对象验证所提方法，实验数据采用美国航空航天局的4组锂离子蓄电池数据，实验数据展示如图3(a)至图3(d)所示。
验证实验首先对4组锂离子电池数据进行预处理，通过数据统计和数据优化得到4个预处理数据集，然后选择第1-3个预处理数据集构造训练样本集，以第4个预处理数据集构造验证样本集和测试样本集。为了验证本文所提健康评估方法对评估与预测效果的影响，验证实验设置两个对比方法分别为LSTM和BP神经网络，具体参数设置如下：
(1)本文方法：采用Adaboosting算法对5类基模型进行集成，包括：支持向量机、长短期记忆神经网络、深度置信网络、无味粒子滤波模型和随机过程模型。支持向量机选择RBF核函数，正则化参数C设置为0.5，Gamma参数设置为1；LSTM模型的结构设置分别为[41001]，学习率0.1，训练采用含动量的梯度下降算法，动量设置为0.8，误差收敛门限为0.01，最大迭代次数为300；DBN模型的结构设置分别为[45005001]，学习率0.1，训练采用含动量的梯度下降算法，动量设置为0.8，误差收敛门限为0.01，最大迭代次数为300；无味例子滤波的离子数设置为10；随机过程采用期望最大化EM算法对基于Gamma过程模型进行估计。
(2)经典LSTM：LSTM结构设置分别为LSTM结构设置为[41001]，学习率0.1，训练采用含动量的梯度下降算法，动量设置为0.8，误差收敛门限为0.01，最大迭代次数为300；
(3)反向传播网络(BP)：神经网络结构为[41001]，同样采用含动量的梯度下降算法，超参数设置与经典LSTM模型一致。
参见图4至图6，图4至图6展示了2种对比方法与本专利方法的预测效果对比，为了提升预测结果可视化效果，图中下半部为真实寿命和预测差 值的5倍。本文通过均方根误差(RMSE)、平均绝对误差(MAE)和相关度(REL)三个尺度衡量预测结果，对比分析结果如表1所示。
表1对比分析结果
在一种可选的实施方式中，步骤01，采集航空电子产品的相关数据包括：
将数据处理及分发层设置在数据采集及预处理层之上；其中，数据处理及分发层与采集及预处理层、数据传输层、显控及存储层共同构成飞机综合状态监测与诊断系统，且数据处理及分发层为机载健康管理域；数据采集及预处理层对机载系统的状态监测数据进行采集、预处理和封装，形成状态监测数据包，并将状态监测数据包通过数据传输层上传至机载健康管理域。
参阅图7，机载健康管理域处于飞机综合状态监测与诊断系统中。飞机综合状态监测与诊断系统与传统BIT系统最大的差异在于：从专业聚焦的角度，在成员级健康数据采集及预处理层之上设计了数据处理及分发层，即机载健康管理域。机载健康管理域集成各类智能推理算法及各种健康评估与诊断模型，最大化利用底层采集的各类状态监测数据，提升系统健康感知与诊断能力，支持传感器智能调度管理与自主式保障。
飞机综合状态监测与诊断系统从逻辑上可分为数据采集与预处理、数据传输、数据处理与分发、显控及存储共4个部分。机载健康管理域主要实现数据处理与分发功能，包括多个成员子系统的健康管理模型及跨子系统健康管理模型软件。健康管理域对各区域管理单元采集的模块、功能及网络的健康信息进行解析后，完成故障综合诊断与功能评估，形成故障日志记录，并按需进行健康信息分发和报告。
机载健康管理域向下通过数据传输部分和数据采集与预处理部分有接口关系。数据采集及预处理层对机载系统的状态监测数据进行采集、预处理和封装，形成状态监测数据包，并将状态监测数据包通过数据传输层上传至机载健康管理域。
数据传输部分包括两种：一种是成员级内部健康数据传输，传输手段包括CAN总线、RS485总线、百兆网、千兆网等，其中，CAN总线用于机架内模块的健康数据传输；RS485总线、百兆网用于各独立设备、天线接口单元或天线的健康数据传输；千兆网用于显控计算机模块的健康数据传输。另一种是用于各成员子系统与系统健康管理域、健康显控及存储服务器之间大规模的健康数据传输，主要采用光纤万兆网。
机载健康管理域向上通过数据传输与显控与存储部分有接口关系。其中，显控单元按空勤、地勤和维护等不同使用对象提供相应的健康信息显示 和控制界面，完成功能、模块、总线网络健康状态信息显示及维护自检过程控制功能。存储单元采用数据库完成各层次健康状态信息的存储。
在一种可选的实施方式中，参阅图8，多模型融合的航空电子产品健康评估方法还包括：
将机载健康管理域的状态划分为四种：上电自检状态、周期自检状态、维护自检状态和故障状态；
其中，上电自检状态是机载健康管理域在飞机上电启动后自动进入的一种自检状态，在该状态下机载系统进行模块上电自检、总线网络节点状态检测和功能线程链路资源自检。
其中，周期自检状态是机载健康管理域在初始化成功后自动进入的一种周期性自检状态，在该状态下机载系统在不打断系统自身正常工作的前提下，周期性检测机载总线网络节点状态、关键硬件资源及软件的运行状态；
其中，维护自检状态是在维护模式下进行的一种深度自检状态，在该状态下机载健康管理域拥有全部的资源控制权限，具备完整的故障检测与隔离能力，可分别针对单个或多个功能及模块启动维护自检过程，设置自检参数，进行故障门限的查询和设置。
其中，系统故障状态是机载系统初始化失败或发生致命故障后进入的一种状态，在该状态下可通过维护自检转换到维护自检状态，若维护自检失败则回归到系统故障状态。
若上电初始化成功，则由上电自检状态转换到周期自检状态；若上电初始化失败，则由上电自检状态转换到故障状态；
在周期自检状态下，若发生致命故障，则转换到故障状态，若接收到维护自检指令，则进入到维护自检状态；
在故障状态下，若接收到维护自检指令，则进入到维护自检状态；在维护自检状态下，若维护自检失败，则转换到故障状态，若接收到退出维护模式指令，则转换到周期自检状态。
在一种可选的实施方式中，多模型融合的航空电子产品健康评估方法还包括：
划分飞机综合状态监测与诊断系统中的接口消息；
参阅图9，将机载健康管理域的外部接口消息划分为数据采集及预处理层与机载健康管理域之间的接口消息，以及机载健康管理域与显控及存储层之间的接口消息。
更为具体地，数据采集及预处理层与机载健康管理域之间的接口消息细分为子系统健康管理请求、子系统加电自检/启动自检结果概要信息、子系统加电自检/启动自检结果详细信息、子系统周期自检结果概要信息、子系统周期结果详细信息、子系统配置数据报告、子系统静态BIT数据报告和子系统软件故障报告共八种消息类型。
更为具体地，机载健康管理域与显控及存储层之间的接口消息细分为系统健康信息查询、系统健康状态概要信息、系统健康状态详细信息、健康管 理配置参数查询、健康管理配置参数设置、重点健康对象设置共六种消息类型。
在一种可选的实施方式中，多模型融合的航空电子产品健康评估方法还包括：
对机载健康管理域进行模型层级划分；
参阅图10，为便于各类算法模型的开发与集成，机载健康管理域采用松耦合、层次化、模块化的开放式模型架构，并采取如下设计原则：
1)成员子系统的结构、行为模型数据以及接口编码等专用信息应包含在子系统模型中；
2)从子系统模型中分离出通用算法，与子系统的专用诊断知识模型独立开发；
3)子系统特有的算法包含在子系统模型中；
4)基于系统配置加载模型数据库，基于子系统硬件/软件配置报告初始化模型配置。
基于上述原则，机载健康管理域的层次结构划分应该与系统自身的物理或逻辑结构划分一致，分为模块级、功能线程级、子系统级、系统级4个层次，并设置每个层级的任务、输入和输出信息。
本实施例中，模块级的任务为模块内部多通道电路单元之间的测试协同管控、故障时间应力分析与模块健康评估；
模块级的输入包括模块模型更新指令和模块启动自检指令；
模块级的输出包括模块工作与环境应力监测参数、模块健康评估与诊断结果；
功能线程级的任务为多模块测试协同管控、模块间故障关联分析与功能健康状态评估；
功能线程级的输入包括模块级输出的模块工作与环境应力监测参数、模块健康评估与诊断结果，以及子系统级输入的功能线程模型更新指令、功能启动自检指令；
功能线程级的输出包括功能线程状态监测参数、功能健康评估与诊断结果；
子系统级的任务为多线程测试协同管控、多线程故障关联分析与子系统剩余能力评估；
子系统级的输入包括功能线程级输出的功能线程状态监测参数、功能健康评估与诊断结果信息，以及系统级输入的子系统模型更新指令、子系统启动自检指令；
子系统级的输出包括子系统状态监测参数、软件故障报告和子系统健康评估与诊断结果；
系统级的任务为子系统间测试协同管控、跨子系统故障诊断与系统剩余能力评估；
系统级的输入包括子系统输出的子系统状态监测参数、软件故障报告和 子系统健康评估与诊断结果；
系统级的输出包括系统健康状态概要信息和系统健康状态详细信息。
在一种可选的实施方式中，多模型融合的航空电子产品健康评估方法还包括：
建立机载健康管理域的健康评估与诊断模型；
参阅图11，机载健康管理域由不同层次、不同对象的健康评估与诊断模型构成，为便于模型软件的开发，需要定义通用的健康评估与诊断模型结构。健康评估与诊断模型组成包括数据管理、诊断模型、健康评估、增强诊断、故障预测及诊断过程管理单元。
数据管理单元是外部输入数据与诊断模型之间的转换器，对输入的故障报告、测试数据包、配置消息、消耗品、状态参数消息进行响应，根据这些消息更新本地缓存，并将消息映射到诊断模型，通过触发诊断模型的更新来对参数变化以及故障指示做出反应。数据管理单元同时管理对子系统的测试请求，建立数据请求到测试请求之间的映射。
诊断模型作为系统诊断状态的知识库，被定义为一个数据库结构对象或包含相关程序的节点，包括与诊断相关的先验知识，如参数状态、故障历时、先决条件与模型的节点关联等。针对机载系统，诊断模型应能体现故障的随机性与间歇性特点，以及故障在复杂系统层次结构模型中横向与纵向传播的不确定性行为。
健康评估单元实现对功能线程和模块的异常检测与状态评估，以及系统层次的剩余能力评估。针对机载系统的功能线程和模块，健康评估单元通过基于多参数综合的异常检测与故障时间应力分析，识别功能线程或模块的故障模式，同时确定是“硬故障”或某种工作与环境应力条件下的间歇故障。系统剩余能力评估主要综合所有功能线程和模块当前或未来的健康状态(正常、失效或退化)，结合系统现阶段的工作模式和资源配置或未来时刻飞机任务要求装配的系统工作模式和资源配置，评估当前或未来时刻系统的剩余能力。
增强诊断单元采用与诊断模型开发相对独立的通用诊断推理机，实现故障溯源、故障确认以及故障关联分析(关联故障消除、虚警和漏检识别)等功能。针对机载系统，诊断推理机设计需突破测试证据不可靠、故障传播不确定以及多故障共存等工程应用难题。为了减小模糊组，诊断过程可能需要额外的测试数据，从而产生对子系统的测试请求。
故障预测单元采用基于特征趋势的预测方法，针对退化特征明显、故障规律可寻的部分产品或部件开展预测，例如模拟/射频电路性能退化(由于老化、环境或部件制作工艺可能导致其性能漂移的现象)、通信错误(由于退化引起链路非随机或过多的通信信息丢失)、直流电压漂移等情况。机上故障预测主要完成数据收集、参数退化趋势跟踪与预测特征提取，剩余寿命估计在机下完成。
诊断过程管理单元实现对模型输入数据集合以及健康评估、增强诊断和 故障预测过程的协同管理。健康评估单元输出功能线程和模块的健康状态：正常、故障或退化。针对故障状态，启动增强诊断过程，消除关联故障或虚警，定位故障原因；针对退化状态，启动故障预测过程，匹配退化模式，分析退化趋势，预测故障发生时间。增强诊断和故障预测的结果再反馈到健康评估单元，为系统剩余能力评估提供输入。
在一种可选的实施方式中，多模型融合的航空电子产品健康评估方法还包括：
划分机载健康管理域的数据库表，得到多个数据表；
参阅图12，机载健康管理域数据库采用面向对象的设计理念，以系统、子系统、功能线程和模块为对象，设计高内聚、低耦合的数据结构，形成各层级对象的健康数据记录表，满足机载系统多层次的架构特点。将机载健康管理域的数据库表划分为系统表、跨子系统诊断结果表、子系统表、软件故障报告表、网络节点状态表、功能表、功能BIT结果表、功能运行参数表、模块表、模块BIT结果表、模块工作参数表和静态BIT配置表。
其中，系统表包括系统标识、跨子系统诊断结果和子系统健康状态概要信息；
跨子系统诊断结果表包括诊断结果标识、诊断时间、故障隔离结果和功能评估结果信息；
子系统表包括子系统标识、子系统健康状态详细信息、所属功能标识、软件故障报告以及网络节点状态信息；
软件故障报告表包括软件标识、故障时间、故障类型、类标识、处理器节点标识；
网络节点状态表包括网络标识、采集时间、节点个数、节点标识、节点状态；
功能表包括功能标识、功能健康状态、所属模块标识、功能自检结果、功能运行参数信息；
功能BIT结果表包括功能标识、采集时间、测试点数量、测试点标识或ID、测试点状态；
功能运行参数表包括功能标识、采集时间、参数个数、参数标识或ID、参数值；
模块表包括模块标识、模块健康状态、模块BIT结果和模块工作参数；
模块BIT结果表包括模块标识、采集时间、测试点数量、测试点标识、测试点状态；
模块工作参数表包括模块标识、采集时间、参数标识、参数值；
静态BIT配置表包括测试点数量、测试点标识、滤波类型、阈值、测试参数类型。
在一种可选的实施方式中，多模型融合的航空电子产品健康评估方法还包括：
建立数据表之间的关联关系。
系统表通过诊断结果标识与跨子系统诊断结果表关联，通过所属子系统标识与子系统表关联；
子系统表通过软件标识与软件故障报告表关联，通过网络标识与网络节点状态表关联，通过所属功能标识与功能表关联；
功能表通过功能标识与功能BIT结果表、功能运行参数表关联，通过所属模块标识与模块表关联；
模块表通过模块标识与模块BIT结果表、模块工作参数表关联，通过测试点标识与静态BIT配置表关联。
最后应当说明的是：以上实施例仅用以说明本发明的技术方案而非对其限制，尽管参照上述实施例对本发明进行了详细的说明，所属领域的普通技术人员应当理解：依然可以对本发明的具体实施方式进行修改或者等同替换，而未脱离本发明精神和范围的任何修改或者等同替换，其均应涵盖在本发明的权利要求保护范围之内。
工业实用性
本发明实施例提供的方案可应用于健康评估技术领域。在本发明实施例中，采集航空电子产品的相关数据；对相关数据进行数据预处理，其中，预处理后的数据包括第一数据和第二数据；基于第一数据，训练多个基模型；对多个基模型进行量化度量和融合，得到集成模型；将第二数据作为测试样本，输入集成模型中，得到航空电子产品的健康评估结果。通过本发明实施例，可以为基于数据驱动的方法在航空电子产品健康评估、预测与管理上的应用提供参考。

Claims (19)

  1. 一种多模型融合的航空电子产品健康评估方法，其特征在于，包括以下步骤：
    采集所述航空电子产品的相关数据；
    对所述相关数据进行数据预处理，其中，预处理后的数据包括第一数据和第二数据；
    基于所述第一数据，训练多个基模型；；
    对所述多个基模型进行量化度量和融合，得到集成模型；
    将所述第二数据作为测试样本，输入所述集成模型中，得到所述航空电子产品的健康评估结果。
  2. 根据权利要求1所述的方法，其特征在于，所述相关数据为表征航空电子产品健康状态的参数；所述参数至少包括以下之一：航空电子产品的工作电压值、电流值、温度值、加载状态、时钟锁定标志、信号幅值；
    所述数据预处理包括数据统计和数据优化；所述数据统计的统计值至少包括以下之一：平均数、中位数、频繁值；所述数据优化采用的方法至少包括以下之一：奇异值剔除、缺失值填充、数据平滑、数据降维。
  3. 根据权利要求1所述的方法，其特征在于，所述多个基模型至少包括机器学习模型、无味粒子滤波模型和随机过程模型；其中，所述机器学习模型至少包括以下之一：支持向量机、长短期记忆神经网络、深度置信网络。
  4. 根据权利要求3所述的方法，其特征在于，所述支持向量机的训练规则包括：
    训练集构成：由训练数据和标签组成，训练数据矩阵为[m，n]，m表示样本数目，n表示每个样本的特征；标签矩阵为[m，1]，表示m个样本所对应的分类标签值；训练样本量为总样本量的70％；
    测试集构成：数据结构与所述训练集一致，测试样本量为所述总样本量的30％；
    模型超参数至少包括以下之一：核函数选择、惩罚因子、核函数内核系数。
  5. 根据权利要求3所述的方法，其特征在于，所述深度置信网络的训练规则包括：
    训练集：由训练样本及预测标签组成，训练样本的特征数是由采集数据的通道数目以及每个通道所需的样本数量决定的；训练数据矩阵为[m，n]，标签矩阵为[m，1]，表示m个样本的预测标签值；训练样本量为总样本量的70％；
    测试集：测试数据矩阵为[o，1]，表示训练样本之后的o个数据；测试样本量为所述总样本量的30％；
    模型结构参数至少包括以下之一：网络层数、各层属性和各层神经元数；
    模型超参数至少包括以下之一：训练次数、批训练大小、学习率和激活 函数。
  6. 根据权利要求3所述的方法，其特征在于，所述无味粒子滤波模型和所述随机过程模型训练是从历史数据中确定预测模型和先验分布，根据样本不断更新权重，获取状态的后验分布，从而完成更新预测模型。
  7. 根据权利要求1所述的方法，其特征在于，所述基于所述第一数据，训练多个基模型，包括：
    根据训练规则中的模型结构参数和超参数完成模型结构初始化；
    将所述第一数据作为训练样本，对模型进行前馈运算，得到模型对所述训练样本的输出值；
    根据梯度下降算法对所述训练样本的标签与所述输出值的误差进行反向修正，更新模型内部权重和偏置；
    循环输入所述训练样本直到误差收敛或达到最大循环次数停止。
  8. 根据权利要求1所述的方法，其特征在于，所述量化度量分为健康评估和趋势预测两类；其中，所述健康评估的量化度量参数至少包括以下之一：准确率和精确率；其中，所述准确率为所有正确分类结果的占总比；所述精确率为正确判别为某类占全部判别为该类的比例；
    所述趋势预测的量化度量参数至少采用以下之一：均方根误差、平均绝对误差和相关系数；其中，所述均方根误差为观测值与真值偏差的平方和与观测次数比值的平方根；所述平均绝对误差为观测值与真值偏差绝对值求平均值；所述相关系数的计算公式为：
    其中，X为观测值，Y为真实值，Cov(X，Y)为X与Y的协方差，Var|X|为X的方差，Var|Y|为Y的方差。
  9. 根据权利要求1所述的方法，其特征在于，在对所述多个基模型进行量化度量和融合，得到集成模型中：
    采用Adaboosting算法对组合策略进行优化，采用线性误差函数对每个基模型进行循环迭代得到一系列弱模型的权重系数；
    若集成样本集基模型数量为T；其中，xi为第i个样本的训练数据，yi为第i个样本的标签，N为训练样本数量。
  10. 根据权利要求9所述的方法，其特征在于，所述集成模型的获取过程为：
    初始化，对目标场景的集成样本集的误差权重W1进行初始化，初始化公式如下：
    其中，w1，i为第i个样本对最终误差影响的误差权重，w1，N为第N个样本对最终误差影响的误差权重；
    计算基模型误差，根据如下线性损失函数公式，依次计算基模型输出ht、样本最大误差Et、第i个样本的相对误差eti、基学习器的预测误差率εt；
    Et＝max|yi-ht(xi)|，i＝1，2...N   (3)

    其中，ht(xi)为第t个基模型对第i个样本的计算结果，wt，i为第t个基模型的第i个样本对最终误差影响的影响权重；
    计算基模型权重系数αt，计算公式如下：
    更新样本集的误差权重，依次计算泛化因子Zt，样本的新误差权重wt+1，i和样本集的新误差权重Wt+1，计算公式如下：


    Wt+1＝(wt+1，1，wt+1，2...wt+1，N)  (9)
    其中，wt+1，N为第N个样本的新误差权重；
    循环执行步骤42至步骤44，直到预测误差率为0或基模型数量达到T；
    计算基模型的集成输出，计算公式如下：
    其中，K表示所有基模型输出的中位数。
  11. 根据权利要求1所述的方法，其特征在于，所述采集所述航空电子产品的相关数据，包括：
    将数据处理及分发层设置在数据采集及预处理层之上；其中，所述数据处理及分发层与采集及预处理层、数据传输层、显控及存储层共同构成飞机综合状态监测与诊断系统，且所述数据处理及分发层为所述机载健康管理域；
    所述数据采集及预处理层对机载系统的状态监测数据进行采集、预处理和封装，形成状态监测数据包，并将所述状态监测数据包通过所述数据传输层上传至所述机载健康管理域，得到所述航空电子产品的相关数据。
  12. 根据权利要求11所述的方法，其特征在于，所述方法还包括：
    将所述机载健康管理域的状态划分为四种：上电自检状态、周期自检状态、维护自检状态和故障状态；
    若上电初始化成功，则由所述上电自检状态转换到所述周期自检状态；若上电初始化失败，则由所述上电自检状态转换到所述故障状态；
    在所述周期自检状态下，若发生致命故障，则转换到所述故障状态，若接收到维护自检指令，则进入到所述维护自检状态；
    在所述故障状态下，若接收到维护自检指令，则进入到所述维护自检状态；在所述维护自检状态下，若维护自检失败，则转换到所述故障状态，若接收到退出维护模式指令，则转换到所述周期自检状态。
  13. 根据权利要求11所述的方法，其特征在于，所述方法还包括：
    将所述机载健康管理域的外部接口消息划分为所述数据采集及预处理层与所述机载健康管理域之间的接口消息，以及所述机载健康管理域与所述显控及存储层之间的接口消息。
  14. 根据权利要求11所述的方法，其特征在于，所述方法还包括：
    将所述机载健康管理域划分为模块级、功能线程级、子系统级、系统级共四个层级，并设置每个层级的任务、输入和输出信息。
  15. 根据权利要求14所述的方法，其特征在于，所述模块级的任务为模块内部多通道电路单元之间的测试协同管控、故障时间应力分析与模块健康评估；
    所述模块级的输入至少包括以下之一：模块模型更新指令和模块启动自检指令；
    所述模块级的输出至少包括以下之一：模块工作与环境应力监测参数、模块健康评估与诊断结果；
    所述功能线程级的任务为多模块测试协同管控、模块间故障关联分析与 功能健康状态评估；
    所述功能线程级的输入至少包括以下之一：模块级输出的模块工作与环境应力监测参数、模块健康评估与诊断结果，以及子系统级输入的功能线程模型更新指令、功能启动自检指令；
    所述功能线程级的输出至少包括以下之一：功能线程状态监测参数、功能健康评估与诊断结果；
    所述子系统级的任务为多线程测试协同管控、多线程故障关联分析与子系统剩余能力评估；
    所述子系统级的输入至少包括以下之一：功能线程级输出的功能线程状态监测参数、功能健康评估与诊断结果信息，以及系统级输入的子系统模型更新指令、子系统启动自检指令；
    所述子系统级的输出至少包括以下之一：子系统状态监测参数、软件故障报告和子系统健康评估与诊断结果；
    所述系统级的任务为子系统间测试协同管控、跨子系统故障诊断与系统剩余能力评估；
    所述系统级的输入至少包括以下之一：子系统输出的子系统状态监测参数、软件故障报告和子系统健康评估与诊断结果；
    所述系统级的输出至少包括以下之一：系统健康状态概要信息和系统健康状态详细信息。
  16. 根据权利要求11所述的方法，其特征在于，所述方法还包括：
    基于构建的数据管理单元、诊断模型、健康评估单元、增强诊断单元、故障预测单元和诊断过程管理单元，建立机载健康管理域的健康评估与诊断模型。
  17. 根据权利要求16所述的方法，其特征在于，所述数据管理单元对关键信息进行响应，更新本地缓存，将外部输入信息映射到所述诊断模型，完成外部输入数据与诊断模型之间的转换；所述关键信息至少包括以下之一：外部输入的故障报告、测试数据包、配置消息、消耗品、状态参数；
    所述诊断模型对系统诊断状态相关的先验知识进行管理；
    所述健康评估单元进行功能线程和模块异常检测、系统级剩余能力评估；
    所述增强诊断单元采用与所述诊断模型相对独立的通用诊断推理机，进行故障溯源、故障确认以及故障关联分析；
    所述故障预测单元采用基于特征趋势的预测方法，针对退化特征明显、故障规律可寻的产品或部件，进行数据收集、参数退化趋势跟踪与预测特征提取；
    所述诊断过程管理单元对模型输入数据集合以及健康评估、增强诊断和故障预测过程进行协同管理，将所述健康评估单元输出的功能线程和模块的故障或退货状态传输至所述增强诊断单元或所述故障预测单元，消除关联故障，匹配退化模式，预测故障发生时间；
    所述诊断过程管理单元将所述增强诊断单元和所述故障预测单元输出的 结果再反馈至所述健康评估单元，为机载系统剩余能力评估提供输入。
  18. 根据权利要求11所述的方法，其特征在于，所述方法还包括：
    将所述机载健康管理域的数据库表划分为系统表、跨子系统诊断结果表、子系统表、软件故障报告表、网络节点状态表、功能表、功能BIT结果表、功能运行参数表、模块表、模块BIT结果表、模块工作参数表和静态BIT配置表；
    所述系统表至少包括以下之一：系统标识、跨子系统诊断结果和子系统健康状态概要信息；
    跨子系统诊断结果表至少包括以下之一：诊断结果标识、诊断时间、故障隔离结果和功能评估结果信息；
    所述子系统表至少包括以下之一：子系统标识、子系统健康状态详细信息、所属功能标识、软件故障报告以及网络节点状态信息；
    所述软件故障报告表至少包括以下之一：软件标识、故障时间、故障类型、类标识、处理器节点标识；
    所述网络节点状态表至少包括以下之一：网络标识、采集时间、节点个数、节点标识、节点状态；
    所述功能表至少包括以下之一：功能标识、功能健康状态、所属模块标识、功能自检结果、功能运行参数信息；
    所述功能BIT结果表至少包括以下之一：功能标识、采集时间、测试点数量、测试点标识或ID、测试点状态；
    所述功能运行参数表至少包括以下之一：功能标识、采集时间、参数个数、参数标识或ID、参数值；
    所述模块表至少包括以下之一：模块标识、模块健康状态、模块BIT结果和模块工作参数；
    所述模块BIT结果表至少包括以下之一：模块标识、采集时间、测试点数量、测试点标识、测试点状态；
    所述模块工作参数表至少包括以下之一：模块标识、采集时间、参数标识、参数值；
    所述静态BIT配置表至少包括以下之一：测试点数量、测试点标识、滤波类型、阈值、测试参数类型。
  19. 根据权利要求18所述的方法，其特征在于，所述方法还包括：
    所述系统表通过诊断结果标识与跨子系统诊断结果表关联，通过所属子系统标识与子系统表关联；
    所述子系统表通过软件标识与软件故障报告表关联，通过网络标识与网络节点状态表关联，通过所属功能标识与功能表关联；
    所述功能表通过功能标识与功能BIT结果表、功能运行参数表关联，通过所属模块标识与模块表关联；
    所述模块表通过模块标识与模块BIT结果表、模块工作参数表关联，通过测试点标识与静态BIT配置表关联。
PCT/CN2023/096340 2022-05-25 2023-05-25 一种多模型融合的航空电子产品健康评估方法 WO2023227071A1 (zh)

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
CN202210572890.8 2022-05-25
CN202210572890 2022-05-25
CN202210922051.4 2022-08-02
CN202210922051.4A CN114969990B (zh) 2022-08-02 2022-08-02 一种多模型融合的航空电子产品健康评估方法

Publications (1)

Publication Number Publication Date
WO2023227071A1 true WO2023227071A1 (zh) 2023-11-30

Family

ID=88918566

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/096340 WO2023227071A1 (zh) 2022-05-25 2023-05-25 一种多模型融合的航空电子产品健康评估方法

Country Status (1)

Country Link
WO (1) WO2023227071A1 (zh)

Cited By (21)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117389202A (zh) * 2023-12-13 2024-01-12 南京德克威尔自动化有限公司 一种用于智能工厂总线控制io模块
CN117409976A (zh) * 2023-12-15 2024-01-16 深圳市微克科技有限公司 一种基于智能穿戴设备的用户健康监测方法、系统及介质
CN117572295A (zh) * 2024-01-12 2024-02-20 山东和兑智能科技有限公司 一种高压套管多模态在线监测及预警方法
CN117633636A (zh) * 2024-01-25 2024-03-01 江苏省特种设备安全监督检验研究院 一种云端互联的特检数据处理系统及其处理设备
CN117724433A (zh) * 2024-02-07 2024-03-19 东北大学 基于多通道分布式深度集成预测的冷轧生产前馈控制方法
CN117740072A (zh) * 2023-12-19 2024-03-22 深圳市祥为测控技术有限公司 一种基于多传感器的水浸感应方法
CN117909852A (zh) * 2024-03-19 2024-04-19 山东省地矿工程勘察院(山东省地质矿产勘查开发局八〇一水文地质工程地质大队) 一种用于水工环生态数据分析的监测数据状态划分方法
CN117993776A (zh) * 2024-02-01 2024-05-07 北京热力智能控制技术有限责任公司 一种热网能耗指标监测装置及考核评价系统
CN117992889A (zh) * 2024-01-15 2024-05-07 国网江苏省电力有限公司营销服务中心 一种基于大数据分析的中高压失准监测系统
CN118035848A (zh) * 2024-04-10 2024-05-14 深圳市博硕科技股份有限公司 一种电子元件智能老化测试系统
CN118034751A (zh) * 2024-04-15 2024-05-14 深圳市天玑互联科技有限公司 基于应用程序的mcu动态控制方法及系统
CN118091406A (zh) * 2024-04-19 2024-05-28 楷维工业服务(上海)有限公司 电机检测修复方法、装置、电子设备及存储介质
CN118152884A (zh) * 2024-03-21 2024-06-07 北京溥络数智科技有限责任公司 化工罐量预测方法、装置和化工罐实时监测系统
CN118229091A (zh) * 2024-05-14 2024-06-21 江苏森基建筑工程科技有限公司 基于数据分析的抗震成品支架安全测试管理系统
CN118227420A (zh) * 2024-04-07 2024-06-21 上海灵娃智能科技有限公司 一种可灵活配置的低代码联动系统
CN118277214A (zh) * 2024-06-04 2024-07-02 无锡雷华网络技术有限公司 一种用于数据处理系统的数据模块管理方法
CN118316271A (zh) * 2024-03-27 2024-07-09 广东敏卓机电股份有限公司 一种高性能风机电机结构
CN118378054A (zh) * 2024-06-27 2024-07-23 青岛哈尔滨工程大学创新发展中心 一种潜射无人机的实时可靠性评估系统及方法
CN118425762A (zh) * 2024-07-03 2024-08-02 斯普屹科技(北京)有限公司 一种用于智能断路器数据采集的检测系统
CN118503884A (zh) * 2024-07-17 2024-08-16 浪潮通用软件有限公司 一种设备状态识别方法、设备及介质
CN118551331A (zh) * 2024-07-30 2024-08-27 国网山东省电力公司电力科学研究院 基于鸟类定位的运行状态数据监测方法及设备

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111175054A (zh) * 2020-01-08 2020-05-19 沈阳航空航天大学 一种基于数据驱动的航空发动机故障诊断方法
CN114969990A (zh) * 2022-08-02 2022-08-30 中国电子科技集团公司第十研究所 一种多模型融合的航空电子产品健康评估方法
CN115017016A (zh) * 2022-08-09 2022-09-06 中国电子科技集团公司第十研究所 一种多层次模型融合的机载健康管理域设计方法

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111175054A (zh) * 2020-01-08 2020-05-19 沈阳航空航天大学 一种基于数据驱动的航空发动机故障诊断方法
CN114969990A (zh) * 2022-08-02 2022-08-30 中国电子科技集团公司第十研究所 一种多模型融合的航空电子产品健康评估方法
CN115017016A (zh) * 2022-08-09 2022-09-06 中国电子科技集团公司第十研究所 一种多层次模型融合的机载健康管理域设计方法

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
GAO FENG; WU LINBO; YUE YANG; LI HAIFENG: "Software Reliability Combination Model Based on AHP and AdaBoosting", COMPUTER ENGINEERING, SHANGHAI JISUANJI XUEHUI, CN, vol. 43, no. 12, 31 December 2017 (2017-12-31), CN , pages 69 - 72, XP009550631, ISSN: 1000-3428 *
YAO, BIN ET AL.: "Research on Architecture Design of Civil Aircraft Airborne Health Management System (AHMS)", PROCEEDINGS OF THE 2018 (7TH) INTERNATIONAL FORUM ON CIVIL AIRCRAFT AVIONICS, 17 April 2018 (2018-04-17) *

Cited By (25)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117389202A (zh) * 2023-12-13 2024-01-12 南京德克威尔自动化有限公司 一种用于智能工厂总线控制io模块
CN117409976A (zh) * 2023-12-15 2024-01-16 深圳市微克科技有限公司 一种基于智能穿戴设备的用户健康监测方法、系统及介质
CN117740072A (zh) * 2023-12-19 2024-03-22 深圳市祥为测控技术有限公司 一种基于多传感器的水浸感应方法
CN117572295B (zh) * 2024-01-12 2024-04-12 山东和兑智能科技有限公司 一种高压套管多模态在线监测及预警方法
CN117572295A (zh) * 2024-01-12 2024-02-20 山东和兑智能科技有限公司 一种高压套管多模态在线监测及预警方法
CN117992889A (zh) * 2024-01-15 2024-05-07 国网江苏省电力有限公司营销服务中心 一种基于大数据分析的中高压失准监测系统
CN117633636A (zh) * 2024-01-25 2024-03-01 江苏省特种设备安全监督检验研究院 一种云端互联的特检数据处理系统及其处理设备
CN117633636B (zh) * 2024-01-25 2024-05-03 江苏省特种设备安全监督检验研究院 一种云端互联的特检数据处理系统及其处理设备
CN117993776A (zh) * 2024-02-01 2024-05-07 北京热力智能控制技术有限责任公司 一种热网能耗指标监测装置及考核评价系统
CN117724433B (zh) * 2024-02-07 2024-04-19 东北大学 基于多通道分布式深度集成预测的冷轧生产前馈控制方法
CN117724433A (zh) * 2024-02-07 2024-03-19 东北大学 基于多通道分布式深度集成预测的冷轧生产前馈控制方法
CN117909852A (zh) * 2024-03-19 2024-04-19 山东省地矿工程勘察院(山东省地质矿产勘查开发局八〇一水文地质工程地质大队) 一种用于水工环生态数据分析的监测数据状态划分方法
CN117909852B (zh) * 2024-03-19 2024-05-24 山东省地矿工程勘察院(山东省地质矿产勘查开发局八〇一水文地质工程地质大队) 一种用于水工环生态数据分析的监测数据状态划分方法
CN118152884A (zh) * 2024-03-21 2024-06-07 北京溥络数智科技有限责任公司 化工罐量预测方法、装置和化工罐实时监测系统
CN118316271A (zh) * 2024-03-27 2024-07-09 广东敏卓机电股份有限公司 一种高性能风机电机结构
CN118227420A (zh) * 2024-04-07 2024-06-21 上海灵娃智能科技有限公司 一种可灵活配置的低代码联动系统
CN118035848A (zh) * 2024-04-10 2024-05-14 深圳市博硕科技股份有限公司 一种电子元件智能老化测试系统
CN118034751A (zh) * 2024-04-15 2024-05-14 深圳市天玑互联科技有限公司 基于应用程序的mcu动态控制方法及系统
CN118091406A (zh) * 2024-04-19 2024-05-28 楷维工业服务(上海)有限公司 电机检测修复方法、装置、电子设备及存储介质
CN118229091A (zh) * 2024-05-14 2024-06-21 江苏森基建筑工程科技有限公司 基于数据分析的抗震成品支架安全测试管理系统
CN118277214A (zh) * 2024-06-04 2024-07-02 无锡雷华网络技术有限公司 一种用于数据处理系统的数据模块管理方法
CN118378054A (zh) * 2024-06-27 2024-07-23 青岛哈尔滨工程大学创新发展中心 一种潜射无人机的实时可靠性评估系统及方法
CN118425762A (zh) * 2024-07-03 2024-08-02 斯普屹科技(北京)有限公司 一种用于智能断路器数据采集的检测系统
CN118503884A (zh) * 2024-07-17 2024-08-16 浪潮通用软件有限公司 一种设备状态识别方法、设备及介质
CN118551331A (zh) * 2024-07-30 2024-08-27 国网山东省电力公司电力科学研究院 基于鸟类定位的运行状态数据监测方法及设备

Similar Documents

Publication Publication Date Title
WO2023227071A1 (zh) 一种多模型融合的航空电子产品健康评估方法
Cheng et al. Remaining useful life prognosis based on ensemble long short-term memory neural network
CN102637019B (zh) 一种工业生产过程智能集成故障诊断方法及装置
CN114969990B (zh) 一种多模型融合的航空电子产品健康评估方法
Chen et al. Dynamic predictive maintenance scheduling using deep learning ensemble for system health prognostics
CN107807860B (zh) 一种基于矩阵分解的电力故障分析方法及系统
CN112083244A (zh) 综合化航空电子设备故障智能诊断系统
Wang et al. A remaining useful life prediction model based on hybrid long-short sequences for engines
Kang et al. A method of online anomaly perception and failure prediction for high-speed automatic train protection system
Chernov et al. Intelligent decision support for power grids using deep learning on small datasets
CN109523030A (zh) 一种基于机器学习的遥测参数异常监测系统
Chen et al. Exploring equipment electrocardiogram mechanism for performance degradation monitoring in smart manufacturing
Hu et al. Online updating with a probability-based prediction model using expectation maximization algorithm for reliability forecasting
CN115017016B (zh) 一种多层次模型融合的机载健康管理域设计方法
Min et al. Fault prediction for distribution network based on CNN and LightGBM algorithm
Lyu et al. Examination on avionics system fault prediction technology based on ashy neural network and fuzzy recognition
Li et al. Method for predicting failure rate of airborne equipment based on optimal combination model
Zeng et al. Research on indexes and verification technology of airborne PHM system
CN114580705A (zh) 航电产品剩余寿命预测方法
Jingyu et al. Statistical analysis of distribution network fault information based on multi-source heterogeneous data mining
Natsumeda et al. RULENet: end-to-end learning with the dual-estimator for remaining useful life estimation
Zhang et al. Research on intelligent health management technology of opto-electronic equipment
Khan et al. Outliers detection and repairing technique for measurement data in the distribution system
Liu et al. Model-based simulation and verification of aircraft prognostics and health management system
Zhang et al. Analysis and Design of Fault Prediction and Health Assessment System for the optoelectronic Equipment

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23811141

Country of ref document: EP

Kind code of ref document: A1