CN110826600B - Engine surge prediction method based on adaptive resonance network online incremental learning - Google Patents

Engine surge prediction method based on adaptive resonance network online incremental learning Download PDF

Info

Publication number
CN110826600B
CN110826600B CN201910995304.9A CN201910995304A CN110826600B CN 110826600 B CN110826600 B CN 110826600B CN 201910995304 A CN201910995304 A CN 201910995304A CN 110826600 B CN110826600 B CN 110826600B
Authority
CN
China
Prior art keywords
engine
surge
network
training
state
Prior art date
Legal status (The legal status 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 status listed.)
Active
Application number
CN201910995304.9A
Other languages
Chinese (zh)
Other versions
CN110826600A (en
Inventor
杨顺昆
李红曼
张宇涵
苟晓冬
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
China Aviation Launch Control System Research Institute
Beihang University
Original Assignee
Beihang University
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
Application filed by Beihang University filed Critical Beihang University
Priority to CN201910995304.9A priority Critical patent/CN110826600B/en
Publication of CN110826600A publication Critical patent/CN110826600A/en
Application granted granted Critical
Publication of CN110826600B publication Critical patent/CN110826600B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01M—TESTING STATIC OR DYNAMIC BALANCE OF MACHINES OR STRUCTURES; TESTING OF STRUCTURES OR APPARATUS, NOT OTHERWISE PROVIDED FOR
    • G01M15/00—Testing of engines
    • G01M15/14—Testing gas-turbine engines or jet-propulsion engines
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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
    • G06N3/045—Combinations of networks
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • Computing Systems (AREA)
  • Software Systems (AREA)
  • Molecular Biology (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Combustion & Propulsion (AREA)
  • Chemical & Material Sciences (AREA)
  • Testing And Monitoring For Control Systems (AREA)

Abstract

The invention discloses an engine surge prediction method based on adaptive resonance network online incremental learning, which comprises the following steps of: acquiring a current engine component parameter data set in real time, inputting the data into a trained adaptive resonance network to further obtain an output mode, and judging the current state of the engine by combining the output mode; the specific steps of training the adaptive resonance network comprise: step 1, collecting data; step 2, data are sorted; step 3, training the network: inputting the training data set in the step 2 into the self-adaptive resonance network, and training the self-adaptive resonance network; and 4, defining an output mode. The invention can train the self-adaptive resonance network to adjust and classify the self-adaptive resonance network in time along with the change of the engine so as to achieve more real-time and accurate monitoring effect, timely evaluate the current engine state according to the output mode at the current moment, and accurately predict whether surge happens at the next moment.

Description

Engine surge prediction method based on adaptive resonance network online incremental learning
Technical Field
The invention relates to the technical field of engine surge prediction, in particular to an engine surge prediction method based on adaptive resonance network online incremental learning.
Background
The engine is a key component of an aircraft, and the performance and the current working state of the engine directly influence the running condition of the whole aircraft. The aircraft engine surge fault is one of main factors influencing the performance and safety of the aircraft engine, and directly causes damage to the engine when the aircraft engine surge fault is serious, and poses a great threat to the safety of the aircraft, so that the diagnosis and the prediction of the surge fault are particularly important.
In the method for predicting the surge of the engine based on data analysis, technicians establish a nonlinear model (support vector machine prediction model)/linear model (Kalman filtering model) by collecting the rotating speed, the pressure ratio and the fuel flow of a gas compressor before and after the occurrence of the surge to monitor the rotating speed, the pressure ratio and the fuel flow of the gas compressor in the engine, calculate the difference between a real value and a predicted value and use the difference as a basis for predicting the surge, and predict the surge; technicians set performance degradation factors of components such as a gas compressor, a turbine, a fan, a rotor and the like to simulate surge, acquire a component parameter data set in the engine operation period, and perform network training by adopting a deep confidence network/feedforward neural network/radial basis function neural network to perform surge diagnosis and prediction; and technicians carry out fault diagnosis or prediction on the gas circuit of the engine by a Bayesian network/expert system/fuzzy logic/data fusion/hybrid method. However, the boundary between the normal state and the surge state of the engine in the real working process is not completely clear, and the working state of the engine cannot be finely classified in the training process by the above model/network/method, so that the accuracy of the model or network for predicting the surge state is relatively low; moreover, the training process of the above model or network is an off-line process, and the related method is only limited to the analysis and summary of historical data, and the real-time state features of the engine during monitoring cannot be extracted and added to the existing model or network, so the above model/network/method also lacks accuracy for the state monitoring and prediction of the engine running in real time.
Therefore, how to design an engine surge prediction method which has strong accuracy and can realize online incremental learning is a problem which needs to be solved urgently by the technical personnel in the field.
Disclosure of Invention
In view of the above, the invention provides an engine surge prediction method based on adaptive resonance network online incremental learning, which can effectively and accurately classify the state of the currently-transmitted engine performance parameter data in the monitoring process, and continuously train the adaptive resonance network in the monitoring process, thereby improving the accuracy of surge prediction.
In order to achieve the purpose, the invention adopts the following technical scheme:
the engine surge prediction method based on the online incremental learning of the adaptive resonance network comprises the following steps:
acquiring parameter data of current engine components in real time, inputting the data into a trained adaptive resonance network to obtain an output mode, and judging the current state of the engine by combining the output mode;
the specific steps of training the adaptive resonance network comprise:
step 1, collecting data: carrying out an engine surge approaching test, and collecting engine part parameter data in a surge period;
step 2, data arrangement: the parameter data collected in the step 1 are arranged into a form capable of being processed by a self-adaptive resonance network and used as a training data set;
step 3, training the network: inputting the training data set in the step 2 into the self-adaptive resonance network, and training the self-adaptive resonance network;
and 4, defining an output mode: inputting the parameter data of the engine components into the trained adaptive resonance network to obtain different output modes, analyzing the output modes to generate an output mode definition method, and thus obtaining the trained adaptive resonance network.
Preferably, the data collection mode in step 1 includes but is not limited to: an engine surge test was performed.
It should be noted that, the adaptive resonance network can not only collect data through a surge test in the training process, but also collect data through various ways to realize online incremental learning.
Preferably, step 1 specifically includes the following:
carrying out an engine surge-approaching test, and acquiring data according to a fixed frequency in a surge period, wherein the acquired data types comprise: the rotation speed, pressure ratio, flow rate and heat insulation efficiency of the compressor.
Preferably, the specific method of the surge test is as follows:
starting the engine to a slow running mode, adjusting the rotating speed to the rotating speed corresponding to the surge state in the characteristic curve, keeping the rotating speed of the engine unchanged, gradually increasing the air pressure of the engine, enabling the gas to overcome the high-pressure gas of the engine and be sprayed into the combustion chamber, improving the front pressure of the guide vane of the gas turbine, further enabling the flow path of the gas compressor to send surge sound after the flow path of the gas compressor is subjected to the processes of blocking, interrupting and backflow, emergently stopping, and cutting off the power supply to enable the engine to relieve surge.
Preferably, step 2 specifically includes the following:
parameter data acquired at different moments in different asthma-forcing tests are combined into parameter data groups, numerical values of all parameters in the parameter data groups are floating point type, and the parameter data groups are combined into the training data set.
Preferably, the training of the adaptive resonance network in step 3 specifically includes the following three processes in sequence: an initialization process, a network input process and a training learning process:
an initialization process: for the bottom-up connection weight omega in the networkijTop-down connection weight tijInitializing and setting a warning parameter rho;
wherein d is a constant, n is the number of training data sets, and 0< rho < 1;
the network input process comprises the following steps: receiving any one data group in the training data set, preprocessing the data group through an input layer, sending a preprocessed input signal to an output layer after the preprocessed input signal reaches a stable state, and entering a training and learning process;
the training and learning process is mainly divided into four stages: the method comprises an identification stage, a comparison stage, a learning stage and a searching stage;
in the recognition stage, according to the formula
Figure BDA0002239541750000041
And hj＝max{hkIdentifying a winning neuron of the output layer, and recording the winning neuron as j;
wherein h iskRepresenting the input from the input layer to the kth node of the output layer, hjIs the largest value in the input, hjCorresponding output value yjIs a winning neuron j, omegaikFor neural network from bottom to topThe connection weight coefficient;
in the comparison stage, carrying out similarity R test on the winning neuron j, giving up the current classification result when the winning neuron j fails the similarity test, entering the search stage, taking the winning neuron j as a real winning neuron when the winning neuron j passes the test, and entering the learning stage by the network;
wherein, when | R | < ρ, the winning neuron j fails the similarity test; when R is greater than rho, the winning neuron j passes similarity test;
in the learning stage, according to the formula
Figure BDA0002239541750000042
And
Figure BDA0002239541750000043
the long-term memory connection weight vector of the self-adaptive resonance network is modified, resonance between an input data set and an output mode is realized, and network memory is enhanced;
wherein Δ represents a variation, d is a constant, uiAn output value representing the intermediate node is shown,
in the search phase, the output value of the winning neuron j in the output layer is reset to 0, and does not participate in competition during the current signal output period, and other neurons in the output layer are further compared to find the next winning neuron.
Preferably, the definition of the output mode in step 4 specifically includes the following contents:
an output mode containing only normal data sets is defined as a normal state;
defining an output mode containing that the normal data set occupation ratio is larger than the surge data set as a light surge state;
defining an output mode containing that the proportion of a surge data set is more than or equal to that of a normal data set as an initial surge state;
an output mode that contains only the surge data set is defined as a surge condition.
It should be noted that the normal data set indicates a data set composed of parameter data of the engine in a normal state, and the surge data set indicates a data set composed of parameter data of the engine in a surge period.
Preferably, the specific judgment method for judging the current state of the engine according to the output mode includes the following steps:
when the output mode is in a normal state, indicating that the engine is in a normal state;
when the output mode is in a light surge state, the engine is in a normal state but tends to enter an abnormal state, and the next time state of the engine needs to be paid attention to in time;
when the output mode is in an initial surge state, the engine is indicated to be abnormal, and measures are needed to prevent surge in time;
when the output mode is in a surge state, the engine is in the surge state, and the surge needs to be relieved in time.
According to the technical scheme, compared with the prior art, the adaptive resonance network adopted by the invention can overcome the defect of low prediction accuracy caused by the fact that the network or the model cannot be trained online in the existing surge prediction method based on data analysis, can actively adjust the connection weight and the output mode of the network according to the input engine component parameters and the inherent parameters of the network, and enhances the network memory, so that the resonance network has good noise immunity and stability. The network adjusts the classification precision of the network by manually adjusting the inherent parameters of the network, and divides the output mode into surge modes with different severity degrees, thereby realizing the surge prediction of the engine; and the network can continue to train in the monitoring process, so that the current input engine performance parameter data can be more accurately classified in the monitoring process, and the surge prediction is more accurate.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the provided drawings without creative efforts.
FIG. 1 is a flow chart of a method for training an adaptive resonance network and using the trained adaptive resonance network for engine prediction according to the present invention;
FIG. 2 is a flow chart of a method for training an adaptive resonant network according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a specific process of a method for training an adaptive resonance network and predicting an engine by using the trained adaptive resonance network according to the present invention.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The embodiment of the invention discloses an engine surge prediction method based on adaptive resonance network online incremental learning.
As shown in fig. 1 and 3, a plurality of engine surge events are performed 101, and engine component parameter data is collected over a surge cycle.
The specific method can be divided into the following two steps.
First, the surge test: starting the engine to a slow vehicle, adjusting the rotating speed to the rotating speed corresponding to the surge state in the characteristic curve, keeping the rotating speed of the engine unchanged, gradually increasing the air pressure of the engine, enabling the air to overcome the high-pressure air of the engine and spray the air into a combustion chamber, improving the front pressure of a gas turbine guide vane, further enabling a compressor flow path to send obvious surge blow-out sound after undergoing the processes of blocking, interrupting and backflow, emergently stopping the engine, and cutting off a power supply to enable the engine to relieve surge, wherein the process is a single surge forcing test, and the multiple surge forcing is carried out, so that the original data required by a subsequent training network can be generated.
Secondly, collecting data: the vibration is increased when the compressor surges, the total pressure and flow at the outlet of the compressor fluctuate greatly, and the exhaust of the engine is increased, so that parameter data which can obviously represent the surge characteristic of the engine, such as the rotating speed, the pressure ratio, the flow and the adiabatic efficiency of the compressor in the engine, have certain explanatory and training significance. In the time range of the surge period, the data acquisition period is 1 time/second, and the types of the acquired data are as follows: the rotation speed, pressure ratio, flow rate and heat insulation efficiency of the compressor. In order to ensure that the trained network has better classification and identification performance, 1000 groups of data sets are selected. The rotating speed of the test piece is measured by a magnetoelectric sensor and a rotating speed measuring instrument; the pressure ratio is measured and calculated by a pressure measuring system; the flow is measured and calculated by pressure and temperature at different points; the adiabatic efficiency is obtained by the correlation calculation of parameters such as atmospheric temperature, coefficient of heat transfer ratio, inlet flow of the compressor and the like.
102. Data arrangement: and (4) arranging the parameter data into a form which can be processed by the self-adaptive resonance network and using the form as a training data set.
The adaptive resonance network has no guiding learning capability, so that the parameter data only need to comprise engine component parameter data sets under different surge tests at different moments, and the form of the data sets is as follows: the numerical values of each parameter are floating point type [ compressor rotation speed, pressure ratio, flow rate, adiabatic efficiency ]. The self-adaptive resonant network selects the second type of resonant network, the basic idea adopts a competitive learning mechanism, the input requirement is wide, binary discrete quantity can be allowed to serve as an input mode, any analog quantity can be allowed to serve as the input mode, the internal learning algorithm is more complex, and the capability of extracting and enhancing signals from complex and changeable background noise is achieved.
As shown in fig. 2, 103, training network: and (4) inputting the data set in the step (2) into a network, and training the network.
Firstly, setting network related constants such as a warning parameter rho and the like to initialize the network, inputting a data set into the network, continuously identifying, comparing, learning and searching, adjusting the warning parameter rho during the period to enable output modes to be divided into four types, and then stopping training. The network is divided into two layers, the input layer has the capabilities of inhibiting input noise and enhancing characteristic signals, and the key role of the output layer is to improve the contrast of the bottom-up filtering input mode and send out a reset signal in a competitive way. The training process of the network can be illustrated in detail by fig. 3.
The specific method comprises an initialization process, a network input process and a training process.
Wherein, the initialization process: according to the formula:
Figure BDA0002239541750000081
and tij(0) For the bottom-up connection weight omega in the networkijTop-down connection weight tijInitializing and setting warning parameters rho (0)<ρ<1) Wherein d is a constant and n is the number of training data sets.
The network input process comprises the following steps: any one data set in the training data set is received, the characteristic signals of the input data set are enhanced through data processing processes of receiving, converting, storing and the like of three sublayers in the input layer, the input signals preprocessed by the input layer reach a stable state and are sent to the output layer, and a training learning stage is started.
The training is mainly divided into four stages: an identification phase, a comparison phase, a learning phase and a search phase.
In the recognition stage, according to the formula
Figure BDA0002239541750000082
hj＝max{hkWhere hkRepresenting the input from the input layer to the kth bit node of the output layer, hjIs the largest value in the input, hjCorresponding output value yjThen the winning neuron j, omegaikFor the weight coefficient of the neural network from bottom to top) to identify the winning neuron of the output layer, and remember the winning godThe warp element is j.
In the comparison stage, the similarity R inspection is carried out on the winning neuron j in the identification stage, when the winning neuron j does not pass the similarity inspection, namely | R | < rho, the classification result is abandoned and the search stage is entered, when the winning neuron j passes the inspection, namely | R | > rho, the winning neuron j is taken as a real winning neuron, and the network enters the learning stage.
In the learning stage, according to the formula
Figure BDA0002239541750000083
And
Figure BDA0002239541750000084
(wherein the variable Δ represents a variation, d is a constant, uiThe output value representing the intermediate node) modifies the long-term memory connection weight vector thereof, realizes the resonance between the input data set and the output mode, and enhances the network memory.
In the search phase, the output value of the neuron j in the output layer is reset to 0, and does not participate in competition during the signal output period, so that a true winning neuron is found in the output layer.
104. Detecting a network: and performing the surge-approaching test again, collecting the parameter data of the engine parts, sorting the parameter data into a normal data set and a surge data set, inputting the normal data set and the surge data set into the trained network, and observing the output condition of the network.
And performing the surge approaching test again, and acquiring parameter data of the engine part according to the frequency of 1 time/second to ensure that the surge data set and the normal data set which are equal in quantity can be collected to be respectively 500 groups or more. The purpose of the detection network is to define labels of different surge levels for the output pattern based on the distribution of the normal data set and the surge data set in the network output. Defining an output mode only containing a normal data set as a normal state, defining an output mode containing the normal data set with the ratio greater than that of a surge data set as a light surge state, defining an output mode containing the surge data set with the ratio greater than or equal to that of the normal data set as an initial surge state, and defining an output mode only containing the surge data set as a surge state.
In this embodiment, the output patterns trained in step 3 are set as 4 classes, which are respectively labeled as 0,1,2, and 3, but the specific meaning represented by the 4 classes of output patterns is unknown, and at this time, the collected surge data sets and normal data sets of equal number are labeled as red and green and input into the trained network, and labels are defined for 0,1,2, and 3 according to the distribution of the input data sets of red and green. If the proportion of the data set marked as red is greater than or equal to green in the output mode 0, defining 0 as an initial surge state; if in output mode 0, the data set marked red is less than green in duty, then 0 is defined as a light surge condition; if in output mode 0, only the data set with the red marker is contained, then 0 is defined as a surge condition; if in output mode 0, only the green labeled data set is contained, then 0 is defined as the normal state; and in the same way, the setting of the state label of the 0,1,2 and 3 output mode is completed.
105. And (3) surge prediction: inputting the parameter data set of the current engine component, judging the current state of the engine by combining the output of the network, and predicting the engine state at the next moment.
The method comprises the steps that a network with a defined label is used for monitoring an engine in real time, a current engine component parameter data set is input into the network, the output mode of the network is observed, and when the output mode is in a normal state, the engine is indicated to be in the normal state; when the output mode is in a light surge state, the engine is in a normal state, but the engine tends to enter an abnormal state, and the next time state of the engine needs to be paid attention to in time; when the output mode is in an initial surge state, the engine is indicated to be abnormal, and measures are needed to prevent surge in time; when the output mode is in a surge state, the engine is in the surge state, and the surge needs to be relieved in time. Therefore, the network can continue to train the network at the current moment, the classification can be adjusted in time along with the change of the engine so as to achieve a more real-time and accurate monitoring effect, the current engine state can be evaluated in time according to the output mode at the current moment, and whether the surge at the next moment is more accurately predicted.
It should be further noted that the step 101-104 in the present application is not required to be performed each time a specific engine prediction is performed. After the engine needing prediction is determined, the adaptive vibration network is trained for the engine. And in the network training process, real-time data can be added at any time for real-time training.
After the adaptive vibration network is trained, the prediction of the engine state can be realized by utilizing the network.
The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other. The device disclosed by the embodiment corresponds to the method disclosed by the embodiment, so that the description is simple, and the relevant points can be referred to the method part for description.
The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims (3)

1. The engine surge prediction method based on the online increment learning of the adaptive resonance network is characterized by comprising the following steps of:
acquiring parameter data of current engine components in real time, inputting the data into a trained adaptive resonance network to obtain an output result, and judging the current state of the engine by combining the output result;
the specific steps of training the adaptive resonance network comprise:
step 1: collecting data: collecting engine component parameter data; the method specifically comprises the following steps: carrying out an engine surge-approaching test, and acquiring data according to a fixed frequency in a surge period, wherein the acquired data types comprise: the rotating speed, the pressure ratio, the flow rate and the heat insulation efficiency of the gas compressor are improved; in the time range of the surge period, the data acquisition period is 1 time/second;
step 2: data arrangement: the parameter data collected in the step 1 are arranged into a form capable of being processed by a self-adaptive resonance network and used as a training data set;
and step 3: training a network: inputting the training data set in the step 2 into the self-adaptive resonance network, and training the self-adaptive resonance network;
and 4, step 4: defining an output mode: inputting engine component parameter data into the trained adaptive resonance network to obtain different output modes, analyzing the output modes to generate an output mode definition method, and thus obtaining the trained adaptive resonance network;
the training of the adaptive resonance network in the step 3 specifically comprises the following three processes in sequence: an initialization process, a network input process and a training learning process:
an initialization process: for the bottom-up connection weight omega in the networkijTop-down connection weight tijInitializing and setting a warning parameter rho;
wherein d is a constant, n is the number of training data sets, and 0< rho < 1;
the network input process comprises the following steps: receiving any one data group in the training data set, preprocessing the data group through an input layer, sending a preprocessed input signal to an output layer after the preprocessed input signal reaches a stable state, and entering a training and learning process;
the training and learning process is mainly divided into four stages: the method comprises an identification stage, a comparison stage, a learning stage and a searching stage;
in the recognition stage, according to the formula
Figure FDF0000013919120000021
And hj＝max{hkIdentifying a winning neuron of the output layer, and recording the winning neuron as j;
wherein h iskRepresenting the input from the input layer to the kth node of the output layer, piRepresents the node after the input signal processed by the input layer reaches the steady state, hjIs the largest value in the input, hjCorresponding output value yjIs a winning neuron j, omegaikThe weight coefficient of the connection of the neural network from bottom to top;
in the comparison stage, carrying out similarity R test on the winning neuron j, giving up the current classification result when the winning neuron j fails the similarity test, entering the search stage, taking the winning neuron j as a real winning neuron when the winning neuron j passes the test, and entering the learning stage by the network;
wherein, when | R | < ρ, the winning neuron j fails the similarity test; when R is greater than rho, the winning neuron j passes similarity test;
in the learning stage, according to the formula
Figure FDF0000013919120000022
And
Figure FDF0000013919120000023
the long-term memory connection weight vector of the self-adaptive resonance network is modified, resonance between an input data set and an output mode is realized, and network memory is enhanced;
wherein Δ represents a variation, d is a constant, uiAn output value representing the intermediate node is shown,
in the searching stage, the output value of a winning neuron j in the output layer is reset to 0, and does not participate in competition in the current signal output period, and other neurons in the output layer are further compared to find out the next winning neuron;
the definition of the output mode in step 4 specifically includes the following contents:
an output mode containing only normal data sets is defined as a normal state;
defining an output mode containing that the normal data set occupation ratio is larger than the surge data set as a light surge state;
defining an output mode containing that the proportion of a surge data set is more than or equal to that of a normal data set as an initial surge state;
defining an output mode containing only a surge data set as a surge state;
the specific judgment method for judging the current state of the engine according to the output mode comprises the following steps:
when the output mode is in a normal state, indicating that the engine is in a normal state;
when the output mode is in a light surge state, the engine is in a normal state but tends to enter an abnormal state, and the next time state of the engine needs to be paid attention to in time;
when the output mode is in an initial surge state, the engine is indicated to be abnormal, and measures are needed to prevent surge in time;
when the output mode is in a surge state, the engine is in the surge state, and the surge needs to be relieved in time.
2. The engine surge prediction method based on the online incremental learning of the adaptive resonance network as claimed in claim 1, wherein the specific method of the surge-forcing test is as follows:
starting the engine to a slow running mode, adjusting the rotating speed to the rotating speed corresponding to the surge state in the characteristic curve, keeping the rotating speed of the engine unchanged, gradually increasing the air pressure of the engine, enabling the gas to overcome the high-pressure gas of the engine and be sprayed into the combustion chamber, improving the front pressure of the guide vane of the gas turbine, further enabling the flow path of the gas compressor to send surge sound after the flow path of the gas compressor is subjected to the processes of blocking, interrupting and backflow, emergently stopping, and cutting off the power supply to enable the engine to relieve surge.
3. The adaptive resonance network online incremental learning-based engine surge prediction method according to claim 1, wherein the step 2 specifically comprises the following steps:
parameter data acquired at different moments in different asthma-forcing tests are combined into parameter data groups, numerical values of all parameters in the parameter data groups are floating point type, and the parameter data groups are combined into the training data set.
CN201910995304.9A 2019-10-18 2019-10-18 Engine surge prediction method based on adaptive resonance network online incremental learning Active CN110826600B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910995304.9A CN110826600B (en) 2019-10-18 2019-10-18 Engine surge prediction method based on adaptive resonance network online incremental learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910995304.9A CN110826600B (en) 2019-10-18 2019-10-18 Engine surge prediction method based on adaptive resonance network online incremental learning

Publications (2)

Publication Number Publication Date
CN110826600A CN110826600A (en) 2020-02-21
CN110826600B true CN110826600B (en) 2022-03-11

Family

ID=69549639

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910995304.9A Active CN110826600B (en) 2019-10-18 2019-10-18 Engine surge prediction method based on adaptive resonance network online incremental learning

Country Status (1)

Country Link
CN (1) CN110826600B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111767610B (en) * 2020-05-22 2022-07-15 成都飞机工业（集团）有限责任公司 Airplane flight performance calculation method based on linearized incremental equation
CN112906969B (en) * 2021-03-01 2024-06-14 盛景智能科技(嘉兴)有限公司 Engine fault prediction method and device, electronic equipment and storage medium

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP4676149B2 (en) * 2004-03-10 2011-04-27 株式会社日立製作所 Data classification device
CN108182328A (en) * 2018-01-05 2018-06-19 北京航空航天大学 A kind of big angle of attack Nonlinear Aerodynamic reduced-order model suitable for stall flutter
CN109297699A (en) * 2018-12-07 2019-02-01 中南大学 A kind of intelligent rotating mechanical failure diagnostic method of mixed decomposition and extraction

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
发动机喘振故障检测的神经网络免疫识别模型;侯胜利等;《振动与冲击》;20100125;第29卷(第1期);第1章节 *
基于改进的ART2 神经网络的滚动轴承故障诊断;吴肇中等;《济南大学学报(自然科学版)》;20191017;第33卷(第6期);第537-543页 *

Also Published As

Publication number Publication date
CN110826600A (en) 2020-02-21

Similar Documents

Publication Publication Date Title
CN111598438B (en) Civil aviation engine gas circuit anomaly detection method based on piecewise fitting analysis and evaluation
CN109164343B (en) Transformer fault diagnosis method based on characteristic information quantization and weighted KNN
CN105677791B (en) For analyzing the method and system of the operation data of wind power generating set
CN108897954A (en) Wind turbines temperature pre-warning method and its system based on BootStrap confidence calculations
EP0637739A2 (en) Testing an internal combustion engine
CN110826600B (en) Engine surge prediction method based on adaptive resonance network online incremental learning
CN111160457B (en) Scroll engine fault detection method based on soft-class extreme learning machine
CN110826587B (en) Vortex shaft engine fault detection method based on improved weighted support vector machine
CN111506049A (en) Multiple fault diagnosis method for aero-engine control system based on AANN network system
CN110794683A (en) Wind power gear box state evaluation method based on deep neural network and kurtosis characteristics
CN112381965A (en) Aeroengine health state identification system and method based on data mining
CN115375026A (en) Method for predicting service life of aircraft engine in multiple fault modes
CN116204825A (en) Production line equipment fault detection method based on data driving
CN115130731A (en) Method, equipment and medium for detecting fault of wind driven generator
CN118501692A (en) Servo motor fault diagnosis method and system
CN113204894B (en) Construction method and application of electric energy metering abnormality diagnosis model
CN117192371B (en) Test method and system for motor driver of new energy automobile
US20220364478A1 (en) Training of machine learning models for data-driven decision-making
CN114764867A (en) Fan fault diagnosis system and method based on image main feature extraction and application
US11339763B2 (en) Method for windmill farm monitoring
CN115618506A (en) Method for predicting power of single-shaft combined cycle gas turbine
CN115310693A (en) Comprehensive energy data cleaning method, system, equipment and medium
CN117009791B (en) Method and system for identifying faults of aeroengine
CN110874088B (en) Monitoring method of ship key equipment system based on multi-dimensional vector model
Slimene et al. Anomaly Detection Method of Aircraft System using Multivariate Time Series Clustering and Classification Techniques

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
TR01 Transfer of patent right

Effective date of registration: 20220328

Address after: 100191 No. 37, Haidian District, Beijing, Xueyuan Road

Patentee after: BEIHANG University

Patentee after: China Aviation Launch Control System Research Institute

Address before: 100191 No. 37, Haidian District, Beijing, Xueyuan Road

Patentee before: BEIHANG University

TR01 Transfer of patent right