CN111625995A - Online time-space modeling method integrating forgetting mechanism and double ultralimit learning machines - Google Patents
Online time-space modeling method integrating forgetting mechanism and double ultralimit learning machines Download PDFInfo
- Publication number
- CN111625995A CN111625995A CN202010450196.XA CN202010450196A CN111625995A CN 111625995 A CN111625995 A CN 111625995A CN 202010450196 A CN202010450196 A CN 202010450196A CN 111625995 A CN111625995 A CN 111625995A
- Authority
- CN
- China
- Prior art keywords
- learning machine
- model
- ultralimit
- online
- ultralimit learning
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 72
- 230000007246 mechanism Effects 0.000 title claims abstract description 44
- 238000012549 training Methods 0.000 claims abstract description 38
- 230000008569 process Effects 0.000 claims abstract description 26
- 238000004422 calculation algorithm Methods 0.000 claims abstract description 12
- 230000036961 partial effect Effects 0.000 claims abstract description 11
- 230000014509 gene expression Effects 0.000 claims description 27
- 239000011159 matrix material Substances 0.000 claims description 26
- 239000013598 vector Substances 0.000 claims description 17
- 230000006870 function Effects 0.000 claims description 15
- 230000004913 activation Effects 0.000 claims description 12
- 230000009977 dual effect Effects 0.000 claims description 7
- 238000012886 linear function Methods 0.000 claims description 6
- 210000002569 neuron Anatomy 0.000 claims description 5
- 230000010354 integration Effects 0.000 claims 1
- 238000009826 distribution Methods 0.000 abstract description 21
- 238000004364 calculation method Methods 0.000 abstract description 6
- 230000008859 change Effects 0.000 abstract description 4
- 238000010586 diagram Methods 0.000 description 12
- 230000008878 coupling Effects 0.000 description 4
- 238000010168 coupling process Methods 0.000 description 4
- 238000005859 coupling reaction Methods 0.000 description 4
- 238000000513 principal component analysis Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 230000000670 limiting effect Effects 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000002829 reductive effect Effects 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 230000002123 temporal effect Effects 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000002411 adverse Effects 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000003889 chemical engineering Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 239000012530 fluid Substances 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000004806 packaging method and process Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000012706 support-vector machine Methods 0.000 description 1
- 238000001308 synthesis method Methods 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
- 230000036962 time dependent Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- 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
- 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
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Computer Hardware Design (AREA)
- Geometry (AREA)
- Feedback Control In General (AREA)
Abstract
The invention discloses an online time-space modeling method integrating a forgetting mechanism and a double-overrun learning machine, which comprises the following steps: decoupling the time-space output variable into a time coefficient sequence and a space sequence according to a nonlinear partial differential equation of a curing process of a curing oven of a nonlinear distribution parameter system; establishing a first ultralimit learning machine model and a second ultralimit learning machine model based on the time coefficient sequence; updating model parameters of the first and second ultralimit learning machine models by using an online sequential learning algorithm of an integrated forgetting mechanism; the space sequence and the updated first and second ultralimit learning machine models are integrated to reconstruct an online space-time model, so that the problems that the existing modeling method of a nonlinear distribution parameter system such as a curing furnace is mostly developed in an offline environment, the timeliness of online sequence training data cannot be reflected, and the calculation efficiency is low are solved, and the online prediction of the temperature of the curing furnace is more matched with the dynamic change of the actual temperature.
Description
Technical Field
The invention relates to the technical field of space-time modeling of a nonlinear distributed parameter system, in particular to an online space-time modeling method integrating a forgetting mechanism and a double-overrun learning machine.
Background
Many industrial processes, such as thermal processes, fluid flows, chemical engineering, etc., are not only time-dependent but also space-dependent, and these systems are typically nonlinear Distributed Parametric Systems (DPSs), which, unlike Lumped Parameter Systems (LPSs), are usually described by one or a set of Partial Differential Equations (PDEs) with corresponding initial and boundary conditions, and the inputs, outputs, and even state parameters of nonlinear distributed parametric systems are coupled in space-time, resulting in infinite dimensional characteristics of such systems, and thus, modeling of such systems is very difficult.
The curing oven is important equipment for providing required temperature distribution in the curing process in the semiconductor rear-end packaging industry, belongs to a nonlinear distribution parameter system, and is difficult to obtain accurate partial differential equation description in the curing process due to very complex boundary conditions and influence of unknown disturbance inside the curing process, but the distribution of the temperature of the curing oven directly influences the curing quality of a chip, and modeling of the nonlinear distribution parameter system such as the curing oven has very important significance for online prediction of the temperature of the curing oven.
Chinese patent publication No. CN109145346A, published as 2019, 1 month and 4 days, proposes a curing thermal process space-time modeling method based on a dual least squares support vector machine, which is used for online prediction and control of the thermal process of a chip curing oven, and is based on Principal Component Analysis (PCA) space-time modeling, but the space-time model based on principal component analysis is developed in an offline environment, i.e., all training data is collected and prepared before modeling, the modeling method cannot reflect the timeliness of online sequence training data, and in the actual chip curing process, the system usually has large-scale time-varying characteristics, which requires online update of the space-time model with new samples to maintain satisfactory performance, although the conventional space-time modeling method has satisfactory modeling performance for the thermal process of the curing oven, there are still some problems in online implementation of the model, the method mainly comprises the following aspects:
1) and (3) online updating: spatio-temporal models based on principal component analysis are model drifts that often lead to large-scale, time-varying systems developed in an off-line environment.
2) Calculating efficiency: for large scale time varying systems, there is a large difference between old and new samples. If the spatio-temporal model only continuously adds new samples in the training sample set without processing the old samples, the learning capability of the spatio-temporal model on the new training samples is limited, so that the characteristics of the time-varying system are difficult to accurately describe. In addition, the increase in the number of samples increases the computational burden on the system and occupies a large amount of memory space.
Disclosure of Invention
In order to overcome the defects that the existing modeling methods of nonlinear distributed parameter systems such as curing furnaces are mostly developed in an offline environment and cannot reflect the timeliness of online sequence training data, the online updating capability of the models is poor, and the calculation efficiency is low, the invention provides the online time-space modeling method integrating the forgetting mechanism and the double-overrun learning machine, the time-varying characteristics of the system are considered, the online updating capability of the established models is strong, the accuracy is high, and the online prediction of the temperature of the curing furnaces is more matched with the dynamic change of the actual temperature.
In order to achieve the above purpose, the technical scheme of the invention is as follows:
an online space-time modeling method integrating a forgetting mechanism and a double-overrun learning machine at least comprises the following steps:
s1, decoupling a space-time output variable into a time coefficient sequence and a space sequence according to a nonlinear partial differential equation of a curing heat process of a curing oven of a nonlinear distributed parameter system;
s2, establishing a first ultralimit learning machine model and a second ultralimit learning machine model based on the time coefficient sequence;
s3, updating model parameters of the first ultralimit learning machine model and the second ultralimit learning machine model by utilizing an online sequential learning algorithm of an integrated forgetting mechanism;
and S4, integrating the spatial sequence with the updated first and second ultralimit learning machine models to reconstruct an online space-time model.
Preferably, the nonlinear partial differential equation of the curing process of the curing oven with the nonlinear distributed parameter system in the step S1 is:
where T (S, T) represents (x, y, z) ∈ (0, S) at time T and position S ═ x, y, z0) P represents a density, k represents a thermal coefficient, c represents a specific heat coefficient, f (T (S, T)) is an unknown nonlinear thermal dynamic associated with the space-time output variable T, Q (S, T) represents a heat source,representing the Laplace space operator, T0(S,0) representing an initial spatiotemporal output variable;
the spatial-temporal output variable T (S, T) is decoupled into a time coefficient sequence and a space sequence, and the expression is as follows:
wherein, ai(t) the ith time coefficient of the time coefficient series ofφi(S) denotes the ith spatial position, the spatial sequence BFSs is
The expression for the ith time coefficient is:
ai(t)=gi(ai(t-1))+hi(u(t-1))
wherein, gi(. and h)iBoth refer to non-linear functions.
Therein, fromThe nonlinear partial differential equation in the curing heat process of the curing oven of the nonlinear distribution parameter system cannot be directly used for space/time coupling of online prediction and control, a space/time separation method is adopted to decouple space-time output variables, a space sequence BFs can be learned by a KL method for collecting space-time distribution data, and the space sequence BFs can be obtained by a KL method for collecting space-time distribution dataThe first n-th order of (A) can capture the dominant dynamic behavior of the DPS, in addition, a plurality of data-based identification methods are used for estimating a low-order time model at present, and after the time model is trained, a foundation can be laid for reconstructing a space-time model through space/time synthesis.
Preferably, the expression of the first overrun learning machine model in step S2 is:
the expression of the second ultralimit learning machine model is:
wherein, βσOutput weight representing a connection output node and a hidden node in the first ultralimit learner model, β'Output weight, ω, representing the connection of output node and hidden node in the second ultralimit learning machine modelσRepresenting input weights, ω 'connecting the output node and the hidden node in the first ultralimit learner model'Representing input weights connecting the output node and the hidden node in the second ultralimit learning machine model, sigma representing the sigma-th hidden node in the first ultralimit learning machine model, N1Representing the number of hidden nodes in the first ultralimit learning machine model, N2Representing the number of hidden nodes in the second ultralimit learning machine model, G1Representing an activation function of a hidden layer in the first ultralimit learning machine model; g2Representing the activation function of the hidden layer in the second ultralimit learning machine model, ησIndicating first ultralimitThreshold of hidden node in learner model, η'A threshold value representing a hidden node in the second ultralimit learning machine model, u (t) a time coefficient in a time coefficient sequence based on which the second ultralimit learning machine model is established, gi(. and h)i(. two non-linear functions satisfy: gi(·)=g(·),hi(·)=h(·)。
Preferably, based on the expressions of the first and second ultralimit learning machine models, the expression of the time coefficient a (t) is further expressed as:
at this time, the input weight ω connecting the output node and the hidden node in the first ultralimit learning machine modelσInput weight ω 'connecting output node and hidden node in second ultralimit learner model'Threshold η for hidden nodes in the first ultralimit learning machine modelσAnd a threshold η 'of hidden nodes in the second ultralimit learner model'Are randomly generated, independent of the training data and independent of each other, and output weights β connecting the output nodes and hidden nodes in the first ultralimit learning machine modelσAnd β 'as output weights connecting the output node and the hidden node in the second ultralimit learner model'Determining according to the input and output data; the further expression of the time coefficient a (t) is omega generated randomly according to the mathematical description of a double-overrun learning machine designed by coupling double nonlinear structuresσ,ω′,ησAnd η'Once estimated, these values will be fixed during subsequent learning.
Further expressed as: a (t) ═ hT(t)θ
Wherein h (t) represents parameter vectors of the first and second ultralimit learning machine models related to the input-output data, and the expression is as follows:
theta represents unknown parameter vectors of the first and second ultralimit learning machine models to be identified,the matrix form of the time coefficients a (t) is:
A=Hθ
wherein a ═ a (2), a (3),.., a (l)]TAn output vector form representing the time coefficient a (t), L representing the data length of the phasor; h ═ HT(2)L hT(L)]TRepresenting a regression matrix;
Wherein,is the Moore-Penrose generalized inverse of matrix H,representing the unknown parameter vector taken from a and H.
Since the output vector form A and the regression matrix H of the time coefficients a (t) are known, only H is considered hereTH is a non-singular condition, satisfies
Preferably, the initial training set block is:the initial model parameters of the first and second ultralimit learning machines are Subscript 0 denotes the block with the initial training setThe associated initial vector or matrix, set to input/output data { z (L)0+1),a(L0+1) } arrive, then:
updated parameter θ using MP generalized inverse matrix1To write as:
wherein,will theta0Is denoted as P1、h(L0+1) and a (L)0+1), then:wherein,by using the formula of the Woodbury,
P1=((P0)-1+h(L0+1)hT(L0+1))-1
=P0-P0h(L0+1)(I+hT(L0+1)P0h(L0+1))-1hT(L0+1)P0
when the m +1 th data { z (L)0+m+1),a(L0+ m +1) } arrive,
wherein,i.e. according to a known parameter thetamAnd newly arrived data z (L)0+m+1),a(L0+ m +1) }, update the model parameter θm+1。
Here, online data usually arrives one by one or block by block, and because the model structure of the online data matches with that of a general thermal system, the established first and second ultralimit learning models can well simulate the spatio-temporal dynamic characteristics of the DPSs, however, for a large time-varying system, the model developed in an offline environment should be updated online to maintain satisfactory model performance, and the conventional method for realizing online updating is to combine the existing data with the new data and train the model repeatedly from zero, which brings a great computational burden to practical application. Thus, online updating of spatio-temporal models requires an online sequential learning algorithm that ensures that only learned model parameters and new data are used.
Preferably, the new input-output data is set as: { z (L)0+1),a(L0+1), { z (2), a (2) } indicates old input-output data { z (2), a (2) } with respect to new input-output data, then:
based on MP generalized inverse matrix and forgetting mechanism, parameter theta1Comprises the following steps:
when the m +1 th data { z (L)0+m+1),a(L0+ m +1) } arrives, discarding the oldest data { z (m +2), a (m +2) }, then:
here, for a large-scale time-varying system, the training samples are usually time-efficient, that is, the training samples have a certain period of validity, and the learning process is continuous, and since the samples are continuously added, the number of the trained samples in the whole system will be continuously increased, which will increase the dimension of the regression matrix H, thereby increasing the computational burden of the system and occupying a large amount of memory space. Therefore, in the online sequential learning process of the DPSs, the forgetting mechanism is embedded into the online learning process, the learning effect can be improved by giving up outdated training data, so that the adverse effect of the outdated training data on subsequent learning is reduced, and the embedding of the forgetting mechanism can ensure that online modeling has the advantages of improving the calculation work efficiency, saving a large amount of calculation work time and not needing to store outdated data.
Preferably, in step S3, the process of updating the model parameters of the first and second ultralimit learning machine models by using the online sequential learning algorithm with the forgetting mechanism is as follows:
s301, initialization: setting the sequence m of the number of training data as zero, selecting the activation function G of the hidden layer in the first ultralimit learning machine model1Activation function G of hidden layer in second ultralimit learning machine model2The number N of hidden nodes in the first ultralimit learning machine model1The number N of hidden nodes in the second ultralimit learning machine model2Randomly generating input weights omega for connecting the output node and the hidden node in the first ultralimit learning machine modelσConnecting output nodes and hiding in a second ultralimit learning machine modelInput weight of node ω'Threshold η for hidden nodes in the first ultralimit learning machine modelσThreshold η 'of hidden node in second ultralimit learner model';
Calculating an initial output matrix of a hidden layer of the first ultralimit learning machine model and the second ultralimit learning machine model:
H0=[hT(2) L hT(L0)]Tcalculating initial unknown model parameters: theta0=P0H0 TA0
Wherein the initial unknown model parameter is the initial output weight of the model, P0=(H0 TH0)-1,A0=[a(2),a(3),...,a(L0)]T;
S302, performing online sequential learning of an integrated forgetting mechanism:
when the m +1 th data { z (L)0+m+1),a(L0+ m +1) } arrive, the m +1 activation function is computed:
s304, utilizing the calculated model output weight to evaluate the first ultralimit learning machine model and the second ultralimit learning machine model according to the condition that A is H theta;
s305, increasing the value of m by 1, and returning to the step S302.
Here, no analytical expressions of DPSs are required in the implementation process, and this is also very suitable for general industrial applications.
Preferably, when the model parameters of the first and second ultralimit learning machine models are updated by using an online sequential learning algorithm integrated with a forgetting mechanism, the training data is received one by one or block by block.
Here, since the training data is received one by one or block by block, if the training data arrives block by block, it is possible to prevent the occurrence of a failure in the reception of the training data
Preferably, the training data L is initialized0The number of the first overrun learning model and the number of the hidden neurons of the second overrun learning model are not less than the sum of the numbers of the hidden neurons of the first overrun learning model and the second overrun learning model, so that the initial regression matrix H is ensured0The rank of (2) is the sum of the number of hidden layers of the two over-limit learning machine models.
Preferably, the expression of the reconstructed online spatio-temporal model in step S4 is:
wherein,representing a time coefficient, phi, in a time series after updating by integrating a forgetting mechanism and a dual-overrun learning machinei(S) represents a spatial sequence;and representing the space-time output variable of the reconstructed online space-time model.
Compared with the prior art, the technical scheme of the invention has the beneficial effects that:
the invention provides an online space-time modeling method integrating a forgetting mechanism and a double-overrun learning machine, which considers the limitation that a nonlinear partial differential equation of a curing furnace in a nonlinear distributed parameter system curing heat process cannot be used for online prediction and space/time coupling control, decouples a space-time output variable into a time coefficient sequence and a space sequence, establishes a first overrun learning machine model and a second overrun learning machine model based on the time coefficient sequence, updates model parameters of the first overrun learning machine model and the second overrun learning machine model by utilizing an online sequential learning algorithm of the integrated forgetting mechanism, embeds the forgetting mechanism into the online learning process, can improve the learning effect by giving up outdated training samples, updates the training data online, improves the online updating capability of the models, ensures that the online modeling has higher calculation work efficiency, A large amount of calculation working time is saved, outdated data does not need to be stored, and the reconstructed online space-time model is closer to the actual curing thermal process of the curing oven, so that the online prediction of the temperature of the curing oven is more matched with the dynamic change of the actual temperature.
Drawings
FIG. 1 is a flow chart of an online spatiotemporal modeling method integrating a forgetting mechanism and a dual ultralimit learning machine according to the present invention;
FIG. 2 is a schematic frame structure diagram of an internal structure of an actual curing oven according to an embodiment of the present invention;
FIG. 3 is a view showing an internal structure of an actual rapid curing oven system according to an embodiment of the present invention;
FIG. 4 is a layout diagram of sensors on a lead frame for data acquisition as set forth in an embodiment of the present invention;
FIG. 5 is a schematic diagram illustrating a comparison between a predicted time coefficient and an actual time coefficient according to an embodiment of the present invention;
FIG. 6 is a graph showing the actual temperature profile of a curing oven system according to an embodiment of the present invention;
FIG. 7 is a distribution diagram of the absolute relative error ARE in the embodiment of the present invention;
FIG. 8 is a graph comparing the predicted and actual temperature T distribution at the sensor 6 using an initial off-line model in an embodiment of the present invention;
FIG. 9 is a graph of the predicted and actual absolute relative error ARE at the sensor 6 using an initial off-line model in an embodiment of the present invention;
FIG. 10 is a graph comparing predicted and actual temperature distributions on the sensor 6 using an online update model in an embodiment of the present invention;
FIG. 11 is a graph of the predicted and actual absolute relative error ARE at the sensor 6 using an online update model in an embodiment of the present invention;
FIG. 12 is a graph comparing predicted and actual temperature distributions on the sensor 11 using an initial off-line model in an embodiment of the present invention;
FIG. 13 is a graph of the predicted and actual absolute relative error ARE at the sensor 11 using an initial off-line model in an embodiment of the present invention;
FIG. 14 is a graph comparing predicted and actual temperature distributions on the sensor 11 using an online update model in an embodiment of the present invention;
FIG. 15 is a schematic diagram of the absolute relative error ARE predicted at the sensor 11 using the online update model of the present application in an embodiment of the present invention;
FIG. 16 is a schematic diagram of the absolute relative error ARE predicted by the online update model according to the embodiment of the present invention;
FIG. 17 is a schematic representation of the absolute relative error ARE of the OS-ELM based spatio-temporal model prediction in an embodiment of the present invention;
FIG. 18 is a comparison between simulation times using the method of the present application and the OS-ELM-based method in an embodiment of the present invention.
Detailed Description
The drawings are for illustrative purposes only and are not to be construed as limiting the patent;
for the purpose of better illustrating the embodiments, certain features of the drawings may be omitted, enlarged or reduced, and do not represent the size of an actual product; it will be understood by those skilled in the art that certain well-known structures in the drawings and descriptions thereof may be omitted. The technical solution of the present invention is further described below with reference to the accompanying drawings and examples.
Example 1
An online spatiotemporal modeling method integrating a forgetting mechanism and a double-overrun learning machine is disclosed, a flow diagram is shown in figure 1, and the online spatiotemporal modeling method comprises the following steps:
s1, decoupling a space-time output variable into a time coefficient sequence and a space sequence according to a nonlinear partial differential equation of a curing heat process of a curing oven of a nonlinear distributed parameter system; the method specifically comprises the following steps: the nonlinear partial differential equation of the curing process of the curing oven of the nonlinear distributed parameter system is as follows:
where T (S, T) represents (x, y, z) ∈ (0, S) at time T and position S ═ x, y, z0) P represents a density, k represents a thermal coefficient, c represents a specific heat coefficient, f (T (S, T)) is an unknown nonlinear thermal dynamic associated with the space-time output variable T, Q (S, T) represents a heat source,representing the Laplace space operator, T0(S,0) representing an initial spatiotemporal output variable;
the spatial-temporal output variable T (S, T) is decoupled into a time coefficient sequence and a space sequence, and the expression is as follows:
wherein, ai(t) the ith time coefficient of the time coefficient series ofφi(S) denotes the ith spatial position, the spatial sequence BFSs is
The expression for the ith time coefficient is:
ai(t)=gi(ai(t-1))+hi(u(t-1))
wherein, gi(. and h)iBoth refer to non-linear functions.
S2, establishing a first ultralimit learning machine model and a second ultralimit learning machine model based on the time coefficient sequence; the expression of the first overrun learning machine model is as follows:
the expression of the second ultralimit learning machine model is:
wherein, βσOutput weight representing a connection output node and a hidden node in the first ultralimit learner model, β'Output weight, ω, representing the connection of output node and hidden node in the second ultralimit learning machine modelσRepresenting input weights, ω 'connecting the output node and the hidden node in the first ultralimit learner model'Representing input weights connecting the output node and the hidden node in the second ultralimit learning machine model, sigma representing the sigma-th hidden node in the first ultralimit learning machine model, N1Representing the number of hidden nodes in the first ultralimit learning machine model, N2Representing the number of hidden nodes in the second ultralimit learning machine model, G1Representing an activation function of a hidden layer in the first ultralimit learning machine model; g2Representing the activation function of the hidden layer in the second ultralimit learning machine model, ησThreshold representing hidden node in first ultralimit learner model, η'A threshold value representing a hidden node in the second ultralimit learning machine model, u (t) a time coefficient in a time coefficient sequence based on which the second ultralimit learning machine model is established, gi(. and h)i(. two non-linear functions satisfy: gi(·)=g(·),hi(·)=h(·)。
Based on the expressions of the first and second ultralimit learning machine models, the expression of the time coefficient a (t) is further expressed as:
at this time, the input weight ω connecting the output node and the hidden node in the first ultralimit learning machine modelσInput weight ω 'connecting output node and hidden node in second ultralimit learner model'Threshold η for hidden nodes in the first ultralimit learning machine modelσAnd a threshold η 'of hidden nodes in the second ultralimit learner model'Are randomly generated, independent of the training data and independent of each other, and output weights β connecting the output nodes and hidden nodes in the first ultralimit learning machine modelσAnd β 'as output weights connecting the output node and the hidden node in the second ultralimit learner model'Determining according to the input and output data; the further expression of the time coefficient a (t) is omega generated randomly according to the mathematical description of a double-overrun learning machine designed by coupling double nonlinear structuresσ,ω′,ησAnd η'Once estimated, these values will be fixed during subsequent learning.
Further expressed as: a (t) ═ hT(t)θ
Wherein h (t) represents parameter vectors of the first and second ultralimit learning machine models related to the input-output data, and the expression is as follows:
theta represents unknown parameter vectors of the first and second ultralimit learning machine models to be identified,the matrix form of the time coefficients a (t) is:
A=Hθ
wherein a ═ a (2), a (3),.., a (l)]TAn output vector form representing the time coefficient a (t), L representing the data length of the phasor; h ═ HT(2) L hT(L)]TRepresenting a regression matrix;
Wherein,is the Moore-Penrose generalized inverse of matrix H,representing the unknown parameter vector taken from a and H.
Since the output vector form A and the regression matrix H of the time coefficients a (t) are known, only H is considered hereTH is a non-singular condition, satisfies
S3, updating model parameters of the first ultralimit learning machine model and the second ultralimit learning machine model by utilizing an online sequential learning algorithm of an integrated forgetting mechanism; first, the initial training set block is:the initial model parameters of the first and second ultralimit learning machines areSubscript 0 denotes the block with the initial training setThe associated initial vector or matrix, set to input/output data { z (L)0+1),a(L0+1) } arrive, then:
updated parameter θ using MP generalized inverse matrix1To write as:
wherein,will theta0Is denoted as P1、h(L0 +1) And a (L)0+1), then:wherein,by using the formula of the Woodbury,
P1=((P0)-1+h(L0+1)hT(L0+1))-1
=P0-P0h(L0+1)(I+hT(L0+1)P0h(L0+1))-1hT(L0+1)P0
when the m +1 th data { z (L)0+m+1),a(L0+ m +1) } arrive,
wherein,i.e. according to a known parameter thetamAnd newly arrived data z (L)0+m+1),a(L0+ m +1) }, update the model parameter θm+1。
Setting the new input-output data as: { z (L)0+1),a(L0+1), { z (2), a (2) } indicates old input-output data { z (2), a (2) } with respect to new input-output data, then:
based on MP generalized inverse matrix and forgetting mechanism, parameter theta1Comprises the following steps:
when the m +1 th data { z (L)0+m+1),a(L0+ m +1) } arrives, discarding the oldest data { z (m +2), a (m +2) }, then:
based on the deduction process, the step of updating the model parameters of the first and second ultralimit learning machine models by using an online sequential learning algorithm integrated with a forgetting mechanism specifically comprises the following steps:
s301, initialization: setting the sequence m of the number of training data as zero, selecting the activation function G of the hidden layer in the first ultralimit learning machine model1Activation function G of hidden layer in second ultralimit learning machine model2The number N of hidden nodes in the first ultralimit learning machine model1The number N of hidden nodes in the second ultralimit learning machine model2Randomly generating a connection output in a first ultralimit learning machine modelInput weights ω of nodes and hidden nodesσInput weight ω 'connecting the output node and the hidden node in the second ultralimit learner model'Threshold η for hidden nodes in the first ultralimit learning machine modelσThreshold η 'of hidden node in second ultralimit learner model';
Calculating an initial output matrix of a hidden layer of the first ultralimit learning machine model and the second ultralimit learning machine model:
H0=[hT(2) L hT(L0)]Tcalculating initial unknown model parameters: theta0=P0H0 TA0
Wherein the initial unknown model parameter is the initial output weight of the model, P0=(H0 TH0)-1,A0=[a(2),a(3),...,a(L0)]T;
S302, performing online sequential learning of an integrated forgetting mechanism:
when the m +1 th data { z (L)0+m+1),a(L0+ m +1) } arrive, the m +1 activation function is computed:
s304, utilizing the calculated model output weight to evaluate the first ultralimit learning machine model and the second ultralimit learning machine model according to the condition that A is H theta;
s305, increasing the value of m by 1, and returning to the step S302.
And S4, integrating the spatial sequence with the updated first and second ultralimit learning machine models to reconstruct an online space-time model.
The expression of the reconstructed online spatio-temporal model is:
wherein,representing a time coefficient, phi, in a time series after updating by integrating a forgetting mechanism and a dual-overrun learning machinei(S) represents a spatial sequence;and representing the space-time output variable of the reconstructed online space-time model.
In addition, when the model parameters of the first ultralimit learning machine model and the second ultralimit learning machine model are updated by utilizing the online sequential learning algorithm of the integrated forgetting mechanism, the training data are received one by one or block by block, and because the training data are received one by one or block by block, if the training data arrive one by block, the training data arrive
Initializing training data L0The number of the first overrun learning model and the number of the hidden neurons of the second overrun learning model are not less than the sum of the numbers of the hidden neurons of the first overrun learning model and the second overrun learning model, so that the initial regression matrix H is ensured0The rank of (2) is the sum of the number of hidden layers of the two over-limit learning machine models.
Next, a real-time test is performed on the curing oven by combining the above-mentioned specific online space-time modeling method, such as a structure diagram of an actual rapid curing oven system shown in fig. 2, and fig. 3 is a structure diagram of a corresponding internal principle frame of the curing oven, where 1 denotes a heater, 2 denotes a sensor, 3 denotes a chamber, and 4 denotes a lead frame. Referring to FIG. 3, the top of the curing oven has four identical heaters 1, pulse width modulated(PWM) signal control, see FIG. 4, will S1~S16A total of 16 thermocouples were placed uniformly on the lead frame 4 as sensors to collect data, and these sensors 2 collected about 2100 time-series data for model training and online learning within a fixed sampling interval Δ t of 10 s.
In order to facilitate the verification of the model established by the online spatio-temporal modeling method proposed by the application and the comparison of other modeling methods, the following error indexes are given:
1) space-time prediction error (e)
2) Absolute Relative Error (ARE)
3) Root Mean Square Error (RMSE)
4) Time Normalized Absolute Error (TNAE)
In the online spatio-temporal modeling, collected samples are divided into two parts, wherein the first 800 samples are used for constructing an initial spatio-temporal model, and the last 1300 samples are used for online learning.
First, an initial spatio-temporal model is constructed. Calculating a space sequence BFs by using a KL method, selecting the order to be 3, and then projecting space-time samples onto BFs to obtain a time coefficientThe low order temporal model is estimated using the temporal coefficients and the corresponding input signals. And finally, reconstructing an initial space-time model by using a space-time synthesis method.
To evaluate the performance of the initial model, the last 1300 samples were tested and the actual time coefficient and the predicted time coefficient were compared in fig. 5, which showed very satisfactory agreement between the predicted and the measured dynamics. In addition, the predicted temperature distribution and the corresponding ARE index for the 2100 th sample were simulated, the temperature profile is shown in fig. 6, and the profile of the ARE index is shown in fig. 7. As can be seen from fig. 6 and 7, the initial spatio-temporal model is a good approximation of the actual DPS.
In the online learning process, 1300 samples are assumed to arrive block by block, and the size is fixed and set to 10, so that 1300 samples can be divided into 130 continuous blocks. The fixed step is set to 1400 samples, i.e. once the number of training samples reaches 1400, the oldest sample is deleted using a forgetting mechanism, leaving the number of samples equal to 1400. In order to evaluate the model performance of the proposed online method and the initial offline model, in a specific implementation, it was simulated that the distribution of predicted and actual temperature T at the sensor 6 using the initial offline model is shown in fig. 8, the absolute relative error ARE is shown in fig. 9, the distribution of predicted and actual temperature T at the sensor 6 using the online update model is shown in fig. 10, the distribution of predicted and actual absolute relative error ARE at the sensor 6 using the online update model is shown in fig. 11, the distribution of predicted and actual temperature T at the sensor 11 using the initial offline model is shown in fig. 12, the absolute relative error ARE predicted and actual at the sensor 11 using the initial offline model is shown in fig. 13, the distribution of predicted and actual temperature T at the sensor 11 using the online update model is shown in fig. 14, a graph of the absolute relative error ARE predicted and actual on the sensor 11 by using the online updating model is shown in fig. 15, and by comparing the curves, it can be seen that the online updating model provided by the present application can better match the dynamic change of the actual temperature, and the effectiveness of the method provided by the present application is also verified.
In order to verify the superiority of the model establishment method of the integrated forgetting mechanism proposed in the present application, fig. 16 shows an error index ARE distribution diagram of a 2100 th test sample predicted by using the online update model of the present application, and fig. 17 shows a schematic diagram of an absolute relative error ARE predicted by using the spatio-temporal model based on OS-ELM, as can be seen from fig. 16 and 17, both online models exhibit better model performance than the initial offline model, however, the model established by the method of the present application has higher model accuracy than the spatio-temporal model based on OS-ELM.
To further compare the performance of the models constructed by the method of the present application with the OS-ELM model, table 1 shows the performance comparison data of the models constructed by the method of the present application based on TNAE with the OS-ELM model, and table 2 shows the performance comparison data of the models constructed by the method of the present application based on RMSE with the OS-ELM model.
TABLE 1
TABLE 2
As can be seen from tables 1 and 2, the model constructed by the method provided by the present application has better model performance. Fig. 18 is a schematic diagram comparing simulation time between the method of the present application and the OS-ELM-based method, and it can be seen from fig. 18 that embedding a forgetting mechanism into online space-time learning can improve learning efficiency, save computation, and release memory space.
The terms describing positional relationships in the drawings are for illustrative purposes only and are not to be construed as limiting the patent;
it should be understood that the above-described embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Other variations and modifications will be apparent to persons skilled in the art in light of the above description. And are neither required nor exhaustive of all embodiments. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention should be included in the protection scope of the claims of the present invention.
Claims (10)
1. An online space-time modeling method integrating a forgetting mechanism and a double-overrun learning machine is characterized by at least comprising the following steps:
s1, decoupling a space-time output variable into a time coefficient sequence and a space sequence according to a nonlinear partial differential equation of a curing heat process of a curing oven of a nonlinear distributed parameter system;
s2, establishing a first ultralimit learning machine model and a second ultralimit learning machine model based on the time coefficient sequence;
s3, updating model parameters of the first ultralimit learning machine model and the second ultralimit learning machine model by utilizing an online sequential learning algorithm of an integrated forgetting mechanism;
and S4, integrating the spatial sequence with the updated first and second ultralimit learning machine models to reconstruct an online space-time model.
2. The online space-time modeling method integrating forgetting mechanism and dual-overrun learning machine as claimed in claim 1, wherein the nonlinear partial differential equation of the curing heat process of the nonlinear distributed parameter system curing oven in step S1 is:
where T (S, T) represents (x, y, z) ∈ (0, S) at time T and position S ═ x, y, z0) P represents a density, k represents a thermal coefficient, c represents a specific heat coefficient, f (T (S, T)) is an unknown nonlinear thermal dynamic associated with the space-time output variable T, Q (S, T) represents a heat source,representing the Laplace space operator, T0(S,0) representing an initial spatiotemporal output variable;
the spatial-temporal output variable T (S, T) is decoupled into a time coefficient sequence and a space sequence, and the expression is as follows:
wherein, ai(t) the ith time coefficient of the time coefficient series ofφi(S) denotes the ith spatial position, the spatial sequence BFSs is
The expression for the ith time coefficient is:
ai(t)=gi(ai(t-1))+hi(u(t-1))
wherein, gi(. and h)iBoth refer to non-linear functions.
3. The online spatiotemporal modeling method integrating a forgetting mechanism and a dual-ultralimit learning machine according to claim 2, characterized in that the expression of the first ultralimit learning machine model in step S2 is:
the expression of the second ultralimit learning machine model is:
wherein, βσOutput weights representing connecting output nodes and hidden nodes in the first ultralimit learning machine model,β′Output weight, ω, representing the connection of output node and hidden node in the second ultralimit learning machine modelσRepresenting input weights, ω 'connecting the output node and the hidden node in the first ultralimit learner model'Representing input weights connecting the output node and the hidden node in the second ultralimit learning machine model, sigma representing the sigma-th hidden node in the first ultralimit learning machine model, N1Representing the number of hidden nodes in the first ultralimit learning machine model, N2Representing the number of hidden nodes in the second ultralimit learning machine model, G1Representing an activation function of a hidden layer in the first ultralimit learning machine model; g2Representing the activation function of the hidden layer in the second ultralimit learning machine model, ησThreshold representing hidden node in first ultralimit learner model, η'A threshold value representing a hidden node in the second ultralimit learning machine model, u (t) a time coefficient in a time coefficient sequence based on which the second ultralimit learning machine model is established, gi(. and h)i(. two non-linear functions satisfy: gi(·)=g(·),hi(·)=h(·)。
4. The online spatiotemporal modeling method integrating a forgetting mechanism and a dual ultralimit learning machine according to claim 3, characterized in that based on the expressions of the first ultralimit learning machine model and the second ultralimit learning machine model, the expression of the time coefficient a (t) is further expressed as:
at this time, the input weight ω connecting the output node and the hidden node in the first ultralimit learning machine modelσInput weight ω 'connecting output node and hidden node in second ultralimit learner model'Threshold η for hidden nodes in the first ultralimit learning machine modelσAnd a threshold η 'of hidden nodes in the second ultralimit learner model'Are randomly generated, are independent of training data and are independent of each other, and the output of the first ultralimit learning machine model is connected with the output node and the hidden nodeOut weight βσAnd β 'as output weights connecting the output node and the hidden node in the second ultralimit learner model'Determining according to the input and output data;
further expressed as: a (t) ═ hT(t)θ
Wherein h (t) represents parameter vectors of the first and second ultralimit learning machine models related to the input-output data, and the expression is as follows:
theta represents unknown parameter vectors of the first and second ultralimit learning machine models to be identified,the matrix form of the time coefficients a (t) is:
A=Hθ
wherein a ═ a (2), a (3),.., a (l)]TAn output vector form representing the time coefficient a (t), L representing the data length of the phasor; h ═ HT(2) L hT(L)]TRepresenting a regression matrix;
5. The online spatiotemporal modeling method integrating the forgetting mechanism and the dual ultralimit learning machine according to claim 4, characterized in that the initial training set block is:the initial model parameters of the first and second ultralimit learning machines areSubscript 0 denotes the block with the initial training setThe associated initial vector or matrix, set to input/output data { z (L)0+1),a(L0+1) } arrive, then:
updated parameter θ using MP generalized inverse matrix1To write as:
wherein,will theta0Is denoted as P1、h(L0+1) and a (L)0+1), then:wherein,by using the formula of the Woodbury,
P1=((P0)-1+h(L0+1)hT(L0+1))-1
=P0-P0h(L0+1)(I+hT(L0+1)P0h(L0+1))-1hT(L0+1)P0
when the m +1 th data { z (L)0+m+1),a(L0+ m +1) } arrive,
6. The online spatiotemporal modeling method integrating the forgetting mechanism and the dual-ultralimit learning machine according to claim 5, characterized in that the new input-output data are set as follows: { z (L)0+1),a(L0+1), { z (2), a (2) } indicates old input-output data { z (2), a (2) } with respect to new input-output data, then:
based on MP generalized inverse matrix and forgetting mechanism, parameter theta1Comprises the following steps:
when the m +1 th data { z (L)0+m+1),a(L0+ m +1) } arrives, discarding the oldest data { z (m +2), a (m +2) }, then:
7. the online spatiotemporal modeling method integrating the forgetting mechanism and the dual ultralimit learning machine as claimed in claim 6, wherein the step S3 of updating the model parameters of the first ultralimit learning machine model and the second ultralimit learning machine model by using the online sequential learning algorithm of the forgetting mechanism comprises:
s301, initialization: setting the sequence m of the number of training data as zero, selecting the activation function G of the hidden layer in the first ultralimit learning machine model1Activation function G of hidden layer in second ultralimit learning machine model2The number N of hidden nodes in the first ultralimit learning machine model1The number N of hidden nodes in the second ultralimit learning machine model2Randomly generating input weights omega for connecting the output node and the hidden node in the first ultralimit learning machine modelσInput weight ω 'connecting the output node and the hidden node in the second ultralimit learner model'Threshold η for hidden nodes in the first ultralimit learning machine modelσThreshold η 'of hidden node in second ultralimit learner model';
Calculating an initial output matrix of a hidden layer of the first ultralimit learning machine model and the second ultralimit learning machine model:
H0=[hT(2) L hT(L0)]Tcalculating the initial unknownsModel parameters: theta0=P0H0 TA0
Wherein the initial unknown model parameter is the initial output weight of the model, P0=(H0 TH0)-1,A0=[a(2),a(3),...,a(L0)]T;
S302, performing online sequential learning of an integrated forgetting mechanism:
when the m +1 th data { z (L)0+m+1),a(L0+ m +1) } arrive, the m +1 activation function is computed:
s304, utilizing the calculated model output weight to evaluate the first ultralimit learning machine model and the second ultralimit learning machine model according to the condition that A is H theta;
s305, increasing the value of m by 1, and returning to the step S302.
8. The online spatiotemporal modeling method integrating a forgetting mechanism and a dual ultralimit learning machine according to claim 7, characterized in that training data is received one by one or block by block when updating model parameters of the first ultralimit learning machine model and the second ultralimit learning machine model by an online sequential learning algorithm of the forgetting mechanism.
9. The method of claim 8, wherein the training data L is initialized0The number of the first and second overrun learning models is not less than the sum of the number of hidden neurons of the first and second overrun learning models。
10. The method for online spatiotemporal modeling with integration of forgetting mechanism and dual ultralimit learning machine according to claim 9, wherein the expression of the reconstructed online spatiotemporal model of step S4 is:
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010450196.XA CN111625995B (en) | 2020-05-25 | 2020-05-25 | Online time-space modeling method integrating forgetting mechanism and double ultralimit learning machines |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202010450196.XA CN111625995B (en) | 2020-05-25 | 2020-05-25 | Online time-space modeling method integrating forgetting mechanism and double ultralimit learning machines |
Publications (2)
Publication Number | Publication Date |
---|---|
CN111625995A true CN111625995A (en) | 2020-09-04 |
CN111625995B CN111625995B (en) | 2022-06-24 |
Family
ID=72259050
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202010450196.XA Active CN111625995B (en) | 2020-05-25 | 2020-05-25 | Online time-space modeling method integrating forgetting mechanism and double ultralimit learning machines |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN111625995B (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114548368A (en) * | 2022-01-20 | 2022-05-27 | 广东产品质量监督检验研究院(国家质量技术监督局广州电气安全检验所、广东省试验认证研究院、华安实验室) | Modeling method, prediction model and prediction method of lithium battery temperature field prediction model based on multilayer nuclear overrun learning machine |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108717505A (en) * | 2018-05-29 | 2018-10-30 | 广东工业大学 | A kind of solidification thermal process space-time modeling method based on K-RVFL |
CN108763759A (en) * | 2018-05-29 | 2018-11-06 | 广东工业大学 | A kind of solidification thermal process space-time modeling method based on ISOMAP |
CN109145346A (en) * | 2018-05-29 | 2019-01-04 | 广东工业大学 | Solidification thermal process space-time modeling method based on dual least square method supporting vector machine |
CN110377942A (en) * | 2019-06-10 | 2019-10-25 | 广东工业大学 | A kind of multi-model space-time modeling method based on limited gauss hybrid models |
-
2020
- 2020-05-25 CN CN202010450196.XA patent/CN111625995B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108717505A (en) * | 2018-05-29 | 2018-10-30 | 广东工业大学 | A kind of solidification thermal process space-time modeling method based on K-RVFL |
CN108763759A (en) * | 2018-05-29 | 2018-11-06 | 广东工业大学 | A kind of solidification thermal process space-time modeling method based on ISOMAP |
CN109145346A (en) * | 2018-05-29 | 2019-01-04 | 广东工业大学 | Solidification thermal process space-time modeling method based on dual least square method supporting vector machine |
CN110377942A (en) * | 2019-06-10 | 2019-10-25 | 广东工业大学 | A kind of multi-model space-time modeling method based on limited gauss hybrid models |
Non-Patent Citations (3)
Title |
---|
张兴宇等: "基于ELM的芯片固化炉炉温建模方法", 《制造业自动化》 * |
郭威等: "具有广义正则化与遗忘机制的在线贯序超限学习机", 《控制与决策》 * |
马致远等: "在线增量极限学习机及其性能研究", 《计算机应用研究》 * |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN114548368A (en) * | 2022-01-20 | 2022-05-27 | 广东产品质量监督检验研究院(国家质量技术监督局广州电气安全检验所、广东省试验认证研究院、华安实验室) | Modeling method, prediction model and prediction method of lithium battery temperature field prediction model based on multilayer nuclear overrun learning machine |
CN114548368B (en) * | 2022-01-20 | 2022-10-11 | 广东产品质量监督检验研究院(国家质量技术监督局广州电气安全检验所、广东省试验认证研究院、华安实验室) | Modeling method and prediction method of lithium battery temperature field prediction model based on multilayer nuclear overrun learning machine |
Also Published As
Publication number | Publication date |
---|---|
CN111625995B (en) | 2022-06-24 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111090945B (en) | Actuator and sensor fault estimation design method for switching system | |
US9489620B2 (en) | Quick analysis of residual stress and distortion in cast aluminum components | |
CN110377942B (en) | Multi-model space-time modeling method based on finite Gaussian mixture model | |
WO2017031718A1 (en) | Modeling method of deformation motions of elastic object | |
CN111539132B (en) | Dynamic load time domain identification method based on convolutional neural network | |
CN114548368B (en) | Modeling method and prediction method of lithium battery temperature field prediction model based on multilayer nuclear overrun learning machine | |
CN109709934B (en) | Fault diagnosis redundancy design method for flight control system | |
CN109885916B (en) | Mixed test online model updating method based on LSSVM | |
JP6730340B2 (en) | Causal estimation device, causal estimation method, and program | |
CN114117840A (en) | Structural performance prediction method based on simulation and test data hybrid drive | |
Castañeda et al. | Decentralized neural identifier and control for nonlinear systems based on extended Kalman filter | |
Fu et al. | Improved unscented Kalman filter algorithm-based rapid identification of thermal errors of machine tool spindle for shortening thermal equilibrium time | |
CN111625995B (en) | Online time-space modeling method integrating forgetting mechanism and double ultralimit learning machines | |
CN115972211A (en) | Control strategy offline training method based on model uncertainty and behavior prior | |
CN110007617B (en) | Uncertainty transmission analysis method of aircraft semi-physical simulation system | |
WO2024139626A1 (en) | Deformation state analysis method for flexible linear body, and related apparatus | |
CN113570165A (en) | Coal reservoir permeability intelligent prediction method based on particle swarm optimization | |
CN116070437B (en) | Lithium battery surface temperature modeling method, device, computer equipment and storage medium | |
CN110084324B (en) | Kalman filtering parameter self-adaptive updating method based on extreme learning machine | |
CN116362140A (en) | Sensor error compensation method, system, device and medium based on fusion of transducer and GRU | |
JP6919856B2 (en) | Reinforcement learning programs, reinforcement learning methods, and reinforcement learning devices | |
CN115481554A (en) | Thermal diffusion digital twin model, temperature field real-time optimization control model and method in explosive fusion casting and curing process | |
CN114880921A (en) | Heat source system temperature field real-time reconstruction method based on limited measuring points | |
CN114186477A (en) | Elman neural network-based orbit prediction algorithm | |
CN107305348B (en) | Dynamic system delay calculation method based on dependency measurement |
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 |