Device and method to predict the onset of oscillatory instabilities in systems with turbulent flow

Patent metadata

Field	Value
Publication number	US-2022018731-A1
Application number	US-202117377918-A
Country	US
Kind code	A1
Filing date	Jul 16, 2021
Priority date	Jul 16, 2020
Publication date	Jan 20, 2022
Grant date	—

Abstract

Disclosed are devices and methods of detecting and mitigating oscillatory instabilities in systems with turbulent flow, such as thermoacoustic, aeroacoustic or aeroelastic equipment. The system includes a sensor array for measuring one or more parameters of an operating equipment S, and an analysis and prediction unit. The analysis and prediction unit is configured to estimate a state tensor to identify a state of the equipment indicating stable operation or impending oscillatory instability. The system further includes an actuator array configured to implement a control action to promote stable operation of the equipment. Methods for robust prediction of the state of stability are also disclosed. A neural ordinary differential equation (ODE) method of predicting stability or instability is disclosed, involving forming a neural network that incorporates an equation characteristic of the operational state. The invention further discloses a hybrid convolutional neural network based prediction method for stability.

First claim

What is claimed is:

1 . A system for predicting and controlling onset of oscillatory instability in a turbulent flow equipment, the system comprising:
a sensor array comprising at least one sensor of a piezoelectric pressure transducer, a microphone or a strain gauge to measure analog signals; an analog to digital converter for converting the analog signals to a digital signal; an analyzing unit, configured to estimate a state tensor including multidimensional data from the digital signal, or to implement a control action to promote stable operation of the equipment; and a learning and prediction unit configured for classifying and recognizing patterns present in multidimensional data to identify potential onset of an unstable operating state of the equipment.

2 . The system of claim 1 , further comprising an actuator array present on the equipment, the actuator array configured to implement the control action to promote a stable operating state of the equipment.

3 . The system of claim 2 , wherein the equipment is a thermoacoustic equipment, an aeroacoustic equipment or an aeroelastic equipment.

4 . The system of claim 3 , wherein the equipment is a thermoacoustic or an aeroacoustic equipment, and the actuator array comprises one or more of cooling holes, staged fuel injectors, micro swirlers, water injectors, or dampers.

5 . The system of claim 3 , wherein the equipment is an aeroelastic equipment, and the actuator array comprises counterweights, fins, micothrustors, flywheels or gyros.

6 . A method of predicting and mitigating occurrence of a potentially unstable operating state of an equipment, the method comprising:
measuring, by one or more sensors in a sensor array, a signal indicative of operating state of the equipment to obtain a measured signal; converting, by analog to digital converter the measured signal into a digital signal; estimating, by an analyzing unit, a state tensor from the digital signal; receiving the state tensor at a learning and prediction unit, the unit having a neural network prediction module with a dynamic model of the operating state of the equipment; calculating a stability parameter corresponding to the received state tensor using the dynamic model; updating the dynamic model with the state tensor; comparing the stability parameter against a threshold; and identifying an impending onset of instability in the operating state of the equipment.

7 . The method of claim 6 , further comprising:
sending the updated dynamic model to the analyzing unit; determining, by the analyzing unit an optimal path for control and control action to mitigate the impending stability; sending the control action to a control unit; and modifying, by the control unit via an actuator array, the operating state of the equipment to mitigate the impending instability.

8 . The method of claim 6 , wherein the neural network method is one of a neural ordinary differential equation (ODE) method, or a hybrid convolutional neural network method.

9 . The method of claim 8 , wherein the neural network is an ODE neural network and the method comprises:
a) constructing a neural network (NN) with at least 3 layers, that maps state variables p(t) of an operational state to a dynamic state variable z(t) and to its time derivative ż, wherein ż=f (z,t,w); b) defining the NN with nodes in each layer and specifying a corresponding activation function; c) receiving weighted inputs corresponding to time series data at times t; d) computing sum of the weighted inputs at each node for initial time t 0 , to a final time instance t 1 for the corresponding operational states to obtain a weight matrix; e) computing a distance in the weight matrix defined as d (i) =| w (i) −w (r) |, wherein w (i) is an averaged weight over a window of n, samples, w is an average weight corresponding to an unstable state; and f) computing a sum of the distances and obtain a parameter μ 1 that varies inversely as the sum, to obtain a measure of stability of the operational state.

10 . The method of claim 9 , wherein in step d) the final state predicted by neural ODE, z p (t 1 ) is given by:
z p ( t 1 )=∫ t 0 t 1 f ( z ( t ), t,w ) dt , wherein,
an error in the prediction is given by:
J =( z ( t 1 )− z p ( t 1 )) 2 .

11 . The method of claim 8 , comprising, prior to step d) training the neural network using weights learned from a first sample data w (1) as initial guess to find weights for subsequent samples;
training the neural network on i th sample data to get corresponding weights of NN as w (i) ; and averaging the weights over a window of n w samples,

w
_

(
i
)

=

1

n
w

⁢

∑

j
=

(

i
-

n
w

+
1

)

j
=
i

⁢

w

(
j
)

.

12 . The method of claim 8 , wherein the neural network is a hybrid convolutional neural network (CNN) and the method comprises:
providing an input adjacency matrix of nodes to construct the convolutional neural network; providing two or more convolutional layers in the CNN with corresponding activation functions; reconstructing a phase space corresponding to P(t) l corresponding to time series data segment of length l to obtain a recurrence plot (RP); obtaining a recurrence matrix, R ij defined as R ij =∥X i −X j |, where X i and X j represent state points of the operational state in reconstructed phase space at i th and j th instances in time; obtaining RPs corresponding to stable and unstable operational states; training the convolutional neural network (CNN) using the RPs for known stable and unstable states; and using the trained CNN to assign probabilities to unknown time series data

Assignees

• Dhadphale Jayesh
• Ruiz Eustaquio Aguilar
• Unni Vishnu Rajasekharan
• Saha Abhishek
• Nair Sujith Indusekharan
• Indian Inst Tech Madras
• Univ Of California San Diego

Inventors

• Dhadphale Jayesh
• Ruiz Eustaquio Aguilar
• Unni Vishnu Rajasekharan
• Saha Abhishek
• Nair Sujith Indusekharan

Classifications

• G06N3/048

  Activation functions · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title

• G06N3/09

  Supervised learning · CPC title