| CPC G01K 11/22 (2013.01) [G01K 13/00 (2013.01)] | 8 Claims |
|
1. An application specific integrated circuit (ASIC) for a preset deep learning network, the ASIC comprising:
a processor including a plurality of neurons, wherein each neuron comprises a microprocessor;
a memory for storing a program, when executed by the processor, to perform a method for detecting temperature rise inside a superconducting levitation device based on deep learning;
a multimedia component;
an I/O interface; and
a communication component configured for wired or wireless communication in the ASIC;
wherein the method for detecting temperature rise inside a superconducting levitation device based on deep learning comprises:
a) collecting, by a vibration acceleration sensor, an initial vibration acceleration information and a vibration acceleration detection information of the superconducting levitation device, and collecting, by a temperature sensor, an initial temperature rise information of a superconductor provided in the superconducting levitation device;
b) extracting, by the processor, features from the initial vibration acceleration information to obtain a high-frequency feature parameter set and a low-frequency feature parameter set, comprising:
obtaining a scale factor and a translation factor of a wavelet transform;
generating a plurality of decomposition feature parameters based on the scale factor, the translation factor and the initial vibration acceleration information through a preset wavelet decomposition model, wherein the preset wavelet decomposition model is represented by:
 wherein wf(a, b) represents a single decomposition feature parameter, a is the scale factor, b is the translation factor, f(t) represents the initial vibration acceleration information, ψ is a complex conjugate of ψ, ψ represents a preset basic wavelet, and t is a time variable; and
sorting the plurality of decomposition feature parameters in sequence to obtain the high-frequency feature parameter set and the low-frequency feature parameter set by:
obtaining a number of branches and a number of levels of a complete binary tree constructed from the plurality of decomposition feature parameters;
configuring a first decomposition feature parameter among the plurality of decomposition feature parameters as a heap top;
inserting remaining decomposition feature parameters among the plurality of decomposition feature parameters sequentially based on the number of branches and the number of levels to obtain an unsorted decomposition feature parameter heap;
sorting the unsorted decomposition feature parameter heap based on a preset min-heap model to obtain a sorted decomposition feature parameter heap; and
dividing the sorted decomposition feature parameter heap based on a preset frequency threshold to obtain the high-frequency feature parameter set and the low-frequency feature parameter set;
c) generating, by the processor, a wavelet band energy information using the high-frequency feature parameter set and the low-frequency feature parameter set by:
creating a plurality of first wavelet basis points respectively corresponding to a plurality of high-frequency feature parameters in the high-frequency feature parameter set using a first preset wavelet basis function;
obtaining a high-frequency band energy set based on the plurality of first wavelet basis points through the following formula:
 wherein EH is the high-frequency band energy set, and {b1, b2, . . . , br} represent the plurality of first wavelet basis points, respectively, and b is an integer greater than 2;
creating a plurality of second wavelet basis points respectively corresponding to a plurality of low-frequency feature parameters in the low-frequency feature parameter set using a second preset wavelet basis function;
obtaining a low-frequency band energy set based on the plurality of second wavelet basis points through the following formula:
 wherein EL is the low-frequency band energy set, and {a1, a2, . . . , as} respectively represent the plurality of second wavelet basis points, and s is an integer greater than 2; and
generating the wavelet band energy information based on the high-frequency band energy set and the low-frequency band energy set using a preset normalization model, wherein the preset normalization model is represented by:
 wherein An is the wavelet band energy information corresponding to n feature parameters, and En2 is a band energy corresponding to n feature parameters;
d) training, by the processor, the preset deep learning network based on the wavelet band energy information and the initial temperature rise information to generate an internal temperature rise detection model of the superconducting levitation device, wherein the preset deep learning network consists of multiple error back propagation neural networks, the training comprising:
training each of the error back propagation neural networks error back propagation neural networks to obtain prediction error information;
producing a predicted sequence weight information corresponding to the prediction error information using a preset weight prediction model; and
constructing a classification model in the internal temperature rise detection model of the superconducting levitation device based on the error back propagation neural networks and the predicted sequence weight information; and
e) generating, by the processor, an internal temperature rise prediction information of the superconducting levitation device using the internal temperature rise detection model based on the vibration acceleration detection information, wherein the internal temperature rise prediction information shows a real-time temperature rise of the superconductor.
|