Adaptive filter based learning model for time series sensor signal classification on edge devices

Patent metadata

Field	Value
Publication number	US-12462199-B2
Application number	US-202117156821-A
Country	US
Kind code	B2
Filing date	Jan 25, 2021
Priority date	Apr 17, 2020
Publication date	Nov 4, 2025
Grant date	Nov 4, 2025

Abstract

This disclosure relates generally to method and system for an adaptive filter based learning model for time series sensor signal classification on edge devices. The adaptive filter based learning model for time series sensor signal classification enables automated-computationally lightweight learning (significant reduction in computational resources) and inferring/classification in real-time or near-real-time on CPU/memory/battery life constrained edge devices. The disclosed techniques for time series sensor signal classification on edge devices characterizes the intrinsic signal processing properties of the input time series sensor signals using linear adaptive filtering and derivative spectrum to efficiently construct the adaptive filter based learning model based on standard classification algorithms for time series sensor signal classification.

First claim

We claim:

1 . A processor-implemented method ( 700 ) for an adaptive filter based learning model for time series sensor signal classification on edge devices, the method comprising:
receiving a plurality of training data that comprises of univariate time series data from a plurality of edge devices of Internet of things (IoT) applications including wearable and implantable for predictive healthcare, low-cost edge processors for machine diagnostics, wherein the plurality of time series input data includes at least one of time series signals or time series classes ( 702 ), and wherein the time series data analysis is performed by obtaining dataset pertinent to faulty automobile engines, wafer representing semiconductor process control measurements of normal and abnormal fabrications, Earthquake dataset consists of vibration signals correspond to earthquake events, and a set of Electrocardiogram (ECG) signals related to cardiac activity problem; constructing a first plurality of prediction filters based on an adaptive filtering techniques using the training data and a time delayed version of the training data to predict an optimal filter co-efficient, wherein the prediction filter acts as a learner of a structure of the input data to predict the optimal filter co-efficient for each of the input data, wherein a first time series signal prediction filter is constructed for each instance of the time series signals or a first time series class prediction filter is constructed for each of the time series classes ( 704 ), wherein the time series input is represented as u(i) is a time series [u 1 , u 2 , u 3 , . . . , u T ] ∈ T is an ordered set of real values, and a response d(i) is a time-advanced version of u(i), advanced by δ samples, such that: d(i)=u(i+δ) with i≥0, wherein when be the training dataset, consists of {u 1 (i), l 1 }, {u 2 (i), l 2 }, . . . , {u N (i), l N }, where N is the number of training instances and l n is corresponding class labels, wherein binary classification, l n ∈+1, the filter coefficients, ω n (k), for k=0,1,2, . . . , p of p th order FIR filter are derived by the adaptive filter constructed based on Finite Impulse Response Recursive Least Squares (FIR-RLS) filter techniques, wherein for each of u(i), an adaptation to create the first plurality of prediction filters is constructed for each instance of time series input signal, which results in deriving the optimal filter coefficients (W n ) for each of the training instances, results in W n , n=1,2, . . . N filter coefficient sets are generated, N=number of input instances, where W n ∈ p , p is the filter order, from u 1 (i), we get W 1 filter coefficients is obtained and subsequently, from u 2 (i), . . . , u N (i), W 2 , . . . , W N are computed and complete coefficient matrix is: W∈ N×p ; constructing a first linear filter using the predicted optimal filter co-efficient to obtain a first deterministic component and a first stochastic component for each of the training data with decomposition of the time series input data into orthogonal signals û and ε and thereby possessing potential to reveal distinct features and the deterministic component and the stochastic component ensures availability of a distinct feature set ( 706 ), wherein each of the u(i) is filtered by respective prediction filters constructed using a corresponding prediction filter coefficients as: u(i) û(i), û and ε are the transformation from each of the input data by the prediction filter is used for determining ε(i)=u(i)−û(i), wherein u(i) decompose into two components as: u(i)=u d (i)+u s (i), where, u d (i) denotes the deterministic or regular part of u(i), and u s (i) is the stochastic or random part of u(i), wherein û(i) represents the deterministic component (u d (i)) and ε(i) is the stochastic component (u s (i)); deriving a first derivative spectrum from the first deterministic component and the first stochastic component based on fast Fourier transform techniques to obtain a complete feature set of the training data, wherein the first derivative spectrum accentuates small structural differences between nearly identical spectra and enhances spectral resolution ( 708 ), wherein spectral derivatives of {û, ε} are computed and the first derivative spectrum { , } is derived and expressed as: {û, ε} { , }; selecting a first subset of features from the complete feature set of the training data based on a first set of pre-defined cut off frequencies are represented by a lower cut off frequency (Ω l ) and a higher cut off frequency (Ω h ) determined using human-driven meta-information infusion, wherein the first set of pre-defined cut off frequencies provides impetus to a learning process through the human-driven meta-information infusion ( 710 ); generating a training data feature vector from the selected sub-set of features from training data ( 712 ); generating a time series sensor signal classification learning model using the training data feature vector based on classification techniques for classification of time series signal on Central Processing Unit (CPU) or memory or battery life constrained edge devices in real-time and the time series sensor signal classification learning model for the time series signal classification provides automated-computationally lightweight learning with reduction in computational resources ( 714 ), wherein for time series sensor signal classification on the edge devices characterizes intrinsic signal processing properties of input time series sensor signals using a linear adaptive filtering and the first derivative spectrum to efficiently construct the adaptive filter based learning model based on classification algorithms for the time series sensor signal classification, wherein the time series sensor signal classification on the edge devices comprises classifying a testing data using the time series sensor signal classification learning model based on comparison between a testing feature vector and the time series sensor signal classification learning model by an inference classifier; and displaying the time series sensor signal classification on the edge devices as received from the inference classifier, wherein the method performs the time series data analysis on the dataset to identify the faulty automobile engines analyzing an engine noise, the semiconductor process control measurements of normal and abnormal fabrication, the vibration signals corresponding to the earthquake events, and the ECG signals related to the cardiac activity problem, with improved classification performance accuracy and reduced training time.

2 . The method of claim 1 , further comprising the step of classification of time series signal on edge devices as follows:
receiving a plurality of testing data that comprises of univariate time series data from the plurality of edge devices, wherein the plurality of time series input data includes at least one of time series signals or time series classes ( 602 ); constructing a second plurality of prediction filters based on adaptive filtering techniques using the testing data and a time delayed version of the testing data to predict an optimal filter co-efficient, wherein a second time series signal prediction filter is constructed for each instance of the time series signals or a second time series class prediction filter is constructed for each instance of the time series classes ( 604 ); constructing a second linear filter using the predicted optimal filter co-efficient to obtain a second deterministic component and a second stochastic component for each of the testing data ( 606 ); deriving a second derivative spectrum from the second deterministic component and the second stochastic component based on fast Fourier transform techniques to obtain a complete feature set of the testing data ( 608 ); selecting a second subset of features from the complete feat

Assignees

• Tata Consultancy Services Ltd

Inventors

• Ukil Arijit
• Pal Arpan
• Bandyopadhyay Soma
• Sahu Ishan
• Deb Trisrota

Classifications

• G06F18/24

  Classification techniques · CPC title

• G06F17/16

  Matrix or vector computation {, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization (matrix transposition G06F7/78)} · CPC title

• G06N20/10

  using kernel methods, e.g. support vector machines [SVM] · CPC title

• G06N5/01

  Dynamic search techniques; Heuristics; Dynamic trees; Branch-and-bound · CPC title

• G06N20/20 Primary

  Ensemble learning · CPC title