Apparatus and method for detecting anomaly in plant pipe using multiple meta-learning
Provided are an apparatus and method for detecting an anomaly in a plant pipe using multiple meta-learning. When a multi-sensor data stream about a plant pipe is received, each of a plurality of meta-learning modules for processing different packet section ranges, extracts one or more preset types of features from sensor data of packet section ranges set according to trend from an arbitrary reception time point, generates 2D image features of the features according to multi-sensor-specific times, generates 3D volume features by accumulating the 2D image features in a depth direction according to multiple sensors, and learns the 3D volume features in parallel through multi-sensor-specific learning modules. Results of the learning of the meta-learning modules are aggregated, and it is determined whether there is an anomaly in a plant pipe according to a learning result selected based on an optimal combination of multiple features, multiple sensors, and multiple packet sections.
1. An apparatus for detecting an anomaly in a plant pipe using multiple meta-learning, the apparatus comprising:
a plurality of meta-learning modules configured to separately receive multi-sensor data streams from a plurality of sensors for processing sensing of a plant pipe and to process meta-learning with sensor data of packet section ranges set according to trend from an arbitrary reception time point; and
a multi-trend meta-evaluator configured to detect an anomaly in the plant pipe by aggregating processing results of the meta-learning modules and to output a detection result,
wherein the plurality of meta-learning modules process multi-sensor data of packet sections having different lengths from the arbitrary reception time point,
each of the meta-learning modules generates two-dimensional (2D) image features by extracting one or more preset types of features from multi-sensor data collected from a set packet section range, generates three-dimensional (3D) volume features by accumulating the 2D image features in a depth direction according to the plurality of sensors, and learns the 3D volume features in parallel through learning modules separately corresponding to the sensors, and
the multi-trend meta-evaluator determines whether there is an anomaly in the plant pipe by aggregating results learned based on optimal combinations of multiple features, multiple sensors, and multiple packet sections according to the plurality of meta-learning modules.
2. The apparatus of claim 1 , wherein the plurality of meta-learning modules comprise:
a short-term trend meta-learning module configured to process sensor data of a first packet section range;
an intermediate-term trend meta-learning module configured to process sensor data of a second packet section range exceeding the first packet section range; and
a long-term trend meta-learning module configured to process sensor data of a third packet section range exceeding the second packet section range.
3. The apparatus of claim 1 , further comprising a plurality of optimizers separately corresponding to the plurality of meta-learning modules and configured to perform control so that optimal combinations are detected for a setting of packet section ranges according to the meta-learning modules, a setting of a combination of the one or more preset types of features, and a setting of a sensor combination resulting from learning costs of the individual learning modules,
wherein the optimizers detect the optimal combinations using a global optimization technique.
4. The apparatus of claim 3 , wherein the optimizers detect the optimal combination using a genetic algorithm technique.
5. The apparatus of claim 1 , wherein each of the meta-learning modules extracts features regarding at least one of a root mean square (RMS) magnitude, a spectrum magnitude, and a frequency feature according to time from the multi-sensor data.
6. The apparatus of claim 5 , wherein each of the meta-learning modules extracts the RMS magnitude according to time by performing sliding window, filtering, resampling, and magnitude quantization on the multi-sensor data, and generates an RMS magnitude image feature by mapping the extracted RMS magnitude to a time-RMS magnitude domain.
7. The apparatus of claim 5 , wherein each of the meta-learning modules extracts the spectrum magnitude according to time by performing sliding window, Fourier transform, filtering, representative value extraction, and magnitude quantization on the multi-sensor data, and generates a spectrum magnitude image feature by mapping the extracted spectrum magnitude to a time-spectrum magnitude domain.
8. The apparatus of claim 5 , wherein features extracted by each of the meta-learning modules only have values along a magnitude pixel trajectory according to time in an image domain, and
the image domain other than the magnitude pixel trajectory is filled with zero such that the features have a sparse density.
9. The apparatus of claim 1 , wherein the multi-trend meta-evaluator processes the aggregation according to any one method among majority voting, a process of making a final determination on learning results of the individual meta-learning modules according to a preset rule, a process of applying different weights to the meta-learning modules and summing weighted trends to make a determination, and ensemble learning.
10. A method of detecting an anomaly in a plant pipe through an apparatus for detecting an anomaly in a plant pipe using multiple meta-learning, the method comprising:
receiving, by each of a plurality of meta-learning modules configured to process different packet section ranges, a multi-sensor data stream about a plant pipe;
extracting, by each of the plurality of meta-learning modules, one or more preset types of features from sensor data of packet section ranges set according to trend from an arbitrary reception time point;
generating, by each of the plurality of meta-learning modules, two dimensional (2D) image features of features according to multi-sensor-specific times;
generating, by each of the plurality of meta-learning modules, three dimensional (3D) volume features by accumulating 2D image features in a depth direction according to multiple sensors;
learning, by each of the plurality of meta-learning modules, 3D volume features in parallel through multi-sensor-specific learning modules; and
aggregating results of the learning of the plurality of meta-learning modules and determining whether there is an anomaly in a plant pipe according to a result of learning selected based on an optimal combination of multiple features, multiple sensors, and multiple packet sections.
11. The method of claim 10 , wherein the plurality of meta-learning modules comprise:
a short-term trend meta-learning module configured to process sensor data of a first packet section range;
an intermediate-term trend meta-learning module configured to process sensor data of a second packet section range exceeding the first packet section range; and
a long-term trend meta-learning module configured to process sensor data of a third packet section range exceeding the second packet section range.
12. The method of claim 10 , wherein the optimal combination is detected for a setting of packet section ranges according to the meta-learning modules, a setting of a combination of the one or more preset types of features, and a setting of a sensor combination resulting from learning costs of the individual learning modules through a global optimization technique.
13. The method of claim 12 , wherein a genetic algorithm technique is used as the global optimization technique to detect the optimal combination.
14. The method of claim 10 , wherein the extracting of the one or more preset types of features comprises extracting features regarding at least one of a root mean square (RMS) magnitude, a spectrum magnitude, and a frequency feature according to time from the multi-sensor data.
15. The method of claim 14 , wherein when the extracting of the one or more preset types of features comprises extracting the RMS magnitude, the generating of the 2D image features comprises extracting the RMS magnitude according to time by performing sliding window, filtering, resampling, and magnitude quantization on the multi-sensor data, and generating an RMS magnitude image feature by mapping the extracted RMS magnitude to a time-RMS magnitude domain.
16. The method of claim 14 , wherein when the extracting of the one or more preset types of features comprises extracting the spectrum magnitude, the generating of the 2D image features comprises extracting the spectrum magnitude according to time by performing sliding window, Fourier transform, filtering, representative value extraction, and magnitude quantization on the multi-sensor data, and generating a spectrum magnitude image feature by mapping the extracted spectrum magnitude to a time-spectrum magnitude domain.
17. The method of claim 10 , wherein the determining of whether there is an anomaly in the plant pipe comprises processing the aggregation according to any one method among majority voting, a process of making a final determination on learning results of the individual meta-learning modules according to a preset rule, a process of applying different weights to the meta-learning modules and summing weighted trends to make a determination, and ensemble learning.