Systems and methods for detecting and classifying anomalous features in one-dimensional data
The present disclosure generally relates to apparatus, software and methods for detecting and classifying anomalous features in one-dimensional data. The apparatus, software and methods disclosed herein use a YOLO-type algorithm on one-dimensional data. For example, the data can be any one-dimensional data or time series, such as but not limited to be power over time data, signal to noise ratio (SNR) over time data, modulation error ratio (MER) data, full band capture data, radio frequency data, temperature data, stock data, or production data. Each type of data may be susceptible to repeating phenomena that produce recognizable anomalous features. In some embodiments, the features can be characterized or labeled as known phenomena and used to train a machine learning model via supervised learning to recognize those features in a new data series.
1. A method for detecting and classifying an anomalous feature in one-dimensional data, the method comprising:
using a one-dimensional convolutional layer to generate a bounding region containing a single spatial variable around an anomalous feature within one-dimensional test data and to determine whether the anomalous feature in the one-dimensional test data falls into a class of anomalous features taught to a neural network using one-dimensional training data.
2. The method of claim 1 further comprising training the neural network to detect and classify anomalous features in the one-dimensional training data.
3. The method of claim 2 , wherein the step of training the neural network maintains the width of an anomalous feature in the training data relative to the width of the training data at all times.
4. The method of claim 1 , wherein the neural network is trained on one-dimensional data that has been smoothed, value scale shifted, flipped, floor shifted, noise injected and/or synthesized.
5. The method of claim 4 , wherein the synthesized data is fully fabricated or injected with an anomalous feature.
6. The method of claim 1 , wherein the one-dimensional test data is time series data.
7. The method of claim 1 , wherein the one-dimensional test data is power over time data, signal to noise ratio (SNR) over time data, modulation error ratio (MER) data, full band capture data, radio frequency data, temperature data, vibration sensor data, humidity data, voltage data, electrical current data, pH value over time data or motion sensor data.
8. The method of claim 1 , wherein the step of using the one-dimensional convolutional layer to generate a bounding region is carried out by a YOLO-type algorithm.
9. The method of claim 1 , wherein the bounding region comprises x, w, class, and a confidence value indicating a likelihood of an anomalous feature being present, where x is the center of the bounding region, w is the width of the bounding region, and class is an ID number associated with the class of the anomalous feature.
10. The method of claim 1 further comprising localizing the anomalous feature within the one-dimensional test data.
11. The method of claim 10 further comprising identifying one or more devices generating the one-dimensional test data comprising the anomalous feature.
12. The method of claim 11 , wherein the anomalous feature is selected from the group consisting of LTE ingress, a MER spike, amplifier suckout, frequency roll-off, waves, wireless interference patterns, tilt patterns, sweep signal patterns and combinations thereof.
13. The method of claim 1 further comprising scaling the one-dimensional test data while preserving anomalous features by:
increasing a number of data points within the one-dimensional test data to generate up-sampled one-dimensional data that is divisible by a predetermined integer;
dividing the up-sampled data into a number of bins equal to the predetermined integer;
calculating a single value for each bin; and
replacing each bin with the single value to obtain down-sampled data.
14. The method of claim 13 , wherein the single value for each bin is the mean, median, mode, average, minimum, maximum, or quantile of data points within the bin.
15. A non-transitory computer-readable medium having a plurality of non-transitory instructions executable with a processor for detecting and classifying anomalous features in one-dimensional data, the plurality of non-transitory instructions being executable for:
using a one-dimensional convolutional layer to generate a bounding region containing a single spatial variable around an anomalous feature within one-dimensional test data and to determine whether the anomalous feature in the one-dimensional test data falls into a class of anomalous features taught to a neural network using one-dimensional training data.
16. The non-transitory computer-readable medium of claim 15 further comprising the plurality of non-transitory instructions being executable for training the neural network to detect and classify anomalous features in the one-dimensional training data.
17. The non-transitory computer-readable medium of claim 16 , wherein the step of training the neural network maintains the width of the anomalous feature in the training data relative to the width of the training data at all times.
18. The non-transitory computer-readable medium of claim 15 further comprising the plurality of non-transitory instructions being executable for localizing the anomalous feature within the one-dimensional test data.
19. The non-transitory computer-readable medium of claim 18 further comprising the plurality of non-transitory instructions being executable for identifying one or more devices generating the one-dimensional test data comprising the anomalous feature.
20. The non-transitory computer-readable medium of claim 15 further comprising the plurality of non-transitory instructions being executable for scaling the one-dimensional test data while preserving anomalous features by:
increasing a number of data points within the one-dimensional test data to generate up-sampled one-dimensional data that is divisible by a predetermined integer;
dividing the up-sampled data into a number of bins equal to the predetermined integer;
calculating a single value for each bin; and
replacing each bin with the single value to obtain down-sampled data.