| CPC G06N 3/0455 (2023.01) [G06N 3/049 (2013.01); G06N 3/063 (2013.01)] | 11 Claims |
|
1. An equipment anomaly detection system (100) comprising:
a signal conversion module (120) for receiving, filtering, rectifying, and analog to digital converting N nominal sensor signal streams into N nominal sensor event streams, and N monitored sensor signal streams into N monitored sensor event streams;
wherein the signal conversion module (120) comprises a bandpass filter (1221) to filter an input signal, a low-pass filter (1223) to rectify and smooth the bandpass-filtered input signal, and a PDM encoder (1224) to encode the smoothed, rectified input signal into an event stream;
a cascade of neural networks for receiving and learning N nominal sensor event streams (190a) from N event stream input connections in a learning state to form a learned state of the cascade of neural networks, and
a detection module (150, 150a) for calculating difference between the N monitored sensor event streams (190b) and the N monitored reconstructed event streams (149b);
wherein the equipment anomaly detection system (100) is characterized by:
the cascade of neural networks in the learned state receives N monitored sensor event streams (190b) from the N event stream input connections as input of the cascade of neural networks in an operation state, and outputs N monitored reconstructed event streams (149b) as output of the cascade of neural networks based on the N monitored sensor event streams (190b), and the N monitored reconstructed event streams (149b) represent a reconstructed version of the N monitored sensor event streams (190b); and
the detection module (150, 150a) sends a detection trigger (180) when the difference between N monitored sensor event streams (190b) and N monitored reconstructed event streams (149b) exceeds a threshold difference value;
wherein the cascade of neural networks comprises: an embedding network (130) being a spiking neural network for receiving the N nominal sensor event streams (190a) from the N event stream input connections as input of the embedding network (130) in the learning state, and outputting M nominal embedded encoding streams (139) as output of the embedding network (130) based on the N nominal sensor event streams (190a), wherein N is an integer greater than the integer M, and the embedding network (130) learns to generate the output of the embedding network (130) from the input of the embedding network (130) and to form a trained embedding network state of the embedding network (130) using spike-timing dependent plasticity (STDP); and a decoding network (140) being a neural network for receiving the M nominal embedded encoding streams (139) as input of the decoding network (140) in the learning state, and outputting N reconstructed event streams as output of the decoding network (140) based on the M nominal embedded encoding streams (139), wherein the decoding network (140) learns to generate the output of the decoding network (140) from the input of the decoding network (140) and to form a trained decoding network state of the decoding network (140) based on a rule to reduce difference between the N reconstructed event streams and the N nominal sensor event streams (190a);
wherein the embedding network (130) in the trained embedding network state outputs M monitored embedded encoding streams (139b) based on the N monitored sensor event streams (190b); and the decoding network (140) in the trained decoding network state outputs the N monitored reconstructed event streams (149b) based on the M monitored embedded encoding streams (139b);
wherein the decoding network filters the M nominal embedded encoding streams by utilizing exponential low-pass filters to obtain a filtered version ŷ (t) of the M nominal embedded encoding streams, filters the N nominal sensor event streams by utilizing exponential low-pass filters to obtain a filtered version x(t) of the N nominal sensor event streams, and generates N nominal reconstructed event streams z(t) as output of the decoding network based on weighted combination of the filtered version ŷ(t) of the M nominal embedded encoding streams with weights Wo, namely z(t)=Wo·ŷ(t), and the decoding network in the learning state performs a self-supervised learning process to adjust the weights Wo to reduce an error E=∥Wo·ŷ(t)−x(t) ∥2+λ81 Wo∥2, where λ represents a regularization parameter for ridge regression;
wherein the decoding network in the trained decoding network state filters the M monitored embedded encoding streams to obtain a filtered version ŷm(t) of the M monitored embedded encoding streams, filters the N monitored sensor event streams to obtain a filtered version zm(t) of the N monitored sensor event streams, and generates N monitored reconstructed event streams zm(t) as output of the decoding network based on weighted combination of the filtered version ŷm(t) of the M monitored embedded encoding streams, and zm(t)=Wo·ŷm(t); the detection module filters the N monitored sensor event streams to obtain a filtered version xm(t) of the N monitored sensor event streams, and the detection module comprises: a subtraction calculator comprising a plurality of subtractors, wherein a subtractor selected from the plurality of subtractors calculates a reconstruction induced difference between a monitored sensor event stream in filtered version xm(t) of the N monitored sensor event streams and a monitored reconstructed event stream in the N monitored reconstructed event streams zm(t), such that the plurality of subtractors calculate reconstruction induced differences stream by stream between the filtered version xm(t) of the N monitored sensor event streams and the N monitored reconstructed event streams zm(t); a summation calculator for calculating summation of the differences between the filtered version xm(t) and the N monitored reconstructed event streams zm(t); and a comparator for comparing the summation with a signal difference threshold, and outputting an alert signal representing anomaly of the N monitored sensor event streams when the summation exceeds the signal difference threshold.
|