Anomaly Detection in Spectrograms Using Neural Network Deviation Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for detecting anomalies in spectrograms, spectra, or signals require prior knowledge of the anomalies, limiting their effectiveness in cases where such knowledge is unavailable.
Innovation Solution
A method utilizing pre-trained artificial neural networks, including an autoencoder and a recurrent neural network with a Long Short-Term Memory block, to generate filtered and predicted data, allowing for anomaly detection without prior knowledge by analyzing deviations in measured data, filtered data, and predicted data using cost functions and thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If prior knowledge of anomalies is used for detection, then detection accuracy is improved, but applicability to unknown anomalies deteriorates
Solution Approach 1:
Instead of training the neural network to detect specific known anomalies, the patent inverts the approach by training it to recognize only normal operation patterns. The anomaly detection is achieved by identifying deviations from the learned normal behavior, allowing the system to detect any type of anomaly without prior knowledge of specific anomaly characteristics.
Solution Approach 2:
The system uses the measured data itself to train the neural network, creating a self-contained detection mechanism. The network learns from the input data what constitutes normal operation and automatically identifies deviations, eliminating the need for external anomaly templates or prior knowledge databases.
2Adaptability or versatility
If complex analysis methods are used to detect anomalies without prior knowledge, then detection capability is improved, but computational complexity increases
Solution Approach 1:
The patent replaces complex manual analysis methods and traditional signal processing techniques with a neural network-based system. The neural network automatically learns patterns and performs anomaly detection through its trained weights and biases, substituting what would otherwise require sophisticated mathematical models and extensive computational resources for feature extraction and analysis.
Solution Approach 2:
The neural network is pre-trained offline using collected normal operation data, performing the complex learning task in advance. This preliminary training phase captures the essential patterns of normal operation, allowing the deployed system to perform rapid real-time anomaly detection without repeating the complex computational work during operation.
3Loss of time
If real-time anomaly detection is implemented, then response time is improved, but processing speed may deteriorate due to complex computations
Solution Approach 1:
The neural network performs the computationally intensive pattern learning during an offline training phase using historical normal operation data. This preliminary action transfers the computational burden from real-time operation to the training phase, enabling fast real-time inference when the network is deployed for actual anomaly detection.
Solution Approach 2:
The trained neural network model serves as a compact representation or copy of the complex knowledge about normal operation patterns. Once trained, this model can be deployed and reused multiple times for rapid anomaly detection without requiring access to the original training data or repeating the complex learning process.
Data Source
Figure 1~2
Figure 3~4
AI summary
A method for detecting anomalies (A) in a spectrogram, spectrum or signal using a detection module (16) having a first machine learning submodule (22), a second machine learning submodule (24) and a comparison submodule (26) is provided. The method comprises the following steps: a) receiving the spectrogram, spectrum or signal as measured data (M) by the detection module (16), b) generating filtered data (F) based on the measured data (M) by the first submodule (22), c) generating predicted data (P) based on the measured data (M) and/or the filtered data (F) by the second submodule (24), and d) determining that the spectrogram, spectrum or signal comprises an anomaly (A) if the filtered data (F) deviates from the measured data (M) and/or the predicted data (P) deviates from the measured data (M) and/or the filtered data (F) by at least a predetermined amount.