Paradigm shift: AI for anomaly detection The inherent shortcomings of static threshold- based systems in detecting impulse line choking necessitate a fundamental shift in our approach to anomaly detection. This paradigm shift is ushered in by AI, offering a transformative capability to learn normal operating patterns. In an AI context, an anomaly is a data point or sequence that is statistically unusual or deviates significantly from the system's learned normal behaviour. The power of AI stems from its ability to adapt and learn from vast amounts of historical operational data. AI models can identify complex correlations between multiple sensor inputs, recognising non-linear relationships and evolving process dynamics. This allows early detection of subtle anomalies such as the gradual drift caused by a partially choked impulse line, which would be completely missed by traditional methods. This leads to predictive capabilities, transforming maintenance from a reactive, breakdown-driven activity to a proactive, condition-based strategy. By embracing these AI methodologies, refineries can shift from merely reacting to threshold breaches to intelligently anticipating and preventing critical transmitter malfunction, particularly those stemming from insidious issues like impulse line choking. This paradigm shift promises a significant leap forward in operational efficiency, safety, and asset integrity. Various anomaly detection techniques1 for flow transmitters were explored, including Multivariate Gaussian distribution (MGD), Mahalanobis Distance (MD), and autoencoders (AE). Python’s scikit-learn library2 provides a comprehensive suite of tools for implementing these methods, including robust algorithms for outlier detection. The AE algorithm was selected for its efficacy in unsupervised learning and its ability to capture complex non-linear relationships in data.
Bottle neck
Encoder
Decoder
Compressed low dimensional representation of inputs
Figure 2 Autoencoders architecture
describes the likelihood of observing a data point x given the mean vector μ and covariance matrix Σ. • Mahalanobis distance : Mahalanobis distance is a measure of the distance between a point and a distribution, considering the correlations between variables. It effectively measures how many standard deviations away a point is from the mean of a distribution. The squared Mahalanobis distance D 2 = (x – m) T ⋅ C - 1 ⋅ (x – m) , where x is the observation vector, m is the mean vector, and C −1 is the inverse covariance matrix. • Autoencoders – the architecture and how it is used for anomaly detection : In this case, AEs, an unsupervised machine learning algorithm, “ By embracing AI methodologies, refineries can shift from merely reacting to threshold breaches to intelligently anticipating and preventing critical transmitter malfunction ” were applied. An AE is a special type of neural network, consisting of an encoder, a bottleneck, and a decoder (see Figure 2 ). The network is trained to reconstruct its input at the output layer. The ‘anomaly’ is then detected when the reconstruction error is high, indicating that the input data do not conform to the patterns learned during training on normal data. This makes AEs particularly powerful for detecting novel or unusual patterns without prior knowledge of what constitutes an anomaly. In an AE, the loss function quantifies the difference between the original input and its reconstruction. Typically, this is measured using
• Multivariate Gaussian : This method assumes that normal data follow a Multivariate Gaussian distribution. Anomalies are points that fall far outside this distribution, typically defined by a low probability density function value. The probability density function We explored various anomaly detection techniques [2] for flow transmitters, including Multivariate Gaussian distribution (MGD), Mahalanobis Distance (MD) and Autoencoders (AE). Python's scikit-learn library [1] provides a comprehensive suite of tools for implementing these methods, including robust algorithms for outlier detection. The AE algorithm was selected for its efficacy in unsupervised learning and its ability to capture complex non-linear relationships in data. • By embracing these AI methodologies, refineries can shift from merely reacting to threshold breaches to intelligently anticipating and preventing critical transmitter malfunction, particularly those stemming from insidious issues like impulse line choking. This paradigm shift promises a significant leap forward in operational efficiency, safety, and asset integrity. ushered in by Artificial Intelligence (AI), offering a transformative capability to learn normal operating patterns. In an AI context, an anomaly is a data point or sequence that is statistically unusual or deviates significantly from the system's learned normal behaviour. The power of AI stems from its ability to adapt and learn from vast amounts of historical operational data. AI models can identify complex correlations between multiple sensor inputs, recognizing non-linear relationships and evolving process dynamics. This allows early detection of subtle anomalies such as the gradual drift caused by a partially choked impulse line that would be completely missed by traditional methods. This leads to predictive capabilities, transforming maintenance from a reactive, breakdown-driven activity to a proactive, condition-based strategy. Multivariate Gaussian : This method assumes that normal data follows a multivariate Gaussian distribution. Anomalies are points that fall far outside this distribution, typically defined by a low probability density function value. The probability density function 𝑃𝑃𝑃𝑃 ( 𝑥𝑥𝑥𝑥 ; 𝜇𝜇𝜇𝜇 , 𝛴𝛴𝛴𝛴 )= 1 ( 2𝛱𝛱𝛱𝛱 ) 𝑛𝑛𝑛�ൗ� 2 (| 𝛴𝛴𝛴𝛴 |) 1∕2 𝑒𝑒𝑒𝑒𝑥𝑥𝑥𝑥𝑒𝑒𝑒𝑒 𝑒𝑒 1 2 ( 𝑥𝑥𝑥𝑥 𝑥 𝑥 ) 𝑇𝑇𝑇𝑇 𝛴𝛴𝛴𝛴 −1 ( 𝑥𝑥𝑥𝑥 𝑥 𝑥 ) ൰ describesthe likelihood of observing a data point x given the mean vector μ and covariance matrix Σ. • Mahalanobis Distance: Mahalanobis distance is a measure of the distance between a point and a distribution, considering the correlations between variables. It effectively measures how many standard deviations away a point is from the mean of a distribution. 𝐷𝐷𝐷𝐷 2 =( 𝑥𝑥𝑥𝑥 𝑥 𝑥𝑥𝑥𝑥 ) 𝑇𝑇𝑇𝑇 ⋅ 𝐶𝐶𝐶𝐶 −1 ⋅ ( 𝑥𝑥𝑥𝑥 𝑥 𝑥𝑥𝑥𝑥 ) Refining India
67
Powered by FlippingBook