Abnormal detection early warning and real-time monitoring system and method for chemical process

Abstract

The invention discloses a chemical process anomaly detection early warning and real-time monitoring system and method. The system comprises a data acquisition module, a feature selection module, an anomaly detection module, an online prediction module and a parameter optimization module. The parameter optimization module adds a sine and cosine search strategy into a Newton-Raphson optimization algorithm NRBO to obtain an improved Newton-Raphson optimization algorithm NRBO, and the parameters of the anomaly detection early warning model CBAM-FiLM-CNN are optimized through the optimization algorithm; the anomaly detection module integrates a CBAM module, a FiLM module and a CNN module, constructs an anomaly detection early warning model CBAM-FiLM-CNN, and outputs an early warning level; according to the method, data from different sensors and different monitoring points are fused, so that the state of the chemical process can be reflected more comprehensively, and the accuracy and reliability of anomaly detection are improved.

Description

Technical Field

[0001] The present invention relates to a monitoring system and method, and in particular to a chemical process abnormality detection, early warning and real-time monitoring system and method. Background Art

[0002] In chemical production, changes in key process parameters such as temperature, pressure, flow, and liquid level have a crucial impact on production safety and product quality. Traditional manual monitoring methods rely on operators observing instrument readings and alarm signals to determine whether the production process is normal and take appropriate measures. However, this approach has significant limitations: the first is that manual judgment is easily influenced by subjective factors and it is difficult to capture complex process changes in real time. In addition, the setting of fixed thresholds often cannot adapt to the dynamic characteristics and complex nonlinear relationships of chemical processes, resulting in frequent false alarms and missed alarms.

[0003] With the development of computer and information technologies, data-driven process monitoring methods have gradually become a research hotspot. For example, traditional methods such as principal component analysis (PCA) and partial least squares (PLS) can improve the accuracy of process state monitoring and anomaly detection through data mining and analysis. Although these methods have achieved significant results in the field of chemical monitoring, they still face certain challenges when processing complex data—particularly in modeling nonlinear relationships and fully considering the interactions between multiple variables.

[0004] In recent years, with the rapid development of artificial intelligence (AI), advanced algorithms such as deep learning and reinforcement learning have begun to emerge. Deep learning methods can automatically learn complex features and nonlinear relationships in data, thereby improving the accuracy and reliability of anomaly detection. Reinforcement learning, on the other hand, enables intelligent control and optimization of chemical processes through interactive learning from the environment. Furthermore, multi-source data fusion technology has become an important research direction. By integrating data from different sensors and monitoring points, it can more comprehensively reflect the status of the chemical process, further improving the accuracy and reliability of anomaly detection. Summary of the Invention

[0005] Purpose of the invention: The purpose of the present invention is to provide a chemical process anomaly detection, early warning and real-time monitoring system to improve the accuracy and reliability of chemical state anomaly detection, on the other hand, to provide a chemical process anomaly detection, early warning and real-time monitoring method.

[0006] Technical solution: The anomaly detection, early warning and real-time monitoring system of the present invention includes:

[0007] The data acquisition module is used to collect temperature, pressure, flow chemical process variable data and operating parameter data through different sensors at a set acquisition frequency, integrate them into a raw data sequence, and construct a sample data set;

[0008] A feature selection module is used to use the AHSIC lasso feature selection algorithm to screen out feature input variables that have a strong correlation with the fault from the sample data set to form a key feature set, and to introduce Gaussian noise into the training set during the data set division process;

[0009] The anomaly detection module is used to build an anomaly detection and warning model CBAM-FiLM-CNN by integrating the CBAM module, the FiLM module, and the CNN module. It performs anomaly detection and warning, and classifies warning levels based on the anomaly detection results, taking into account the severity, impact range, and frequency of the anomaly.

[0010] The online prediction module is used to determine the threshold value based on the prediction error calculation and exponentially weighted average smoothing, combined with the mean, standard deviation and related outlier information of the smoothed error, to determine whether the predicted value is abnormal;

[0011] The parameter optimization module is used to add a sine-cosine search strategy to the Newton-Raphson optimization algorithm NRBO to obtain an improved Newton-Raphson optimization algorithm NRBO, and optimize the parameters of the anomaly detection and early warning model CBAM-FiLM-CNN through the optimization algorithm.

[0012] Preferably, the anomaly detection module includes:

[0013] The CBAM module is used to input the preprocessed data into the CBAM module. CBAM performs attention calculation on the data along the channel and spatial dimensions, multiplying the channel attention weight and spatial attention weight with the original feature map respectively, highlighting important features and suppressing unimportant features, and obtaining data processed by the CBAM module;

[0014] The FiLM module is used to input the data processed by the CBAM module into the FiLM module. The FiLM module processes the data through the LPU and maps the input data to a low-dimensional space using Legendre polynomial projection, retaining the historical information of the data. The data enters the FEL, which uses Fourier transform and low-rank approximation to reduce the dimensionality of the data;

[0015] The CNN module is used to input the data processed by the FiLM module into the CNN module for anomaly detection. In the anomaly detection and early warning model CBAM-FiLM-CNN, the output of the CNN is a probability value, which indicates the possibility that the data is normal or abnormal. By setting an appropriate threshold, the probability value is compared with the threshold. If the probability value is greater than the threshold, it is judged as abnormal and an early warning is triggered; if it is less than the threshold, it is judged as normal.

[0016] The anomaly detection, early warning and real-time monitoring method of the present invention is characterized by comprising the following steps:

[0017] S1. Use different types of sensors to collect temperature, pressure, and flow chemical process variable data and operating parameter data at a set frequency, integrate them into a raw data sequence, and construct a sample data set;

[0018] S2. Using the AHSIC lasso feature selection algorithm, characteristic input variables with strong correlation with the fault are screened from the sample data set to form a key feature set, and Gaussian noise is introduced when dividing the training set;

[0019] S3. Build an anomaly detection and warning model CBAM-FiLM-CNN that integrates the CBAM module, FiLM module, and CNN module;

[0020] S4. Use the improved Newton-Raphson optimization algorithm NRBO to optimize the number of hidden layers and learning rate of the abnormal warning model CBAM-FiLM-CNN;

[0021] S5. Calculate the prediction error and smooth it with an exponentially weighted average, combine the mean and standard deviation of the smoothed error, and related outlier information to determine a threshold value and determine whether the prediction value is abnormal;

[0022] S6. Use the trained anomaly detection and warning model CBAM-FiLM-CNN to perform anomaly detection and warning. According to the anomaly detection results, the warning level is divided into the anomaly severity, impact range and frequency.

[0023] Preferably, the S2 includes:

[0024] S21. Assign an initial dynamic attention weight to each feature in the time series data, and continuously adjust it according to the relationship between the feature and the target variable and the time series characteristics. The initial value of the weight is randomly assigned or set based on simple statistical information;

[0025] S22. Divide the time series data into certain time windows, where each window contains data of several time steps. The window size can be adjusted according to the characteristics of the data and actual needs;

[0026] S23. In each time window, the dynamic time warping (DTW) algorithm is used to calculate the similarity measure between each feature and the target variable as the feature correlation score. Based on the calculated correlation score, the exponential moving average (EMA) is combined to update the dynamic attention weight. The specific formula is:

[0027]

[0028] Among them, W t (i) represents the dynamic attention weight of the i-th feature at time step t, W t-1 (i) represents the weight of the feature at the previous time step, α represents the smoothing factor, S t (i) represents the correlation score of the i-th feature at time step t, and m represents the total number of features.

[0029] S24. Use the generated weight matrix to weight the original feature matrix to obtain a weight matrix G = X⊙A, rescale the original features according to the weights, and map the weighted matrix to a high-dimensional kernel space in the high-dimensional space;

[0030] Where ⊙ represents the Hadamard product.

[0031] S25. Based on the principle of maximizing relevance and minimizing redundancy, the similarity between the feature space and the output space is calculated, and the feature subset with higher similarity to the target variable space is selected as the new feature space;

[0032] S26. After feature extraction of the data, the GDA algorithm is used to enhance the data, which enhances the closely related variables in the abnormal samples. The sample set is divided into a training set and a test set, and Gaussian white noise is introduced into the training set.

[0033] Preferably, the introduction of Gaussian white noise in S26 is performed by the following formula:

[0034]

[0035] X noise =randn(len(X train ))×sqrt(P n );

[0036]

[0037] Among them, len(X train ) represents X train The length of P s Represents X train The power, snr represents the signal-to-noise ratio, P nThe noise power is calculated based on the signal-to-noise ratio. The noise power is obtained by converting the signal-to-noise ratio from decibels to linear and dividing the signal power by this ratio; X noise The generated rule has a mean of 0 and a standard deviation of sqrt(P n ) of Gaussian white noise, randn(len(X train Generate random samples of the same length as the signal and multiply them by the square root of the noise power; Indicates the generated signal after adding noise.

[0038] Preferably, the S3 includes:

[0039] S31, the pre-processed data is input to the CBAM module for processing, then to the FiLM module for processing, and finally to the CNN module for processing;

[0040] S32. Train the anomaly detection and warning model CBAM-FiLM-CNN, select a suitable loss function to measure the difference between the model prediction results and the true labels, and use an optimization algorithm to adjust the model parameters to minimize the loss function.

[0041] S33. Through continuous iterative training, the model learns the characteristic patterns of the data.

[0042] Preferably, the improved Newton-Raphson optimization algorithm NRBO formula described in S4 is as follows:

[0043]

[0044] w(n)=w max *(w max -w min )*(Mit-n) / Mit;

[0045] in, represents the position after individual disturbance, sin(n) represents the sine function, cos(n) represents the cosine function, n represents the number of iterations, w(n) represents the adaptive inertia weight, P represents the random probability, w max represents the maximum inertia weight, w min Represents the minimum inertia weight, and Mit represents the maximum number of iterations.

[0046] Preferably, the warning levels in S6 include:

[0047] Minor anomalies are judged as low-level warnings, and relevant personnel only need to pay close attention to subsequent changes in the data;

[0048] If the abnormality does not threaten the core functions of the system, it is judged as a medium-level warning, and preliminary response measures need to be initiated;

[0049] Abnormal situations that seriously threaten the normal operation of the system and may cause major failures will be judged as high-level warnings, immediately triggering the emergency response mechanism and organizing a professional team to carry out emergency repairs.

[0050] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: 1. Through the improved Newton-Raphson optimization algorithm NRBO, an elite reverse dynamic strategy is adopted in the initialization stage to replace the traditional random initialization method, and an adaptive sine-cosine search strategy is introduced in the optimization process, which effectively enhances the algorithm's perturbation ability for updating the optimal solution position and increases population diversity, thereby successfully avoiding the defect of the traditional Newton-Raphson optimization algorithm that is prone to falling into local optimality, and significantly improving the optimization effect and detection accuracy of the anomaly detection and early warning model; 2. By constructing an anomaly detection and early warning model CBAM-FiLM-CNN that integrates the CBAM module, the FiLM module and the CNN module, the preprocessed data first enters the CBAM module for attention calculation to achieve focusing on key features, and then the data features are further enhanced by the FiLM module and noise interference is effectively suppressed. Finally, the CNN module is used for anomaly detection, which improves the detection accuracy and reliability of chemical process anomalies, and also realizes efficient processing and accurate analysis of complex chemical process data, effectively reduces the risk of false alarms and missed alarms, and provides a strong guarantee for the stable operation of chemical production. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 Schematic diagram of the process of the present invention;

[0052] Figure 2 This is a flow chart of the SPD-Conv module introduced into YOLOv9 to improve the present invention. DETAILED DESCRIPTION

[0053] The technical solution of the present invention is described in detail below with reference to the accompanying drawings.

[0054] Chemical process anomaly detection, early warning and real-time monitoring system, including:

[0055] The data acquisition module is used to collect temperature, pressure and flow chemical process variable data and operation parameter data through different sensors at a set acquisition frequency, integrate them into a raw data sequence, and construct a sample data set.

[0056] The feature selection module is used to use the AHSIC lasso feature selection algorithm to screen out feature input variables that have a strong correlation with the fault from the sample data set to form a key feature set. In the process of dividing the data set, Gaussian noise is introduced into the training set.

[0057] The anomaly detection module is used to build an anomaly detection and warning model CBAM-FiLM-CNN by integrating the CBAM module, the FiLM module, and the CNN module. It performs anomaly detection and warning. Based on the anomaly detection results, it classifies the warning level based on the severity, impact range, and frequency of the anomaly.

[0058] CBAM module: The preprocessed data is input into the CBAM module, which performs attention calculations on the data along both the channel and spatial dimensions. In the channel dimension, global average pooling and maximum pooling operations are performed to generate channel feature description vectors, which are then input into a shared multi-layer perception mechanism (MLP) to obtain channel attention weights. In the spatial dimension, global average pooling and maximum pooling are performed on the feature maps, the results are concatenated, and then a convolutional layer is used to generate spatial attention weights. Finally, the channel attention weights and spatial attention weights are multiplied by the original feature map to highlight important features and suppress unimportant features, resulting in data processed by the CBAM module.

[0059] FiLM module: Data processed by the CBAM module is input into the FiLM module. The FiLM module first processes the data through the Legendre Projection Unit (LPU), using Legendre polynomial projection to preserve historical data information. During this process, the LPU updates the memory unit based on dynamic equations to capture long-term data dependencies. The data then enters the Frequency Enhanced Layer (FEL), which uses Fourier transforms and low-rank approximations to reduce the data's dimensionality and remove noise.

[0060] Specifically, FEL first performs a Fourier transform on the data, converting it to the frequency domain. It then selects frequency patterns in the low-frequency portion and combines them with low-rank approximation techniques to reduce the number of model parameters and computational complexity while preserving the key features of the data. After processing by the FiLM module, the data's features are further enhanced and noise is effectively suppressed.

[0061] CNN module: Data processed by the FiLM module is input into the CNN module for anomaly detection. CNN consists of multiple convolutional layers, pooling layers, and fully connected layers. Convolutional layers perform convolution operations by sliding convolution kernels over the data, extracting local features. Convolution kernels of different sizes and strides can capture feature information at different scales. Pooling layers are used to reduce the dimensionality of the data, reducing computational effort while retaining key features. Common pooling operations include max pooling and average pooling. Fully connected layers concatenate the feature vectors processed by convolution and pooling, mapping them to the final prediction space. In anomaly detection and early warning models, the output of the CNN can be a probability value, indicating the likelihood that the data is normal or abnormal. By setting an appropriate threshold, the probability value is compared with the threshold. If the probability value is greater than the threshold, it is considered an anomaly and triggers an early warning; if it is less than the threshold, it is considered normal.

[0062] The online prediction module is used to determine the threshold value based on the prediction error calculation and exponentially weighted average smoothing, combined with the mean, standard deviation and related outlier information of the smoothed error, to judge whether the predicted value is abnormal.

[0063] The parameter optimization module is used to add the sine and cosine search strategy to the Newton-Raphson optimization algorithm NRBO to obtain an improved Newton-Raphson optimization algorithm NRBO, and optimize the parameters of the anomaly detection and early warning model CBAM-FiLM-CNN through the optimization algorithm.

[0064] The method for detecting, warning and real-time monitoring of abnormalities in a chemical process includes the following steps:

[0065] S1. Use different types of sensors to collect temperature, pressure, flow chemical process variable data and operating parameter data at a set frequency, integrate them into a raw data sequence, and construct a sample data set.

[0066] S2. Using the AHSIC lasso feature selection algorithm, feature input variables that have a strong correlation with the fault are screened out from the sample data set to form a key feature set, and Gaussian noise is introduced when dividing the training set.

[0067] The AHSIC lasso feature selection algorithm filters the feature input variables including:

[0068] (1) Initialize dynamic attention weights:

[0069] An initial dynamic attention weight is assigned to each feature in the time series data. This weight will be continuously adjusted in subsequent steps based on the relationship between the feature and the target variable and the time series characteristics. The initial value of the weight can be randomly assigned or set based on simple statistical information (such as the variance of the feature).

[0070] (2) Dynamic attention weight update:

[0071] Divide the time series data into time windows, where each window contains data from several time steps. The window size can be flexibly adjusted based on the characteristics of the data and actual needs. For example, in stock price time series data, a time window containing 5 trading days of data can be set to analyze short-term price fluctuation characteristics.

[0072] Calculate feature correlation within a window. Within each time window, use the Dynamic Time Warping (DTW) algorithm to calculate the similarity measure between each feature and the target variable (e.g., predicting future stock price trends), which serves as the feature's correlation score. The DTW algorithm effectively handles the scaling and shifting of time series along the time axis, accurately measuring the similarity between different series. For each feature, calculate its DTW distance with the target variable within the window. The smaller the distance, the greater the similarity between the feature and the target variable within that window, and the higher the correlation score.

[0073] Attention weight update: Based on the calculated correlation score, the dynamic attention weight is updated in combination with the idea of ​​exponential moving average (EMA). The specific formula is:

[0074]

[0075] Among them, W t (i) represents the dynamic attention weight of the i-th feature at time step t, W t-1 (i) represents the weight of the feature at the previous time step, α represents the smoothing factor, S t (i) represents the relevance score of the i-th feature at time step t, and m represents the total number of features. In this way, the weights of features with high relevance will gradually increase, while the weights of features with low relevance will decrease.

[0076] Weighting of original features: Use the generated weight matrix A to weight the original feature matrix X to obtain the weighted matrix g = X⊙A (⊙ represents the Hadamard product). Rescale the original features according to the weights so that important features are more prominent in the data, while the influence of unimportant features is weakened.

[0077] (3)HSICLasso feature selection:

[0078] Mapping to a high-dimensional kernel space: Map the weighted matrix G to the high-dimensional kernel space. In high-dimensional space, data features can be more fully expressed and distinguished. Features that may be difficult to distinguish in low-dimensional space may show obvious differences in high-dimensional kernel space. Some features that are linearly inseparable in the original feature space may become separable after mapping to the high-dimensional kernel space.

[0079] Calculate similarity and select features. Based on the principle of maximizing relevance and minimizing redundancy, calculate the similarity between the feature space and the output space (i.e., the space related to the target variable). The optimization formula is:

[0080]

[0081] Among them, L represents the matrix related to the target variable, K (k) represents the kernel matrix associated with the kth feature, a k represents the coefficient corresponding to the kth feature, and λ represents the regularization parameter. By minimizing this objective function, the correlation between the feature and the target variable and the redundancy between the features are balanced, and the most valuable features for the target variable are screened out.

[0082] Determine the new feature space and select a feature subset with higher similarity to the target variable space as the new feature space; these selected features will serve as the input of the subsequent anomaly detection and early warning model, which can improve the performance and efficiency of the model because redundant and irrelevant variables are removed, reducing the complexity of model training while retaining the most important information for the target variable.

[0083] (4) The chemical process is a complex and variable linear system. Anomalies often occur in a short period of time and data collection closely related to the anomalies is difficult, making it difficult to obtain sufficient abnormal samples. Model training, especially deep learning model training, often requires a large number of samples. GDA adds Gaussian noise to the data to expand the abnormal data set while improving the model's resistance to noise. Therefore, after feature extraction of the data, the GDA algorithm is used to enhance the data, enhancing the closely related variables in the abnormal samples and providing sufficient training sets for model training. The specific formula is as follows:

[0084] Divide the sample set X into the training set X train 、Y train and the test set X test 、Y test , for X train Gaussian white noise is added using the following formula:

[0085]

[0086] X noise =randn(len(X train ))×sqrt(P n );

[0087]

[0088] Among them, len(X train ) represents X train The length, Ps Represents X train The power of the signal, where the square of the signal amplitude is used to represent the approximate signal power; SNR represents the signal-to-noise ratio (dB), which represents the proportional relationship between the signal and noise intensity. The higher the value, the stronger the signal is relative to the noise. n The noise power is calculated based on the signal-to-noise ratio. The noise power is obtained by converting the signal-to-noise ratio from decibels to linear and dividing the signal power by this ratio; X noise The generated rule has a mean of 0 and a standard deviation of sqrt(P n ) of Gaussian white noise, randn(len(X train ))Generate random samples of the same length as the signal, and multiply by the square root of the noise power to ensure the noise has the correct power; Indicates the generated signal after adding noise.

[0089] S3. Construct an anomaly detection and warning model CBAM-FiLM-CNN that integrates the CBAM module, FiLM module, and CNN module.

[0090] The preprocessed data is input into the CBAM module for attention calculation, the data processed by the CBAM module is input into the FiLM module for processing, and the data processed by the FiLM module is input into the CNN module for prediction.

[0091] The anomaly detection and early warning model, CBAM-FiLM-CNN, is trained using labeled normal and abnormal sample data. During training, an appropriate loss function (such as the cross-entropy loss function) is selected to measure the difference between the model's predictions and the true labels. An optimization algorithm (such as the Adam optimization algorithm) is used to adjust the model's parameters to minimize the loss function. Through continuous iterative training, the model gradually learns the characteristic patterns of the data, improving the accuracy of anomaly detection and early warning. To prevent overfitting, regularization techniques such as L1 and L2 regularization and Dropout can be used to improve the model's generalization capabilities. Early stopping can also be used during training to determine whether the model has converged based on performance indicators (such as accuracy and recall) on the validation set to avoid overtraining.

[0092] S4. Use the improved Newton-Raphson optimization algorithm NRBO to optimize the number of hidden layers and learning rate of the abnormal warning model CBAM-FiLM-CNN.

[0093] (1) The elite reverse dynamic boundary continuously narrows the search space, improves the search space introduced in the position update, and improves the optimization efficiency of NRBO. The expression is as follows:

[0094]

[0095] Among them, x n+1 Represents the position of the individual after iteration, a random number within r1∈[0,1]; ub represents the upper limit of the optimized parameter; lb represents the lower limit of the optimized parameter; represents the optimal position of an individual.

[0096] (2) The adaptive sine-cosine search strategy includes: adding the sine-cosine search strategy to the late optimization of the Newton-Raphson optimization algorithm to perturb the optimal solution to obtain a new solution, thereby reducing the possibility of the original algorithm falling into the local optimal solution. The formula is as follows:

[0097]

[0098] w(n)=w max *(w max -w min )*(Mit-n) / Mit;

[0099] in, represents the position after individual disturbance, sin(n) represents the sine function, cos(n) represents the cosine function, n represents the number of iterations, w(n) represents the adaptive inertia weight, P represents the random probability, w max represents the maximum inertia weight, w min Represents the minimum inertia weight, and Mit represents the maximum number of iterations.

[0100] S5. A dynamic warning threshold technology is proposed. The prediction error is calculated and smoothed by exponential weighted average. The threshold is determined by combining the mean, standard deviation and related outlier information of the smoothed error to determine whether the predicted value is abnormal.

[0101] Dynamic warning threshold technology includes:

[0102] Assume that the forecast error sequence is e p , where p = 1, 2,…, n.

[0103] First, the forecast error is smoothed by exponential weighted average to obtain the smoothed error sequence

[0104]

[0105] Where j represents the smoothing factor, which is between [0, 1] and is used to control the degree of smoothing. A larger j value gives more weight to recent errors, while a smaller j value makes the smoothed error more sensitive to historical data. The mean and standard deviation of the smoothed error are calculated as follows:

[0106]

[0107] Combine the mean, standard deviation and related abnormal value information of the smoothing error to determine the dynamic warning threshold T t :

[0108]

[0109] Where k′ is a constant used to adjust the looseness of the threshold, IQR is the interquartile range of the smoothed error, i.e., IQR = Q3 - Q1, where Q1 and Q3 are the 25th and 75th percentiles of the smoothed error, respectively, and f(IQR) is used to reflect the impact of outliers.

[0110] When the new prediction error e new When generated, if |e new |>T t , the predicted value is considered abnormal and an early warning is triggered. In this way, by combining the smoothing of prediction errors, statistical characteristics, and outlier information, a dynamic early warning threshold can be determined, allowing for more flexible and accurate judgment of whether the predicted value is abnormal, thereby improving the accuracy and reliability of detection.

[0111] S6. Use the trained anomaly detection and warning model CBAM-FiLM-CNN to perform anomaly detection and warning. The warning level is divided according to the anomaly detection results, and the severity, impact range, and frequency of the anomaly are comprehensively considered when dividing the warning level.

[0112] Minor anomalies are classified as low-level warnings, requiring only relevant personnel to closely monitor subsequent data trends. If an anomaly may have a certain impact on system operation but does not currently endanger core functions, it is set as a medium-level warning, requiring initial response measures to be initiated, such as further investigating the cause of the anomaly and preparing backup plans. For anomalies that seriously threaten the normal operation of the system and may cause major failures, a high-level warning is assigned, immediately triggering an emergency response mechanism and organizing a professional team to carry out repairs to minimize losses. This detailed classification of warning levels allows relevant personnel to quickly and appropriately respond to different levels, effectively ensuring the stable operation of the system and improving overall safety and reliability.

Claims

1. A chemical process abnormality detection, early warning and real-time monitoring system, characterized in that: include: The data acquisition module is used to collect temperature, pressure, flow chemical process variable data and operating parameter data through different sensors at a set frequency, integrate them into a raw data sequence, and construct a sample data set; A feature selection module is used to use the AHSIC lasso feature selection algorithm to screen out feature input variables that have a strong correlation with the fault from the sample data set to form a key feature set, and to introduce Gaussian noise into the training set during the data set division process; The anomaly detection module is used to build an anomaly detection and warning model CBAM-FiLM-CNN by integrating the CBAM module, the FiLM module, and the CNN module. It performs anomaly detection and warning. Based on the anomaly detection results, it classifies the warning level based on the severity, impact range, and frequency of the anomaly. The online prediction module is used to determine the threshold value based on the prediction error calculation and exponentially weighted average smoothing, combined with the mean, standard deviation and related outlier information of the smoothed error, to determine whether the predicted value is abnormal; The parameter optimization module is used to add a sine-cosine search strategy to the Newton-Raphson optimization algorithm NRBO to obtain an improved Newton-Raphson optimization algorithm NRBO, and optimize the parameters of the anomaly detection and early warning model CBAM-FiLM-CNN through the optimization algorithm.

2. The system according to claim 1, wherein: The anomaly detection module includes: The CBAM module is used to input the preprocessed data into the CBAM module. CBAM performs attention calculation on the data along the channel and spatial dimensions, multiplying the channel attention weight and spatial attention weight with the original feature map respectively, highlighting important features and suppressing unimportant features, and obtaining data processed by the CBAM module; The FiLM module is used to input the data processed by the CBAM module into the FiLM module. The FiLM module processes the data through the LPU and maps the input data to a low-dimensional space using Legendre polynomial projection, retaining the historical information of the data. The data enters the FEL, which uses Fourier transform and low-rank approximation to reduce the dimensionality of the data; The CNN module is used to input the data processed by the FiLM module into the CNN module for anomaly detection. In the anomaly detection and early warning model CBAM-FiLM-CNN, the output of the CNN is a probability value, which indicates the possibility that the data is normal or abnormal. By setting an appropriate threshold, the probability value is compared with the threshold. If the probability value is greater than the threshold, it is judged as abnormal and an early warning is triggered; if it is less than the threshold, it is judged as normal.

3. A method for detecting, warning and real-time monitoring of abnormalities in a chemical process, characterized in that: The following steps are involved: S1. Use different types of sensors to collect temperature, pressure, and flow chemical process variable data and operating parameter data at a set frequency, integrate them into a raw data sequence, and construct a sample data set; S2. Using the AHSIC lasso feature selection algorithm, characteristic input variables with strong correlation with the fault are screened from the sample data set to form a key feature set, and Gaussian noise is introduced when dividing the training set; S3. Build an anomaly detection and warning model CBAM-FiLM-CNN that integrates the CBAM module, FiLM module, and CNN module; S4. Use the improved Newton-Raphson optimization algorithm NRBO to optimize the number of hidden layers and learning rate of the abnormal warning model CBAM-FiLM-CNN; S5. Calculate the prediction error and smooth it with an exponentially weighted average, combine the mean and standard deviation of the smoothed error, and related outlier information to determine a threshold value and determine whether the prediction value is abnormal; S6. Use the trained anomaly detection and warning model CBAM-FiLM-CNN to perform anomaly detection and warning. According to the anomaly detection results, the warning level is divided into the anomaly severity, impact range and frequency.

4. The method according to claim 3, characterized in that The S2 includes: S21. Assign an initial dynamic attention weight to each feature in the time series data, and continuously adjust it according to the relationship between the feature and the target variable and the time series characteristics. The initial value of the weight is randomly assigned or set based on simple statistical information; S22. Divide the time series data into certain time windows, where each window contains data of several time steps. The window size can be adjusted according to the characteristics of the data and actual needs; S23. In each time window, the dynamic time warping (DTW) algorithm is used to calculate the similarity measure between each feature and the target variable as the feature correlation score. Based on the calculated correlation score, the exponential moving average (EMA) is combined to update the dynamic attention weight. The specific formula is: Among them, W t (i) represents the dynamic attention weight of the i-th feature at time step t, W t-1 (i) represents the weight of the feature at the previous time step, α represents the smoothing factor, S t (i) represents the correlation score of the i-th feature at time step t, and m represents the total number of features. S24. Use the generated weight matrix to weight the original feature matrix to obtain a weight matrix G = X⊙A, rescale the original features according to the weights, and map the weighted matrix to a high-dimensional kernel space in the high-dimensional space; Where ⊙ represents the Hadamard product. S25. Based on the principle of maximizing relevance and minimizing redundancy, the similarity between the feature space and the output space is calculated, and the feature subset with higher similarity to the target variable space is selected as the new feature space; S26. After feature extraction of the data, the GDA algorithm is used to enhance the data, which enhances the closely related variables in the abnormal samples. The sample set is divided into a training set and a test set, and Gaussian white noise is introduced into the training set.

5. The method according to claim 4, characterized in that S26 introduces Gaussian white noise through the following formula: X noise =randn(len(X train ))×sqrt(P n ); Among them, len(X train ) represents X train The length of P s Represents X train The power, snr represents the signal-to-noise ratio, P n The noise power is calculated based on the signal-to-noise ratio. The noise power is obtained by converting the signal-to-noise ratio from decibels to linear and dividing the signal power by this ratio; X noise The generated rule has a mean of 0 and a standard deviation of sqrt(P n ) of Gaussian white noise, randn(len(X train Generate random samples of the same length as the signal and multiply them by the square root of the noise power; Indicates the generated signal after adding noise.

6. The method according to claim 3, characterized in that The S3 includes: S31, the pre-processed data is input to the CBAM module for processing, then to the FiLM module for processing, and finally to the CNN module for processing; S32. Train the anomaly detection and warning model CBAM-FiLM-CNN, select a suitable loss function to measure the difference between the model prediction results and the true labels, and use an optimization algorithm to adjust the model parameters to minimize the loss function. S33. Through continuous iterative training, the model learns the characteristic patterns of the data.

7. The method according to claim 3, characterized in that The improved Newton-Raphson optimization algorithm NRBO formula described in S4 is as follows: in(n)=in max *(In max -In min )*(Myth-n) / Myth; in, represents the position after individual disturbance, sin(n) represents the sine function, cos(n) represents the cosine function, n represents the number of iterations, w(n) represents the adaptive inertia weight, P represents the random probability, w max represents the maximum inertia weight, w min Represents the minimum inertia weight, and Mit represents the maximum number of iterations.

8. The method according to claim 3, characterized in that The warning levels described in S6 include: Minor anomalies are considered low-level warnings, and subsequent data changes require attention; If the abnormality does not threaten the core functions of the system, it is judged as a medium-level warning and preliminary response measures need to be initiated; Abnormal situations that seriously threaten the normal operation of the system and may cause major failures will be judged as high-level warnings, triggering the emergency response mechanism and emergency repairs.

9. A computer device, characterized in that: The method comprises one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and when the programs are executed by the processors, the steps of the method according to claims 3-8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to claims 3 to 8 are implemented.

Images