Chemical production monitoring system based on time series prediction

Abstract

The present application relates to the technical field of chemical production monitoring, and discloses a kind of chemical production monitoring system based on timing prediction, comprising: S1: information acquisition module gathers multiple data in production process, wherein multiple data is divided into risk timing data and conventional timing data according to process or function class, S2: risk timing data is extracted timing feature by CNN, timing prediction is carried out, S3: the standard range of RMSE in timing prediction is set, the RMSE in risk timing data prediction is obtained multiple times, and the RMSE over-range time period is extracted;Conventional timing data abnormal parameters are stored, and the correlation between the conventional timing data abnormal parameters and the risk timing data abnormal parameters is obtained by experiencing multiple data storage;When conventional timing data is in dynamic change state, conventional timing data is promoted to risk timing data, and coupled monitoring is carried out by introducing model, dynamic computing power is increased according to production line use time length, so that monitoring quantity and production line state are coupled.

Description

Technical Field

[0001] This invention relates to the field of chemical production monitoring technology, specifically a chemical production monitoring system based on time-series prediction. Background Technology

[0002] Chemical process risk early warning technology mainly focuses on process fault monitoring and prediction. Chemical process data has dynamic time series characteristics. Although the use of time series prediction methods as early warning technology has made use of the data accumulated by enterprises, it currently mainly focuses on predicting some core parameters. When there is a large deviation between the predicted value and the actual value, the system often reflects it in the form of anomalies or errors, without performing source tracing prediction.

[0003] For example, in a chemical synthesis production line, the main equipment, such as synthesis towers, heat exchangers, feeders, separators, and furnaces, are used in specific processes for a particular chemical substance. These processes directly impact chemical production. When predicting such parameters using models, a significant deviation between the predicted and actual values ​​is defined as a hazard. However, multiple parameters are interconnected in chemical production. A significant deviation between predicted and actual values ​​may be due to the introduction of new variables, which could then become key parameters affecting chemical production.

[0004] However, in the production process, there are too many nodes that can be controlled or monitored. The synchronous prediction of all nodes leads to too many data categories, and the timing of the occurrence of variables in ordinary nodes is random. Full parameter prediction results in a large waste of computing power. Summary of the Invention

[0005] The purpose of this invention is to provide a monitoring system for chemical production based on time-series prediction. Based on time-series prediction, the information acquisition module synchronously collects risk time-series data and regular time-series data, identifies significant changes in the regular time-series data, and stores abnormal parameters of the regular time-series data. Through multiple data storage processes, the correlation between the abnormal parameters of the regular time-series data and the abnormal parameters of the risk time-series data is obtained. Furthermore, when the regular time-series data is in a dynamic and variable state, it is upgraded to risk time-series data and introduced into a model for coupled monitoring, reducing prediction computational power. Simultaneously, the computational power is dynamically increased based on the production line usage time, thus coupling the monitored quantities with the production line status to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A time-series prediction-based monitoring system for chemical production includes: S1: The information acquisition module collects multiple data points during the production process. These multiple data points are divided into risk time-series data and regular time-series data according to the process or function. The information acquisition module collects risk time-series data and regular time-series data simultaneously. S2: Extract time-series features from risk time-series data using CNN. The CNN output features are used as input to LSTM, which consists of a chain of units including forget gate, input gate, and output gate, for time-series prediction. S3: Set the standard range of RMSE in time series forecasting, obtain the RMSE in risk time series data forecasting multiple times, and extract the time period when the RMSE exceeds the range; S4: Based on the RMSE out-of-range time period, extract regular time data, obtain variables from regular time series data, and based on the frequency of occurrence of the RMSE out-of-range time period and the synchronous change frequency of regular time series data, elevate the correlation data between regular time series data and risk time series data to risk time series data for long-term prediction. S5: Quantitative risk assessment of the predicted data is performed using triangular fuzzy numbers, classifying the risk level of the predicted data into low risk, Level I risk, and Level II risk, and determining the risk level of early warning indicators in future risk time series data.

[0007] As a further aspect of the present invention: in step S1, the risk time series data includes the operating data of multiple core equipment in the production line, and the regular time series data includes other data in the production line besides the risk time series data. The other data includes multiple parameters or data in the production line such as transportation, employees, storage, and environment that do not directly affect the safety of chemical production.

[0008] As a further aspect of this invention: before predicting risk time series data, data normalization is performed, training data is scaled to 0-1, dimensions are removed, consistency analysis is achieved, and the transformation function equation is: in, , These represent the minimum and maximum values ​​of the sample data, respectively. This represents the normalized value of the sample data.

[0009] As a further aspect of the present invention: the updated risk time series data in step S4 is used for time series prediction in step S2.

[0010] As a further aspect of the present invention: In step S2, the CNN convolution operation formula is as follows: This represents a set of feature vectors input to the CNN convolutional layer. Indicates the first l-1 layer network i The output feature vector of each convolutional kernel Indicates the first i The weight values ​​of the l-th convolutional kernel in layer l. The asterisk (*) represents the bias value, and the convolution operation is used to determine the final output value. express.

[0011] As a further aspect of this invention: dimensionality reduction is performed on the vector output by the convolutional layer using a pooling layer. After the convolutional pooling operation, at least one fully connected layer is connected, and the output formula of the fully connected layer is: y represents the output of the fully connected layer. For input to the fully connected layer, These are the weighting coefficients. For the bias term, ( f ·) represents a non-linear activation function.

[0012] As a further aspect of the present invention: LSTM consists of a forget gate, an input gate, and an output gate. The output value of the forget gate determines the proportion of information discarded by the memory unit at the previous time step. The input gate determines which new information the network unit should store at the current time step. The output gate determines the output of the LSTM network unit at the current time step. At the same time, the output of the gate is also used as the input signal for the next time step and transmitted to the next time step.

[0013] As a further aspect of the present invention: an internal memory unit is provided between the input gate and the output gate, and the internal memory unit is used to update and obtain new network unit states.

[0014] Compared with the prior art, the beneficial effects of the present invention are: Based on time series prediction, the information acquisition module synchronously collects risk time series data and regular time series data, and obtains significant changes in regular time series data. It also stores abnormal parameters of regular time series data. After multiple data storage, the correlation between the abnormal parameters of regular time series data and the abnormal parameters of risk time series data is obtained. When regular time series data is in a dynamic and volatile state, it is upgraded to risk time series data and introduced into the model for coupled monitoring, reducing prediction computing power. At the same time, the computing power is dynamically increased according to the production line usage time, so that the monitoring quantity and the production line status are coupled together. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a diagram of the LSTM cell structure. Figure 2 This is a schematic diagram of a time-series prediction-based monitoring system for chemical production. Detailed Implementation

[0017] Please see Figure 1 In this embodiment, the steps of the monitoring system for chemical production based on time-series prediction are as follows: S1: The information acquisition module collects multiple data points during the production process. These data points are categorized into risk time-series data and regular time-series data according to their process or function. The information acquisition module collects both risk time-series data and regular time-series data simultaneously.

[0018] In step S1, multiple monitoring parameters of the intelligent chemical production line are recorded synchronously. Risk time-series data includes the operating data of core equipment in multiple processes of the production line, while regular time-series data includes other data in the production line besides risk time-series data. Other data includes parameters or data related to transportation, employees, storage, and the environment that do not directly affect the safety of chemical production. Only risk time-series data is predicted to reduce computing power. For example, fluctuations in the opening diameter of flow control valves or other factors that do not directly affect risk time-series data are not predicted during the initial stage of equipment operation.

[0019] S2: Extract time-series features from risk time-series data using CNN. The CNN output features are used as input to LSTM, which consists of a chain of units including forget gate, input gate, and output gate, for time-series prediction. Before predicting risk time series data, data normalization is performed, training data is scaled to 0-1, dimensions are removed, consistency analysis is performed, and the transformation function equation is: in, , These represent the minimum and maximum values ​​of the sample data, respectively. This represents the normalized value of the sample data.

[0020] In step S2, the CNN convolution operation formula is as follows: This represents a set of feature vectors input to the CNN convolutional layer. Indicates the first l -1 layer network i The output feature vector of each convolutional kernel Indicates the first i The weight values ​​of the l-th convolutional kernel in layer l. The asterisk (*) represents the bias value, and the convolution operation is used to determine the final output value. express.

[0021] The convolutional layer is the most crucial structural unit in a convolutional neural network (CNN), its function being to extract different local features from the input vector matrix. A complete CNN typically contains multiple convolutional layers, and the weight parameters of these layers are shared. This characteristic helps to mitigate the overfitting problem of traditional neural networks to some extent. The convolution operation can be understood as the convolution kernel sliding across the input signal, taking a segment of data of the same size as the kernel for computation each time, resulting in the output value of the convolution operation.

[0022] Pooling layers perform feature dimensionality reduction on the vectors output by convolutional layers. After the convolutional pooling operation, at least one fully connected layer is connected. Pooling layers, also known as downsampling layers, are another important network structure unit in convolutional neural networks, primarily responsible for performing feature dimensionality reduction on the vectors output by convolutional layers. The operation of pooling layers is similar to that of convolution, calculating and outputting elements of the pooling kernel's receptive field from the input feature vector according to rules. The pooling operation does not require complex learning parameters. Simply specify the pooling type to be performed. Pooling types generally include max pooling and average pooling. Max pooling uses the maximum value of the elements in the receptive field of the pooling kernel as the output, while average pooling uses the average value of the elements in the receptive field of the pooling kernel as the output. The results of these pooling operations will form the feature vector of the next layer in the convolutional neural network. In convolutional neural networks, convolutional layers and pooling layers are generally stacked alternately to "compress and purify" the input data, extract local features, and sample and combine them to generate hierarchical feature samples.

[0023] After convolutional pooling operations, one or more fully connected layers are typically connected. Fully connected layers are the main components in convolutional neural networks. The fully connected layer is responsible for fitting the feature vector output by the pooling layer. It assigns a weight to each feature in the previous layer, which can be interpreted as the contribution rate of that feature to the stability conclusion. The output formula for the fully connected layer is: y represents the output of the fully connected layer. For input to the fully connected layer, These are the weighting coefficients. For the bias term, ( f ·) represents a non-linear activation function.

[0024] LSTM consists of a forget gate, an input gate, and an output gate.

[0025] The output value of the forget gate determines the proportion of information discarded from the memory unit in the previous time step. The specific formula is as follows: The input gate determines what new information the network unit should store at any given time, and the specific formula is as follows: Internal memory unit: This unit is the third step in the LSTM unit's computation process, and its main function is to obtain new... Network unit status, updated information is provided by ⊙ as well as ⊙ The composition, specifically the formula, is as follows: The output gate determines the output of the LSTM network unit at the current time step, and the result of the gate output also serves as the input signal for the next time step. The specific formula is as follows: in, This represents the input value of the process parameters at time t, and the intermediate output. This represents the output value of the hidden state at time t-1, which serves as the input at time t. It's the sigmoid activation function, which updates the value using a number between 0 and 1, where 0 represents discarding completely and 1 represents retaining completely. The cell state is input at time t. This represents the output information of the memory unit at time t-1, and tanh represents the hyperbolic tangent activation function. , , , The matrix weights correspond to the respective gates mentioned above. , , , These are the bias terms for the corresponding gates.

[0026] The collected risk time series data are treated as independent time series. and will As input data for the CNN-LSTM network, a one-dimensional convolutional model is used to predict future early warning indicator data. The hyperparameters of the deep learning time-series prediction model are initially set based on human experience, and then the optimal parameters are selected based on the performance on the validation set. The number of neurons in the CNN input layer is set to 10, the number of convolutional layers is set to 2, and the convolutional kernel size is 2×1. During the model building process, more features can be extracted by changing the number of convolutional kernels. In this study, the number of neurons within the convolutional kernels is set to 16 and 32 respectively. The vector array obtained by feature extraction from the input data through two convolutional kernels is flattened and finally output as a one-dimensional vector array as features. The model uses the ReLU activation function, which solves the gradient vanishing problem.

[0027] S3: Set the standard range of RMSE in time series forecasting, obtain the RMSE in the risk time series data forecast multiple times, and extract the time period when the RMSE exceeds the range. The risk time series data updated in step S4 is used for time series forecasting through step S2.

[0028] In the formula, N represents the total number of predictions. , These represent the predicted value and the actual value at time i, respectively.

[0029] S4: Based on the RMSE out-of-range time period, extract regular time data, obtain variables from regular time series data, and based on the occurrence frequency of the RMSE out-of-range time period and the synchronous change frequency of regular time series data, elevate the correlation data between regular time series data and risk time series data to risk time series data for long-term prediction.

[0030] For example, please refer to the following: (Using a level tank as an example) Figure 2 The liquid level in the tank showed obvious abnormal fluctuations in the red box area, and the fluctuation range was significantly higher than that in other areas. Using this time as a standard, regular time data was extracted, and significant changes in regular time series data were obtained. The abnormal parameters of the regular time series data were stored. After multiple data storage, the correlation between the abnormal parameters of regular time series data and the abnormal parameters of risk time series data was obtained. When the regular time series data was in a dynamic and volatile state, the regular time series data was upgraded to risk time series data and introduced into the model for coupled monitoring.

[0031] S5: Quantitative risk assessment of the predicted data is performed using triangular fuzzy numbers, classifying the risk level of the predicted data into high, medium, and low risks, and determining the risk level of early warning indicators in future risk time series data.

[0032] Triangular fuzzy numbers, as a common type of fuzzy number in fuzzy mathematics theory, are among the most widely used and important fuzzy numbers due to their simple structure and ease of understanding and conversion. In recent years, triangular fuzzy numbers have been widely applied in fuzzy control, fuzzy inference, and fuzzy risk decision-making. In fuzzy control, triangular fuzzy numbers are often used to represent the degree of fuzziness of inputs or outputs, such as temperature and humidity. In fuzzy inference and fuzzy risk decision-making, triangular fuzzy numbers can be used to describe fuzzy conditions or fuzzy outputs in rules.

[0033] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A monitoring system for chemical production based on time-series prediction, characterized in that: include: S1: The information acquisition module collects multiple data points during the production process. These multiple data points are divided into risk time-series data and regular time-series data according to the process or function. The information acquisition module collects risk time-series data and regular time-series data simultaneously. S2: Extract time-series features from risk time-series data using CNN. The CNN output features are used as input to LSTM, which consists of a chain of units including forget gate, input gate, and output gate, for time-series prediction. S3: Set the standard range of RMSE in time series forecasting, obtain the RMSE in risk time series data forecasting multiple times, and extract the time period when the RMSE exceeds the range; S4: Based on the RMSE out-of-range time period, extract regular time data, obtain variables from regular time series data, and based on the frequency of occurrence of the RMSE out-of-range time period and the synchronous change frequency of regular time series data, elevate the correlation data between regular time series data and risk time series data to risk time series data for long-term prediction. S5: Quantitative risk assessment of the predicted data is performed using triangular fuzzy numbers, classifying the risk level of the predicted data into low risk, Level I risk, and Level II risk, and determining the risk level of early warning indicators in future risk time series data.

2. The monitoring system for chemical production based on time-series prediction according to claim 1, characterized in that: In step S1, the risk time series data includes the operating data of core equipment in multiple processes of the production line, and the regular time series data includes other data in the production line besides the risk time series data. The other data includes multiple parameters or data in the production line such as transportation, employees, storage, and environment that do not directly affect the safety of chemical production.

3. The monitoring system for chemical production based on time-series prediction according to claim 1, characterized in that: Before predicting risk time series data, data normalization is performed, training data is scaled to 0-1, dimensions are removed, consistency analysis is performed, and the transformation function equation is: in, , These represent the minimum and maximum values ​​of the sample data, respectively. This represents the normalized value of the sample data.

4. The monitoring system for chemical production based on time-series prediction according to claim 1, characterized in that: The updated risk time series data in step S4 is used for time series prediction in step S2.

5. A monitoring system for chemical production based on time-series prediction according to claim 1, characterized in that: In step S2, the CNN convolution operation formula is as follows: This represents a set of feature vectors input to the CNN convolutional layer. Indicates the first l -1 layer network i The output feature vector of each convolutional kernel Indicates the first i The weight values ​​of the l-th convolutional kernel in layer l. The asterisk (*) represents the bias value, and the convolution operation is used to determine the final output value. express.

6. A monitoring system for chemical production based on time-series prediction according to claim 1, characterized in that: The convolutional layer's output vector is subjected to feature dimensionality reduction via pooling layers. After the convolutional pooling operation, at least one fully connected layer is connected. The output formula of the fully connected layer is: y represents the output of the fully connected layer. For input to the fully connected layer, These are the weighting coefficients. For the bias term, ( f ·) represents a non-linear activation function.

7. A monitoring system for chemical production based on time-series prediction according to claim 1, characterized in that: LSTM consists of a forget gate, an input gate, and an output gate. The output value of the forget gate determines the proportion of information discarded from the memory unit at the previous time step. The input gate determines which new information the network unit should store at the current time step. The output gate determines the output of the LSTM network unit at the current time step, and the output of the gate is also used as the input signal for the next time step and passed to the next time step.

8. A monitoring system for chemical production based on time-series prediction according to claim 1, characterized in that: There is an internal memory unit between the input gate and the output gate, which is used to update and obtain the new network unit state.

Images