Chemical process multi-parameter intelligent monitoring method based on deep learning

Abstract

The invention discloses a chemical process multi-parameter intelligent monitoring method based on deep learning, and the method comprises the following steps: collecting and preprocessing the data of a chemical production process, and obtaining a standardized input sequence; according to the chemical process flow diagram and the equipment topological relation, a standardized process structure matrix is obtained; inputting the standardized input sequence and the standardized process structure matrix into a constraint decoupling echo state network model, and outputting a hidden state sequence; performing multi-scale gating fusion on the hidden state sequence, and outputting a parameter prediction result; calculating a residual error based on a parameter prediction result and an actual observation value, and outputting an anomaly detection result and channel-level attribution information; triggering a self-adaptive correction process to form a monitoring model after self-adaptive correction; and outputting a parameter prediction result, an anomaly detection result and channel-level attribution information based on the monitoring model after adaptive correction. According to the invention, the constraint decoupling echo state network is adopted, and multi-parameter intelligent monitoring of the chemical process is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring and industrial process control technology for chemical processes, and in particular to a method for intelligent monitoring of multiple parameters in chemical processes based on deep learning. Background Technology

[0002] In chemical production processes, numerous operating parameters exist for process units, and complex nonlinear coupling relationships exist between variables. Some key parameters are difficult to obtain in real time through direct measurement methods. Existing chemical process monitoring methods mainly rely on traditional statistical analysis models or neural network models based on fixed structures.

[0003] Statistical modeling methods are poorly adapted to nonlinear dynamic characteristics and cannot accurately reflect the dynamic relationship between multiple parameters. While conventional neural network models have certain feature learning capabilities, they have high training costs and lack interpretability, failing to reflect the actual physical structure and energy transfer path of chemical processes.

[0004] Furthermore, when processing multi-parameter chemical process data, traditional echo state networks typically generate their reservoir connections randomly, failing to constrain them by the topological structure between chemical equipment. This results in the feature propagation direction being inconsistent with the actual process logic, easily generating spurious correlations and affecting the accuracy of anomaly detection.

[0005] Existing methods typically require model retraining when faced with process changes or data distribution drift, resulting in high update costs and making it difficult to meet the real-time monitoring and adaptive needs of chemical plant sites.

[0006] In summary, existing multi-parameter monitoring technologies for chemical processes generally suffer from problems such as low coupling between modeling structure and process characteristics, insufficient anomaly detection accuracy, and high model maintenance costs.

[0007] Therefore, how to provide a deep learning-based intelligent monitoring method for multiple parameters in chemical processes is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0008] One objective of this invention is to propose a deep learning-based intelligent monitoring method for multiple parameters in chemical processes. This invention constructs a constrained decoupled echo state network model, incorporating chemical process structure information into the connection structure between the network input layer and the reservoir, thereby achieving structural constraints and feature decoupling of multi-parameter signals. The constrained decoupled echo state network model limits the information propagation path through a process structure matrix, ensuring that the network state evolution conforms to the actual material flow and energy transfer patterns of chemical processes. A multi-scale gating fusion mechanism is used to extract dynamic features at different time scales, and the output readout layer generates chemical process parameter prediction results. Anomaly detection and channel-level attribution are achieved through structure-weighted residuals and adaptive thresholds. This invention has the advantages of strong structural interpretability, high anomaly detection accuracy, and good model adaptability, enabling real-time monitoring and intelligent early warning of multiple parameters in chemical production processes.

[0009] The intelligent monitoring method for multiple parameters of a chemical process based on deep learning according to embodiments of the present invention includes the following steps: Collect multi-parameter time-series data of chemical production processes, preprocess the multi-parameter time-series data, and obtain standardized input sequences; Based on the chemical process flow diagram and equipment topology, a process structure matrix is ​​constructed, and symmetry and sparsity checks and normalization are performed to obtain a standardized process structure matrix. The echo state network model is decoupled from the input constraints of the standardized input sequence and the standardized process structure matrix. Structural constraint projection and feature decoupling are then performed to output the hidden state sequence. Multi-scale gating fusion is performed on the hidden state sequence to obtain the fused state representation, and the readout layer parameters are trained based on the fused state to output the parameter prediction results; The residuals are calculated based on the parameter prediction results and the actual observations, anomaly scoring results are generated, anomaly judgment boundaries are determined, and anomaly detection results and channel-level attribution information are output. When a process structure adjustment or data distribution change is detected, the adaptive correction process of the constraint decoupling echo state network model is triggered, forming an adaptively corrected monitoring model. The adaptively calibrated monitoring model is deployed on edge computing nodes at the chemical production site to receive new standardized input sequences in real time and output parameter prediction results, anomaly detection results, and channel-level attribution information.

[0010] Optionally, the step of collecting multi-parameter time-series data of the chemical production process and preprocessing the multi-parameter time-series data to obtain a standardized input sequence specifically includes: Sensors deployed in chemical production plants, including reactors, heat exchangers, separation towers, storage tanks, and pipeline nodes, are used to collect six types of chemical process parameters: temperature, pressure, flow rate, liquid level, concentration, and pH value, forming a raw multi-parameter time-series dataset. The original multi-parameter time series dataset is time-synchronized. A uniform sampling time interval is set, and the parameter values ​​collected by each sensor at different time points are rearranged in chronological order so that all chemical process parameters have corresponding data values ​​at each sampling time, forming an aligned data sequence arranged by time steps. Each time step corresponds to a vector containing all chemical process parameters, which is used to represent the chemical production operation status at the current time. Sensor calibration is performed on the time-aligned data sequence, measurement points with calibration errors exceeding the allowable range are removed, and the moving average method is used to smooth the boundary abrupt data to obtain the calibration sequence; Missing values ​​were imputed in the calibration sequence, and the data gaps were filled by linear interpolation or the mean of adjacent time points to obtain a continuous sequence of chemical process parameters. Noise suppression processing is performed on the parameter sequence of the continuous chemical process. High-frequency random noise is eliminated by the sliding window averaging method or wavelet threshold denoising method to obtain the filtered sequence. The filtered sequence is standardized by first calculating the mean and standard deviation of each chemical process parameter over the entire sampling period. Then, the original value of each parameter in each time step is subtracted from the mean value of each parameter, and the difference is divided by the corresponding standard deviation to obtain the standardized value of each parameter at the corresponding time step. After the standardization process is completed, the value range of all chemical process parameters is unified to the same scale. The standardized time series data of multiple parameters of chemical process are arranged in chronological order to form a standardized input sequence.

[0011] Optionally, the step of constructing a process structure matrix based on the chemical process flow diagram and equipment topology, and performing symmetry and sparsity checks and normalization to obtain a standardized process structure matrix specifically includes: Based on the chemical process flow diagram, identify all equipment units in the chemical production system, including reactors, heat exchangers, separation towers, storage tanks, and pipeline nodes. Clarify the material flow and energy flow directions between each equipment unit, define each equipment unit as a node, and define the connection between nodes that has a material flow or energy flow relationship as a directed connection. Construct an equipment topology diagram according to the relationship between nodes and directed connections to fully describe the structural composition of the chemical production system and the material flow and energy flow paths. Based on the equipment topology diagram, a process structure matrix is ​​constructed according to the connection relationship between the nodes in the equipment topology diagram. The rows and columns of the matrix correspond to each equipment node in the chemical production system. When there is a direct material or energy transfer from one node to another, the corresponding row and column position in the matrix is ​​marked as one. When there is no direct transfer relationship between two nodes, the corresponding row and column position in the matrix is ​​marked as zero, thus forming a process structure matrix that represents the direct connection relationship between equipment in the chemical production system. Symmetry verification of the process structure matrix: If there are both positive and negative material transfer relationships or energy transfer relationships between two equipment nodes in a chemical production system, then the two corresponding positions in the process structure matrix are marked as one, so that the process structure matrix has the same value in the row and column positions corresponding to the nodes, thereby forming a symmetric process structure matrix that can reflect bidirectional transfer relationships. The sparsity of the symmetric matrix is ​​checked by removing rows and columns corresponding to isolated nodes that have no connection to any other node, in order to ensure the connectivity of the matrix. The process structure matrix after sparsity verification is normalized by calculating the sum of all elements in each row of the matrix in turn, and dividing the value of each element in the row by the sum of the elements in the row so that the sum of all elements in the row is equal to one. This yields the normalized result of the data in each row at the same scale, forming a standardized process structure matrix.

[0012] Optionally, the step of decoupling the standardized input sequence and the standardized process structure matrix input constraints of the echo state network model, performing structural constraint projection and feature decoupling, and outputting the hidden state sequence specifically includes: A constrained decoupled echo state network model is established, which includes an input mapping layer, a structural constraint projection layer, a reservoir state layer, a state fusion layer, and an output readout layer. The input mapping layer is used to receive the standardized input sequence, the structural constraint projection layer is used to introduce the process structure matrix, the reservoir state layer is used to perform temporal state propagation, the state fusion layer is used for multi-scale gating integration, and the output readout layer is used to generate parameter prediction results. The standardized input sequence and the standardized process structure matrix are input into the structural constraint projection layer. According to the connection relationship of the equipment nodes represented by each element in the process structure matrix, matrix multiplication is performed on the input parameter vector of each time step. The input parameters are weighted and combined with the structural weights of the corresponding nodes to obtain the projected input vector after process structure constraint. The projected input vector and the standardized input vector are concatenated along the feature dimension to form a combined input vector that contains process structure information and original parameter information; A reservoir connection mask matrix is ​​generated based on the standardized process structure matrix. Each element in the process structure matrix is ​​matched with the corresponding connection position in the reservoir. When the value of an element in the process structure matrix is ​​greater than zero, it indicates that there is a connection relationship between the corresponding equipment nodes. Then, the corresponding position in the reservoir connection mask matrix is ​​marked as one. When the value of an element in the process structure matrix is ​​equal to zero, it indicates that there is no connection between the corresponding nodes. In this case, the corresponding position in the reservoir connection mask matrix is ​​marked as zero. Subsequently, a random sparse matrix with the same size as the reservoir connection mask matrix is ​​generated. The element values ​​of the random sparse matrix are decimals randomly distributed within a preset range. The reservoir connection mask matrix is ​​multiplied by the corresponding element of the random sparse matrix to obtain the reservoir connection matrix. This ensures that the connection relationship within the reservoir exists only between nodes allowed by the process structure matrix, thereby forming a reservoir connection topology constrained by the process structure. To adjust the spectral radius of the reservoir connection matrix, firstly, all eigenvalues ​​of the reservoir connection matrix are calculated, and the largest eigenvalue is determined as the current spectral radius. Then, based on the preset echo stability condition, all elements in the reservoir connection matrix are scaled by the same proportion so that the adjusted largest eigenvalue falls within the range between zero and one. This ensures that the reservoir has convergence and dynamic stability during state propagation and avoids numerical divergence in the network during time-series updates. At each time step, a state update is performed based on the combined input vector and the hidden state vector from the previous time step to generate the hidden state vector for the current time step: ; in, This indicates that the constrained decoupled echo state network is at time step The hidden state vector is used to store the network's internal state at the current moment. This represents the hidden state vector from the previous time step, used to record the network's internal state at the previous moment. Indicates at time step The leakage rate parameter, a real number between 0 and 1, is used to control the proportion of inheritance from the previous state to the current state. This represents the hyperbolic tangent activation function, used to perform a nonlinear mapping on the linearly weighted result. This represents the input mapping matrix, used to map the combined input vectors to the reservoir space. Indicates time step The combined input vector, The pool connection matrix is ​​obtained by multiplying the corresponding elements of the pool connection mask matrix and the random sparse matrix, and is used to define the connection topology inside the pool. The hidden state vectors obtained in all time steps are arranged in chronological order to form a complete hidden state sequence.

[0013] Optionally, the step of performing multi-scale gated fusion on the hidden state sequence to obtain a fused state representation, and training the readout layer parameters based on the fused state, specifically includes the following: Multi-scale gated fusion is performed on the hidden state sequence. The hidden state sequence is decomposed according to the pre-set time scale parameters, sparsity parameters and spectral radius parameters. The decomposed hidden state sequence is input into multiple parallel sub-state pools. Each sub-state pool independently performs state propagation and state update operations according to its own parameter configuration, so as to dynamically respond to the input signal at different time scales and generate corresponding hidden state sub-sequences. Each sub-state sub-sequence represents the temporal feature change of the sub-state pool at the corresponding time scale. At each time step, the combined input vector of the current time step and the hidden state vector of the previous time step are input together to the gating weight calculation unit. The gating weight value corresponding to each sub-state pool is calculated through linear weighting and nonlinear transformation. Each gating weight value reflects the degree of participation of the sub-state pool in the overall state fusion at the current time step. The gating weight values ​​of all sub-state pools are normalized so that the sum of all gating weights is one, thereby ensuring that the weight ratio of different sub-state pools in the state fusion process remains a uniform normalization constraint. At each time step, the hidden state vectors output by all sub-state pools are weighted and summed according to the corresponding gating weights. The hidden state vectors of each sub-state pool are multiplied by the corresponding gating weights and then summed to obtain the fused state vector of the current time step. The fused state vector integrates the state information of each sub-state pool at different time scales and is used to characterize the overall dynamic characteristics of the chemical process at the current time step. The fused state vectors are arranged in chronological order to form a fused state sequence. In the output readout layer, set the output weight matrix, keep the parameters of the input mapping matrix and the reservoir connection matrix unchanged, and only train the parameters of the output weight matrix to minimize the prediction error or reconstruction error. The fused state sequence is linearly mapped using the trained output weight matrix, and the parameter prediction vector at each time step is calculated to obtain the parameter prediction result.

[0014] Optionally, the step of calculating the residual based on the parameter prediction results and the actual observed values, generating anomaly scoring results, determining anomaly judgment boundaries, and outputting anomaly detection results and channel-level attribution information specifically includes: At each time step, the difference between the parameter prediction result and the actual observation value at the corresponding time step is calculated. The difference between the predicted value and the actual value is calculated for each parameter component in the multi-parameter of the chemical process to form a residual vector containing the differences of all monitored parameters. The residual vector is weighted according to the standardized process structure matrix. Each element in the process structure matrix and the corresponding parameter component in the residual vector are weighted and summed according to the matrix multiplication relationship. This ensures that each residual component is affected by the weights of the related equipment nodes and process paths during the calculation, thus obtaining a structure-weighted residual vector. The structure-weighted residual vector reflects the residual propagation relationship between the monitoring parameters under the topological constraints of the chemical process, and provides input for the subsequent calculation of the mean squared error. The components of the structural weighted residual vector are squared sequentially. The weighted residual values ​​of each parameter channel are squared and summed, and then divided by the total number of parameters to calculate the average value, thus obtaining the weighted average squared error of the current time step. The weighted average squared error is used to represent the overall error level of multiple parameters in the chemical process under the constraints of the process structure, and provides input data for subsequent adaptive threshold calculation and abnormal state judgment. The statistical characteristics of error distribution are calculated using historical steady-state operation data samples of chemical processes. The mean and standard deviation of the squared mean of errors are obtained respectively. The mean is used as the benchmark error level and the standard deviation is used as the fluctuation range index. An adaptive threshold is calculated based on these two statistics, such that the threshold is equal to the sum of the mean and standard deviation multiplied by the adjustment coefficient, where the adjustment coefficient is a preset constant used to control the sensitivity of the threshold. The adaptive threshold is used to determine the anomaly of the weighted mean squared error of the current time step in subsequent steps. At each time step, the weighted mean square error of the current time step is compared with the corresponding adaptive threshold. When the weighted mean square error is greater than the adaptive threshold, the current time step is determined to be an abnormal state, and the abnormality identifier is recorded. At the same time, the abnormality detection result corresponding to the current time step is generated to indicate that the chemical process deviates from the normal operating pattern at the current time step. After the current time step is in an abnormal state, the structure-weighted residual vector corresponding to the current time step is read. The numerical amplitudes of each parameter component in the residual vector are compared and sorted. The parameter component with the larger amplitude is regarded as the monitoring channel that contributes more to the abnormal state. Based on the sorting result, the channel-level attribution information is output to indicate the specific monitoring parameter channel or related chemical equipment node that caused the abnormality, providing a basis for subsequent abnormality location and process adjustment.

[0015] Optionally, the step of triggering the adaptive correction process of the constraint-decoupled echo state network model when a process structure adjustment or data distribution change is detected, to form an adaptively corrected monitoring model, specifically includes: During the operation of the chemical process, the distribution characteristics of the input data and the structural information of the process parameters are continuously monitored. When a significant shift in the data distribution compared with the historical steady-state distribution is detected, or when adjustments are detected in the process flow diagram or equipment connection relationship, the model adaptive correction process is triggered. When performing adaptive correction, the state pool topology and standardized process structure matrix of the constrained decoupled echo state network model remain unchanged, and the parameters of the reserve pool connection matrix and input mapping matrix remain fixed. Incremental updates are performed on the output weight matrix of the output readout layer and the gate weight parameters in the multi-scale gated fusion to obtain new output weight matrix and gate weight parameters; The new output weight matrix and gating weight parameters are reloaded into the constrained decoupled echo state network model to form an adaptively corrected monitoring model.

[0016] Optionally, the incremental update includes calculating a correction amount based on the difference between the actual observed value and the corresponding prediction result at each time step, adjusting the output weight matrix of the output readout layer according to the set learning rate, so that the new weight matrix is ​​updated in the direction of reducing error on the original basis, and adjusting the gating weight parameters of each sub-state pool according to the difference between the weighted average of the squared error at the current time step and the adaptive threshold, so that the gating weights can adapt to the changes in the current operating state while maintaining the normalization constraint.

[0017] The beneficial effects of this invention are: This invention introduces a constrained decoupling echo state network model, embedding the chemical process structure matrix into a deep learning framework, thus achieving an organic combination of process topology and network structure. This design enables the model to retain the real physical constraints between chemical equipment during feature extraction, improving the accuracy and interpretability of time-series feature representation. Through a multi-scale gating fusion mechanism, this invention can capture the dynamic changes of chemical processes at different time scales, allowing the model to maintain stable predictive capabilities even when facing complex operating conditions and fluctuations.

[0018] Furthermore, this invention utilizes structure-weighted residuals and adaptive threshold strategies in the anomaly detection stage to achieve precise identification of abnormal states. It can quantitatively evaluate multi-parameter anomalies in chemical processes and locate key equipment or parameter channels that may cause anomalies through channel-level attribution methods, thereby improving the accuracy and response speed of anomaly diagnosis.

[0019] This invention further employs an adaptive correction mechanism with a fixed structure and incremental updates to the readout layer. This allows for rapid adaptive updates without retraining the network when processes are adjusted or data distribution changes, ensuring the model's long-term usability and online stability. Overall, this method significantly improves upon existing technologies in terms of model lightweighting, feature interpretability, anomaly detection accuracy, and adaptability, providing a structurally sound and stable solution for multi-parameter intelligent monitoring of chemical production processes. Attached Figure Description

[0020] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0021] Figure 1 This is a flowchart of the intelligent monitoring method for multiple parameters of chemical processes based on deep learning proposed in this invention; Figure 2 This is a schematic diagram of the constrained decoupling echo state network model in the deep learning-based intelligent monitoring method for multi-parameter chemical processes proposed in this invention. Figure 3 This is a schematic diagram illustrating the relationship between the construction of the process structure matrix and structural constraints in the deep learning-based intelligent monitoring method for multiple parameters of chemical processes proposed in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figure 1-3 A deep learning-based intelligent monitoring method for multiple parameters in chemical processes includes the following steps: Collect multi-parameter time-series data of chemical production processes, preprocess the multi-parameter time-series data, and obtain standardized input sequences; Based on the chemical process flow diagram and equipment topology, a process structure matrix is ​​constructed, and symmetry and sparsity checks and normalization are performed to obtain a standardized process structure matrix. The echo state network model is decoupled from the input constraints of the standardized input sequence and the standardized process structure matrix. Structural constraint projection and feature decoupling are then performed to output the hidden state sequence. Multi-scale gating fusion is performed on the hidden state sequence to obtain the fused state representation, and the readout layer parameters are trained based on the fused state to output the parameter prediction results; The residuals are calculated based on the parameter prediction results and the actual observations, anomaly scoring results are generated, anomaly judgment boundaries are determined, and anomaly detection results and channel-level attribution information are output. When a process structure adjustment or data distribution change is detected, the adaptive correction process of the constraint decoupling echo state network model is triggered, forming an adaptively corrected monitoring model. The adaptively calibrated monitoring model is deployed on edge computing nodes at the chemical production site to receive new standardized input sequences in real time and output parameter prediction results, anomaly detection results, and channel-level attribution information.

[0024] In this embodiment, the process of collecting multi-parameter time-series data of the chemical production process and preprocessing the multi-parameter time-series data to obtain a standardized input sequence specifically includes: Sensors deployed in chemical production plants, including reactors, heat exchangers, separation towers, storage tanks, and pipeline nodes, are used to collect six types of chemical process parameters: temperature, pressure, flow rate, liquid level, concentration, and pH value, forming a raw multi-parameter time-series dataset. The original multi-parameter time series dataset is time-synchronized. A uniform sampling time interval is set, and the parameter values ​​collected by each sensor at different time points are rearranged in chronological order so that all chemical process parameters have corresponding data values ​​at each sampling time, forming an aligned data sequence arranged by time steps. Each time step corresponds to a vector containing all chemical process parameters, which is used to represent the chemical production operation status at the current time. Sensor calibration is performed on the time-aligned data sequence, measurement points with calibration errors exceeding the allowable range are removed, and the moving average method is used to smooth the boundary abrupt data to obtain the calibration sequence; Missing values ​​were imputed in the calibration sequence, and the data gaps were filled by linear interpolation or the mean of adjacent time points to obtain a continuous sequence of chemical process parameters. Noise suppression processing is performed on the parameter sequence of the continuous chemical process. High-frequency random noise is eliminated by the sliding window averaging method or wavelet threshold denoising method to obtain the filtered sequence. The filtered sequence is standardized by first calculating the mean and standard deviation of each chemical process parameter over the entire sampling period. Then, the original value of each parameter in each time step is subtracted from the mean value of each parameter, and the difference is divided by the corresponding standard deviation to obtain the standardized value of each parameter at the corresponding time step. After the standardization process is completed, the value range of all chemical process parameters is unified to the same scale. The standardized time series data of multiple parameters of chemical process are arranged in chronological order to form a standardized input sequence.

[0025] In this embodiment, the step of constructing a process structure matrix based on the chemical process flow diagram and equipment topology, and performing symmetry and sparsity checks and normalization to obtain a standardized process structure matrix specifically includes: Based on the chemical process flow diagram, identify all equipment units in the chemical production system, including reactors, heat exchangers, separation towers, storage tanks, and pipeline nodes. Clarify the material flow and energy flow directions between each equipment unit, define each equipment unit as a node, and define the connection between nodes that has a material flow or energy flow relationship as a directed connection. Construct an equipment topology diagram according to the relationship between nodes and directed connections to fully describe the structural composition of the chemical production system and the material flow and energy flow paths. Based on the equipment topology diagram, a process structure matrix is ​​constructed according to the connection relationship between the nodes in the equipment topology diagram. The rows and columns of the matrix correspond to each equipment node in the chemical production system. When there is a direct material or energy transfer from one node to another, the corresponding row and column position in the matrix is ​​marked as one. When there is no direct transfer relationship between two nodes, the corresponding row and column position in the matrix is ​​marked as zero, thus forming a process structure matrix that represents the direct connection relationship between equipment in the chemical production system. Symmetry verification of the process structure matrix: If there are both positive and negative material transfer relationships or energy transfer relationships between two equipment nodes in a chemical production system, then the two corresponding positions in the process structure matrix are marked as one, so that the process structure matrix has the same value in the row and column positions corresponding to the nodes, thereby forming a symmetric process structure matrix that can reflect bidirectional transfer relationships. The sparsity of the symmetric matrix is ​​checked by removing rows and columns corresponding to isolated nodes that have no connection to any other node, in order to ensure the connectivity of the matrix. The process structure matrix after sparsity verification is normalized by calculating the sum of all elements in each row of the matrix in turn, and dividing the value of each element in the row by the sum of the elements in the row so that the sum of all elements in the row is equal to one. This yields the normalized result of the data in each row at the same scale, forming a standardized process structure matrix.

[0026] In this embodiment, the step of decoupling the standardized input sequence and the standardized process structure matrix input constraints of the echo state network model, performing structural constraint projection and feature decoupling, and outputting the hidden state sequence specifically includes: A constrained decoupled echo state network model is established, which includes an input mapping layer, a structural constraint projection layer, a reservoir state layer, a state fusion layer, and an output readout layer. The input mapping layer is used to receive the standardized input sequence, the structural constraint projection layer is used to introduce the process structure matrix, the reservoir state layer is used to perform temporal state propagation, the state fusion layer is used for multi-scale gating integration, and the output readout layer is used to generate parameter prediction results. The standardized input sequence and the standardized process structure matrix are input into the structural constraint projection layer. According to the connection relationship of the equipment nodes represented by each element in the process structure matrix, matrix multiplication is performed on the input parameter vector of each time step. The input parameters are weighted and combined with the structural weights of the corresponding nodes to obtain the projected input vector after process structure constraint. The projected input vector and the standardized input vector are concatenated along the feature dimension to form a combined input vector that contains process structure information and original parameter information; A reservoir connection mask matrix is ​​generated based on the standardized process structure matrix. Each element in the process structure matrix is ​​matched with the corresponding connection position in the reservoir. When the value of an element in the process structure matrix is ​​greater than zero, it indicates that there is a connection relationship between the corresponding equipment nodes. Then, the corresponding position in the reservoir connection mask matrix is ​​marked as one. When the value of an element in the process structure matrix is ​​equal to zero, it indicates that there is no connection between the corresponding nodes. In this case, the corresponding position in the reservoir connection mask matrix is ​​marked as zero. Subsequently, a random sparse matrix with the same size as the reservoir connection mask matrix is ​​generated. The element values ​​of the random sparse matrix are decimals randomly distributed within a preset range. The reservoir connection mask matrix is ​​multiplied by the corresponding element of the random sparse matrix to obtain the reservoir connection matrix. This ensures that the connection relationship within the reservoir exists only between nodes allowed by the process structure matrix, thereby forming a reservoir connection topology constrained by the process structure. To adjust the spectral radius of the reservoir connection matrix, firstly, all eigenvalues ​​of the reservoir connection matrix are calculated, and the largest eigenvalue is determined as the current spectral radius. Then, based on the preset echo stability condition, all elements in the reservoir connection matrix are scaled by the same proportion so that the adjusted largest eigenvalue falls within the range between zero and one. This ensures that the reservoir has convergence and dynamic stability during state propagation and avoids numerical divergence in the network during time-series updates. At each time step, a state update is performed based on the combined input vector and the hidden state vector from the previous time step to generate the hidden state vector for the current time step: ; in, This indicates that the constrained decoupled echo state network is at time step The hidden state vector is used to store the network's internal state at the current moment. This represents the hidden state vector from the previous time step, used to record the network's internal state at the previous moment. Indicates at time step The leakage rate parameter, a real number between 0 and 1, is used to control the proportion of inheritance from the previous state to the current state. This represents the hyperbolic tangent activation function, used to perform a nonlinear mapping on the linearly weighted result. This represents the input mapping matrix, used to map the combined input vectors to the reservoir space. Indicates time step The combined input vector, The pool connection matrix is ​​obtained by multiplying the corresponding elements of the pool connection mask matrix and the random sparse matrix, and is used to define the connection topology inside the pool. The hidden state vectors obtained in all time steps are arranged in chronological order to form a complete hidden state sequence.

[0027] In this embodiment, the step of performing multi-scale gated fusion on the hidden state sequence to obtain a fused state representation, and training the readout layer parameters based on the fused state, specifically includes the following: Multi-scale gated fusion is performed on the hidden state sequence. The hidden state sequence is decomposed according to the pre-set time scale parameters, sparsity parameters and spectral radius parameters. The decomposed hidden state sequence is input into multiple parallel sub-state pools. Each sub-state pool independently performs state propagation and state update operations according to its own parameter configuration, so as to dynamically respond to the input signal at different time scales and generate corresponding hidden state sub-sequences. Each sub-state sub-sequence represents the temporal feature change of the sub-state pool at the corresponding time scale. At each time step, the combined input vector of the current time step and the hidden state vector of the previous time step are input together to the gating weight calculation unit. The gating weight value corresponding to each sub-state pool is calculated through linear weighting and nonlinear transformation. Each gating weight value reflects the degree of participation of the sub-state pool in the overall state fusion at the current time step. The gating weight values ​​of all sub-state pools are normalized so that the sum of all gating weights is one, thereby ensuring that the weight ratio of different sub-state pools in the state fusion process remains a uniform normalization constraint. At each time step, the hidden state vectors output by all sub-state pools are weighted and summed according to the corresponding gating weights. The hidden state vectors of each sub-state pool are multiplied by the corresponding gating weights and then summed to obtain the fused state vector of the current time step. The fused state vector integrates the state information of each sub-state pool at different time scales and is used to characterize the overall dynamic characteristics of the chemical process at the current time step. The fused state vectors are arranged in chronological order to form a fused state sequence. In the output readout layer, set the output weight matrix, keep the parameters of the input mapping matrix and the reservoir connection matrix unchanged, and only train the parameters of the output weight matrix to minimize the prediction error or reconstruction error. The fused state sequence is linearly mapped using the trained output weight matrix, and the parameter prediction vector at each time step is calculated to obtain the parameter prediction result.

[0028] In this embodiment, the step of calculating the residual based on the parameter prediction results and the actual observed values, generating anomaly scoring results, determining anomaly judgment boundaries, and outputting anomaly detection results and channel-level attribution information specifically includes: At each time step, the difference between the parameter prediction result and the actual observation value at the corresponding time step is calculated. The difference between the predicted value and the actual value is calculated for each parameter component in the multi-parameter of the chemical process to form a residual vector containing the differences of all monitored parameters. The residual vector is weighted according to the standardized process structure matrix. Each element in the process structure matrix and the corresponding parameter component in the residual vector are weighted and summed according to the matrix multiplication relationship. This ensures that each residual component is affected by the weights of the related equipment nodes and process paths during the calculation, thus obtaining a structure-weighted residual vector. The structure-weighted residual vector reflects the residual propagation relationship between the monitoring parameters under the topological constraints of the chemical process, and provides input for the subsequent calculation of the mean squared error. The components of the structural weighted residual vector are squared sequentially. The weighted residual values ​​of each parameter channel are squared and summed, and then divided by the total number of parameters to calculate the average value, thus obtaining the weighted average squared error of the current time step. The weighted average squared error is used to represent the overall error level of multiple parameters in the chemical process under the constraints of the process structure, and provides input data for subsequent adaptive threshold calculation and abnormal state judgment. The statistical characteristics of error distribution are calculated using historical steady-state operation data samples of chemical processes. The mean and standard deviation of the squared mean of errors are obtained respectively. The mean is used as the benchmark error level and the standard deviation is used as the fluctuation range index. An adaptive threshold is calculated based on these two statistics, such that the threshold is equal to the sum of the mean and standard deviation multiplied by the adjustment coefficient, where the adjustment coefficient is a preset constant used to control the sensitivity of the threshold. The adaptive threshold is used to determine the anomaly of the weighted mean squared error of the current time step in subsequent steps. At each time step, the weighted mean square error of the current time step is compared with the corresponding adaptive threshold. When the weighted mean square error is greater than the adaptive threshold, the current time step is determined to be an abnormal state, and the abnormality identifier is recorded. At the same time, the abnormality detection result corresponding to the current time step is generated to indicate that the chemical process deviates from the normal operating pattern at the current time step. After the current time step is in an abnormal state, the structure-weighted residual vector corresponding to the current time step is read. The numerical amplitudes of each parameter component in the residual vector are compared and sorted. The parameter component with the larger amplitude is regarded as the monitoring channel that contributes more to the abnormal state. Based on the sorting result, the channel-level attribution information is output to indicate the specific monitoring parameter channel or related chemical equipment node that caused the abnormality, providing a basis for subsequent abnormality location and process adjustment.

[0029] In this embodiment, the step of triggering the adaptive correction process of the constraint-decoupled echo state network model when a process structure adjustment or data distribution change is detected, and forming an adaptively corrected monitoring model, specifically includes: During the operation of the chemical process, the distribution characteristics of the input data and the structural information of the process parameters are continuously monitored. When a significant shift in the data distribution compared with the historical steady-state distribution is detected, or when adjustments are detected in the process flow diagram or equipment connection relationship, the model adaptive correction process is triggered. When performing adaptive correction, the state pool topology and standardized process structure matrix of the constrained decoupled echo state network model remain unchanged, and the parameters of the reserve pool connection matrix and input mapping matrix remain fixed. Incremental updates are performed on the output weight matrix of the output readout layer and the gate weight parameters in the multi-scale gated fusion to obtain new output weight matrix and gate weight parameters; The new output weight matrix and gating weight parameters are reloaded into the constrained decoupled echo state network model to form an adaptively corrected monitoring model.

[0030] In this embodiment, the incremental update includes calculating a correction amount based on the difference between the actual observed value and the corresponding prediction result at each time step, adjusting the output weight matrix of the output readout layer according to the set learning rate, so that the new weight matrix is ​​updated in the direction of reducing error on the original basis, and adjusting the gating weight parameters of each sub-state pool according to the difference between the weighted average of the squared error at the current time step and the adaptive threshold, so that the gating weights can adapt to the changes in the current operating state while maintaining the normalization constraint.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous reaction process monitoring system in a large chemical enterprise. This system includes a reactor, heat exchanger, compressor, pump and valve assembly, and multiple temperature, pressure, flow, and concentration sensors. The production process involves a large number of parameters, high sampling frequency, and complex variable correlations. Traditional monitoring methods mainly rely on fixed threshold alarms or experience-based statistical models, which cannot effectively address anomalies caused by process fluctuations and equipment coupling. In actual production, there are significant time delays and dynamic couplings between the temperature, pressure, and flow rates of different process units. Furthermore, when materials are switched or energy loads are adjusted, parameter distributions drift, easily leading to false alarms or missed alarms.

[0032] When applying the deep learning-based intelligent monitoring method for multi-parameter chemical processes of this invention, firstly, multi-parameter time-series data of the reaction unit during continuous operation are collected, with a sampling period of 1 second and a cumulative collection time exceeding 300 hours, involving 24 monitoring parameters. After noise filtering, missing value repair, and normalization of the raw signals through a data preprocessing module, a standardized input sequence is formed. A process structure matrix is ​​constructed based on the process flow diagram and equipment topology information, with a matrix dimension of 24×24, a sparsity of 0.15, and a symmetry deviation of less than 0.02. This matrix serves as a structural constraint input to the constraint-decoupled echo state network model, used to limit the information propagation path, enabling the model to reflect the actual process coupling relationships during training.

[0033] During the model training phase, 200 hours of historical stable operating data were selected for learning. A constrained decoupled echo state network combined with a multi-scale gating fusion mechanism was used for modeling, with a learning rate set to 0.001 and three gating layers. In the validation phase, subsequent 100 hours of production data were used for prediction and anomaly detection comparison. The results showed that the model's average prediction error for key monitoring parameters was only 1.8%, approximately 36% lower than traditional long short-term memory networks. The anomaly recognition rate during periods of process fluctuation reached 96.3%, significantly higher than the 78.5% of traditional statistical methods. During equipment maintenance, a potential anomaly of decreased efficiency in a primary heat exchanger was detected, triggering an alarm 5 hours in advance and preventing production downtime losses, demonstrating the method's high prediction accuracy and real-time response capability.

[0034] Table 1. Comparison of Experimental Results of Intelligent Monitoring of Multiple Parameters in Chemical Processes

[0035] As can be seen from the table above, this invention significantly improves upon traditional methods in several key performance indicators for multi-parameter monitoring of chemical processes. Firstly, in terms of prediction accuracy, the model of this invention controls the average error of reactor temperature, system pressure, flow rate, concentration, motor current, valve opening, and cooling water temperature difference to within 2%. Specifically, the prediction error for temperature is 1.01%, and the prediction error for pressure is 1.27%, lower than the 3% to 5% error level of traditional methods, resulting in an overall improvement in prediction accuracy of over 35%. This indicates that the constrained decoupled echo state network, after integrating process structure constraints and dynamic characteristics, can more accurately characterize the temporal variation patterns of chemical processes.

[0036] Secondly, regarding anomaly identification capabilities, the anomaly identification rate of the method in this invention remained between 94% and 98% across different monitoring parameters, reaching a maximum of 98.1%, which is a significant improvement over the average identification rate of 78.5% for traditional statistical methods. This result indicates that the synergistic mechanism of structure-weighted residuals and adaptive thresholds can effectively suppress false alarms caused by random fluctuations, improving the sensitivity and stability of identifying real anomalies. Furthermore, in terms of anomaly warning timeliness, the average advance warning time of this invention reaches 4.7 hours, which is a leading advantage compared to traditional systems. Specifically, for anomalies such as decreased heat exchanger efficiency, the model issues an alarm 5 hours before the actual failure occurs, allowing sufficient intervention time for on-site operators and demonstrating the model's real-time response capability.

[0037] Comprehensive comparison shows that this invention, by introducing a constrained decoupled echo state network model and a multi-scale gating fusion mechanism, improves the prediction accuracy, anomaly identification rate, and timely early warning of multi-parameter monitoring in chemical processes. The model maintains structural interpretability and adaptability while being deployed in a lightweight manner, enabling it to adapt to process changes and data drift, achieving stable and reliable monitoring results under complex operating conditions. This result verifies the engineering feasibility and technical superiority of this invention in the field of intelligent monitoring of chemical processes.

[0038] The above description is only 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 deep learning-based intelligent monitoring method for multiple parameters in chemical processes, characterized in that, Includes the following steps: Collect multi-parameter time-series data of chemical production processes, preprocess the multi-parameter time-series data, and obtain standardized input sequences; Based on the chemical process flow diagram and equipment topology, a process structure matrix is ​​constructed, and symmetry and sparsity checks and normalization are performed to obtain a standardized process structure matrix. The echo state network model is decoupled from the input constraints of the standardized input sequence and the standardized process structure matrix. Structural constraint projection and feature decoupling are then performed to output the hidden state sequence. Multi-scale gating fusion is performed on the hidden state sequence to obtain the fused state representation, and the readout layer parameters are trained based on the fused state to output the parameter prediction results; The residuals are calculated based on the parameter prediction results and the actual observations, anomaly scoring results are generated, anomaly judgment boundaries are determined, and anomaly detection results and channel-level attribution information are output. When a process structure adjustment or data distribution change is detected, the adaptive correction process of the constraint decoupling echo state network model is triggered, forming an adaptively corrected monitoring model. The adaptively calibrated monitoring model is deployed on edge computing nodes at the chemical production site to receive new standardized input sequences in real time and output parameter prediction results, anomaly detection results, and channel-level attribution information.

2. The intelligent monitoring method for multiple parameters of chemical processes based on deep learning according to claim 1, characterized in that, The preprocessing includes time synchronization, sensor calibration, missing value completion, and noise filtering.

3. The intelligent monitoring method for multiple parameters of chemical processes based on deep learning according to claim 1, characterized in that, The process structure matrix is ​​constructed based on the chemical process flow diagram and equipment topology, and then subjected to symmetry and sparsity checks and normalization to obtain a standardized process structure matrix. Specifically, this includes: Based on the chemical process flow diagram, identify all equipment units in the chemical production system, define each equipment unit as a node, define the connection between nodes that has a material or energy transfer relationship as a directed connection, and construct an equipment topology diagram. Based on the equipment topology diagram, a process structure matrix is ​​constructed according to the connection relationships between the nodes in the equipment topology diagram; The process structure matrix is ​​symmetry checked to form a symmetric process structure matrix. The symmetric matrix is ​​then sparsity checked to remove rows and columns corresponding to isolated nodes that have no connection to any other node. The process structure matrix after sparsity verification is normalized to form a standardized process structure matrix.

4. The intelligent monitoring method for multiple parameters of chemical processes based on deep learning according to claim 1, characterized in that, The process of decoupling the standardized input sequence and the standardized process structure matrix input constraints of the echo state network model, performing structural constraint projection and feature decoupling, and outputting the hidden state sequence specifically includes: A constraint-decoupled echo state network model is established, which includes an input mapping layer, a structural constraint projection layer, a reservoir state layer, a state fusion layer, and an output readout layer. The standardized input sequence and the standardized process structure matrix are input into the structural constraint projection layer. Matrix multiplication is then performed according to the device node connection relationships represented by each element in the process structure matrix to obtain the projection input vector. The projected input vector and the standardized input vector are concatenated along the feature dimension to form a combined input vector; The reservoir connection mask matrix is ​​generated based on the standardized process structure matrix, and the reservoir connection mask matrix is ​​multiplied by the random sparse matrix using the Hadamard product to obtain the reservoir connection matrix. The spectral radius of the reservoir connection matrix is ​​adjusted by calculating all eigenvalues ​​of the reservoir connection matrix, determining the largest eigenvalue as the current spectral radius, and scaling all elements in the reservoir connection matrix according to the preset echo stability condition. At each time step, a state update is performed based on the combined input vector and the hidden state vector of the previous time step to generate the hidden state vector of the current time step. The hidden state vectors obtained in all time steps are arranged in chronological order to form a hidden state sequence.

5. The intelligent monitoring method for multiple parameters of chemical processes based on deep learning according to claim 1, characterized in that, The process of performing multi-scale gated fusion on the hidden state sequence to obtain a fused state representation, and training readout layer parameters based on the fused state, specifically includes the following parameter prediction results: Multi-scale gated fusion is performed on the hidden state sequence, and the hidden state sequence is input into multiple parallel sub-state pools to obtain hidden state sub-sequences at different time scales. At each time step, the combined input vector of the current time step and the hidden state vector of the previous time step are input together into the gating weight calculation unit to calculate the gating weight corresponding to each sub-state pool. At each time step, the hidden state vectors output by all sub-state pools are weighted and summed according to the corresponding gating weights to obtain the fused state vector of the current time step, and arranged in chronological order to form a fused state sequence. In the output readout layer, an output weight matrix is ​​set, the parameters of the output weight matrix are trained, and the trained output weight matrix is ​​used to perform a linear mapping on the fused state sequence to obtain the parameter prediction results.

6. The intelligent monitoring method for multiple parameters of chemical processes based on deep learning according to claim 1, characterized in that, The process of calculating residuals based on parameter prediction results and actual observations, generating anomaly scoring results, determining anomaly judgment boundaries, and outputting anomaly detection results and channel-level attribution information specifically includes: At each time step, the difference between the parameter prediction result and the actual observation value at the corresponding time step is calculated to form a residual vector; The residual vector is weighted according to the standardized process structure matrix. Each element in the process structure matrix and the corresponding parameter component in the residual vector are weighted and summed according to the matrix multiplication relationship to obtain the structure-weighted residual vector. The weighted residual vector is squared sequentially to obtain the weighted mean square error. The statistical characteristics of the error distribution are calculated using historical steady-state operating data samples of chemical processes. The mean and standard deviation of the squared mean of the error are obtained, and the adaptive threshold is calculated. At each time step, the weighted mean square error of the current time step is compared with the corresponding adaptive threshold. When the weighted mean square error is greater than the adaptive threshold, the running state is determined to be abnormal and the corresponding abnormal detection result is generated. When the running state is abnormal, the structure-weighted residual vector corresponding to the time step is read, the numerical magnitudes of each parameter component in the residual vector are compared and sorted, and channel-level attribution information is output.

7. The intelligent monitoring method for multiple parameters of chemical processes based on deep learning according to claim 1, characterized in that, When a process structure adjustment or data distribution change is detected, the adaptive correction process of the constraint-decoupled echo state network model is triggered to form an adaptively corrected monitoring model, specifically including: When a process structure adjustment or data distribution change is detected, the adaptive correction process of the constraint decoupling echo state network model is triggered. When performing adaptive correction, the state pool topology and standardized process structure matrix of the constrained decoupled echo state network model are kept unchanged, and the parameters of the reserve pool connection matrix and input mapping matrix are kept fixed. Incremental updates are performed on the output weight matrix of the output readout layer and the gate weight parameters in the multi-scale gated fusion to obtain new output weight matrix and gate weight parameters; The new output weight matrix and gating weight parameters are reloaded into the constrained decoupled echo state network model to form an adaptively corrected monitoring model.

8. The intelligent monitoring method for multiple parameters of chemical processes based on deep learning according to claim 7, characterized in that, The incremental update includes calculating a correction amount based on the difference between the actual observation and the corresponding prediction result at each time step, adjusting the output weight matrix of the output readout layer according to the set learning rate, and correcting the gating weight parameters of each sub-state pool according to a preset adjustment coefficient based on the difference between the weighted average error of the current time step and the adaptive threshold.

Images