Real-time reaction monitoring device based on intelligent sensing and feedback optimization method

Abstract

The invention relates to a real-time reaction monitoring device based on intelligent sensing and a feedback optimization method, and aims to solve the problems that multi-source heterogeneous data is difficult to fuse, the dynamic optimization efficiency is low and abnormal working conditions are difficult to diagnose in the industrial reaction process. According to the device, an intelligent sensor array is constructed by integrating various physical and chemical sensors, integrated packaging of the sensors is realized by adopting a micro-fluidic chip technology, and preprocessing and digitization of data are supported. In combination with deep learning and reinforcement learning algorithms, the method realizes feature extraction, fusion analysis and dynamic optimization control of multi-sensor data, and has the functions of abnormal working condition diagnosis, real-time feedback and closed-loop optimization. According to the system architecture, edge computing and cloud service collaboration are adopted, distributed optimization control, multi-device collaboration and industrial interconnection standard integration are supported, and efficient and reliable man-machine interaction and data visualization functions are provided. By constructing a knowledge management mechanism, optimization rules and models accumulated in the production process are precipitated, and continuous improvement and rapid adaptation are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent sensing and industrial process monitoring and control, and particularly to a real-time reaction monitoring device and a feedback optimization method based on intelligent sensing. Background Art

[0002] In the industrial production process, the monitoring and optimization of the reaction process are key factors affecting product quality, production efficiency, and resource utilization rate. However, the existing reaction monitoring and optimization technologies still face the following challenges:

[0003] The industrial reaction process usually involves the real-time monitoring of multiple physical, chemical, and environmental parameters, such as temperature, pressure, pH value, spectral data, etc. These data come from diverse sources, with inconsistent formats and dimensions, posing great challenges to subsequent fusion analysis and comprehensive evaluation. In addition, due to the heterogeneity of sensors, there are differences in the accuracy and reliability of the data, further exacerbating the complexity of data processing.

[0004] Traditional reaction process control mostly relies on empirical rules or static optimization models and is difficult to cope with complex working conditions and dynamic environmental changes. These methods usually require manual intervention and cannot achieve efficient real-time adjustment, resulting in problems such as resource waste, large fluctuations in product quality, and lag in response speed during the production process.

[0005] In actual production, the reaction process may be affected by factors such as equipment failure, raw material fluctuations, or environmental changes, resulting in abnormal working conditions. However, most of the existing anomaly detection methods rely on preset thresholds or simple statistical analysis and lack the ability to accurately identify and diagnose complex dynamic anomalies. This may not only delay the problem handling time but also lead to production interruption or major safety hazards.

[0006] Intelligent and automated technologies are gradually becoming the core direction of industrial development. However, most of the current reaction monitoring and optimization systems still rely on manual intervention in the data analysis and decision-making process and lack intelligent feature extraction, model adaptive optimization, and feedback closed-loop control capabilities, making it difficult to meet the requirements of modern industry for intelligence and automation.

[0007] The optimization process in industrial production often relies on the experience of senior engineers. However, these experiences are difficult to manage systematically and structurally, resulting in the inability to quickly reuse existing knowledge when personnel change or new working conditions appear, affecting production efficiency and the ability to continuously optimize.

[0008] Based on the above technical background, there is an urgent need to develop an intelligent system that can integrate multi-source heterogeneous data monitoring, dynamic optimization, anomaly diagnosis, and knowledge management to improve the monitoring and optimization level of industrial reaction processes and achieve efficient, stable, and sustainable production goals. The present invention is proposed under this technical background, aiming to construct an efficient, reliable, and intelligent real-time reaction monitoring device and feedback optimization method through the combination of intelligent sensors, deep learning, and reinforcement learning technologies to solve the bottleneck problems in the prior art. Summary of the Invention

[0009] The present invention proposes a real-time reaction monitoring device and feedback optimization method based on intelligent sensing, aiming to solve the problems in the prior art such as difficult multi-source heterogeneous data fusion, low dynamic optimization efficiency, and difficult diagnosis of abnormal working conditions. The present invention integrates multiple sensors to construct an intelligent sensor array, combines deep learning and reinforcement learning algorithms, and realizes real-time monitoring, data analysis, and intelligent optimization of industrial reaction processes.

[0010] A real-time reaction monitoring device based on intelligent sensing, comprising:

[0011] Multi-source heterogeneous sensor array: including UV-Vis spectroscopy sensor, NIR spectroscopy sensor, Raman spectroscopy sensor for collecting spectral data; pH sensor, conductivity sensor, and ion-selective electrode for chemical parameter monitoring; thermocouple and infrared temperature sensor for thermal parameter collection; and strain gauge and piezoelectric sensor for mechanical parameter measurement;

[0012] Microfluidic chip: integrating the sensor array, with built-in signal processing circuit and analog-to-digital conversion circuit for preprocessing and digitizing sensor signals;

[0013] Modular interface system: adopting standardized electrical interface (FFC flexible cable connector) and liquid path interface (Luer connector) to achieve modular connection and plug-and-play function;

[0014] Intelligent control unit: adopting a high-performance STM32H7 series microcontroller, integrating ARM Cortex-M7 core, Bluetooth 5.0 module, and Wi-Fi module for data processing, wireless communication, and self-diagnosis functions;

[0015] Edge computing module: for real-time processing of multi-source sensor data, including noise reduction, normalization, feature extraction, and anomaly detection.

[0016] Preferably, the sensor array realizes data normalization processing through the following formula:

[0017]

[0018] where x′ iis the normalized data, x min and x max are the minimum and maximum values of the measured values respectively.

[0019] Preferably, the device uses an attention mechanism to perform weighted fusion on sensor features, and the calculation formula for the fused feature F is:

[0020]

[0021] where, α i is the weight, and the calculation formula is:

[0022]

[0023] f i is the sensor feature value, and e i is the attention score.

[0024] Furthermore, the device uses a Temporal Convolutional Network (TCN) to model the dynamic process, and its convolution formula is:

[0025]

[0026] where, y t is the output value at time t, x t-k is the input value at time t - k, and w k is the convolution kernel weight.

[0027] Furthermore, the feedback optimization method of the device includes the following steps:

[0028] Data normalization: Use the formula to normalize the sensor data;

[0029] Feature extraction and weighted fusion: Extract sensor features through a Convolutional Autoencoder (CAE), and combine the attention mechanism to assign weights to achieve feature fusion;

[0030] Dynamic process modeling: Use a Temporal Convolutional Network (TCN) to perform dynamic modeling on the reaction process;

[0031] Anomaly detection: Generate anomaly samples through a Generative Adversarial Network (GAN) and detect anomaly data.

[0032] Furthermore, optimize the reaction control strategy through the following reward function:

[0033] R = w1·Y + w2·S - w3·C

[0034] where: R is the reward value; Y is the yield; S is the selectivity; C is the cost; w1, w2, w3 are the weighting coefficients.

[0035] Furthermore, the deep reinforcement learning algorithm (DDPG) is used for control optimization, and an active learning mechanism is combined to improve the optimization efficiency.

[0036] Furthermore, in combination with the transfer learning strategy, by fixing the feature extraction network parameters and fine-tuning the control network, fast adaptation to different working conditions is achieved.

[0037] Furthermore, a system architecture that coordinates edge computing and cloud services is adopted. Data processing and anomaly detection are performed at the edge, and global optimization and policy update are performed in the cloud.

[0038] Furthermore, the device follows the industrial interconnection standard OPC-UA communication protocol and uses EtherCAT and PROFINET industrial Ethernet technologies to achieve data interconnection.

[0039] Construction of the intelligent sensor array

[0040] The present invention uses UV-Vis, NIR, Raman spectroscopic sensors, as well as various sensors such as pH and conductivity, and realizes integrated packaging through microfluidic chip technology, supporting data preprocessing and digitization. At the same time, the modular interface design enables the sensor array to have a plug-and-play function, supporting wireless communication and self-diagnosis capabilities.

[0041] For multi-source heterogeneous sensor data, the present invention designs a feature extraction method based on a convolutional autoencoder, combines an attention mechanism to achieve dynamic weight allocation, thereby highlighting the contribution of key features. In addition, the dynamics of the reaction process are modeled through a temporal convolutional network (TCN), and an adversarial generative network (GAN) and an isolation forest algorithm are introduced to realize automatic detection and diagnosis of abnormal working conditions.

[0042] The present invention uses the deep reinforcement learning (DDPG) algorithm, combines an active learning mechanism to dynamically adjust control variables, and realizes online closed-loop optimization of the reaction process. Through transfer learning and a distributed reinforcement learning framework, it quickly adapts to complex and changing production requirements and supports optimized control for multi-device collaboration.

[0043] The present invention adopts a system architecture that coordinates edge computing and cloud services, and realizes efficient communication between multiple devices in combination with the industrial interconnection standard (OPC-UA). At the same time, a friendly human-computer interaction interface 6 and a data visualization engine are developed to provide flexible optimization configuration and real-time monitoring functions for process experts.

[0044] The present invention constructs an enterprise-level knowledge management mechanism to precipitate the rules, models, and strategies formed during the optimization process. Through incremental learning and semantic retrieval, it supports fast adaptation and continuous improvement under new working conditions.

[0045] Advantages of the present invention:

[0046] A multi-source heterogeneous intelligent sensor array integrated on a microfluidic chip realizes all-round, real-time, and high-precision monitoring of the reaction process, laying a foundation for data-driven optimization; a deep learning model combining an autoencoder and an attention mechanism can efficiently fuse multi-source sensor data and extract implicit features crucial for process optimization; a soft sensor model for the reaction process with anomaly detection and diagnosis functions is constructed by combining a temporal convolutional network with isolation forest and GAN; deep reinforcement learning is applied to the optimization control of the reaction process, and mechanisms such as active exploration and transfer learning are introduced to achieve policy self-optimization in a continuous action space; a system architecture that coordinates edge computing and cloud services is adopted to achieve full integration and utilization of data and knowledge while meeting real-time requirements; a process optimization knowledge management mechanism based on ontology is constructed to realize semantic representation, storage, retrieval, and continuous iterative update of optimization knowledge.

[0047] The present invention innovatively applies multi-source intelligent sensing, deep learning, reinforcement learning, and knowledge engineering to reaction process optimization, providing a complete technical framework and solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only those of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0049] Figure 1 It is a schematic diagram of the device hardware structure of an embodiment of the present invention;

[0050] Figure 2 It is a schematic diagram of the module structure of an embodiment of the present invention;

[0051] Figure 3 It is a schematic diagram of the circuit structure of an embodiment of the present invention;

[0052] Figure 4 It is a schematic diagram of the data processing flow of an embodiment of the present invention;

[0053] Figure 5 It is a schematic diagram of the control optimization flow of an embodiment of the present invention.

[0054] In the figure, 1 is an intelligent sensor array; 2 is an edge computing unit; 3 is a communication module; 4 is an actuator; 5 is a cloud service platform; 6 is a human-computer interaction interface. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0055] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. At the same time, it should be noted here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments. For some well-known technologies, those skilled in the art can also adopt other alternative methods for implementation; moreover, the accompanying drawings are only for more specifically describing the embodiments and are not intended to specifically limit the present invention.

[0056] It should be noted that in the specification, the mention of "one embodiment", "embodiment", "exemplary embodiment", "some embodiments", etc. indicates that the described embodiment may include specific features, structures or characteristics, but not necessarily every embodiment includes such specific features, structures or characteristics. Additionally, when combining an embodiment to describe a specific feature, structure or characteristic, implementing such feature, structure or characteristic in combination with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the relevant art.

[0057] Generally, terms can be understood at least in part from their use in the context. For example, at least in part depending on the context, the term "one or more" used herein can be used to describe any feature, structure or characteristic in a singular sense, or can be used to describe a combination of features, structures or characteristics in a plural sense. Additionally, the term "based on" can be understood as not necessarily intended to convey a set of exclusive factors, but rather, at least in part depending on the context, can allow for the existence of other factors that may not be explicitly described.

[0058] Refer to Figures 1 to 5

[0059] I. Construction of the intelligent sensor array 1

[0060] 1.1 Sensor selection and integration

[0061] The present invention adopts multi-source heterogeneous sensors, including UV-Vis, NIR, Raman spectroscopy sensors, pH, conductivity, ion-selective electrode electrochemical sensors, thermocouples, thermal resistors, infrared temperature measurement thermal sensors, and strain gauges, piezoelectric sensors for mechanical sensing. Through MEMS (Micro-Electro-Mechanical Systems) technology, the above-mentioned miniaturized sensors are integrated on a microfluidic chip to achieve the integrated packaging of the sensor array. A miniaturized signal processing circuit and an analog-to-digital conversion (ADC) circuit are also integrated in the microfluidic chip for preprocessing and digitizing the sensor signals.

[0062] 1.2 Modular interface design

[0063] To achieve the modularization and plug-and-play of the sensor array, standardized electrical interfaces and liquid path interfaces are designed on the periphery of the microfluidic chip. The electrical interface uses an FFC flexible cable connector, and sensor signals, power supply, and control signals are led out through a unified interface, facilitating connection to the main controller. The liquid path interface uses standard Luer connectors, which can be conveniently connected to the microfluidic pipeline to realize the import and export of reaction liquids. The modular interface design greatly improves the versatility and scalability of the sensor array.

[0064] 1.3 Implementation of intelligent design

[0065] The sensor array needs to have certain intelligent functions, including data processing, wireless communication module 3, and self-diagnosis. Therefore, a high-performance STM32H7 series microcontroller is selected as the main control chip of the sensor array. STM32H7 uses an ARM Cortex-M7 core, with a main frequency of up to 480 MHz, and is built-in with 2 MB of Flash and 1 MB of SRAM, with excellent computing and storage performance. Bluetooth 5.0 and Wi-Fi wireless communication modules are integrated on the STM32H7 for data transmission and instruction interaction with the host computer. For each type of sensor, corresponding self-diagnosis and self-calibration algorithms are developed and burned into the on-chip Flash of the STM32H7. Through intelligent design, the sensor array can achieve autonomous operation and self-maintenance, greatly improving the reliability and usability of the system.

[0066] II. Data fusion analysis method

[0067] 2.1 Normalization of heterogeneous data

[0068] Since the original data formats and value ranges output by various sensors vary greatly, normalization processing is required to convert them into a unified representation form. For spectral sensor data, wavelet transform is used for noise reduction and feature extraction, and the spectral intensity is normalized to the [0, 1] interval. For temperature sensor data, the Kalman filter algorithm is used for smoothing processing, and the temperature value is unified into degrees Celsius. For pH and conductivity electrochemical sensor data, they are directly normalized to the [0, 1] interval. The normalized multi-source sensor data can be conveniently used for subsequent fusion analysis.

[0069] The normalization method formula is as follows:

[0070]

[0071] Among them, x′ i is the normalized data, x min and x max are the minimum and maximum values of the measured values respectively, and the normalized x′ iUnify it to the interval [0, 1] to provide a standardized input for subsequent analysis.

[0072] 2.2 Feature Extraction Based on Deep Learning

[0073] The normalized sensor time-series data has the characteristics of high dimensionality and strong time-variability. To extract more abstract and robust feature representations from it, the present invention adopts a deep learning method based on a convolutional autoencoder (CAE). The CAE consists of two parts: an encoder and a decoder. The encoder uses a one-dimensional convolutional neural network (Conv1D) to hierarchically abstract the time-series data and map it to a low-dimensional feature space; the decoder uses a one-dimensional transposed convolutional neural network (ConvTranspose1D) to reconstruct the feature vector into the original data. By training the CAE in an unsupervised manner to minimize the reconstruction error, a low-dimensional feature vector that can fully represent the internal structure of the original data can be obtained. To further improve the robustness of the features, a regularization term, the L1 / L2 norm, is introduced into the loss function of the CAE to weaken the correlation between features.

[0074] 2.3 Feature Fusion with Attention Mechanism

[0075] For the multi-sensor fusion scenario, the importance of different sensor features is dynamically changing, and weights need to be adaptively assigned according to the task and environment. For this purpose, the present invention introduces an attention mechanism to automatically learn feature weights. Specifically, the features of each sensor extracted by the CAE are first fed into the attention module, using a Transformer or Self-Attention structure. Through the learning of the weight matrix, the significance weights of each feature are obtained. The weighted fusion feature vector not only contains the complementary information of multiple sensors but also highlights the contribution of key features. The entire attention fusion process can be jointly trained with the CAE to achieve end-to-end optimization of feature extraction and fusion.

[0076] Use the attention mechanism to perform weighted fusion on the features extracted by various sensors. The calculation formula is as follows:

[0077]

[0078] where α i is the weight of each sensor feature, and f i is the feature value. The weight calculation formula is:

[0079]

[0080] e i is the attention score of the feature, indicating the relative importance of the feature to the task.

[0081] 2.4 Process Modeling of Temporal Convolutional Network

[0082] To model the dynamics of the reaction process, the causal dependencies of time series data need to be considered. The present invention uses a temporal convolutional network to characterize the evolution law of features in the time dimension. TCN uses causal convolutions, that is, the output at the current moment is only related to the input at historical moments and has nothing to do with future moments, which conforms to time causality. On the basis of the fused features, multiple causal convolutional layers and residual connection blocks are stacked, and finally mapped to the output variables, reactant concentration and yield, through a fully connected layer. TCN has the advantages of parameter sharing and parallel computing and can efficiently process long sequence data. To address the problem of limited training samples, during the training of the TCN model, a data augmentation strategy is adopted. By adding Gaussian noise and random perturbations to the original data, a large number of virtual samples are generated to improve the generalization ability of the model.

[0083] The causal convolution formula of the temporal convolutional network is as follows:

[0084]

[0085] where y t is the output value at time t, x t-k is the input value at time t - k, and w k is the convolutional kernel weight.

[0086] 2.5 Detection and Diagnosis of Abnormal Conditions

[0087] During the reaction process, abnormal situations such as equipment failures and operation errors may occur and need to be detected and diagnosed in a timely manner. On the basis of the TCN process model, the present invention introduces an adversarial generative network for the generation and recognition of abnormal data. Specifically, a generator network is trained to receive the process features under normal conditions as input and output similar virtual features; at the same time, a discriminator network is trained to distinguish the output of the generator from real abnormal data. Through the adversarial training of the generator and the discriminator, virtual samples with a distribution similar to that of real abnormal data can be obtained. Adding the virtual abnormal samples generated by GAN to the training set of TCN can enhance the discriminative ability of TCN for various abnormal conditions and achieve abnormal detection and diagnosis. The present invention also introduces an unsupervised anomaly detection algorithm based on the isolation forest. By means of statistical modeling, outliers that do not conform to historical laws are automatically discovered as a supplementary means for TCN diagnosis;

[0088] The abnormal score calculation formula is as follows:

[0089]

[0090] where D(x) is the classification probability of the discriminator for the input data x. The closer the abnormal score S is to 1, the more likely the data is to be abnormal.

[0091] III. Intelligent Optimization Based on Reinforcement Learning

[0092] 3.1 Deep Reinforcement Learning Controller Design

[0093] To achieve the online closed-loop optimization of the reaction process, it is necessary to dynamically adjust the control variables, namely the feed rate and heating power, according to the real-time collected process states. In this invention, the DDPG (Deep Deterministic Policy Gradient) algorithm in deep reinforcement learning is adopted to train the control strategy for reaction process optimization. The process state features extracted by TCN are used as the state variables of reinforcement learning, the reaction condition parameters are used as the action variables, and the reward function is designed based on the objective functions (yield, selectivity). DDPG adopts the Actor-Critic architecture. The Actor network is responsible for generating continuous control actions, and the Critic network is responsible for evaluating the state-action value function. During the training process, the Experience Replay mechanism is adopted. By reusing historical data, the impact of data correlation on training stability is alleviated. In addition, the Hindsight Experience Replay (HER) technology is introduced, which can learn from failed experiences, accelerate the learning process, and obtain a better control strategy. By deploying the trained DDPG controller, the intelligent optimization of the reaction process is realized.

[0094] During the optimization process, training is carried out using the reward function, and the objective function is:

[0095] R = w1·Y + w2·S - w3·C

[0096] Where: R is the reward value; Y is the yield; S is the selectivity; C is the cost; w1, w2, and w3 are weighting coefficients, and the specific values are determined according to actual requirements.

[0097] 3.2 Active Learning Exploration Mechanism

[0098] During the sampling and learning process of DDPG, the intelligent agent often faces the dilemma of "exploration - exploitation". Excessive exploration will lead to low optimization efficiency, while overexploitation is likely to fall into local optima. To balance the two, this invention proposes an active learning exploration mechanism. Combining the uncertainty predicted by the TCN process model and the Q-value evaluation of DDPG, an "exploration value" index is defined. Periodically, the "difficult" states with high exploration values are selected, and active sampling is carried out in their neighborhoods to obtain new optimization directions. The samples obtained by active sampling are added to the experience pool for DDPG training, which can specifically improve the model performance. At the same time, during the exploration process, according to the impact of new samples on the model uncertainty, the exploration intensity is adaptively adjusted, reducing ineffective exploration while ensuring the continuous progress of optimization. Active learning makes the exploration of DDPG more efficient and intelligent.

[0099] 3.3 Fast Adaptation Based on Transfer Learning

[0100] In industrial production, the operating conditions and optimization objectives of the reaction process may change, such as fluctuations in raw material composition and adjustment of product indicators. To quickly adapt to new optimization tasks, a transfer learning mechanism is introduced to make full use of existing model knowledge and shorten the learning time. When a new task arrives, the network parameters of the TCN feature extraction part and the attention fusion part are fixed, and only the DDPG decision network is fine-tuned to adapt to the new objective function and reward setting. At the same time, a progressive learning strategy is adopted, first fine-tuning under the original operating conditions and then gradually transitioning to the new operating conditions to avoid performance deterioration caused by sudden changes in the strategy. The introduction of transfer learning enables the intelligent optimization system to quickly adapt to complex and changing production requirements at the lowest cost.

[0101] 3.4 Parallel Optimization of Distributed Architecture

[0102] In the scenarios of large-scale installations and collaborative production of multiple devices, synchronous optimization control of multiple reaction units is required. To meet this demand, the present invention adopts a distributed reinforcement learning framework to achieve parallel optimization. On each control unit, a DDPG agent is equipped to be responsible for local data collection, feature extraction, and policy execution tasks. The agents communicate through the industrial Ethernet and regularly upload local state transition samples to the central node. The central node aggregates multi-source samples, jointly trains the Critic network, updates the global value function, and distributes the updated parameters to each agent. Through this "federated learning" method, shared learning and consistent optimization of models and policies are achieved. While protecting the privacy and independence of each device, the overall value of the data can be fully exploited, improving the efficiency and effectiveness of optimization. In the implementation process, DTCN and IPPO distributed reinforcement learning algorithms are adopted to efficiently handle communication delays and device heterogeneity problems. Through distributed optimization, the present invention can provide strong support for the intelligent control of large-scale complex systems.

[0103] IV. System Integration and Application Implementation

[0104] 4.1 Collaboration between Edge Computing and Cloud Service Platform 5

[0105] The system architecture of the present invention combines edge computing with a cloud service platform 5 to achieve reasonable allocation of data, computing power, and storage. On the side of field devices, an intelligent sensor array 1 and an edge computing unit 2 are equipped. The sensor array is responsible for real-time collection of various process parameters, and performs data preprocessing, feature extraction, and anomaly detection computing tasks through the edge computing unit 2. The key information of the state features and decision instructions formed locally is securely uploaded to the cloud through the industrial Ethernet. A high-performance computing cluster is built in the cloud, which is responsible for global tasks such as DDPG training, multi-device collaboration, and knowledge management. According to needs, the updated optimization model and policy parameters are sent down to the edge unit to guide process control. At the same time, the cloud provides data backup, remote monitoring, and software upgrade services to ensure the stable operation of the system. The edge-cloud collaboration gives full play to the advantages of both ends, meeting the real-time requirements and realizing in-depth exploration of data value.

[0106] 4.2 Integration of Industrial Internet Standards

[0107] To facilitate the interconnection and interoperability of all links in the industrial chain, the present invention strictly follows industrial Internet standards for system integration. In terms of communication protocols, OPC-UA is adopted as the unified data exchange standard. The sensor array, edge computing unit 2, and actuator 4 communicate bidirectionally with MES, DCS, and LIMS production management systems through OPC-UA. Optimization instructions can be sent down to the field control system, and production data can be uploaded to the management level, achieving seamless end-to-end connection. In terms of network interconnection, EtherCAT and PROFINET industrial Ethernet technologies are adopted to build a high-speed, reliable, and real-time data bus to meet the networking requirements of large-scale devices. The programming specifications of IEC-61131 and IEC-61499 are followed to ensure the portability and reusability of control logic. Through the integrated application of industrial Internet standards, the intelligent optimization system is fully integrated with the enterprise's original information system, improving the implementation efficiency.

[0108] 4.3 Human-Machine Interaction and Visualization

[0109] To facilitate process experts to participate in the optimization process, the present invention develops a friendly human-machine interaction interface 6. Through a Web browser, experts can view the key parameters, optimization indicators, and anomaly alarms of the reaction process in real time. For operators in different positions, personalized information display schemes are designed, such as KPI dashboards for workshop directors, equipment status monitoring for operators, and process optimization curves for process engineers.

[0110] In addition, the present invention also provides a series of interactive optimization tools. Experts can flexibly configure optimization objectives and constraints through graphical drag-and-drop operations, such as "increasing the yield by 5% while controlling the cost not to exceed 10,000 yuan / ton". The system automatically generates optimization tasks according to the settings of experts and provides real-time feedback on the optimization process and results. Experts can also conduct retrospective analysis on historical optimization cases through functions such as simulation and playback to extract improvement experiences.

[0111] The present invention also incorporates a data visualization engine, which supports multi-dimensional visual analysis of sensor data, process parameters, and quality indicators. For example, the correlation between parameters is shown through parallel coordinate plots, the product quality of different batches is compared through radar charts, and the optimal process parameter combinations are explored through 3D scatter plots, etc. Visualization greatly improves the efficiency and insight of data analysis.

[0112] 4.4 Knowledge Management and Continuous Optimization

[0113] Industrial big data contains rich process optimization knowledge, but it is easily lost over time. To make full use of this knowledge, the present invention constructs a knowledge management mechanism to precipitate knowledge assets such as rules, models, and strategies obtained during the optimization process, forming an enterprise-specific knowledge base.

[0114] Specifically, we use the ontology method to structurally represent and semantically organize optimization knowledge. For example, "lowering the reaction temperature can increase selectivity" is represented as an ontology statement and associated with relevant process parameters and reaction mechanisms. The knowledge base supports intelligent services such as semantic retrieval and associative reasoning to provide reference for optimization decisions.

[0115] At the same time, we also design a self-update mechanism for knowledge. The system continuously collects data during the optimization execution process and iteratively optimizes the original model through incremental learning to continuously accumulate and improve knowledge. When new working conditions or optimization objectives appear, the system can quickly retrieve relevant examples from the knowledge base to reduce the cost of repeated trial and error.

[0116] By constructing an enterprise-level optimization knowledge management and continuous optimization mechanism, the present invention not only serves current process optimization, but also lays a data and knowledge foundation for long-term process improvement and innovation.

[0117] The present invention covers any substitutions, modifications, equivalent methods, and solutions made within the essence and scope of the present invention. For the public to have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention, and those skilled in the art can fully understand the present invention without these detailed descriptions. In addition, well-known methods, processes, procedures, components, and circuits are not described in detail to avoid unnecessary confusion to the essence of the present invention.

[0118] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A real-time reaction monitoring device based on intelligent sensing, characterized in that, Including: Multi-source heterogeneous sensor array: Including UV-Vis spectroscopy sensor, NIR spectroscopy sensor, Raman spectroscopy sensor for collecting spectral data; pH sensor, conductivity sensor and ion-selective electrode for chemical parameter monitoring; thermocouple and infrared temperature sensor for thermal parameter collection; and strain gauge and piezoelectric sensor for mechanical parameter measurement; Microfluidic chip: integrating the sensor array, with built-in signal processing circuit and analog-to-digital conversion circuit for preprocessing and digitizing sensor signals; Modular interface system: using standardized electrical interface (FFC flexible cable connector) and liquid path interface (Luer connector) to achieve modular connection and plug-and-play function; Intelligent control unit: adopting high-performance STM32H7 series microcontroller, integrating ARM Cortex-M7 core, Bluetooth 5.0 module and Wi-Fi module for data processing, wireless communication and self-diagnosis function; Edge computing module: for real-time processing of multi-source sensor data, including noise reduction, normalization, feature extraction and anomaly detection.

2. The real-time reaction monitoring device based on intelligent sensing according to claim 1, wherein The sensor array realizes data normalization processing through the following formula: where x′ i is the normalized data, and x min and x max are the minimum and maximum values of the measured values, respectively.

3. The real-time reaction monitoring device based on intelligent sensing according to claim 1, characterized in that, The device uses attention mechanism to weight-fuse sensor features, and the calculation formula of the fused feature F is: where α i is the weight, and the calculation formula is: f i is the sensor eigenvalue, and e i is the attention score.

4. The real-time reaction monitoring device based on intelligent sensing according to claim 1, characterized in that, The device uses Temporal Convolutional Network (TCN) to model the dynamic process, and its convolution formula is: Among them, y t is the output value at time t, x t-k is the input value at time t - k, and w k is the convolution kernel weight.

5. A feedback optimization method based on the device described in claim 1, characterized in that, Including the following steps: Data normalization: Using the formula to normalize the sensor data; Feature extraction and weight fusion: extracting sensor features through Convolutional Autoencoder (CAE), and combining attention mechanism to assign weights to achieve feature fusion; Dynamic process modeling: using Temporal Convolutional Network (TCN) to dynamically model the reaction process; Anomaly detection: generating anomaly samples through Generative Adversarial Network (GAN) and detecting anomaly data.

6. The feedback optimization method according to claim 5, wherein Optimizing the reaction control strategy through the following reward function: R = w1·Y + w2·S - w3·C Where: R is the reward value; Y is the yield; S is the selectivity; C is the cost; w1, w2, w3 are weighting coefficients.

7. The feedback optimization method according to claim 5, wherein Adopting Deep Deterministic Policy Gradient (DDPG) algorithm for control optimization, and combining active learning mechanism to improve the optimization efficiency.

8. The feedback optimization method according to claim 5, wherein Combining transfer learning strategy, by fixing the parameters of the feature extraction network and fine-tuning the control network, to achieve rapid adaptation to different working conditions.

9. The device and method according to claims 1 and 5, characterized in that, Adopting a system architecture that coordinates edge computing and cloud service platform 5, performing data processing and anomaly detection at the edge, and performing global optimization and policy update in the cloud.

10. The real-time reaction monitoring device based on intelligent sensing according to claim 1, characterized in that, The device follows the industrial interconnection standard OPC-UA communication protocol, and uses EtherCAT and PROFINET industrial Ethernet technologies to achieve data interconnection.

Images