Method for trace chemical detection of color change in bromination reactor in production of fluorobromoethane

Abstract

The application belongs to the technical field of chemical analysis and trace detection, and specifically discloses a trace chemical detection method for color change of a bromination reactor in fluorobromethane production. The method comprises the following steps: arranging a corrosion-resistant modified fluorescent molecular probe array to obtain specific fluorescence signals; combining a convolutional neural network and a random forest ensemble model to realize intelligent mapping of the fluorescence signals to concentrations; synchronously introducing a bromide ion selective electrode, fusing electrochemical and fluorescence signals through Kalman filtering, and cross-verification; and based on the concentration estimation value, constructing a feedback control and mutation early warning mechanism, dynamically adjusting the feed valve and triggering safety intervention. Through the above technical scheme, the application realizes online detection and closed-loop regulation of trace substances with high sensitivity, wide dynamic range and strong anti-interference, effectively ensuring reaction stability and process safety.

Description

Technical Field

[0001] This invention belongs to the field of chemical analysis and trace detection technology, specifically relating to a trace chemical detection method for color changes in the bromination reactor during the production of fluorobromoethane. Background Technology

[0002] With the increasing demands for process control in fine chemicals, real-time online detection of trace impurities has become a crucial step in ensuring the stability of high-purity halogenated hydrocarbon synthesis processes. As an important fluorine-containing intermediate, the concentration fluctuations of byproducts and unreacted raw materials during the bromination reaction of fluorobromoethane directly affect product purity and equipment corrosion rates. Especially in continuous production, color changes within the reactor often range from ppb levels (10⁻⁶ ppm). -9 Up to 10 -5 The presence of bromine or fluorinated vinyl residues (g / mL) as the triggering agent places extremely high demands on the sensitivity, selectivity, and anti-interference capabilities of the detection method.

[0003] However, traditional UV-Vis spectroscopy or gas chromatography are limited by detection limits and response speeds, making it difficult to dynamically track trace components and resulting in lags in process control. Furthermore, single sensor signals are susceptible to interference from temperature, solvent polarity, and coexisting ions, and lack multi-dimensional verification mechanisms, failing to ensure the reliability of identification results at low concentrations. In addition, existing analytical systems mostly employ static calibration models, which are ill-suited to the nonlinear and time-varying chemical environments in reaction systems, leading to insufficient concentration inversion accuracy and an inability to support closed-loop control requirements.

[0004] Therefore, there is an urgent need for an interdisciplinary trace detection method that integrates chemical sensing, intelligent algorithms and electrochemical verification to achieve high-precision and robust online monitoring of color changes in the bromination reactor during the production of fluorobromoethane. Summary of the Invention

[0005] The purpose of this invention is to provide a trace chemical detection method for color changes in the bromination reactor during the production of fluorobromoethane, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a trace chemical detection method for color changes in the bromination reactor during the production of fluorobromoethane, comprising the following steps:

[0007] A chemical sensing amplification module was constructed, and an array of modified fluorescent molecular probes encapsulated in corrosion-resistant materials was deployed inside the reactor. The fluorescent molecular probes have specific responses to trace amounts of bromine and fluorinated ethylene. When the concentration of the target substance is within the set concentration range, a fluorescent signal is generated. The intensity of the fluorescent signal has a nonlinear monotonic relationship with the concentration of the target substance.

[0008] An intelligent mapping module is constructed to collect the fluorescence signal, preprocess it, and then input it into a one-dimensional convolutional neural network and random forest ensemble algorithm model. The model parameters are dynamically updated through an online incremental learning mechanism to achieve a high-precision mapping between the fluorescence signal and the concentrations of trace bromine and fluorinated ethylene.

[0009] An electrochemical verification module was constructed, and a bromide ion selective electrode was deployed simultaneously to detect the bromide ion concentration in the reaction solution in real time. The bromide ion concentration was converted into the free elemental bromide concentration using a bromide valence equilibrium model. The electrochemical signal and the fluorescence signal were input into an extended Kalman filter algorithm for multi-source data fusion and noise suppression, and the cross-validated concentration estimate was output.

[0010] A feedback control module is constructed, which establishes a concentration-valve opening controller based on the concentration estimate to dynamically adjust the opening of the raw material feed valve to maintain the stability of the reaction system. At the same time, a concentration mutation threshold is set, and an early warning mechanism is triggered when the concentration change rate per unit time exceeds the set limit.

[0011] Preferably, the modified fluorescent molecular probe uses a benzothiazole derivative as the fluorescent group, and its surface is modified with perfluoroalkylsilane to enhance its stability in a highly corrosive bromination environment; the probe array contains at least 8 sensing units with different spatial distributions, and each unit has different response sensitivities to the fluorescence quenching effect of bromine and the fluorescence enhancement effect of fluorinated ethylene, forming a multidimensional response feature vector to achieve signal decoupling between the two substances.

[0012] Preferably, the one-dimensional convolutional neural network includes multiple one-dimensional convolutional layers and fully connected layers. The input dimension corresponds to the real-time fluorescence intensity of each probe unit and is used to extract local correlation features in the probe array channel dimension. The random forest model consists of a predetermined number of decision trees. The outputs of the two models are fused by weighted average, and the weights are dynamically adjusted according to historical prediction errors.

[0013] Preferably, the convolutional neural network includes multiple convolutional layers and fully connected layers, the input dimension corresponds to the real-time fluorescence intensity of each probe unit, the random forest model consists of a predetermined number of decision trees, and the outputs of the two models are fused by weighted average, with the weights dynamically adjusted according to historical prediction errors.

[0014] Preferably, the online incremental learning mechanism operates in a sliding time window manner with a window length of 24 hours. A model fine-tuning is triggered every 1,000 new samples. The fine-tuning process adopts a transfer learning strategy, retains the underlying feature extraction structure, and only updates the top-level mapping parameters.

[0015] Preferably, the bromide ion selective electrode uses a silver sulfide-silver bromide composite sensitive membrane, with a detection limit of 10. -9The response time is less than 5 seconds, and the electrode signal is aligned with the fluorescence signal at the same timestamp after being corrected by the temperature compensation circuit. The bromine valence equilibrium model combines the redox potential, reaction progress and temperature parameters detected in real time in the reaction system, and constructs a nonlinear conversion relationship based on the pre-calibrated bromine dissociation-complexation equilibrium constant.

[0016] Preferably, the state vector of the extended Kalman filter algorithm includes two variables: bromine concentration and fluorinated ethylene concentration. The state equation is constructed based on a nonlinear reaction kinetic model and linearized through a first-order Taylor expansion. The observation equation is composed of fluorescence signal and electrochemical signal, and the process noise covariance matrix is ​​adjusted in real time according to the reactor stirring rate and temperature fluctuation.

[0017] Preferably, the concentration-valve opening controller adopts a proportional-integral-derivative control structure, the valve opening adjustment range is the full stroke range, and the minimum adjustment step is 0.1%.

[0018] Preferably, the concentration mutation threshold is set as follows: when the bromine concentration changes by more than 0.5 × 10⁻⁶ within 10 seconds. -6 g / mL, or a change in fluorinated ethylene concentration exceeding 0.1 × 10 g / mL within 10 seconds. -5 When the concentration reaches g / mL, the system automatically triggers multi-level early warnings, including audible and visual alarms, data logging lock, and sending an interrupt command to the central control system.

[0019] Preferably, the method further includes establishing a historical concentration database, storing no less than 100,000 sets of time series data, each set of data containing fluorescence signals, electrochemical signals, temperature, pressure and valve opening information, for offline training of the initial model and fault mode backtracking analysis.

[0020] Compared with the prior art, the present invention has the following beneficial effects:

[0021] 1. Breaking through the detection limits of traditional spectroscopic methods, this technology achieves reliable identification of trace bromine and fluorinated ethylene through chemical sensing amplification and multi-probe array design. The detection limit reaches the set low concentration level, and the dynamic range spans multiple orders of magnitude, effectively capturing the critical concentration fluctuations that cause color changes in the reactor.

[0022] 2. By abandoning reliance on a single sensor, the fluorescence chemical sensing and electrochemical detection signals are simultaneously integrated. Kalman filtering is used to achieve spatiotemporal alignment and noise suppression of multi-source data. Even under strong interference, the concentration inversion error remains below the error threshold, significantly improving the reliability and anti-interference capability of low concentration identification.

[0023] 3. By adopting an integrated architecture of convolutional neural networks and random forests, combined with an online incremental learning mechanism, the model can continuously adapt to the nonlinear and time-varying chemical environment in the reaction system, avoiding the problem of static calibration failure. Under long-term operation, the concentration prediction correlation coefficient is higher than the correlation threshold, meeting the stringent accuracy requirements of closed-loop control.

[0024] 4. Not only does it achieve real-time linkage control of concentration and valve opening to maintain reaction stability, but it also introduces a sudden change early warning mechanism based on the rate of change, which can intervene in advance at the initial stage of side reaction outbreaks to prevent equipment corrosion from intensifying or product purity from dropping sharply, thereby improving process safety and continuous operation time. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0026] Figure 2 This is a schematic diagram of the core principle framework of the collaborative operation of the chemical sensing amplification module and the intelligent mapping module in this invention;

[0027] Figure 3 This is a flowchart illustrating the logical flow of fluorescence signal acquisition, preprocessing, and ensemble mapping of convolutional neural networks and random forests in this invention.

[0028] Figure 4 This is a flowchart illustrating the logical flow of the electrochemical verification module and the Kalman filter multi-source data fusion in this invention.

[0029] Figure 5 This is a flowchart illustrating the logical flow of the feedback control module in this invention, which implements closed-loop regulation of concentration-valve opening and early warning of sudden changes.

[0030] Figure 6 This is a schematic diagram illustrating the overall coordination of multi-level interaction relationships and data flow in the distributed control system of this invention. Detailed Implementation

[0031] Example 1: Reference Figures 1 to 6 The trace chemical detection method for color changes in the bromination reactor during the production of fluorobromoethane proposed in this invention focuses on trace detection, breaks down multidisciplinary boundaries, and includes keywords such as trace detection and multidisciplinary integration. Following a logical chain of "trace concentration - characteristic signal - multi-source verification," it captures 10... -9 Up to 10 -5The study investigated the concentration changes of trace bromine and fluorinated vinyl compounds at the g / mL level. It employed four interdisciplinary modules to construct the system: a chemical sensing amplification module using a corrosion-resistant probe array modified with fluorescent molecules to convert trace concentrations into quantified fluorescence signals; an intelligent mapping module using a CNN+RF integrated algorithm combined with online incremental learning to achieve precise signal-concentration mapping; an electrochemical verification module using bromide ion selective electrode detection and a Kalman filter algorithm to integrate multi-source data and improve reliability; and a feedback control module establishing a concentration-valve opening controller, coupled with a sudden change warning to ensure process stability.

[0032] The core innovation lies in interdisciplinary signal amplification, multi-source trace verification, and dynamic adaptive mapping, achieving ppb-level detection accuracy, solving the pain points of traditional trace identification, and being applied to a trace chemical detection method for color changes in the bromination reactor in the production of fluorobromoethane.

[0033] In the above-mentioned trace chemical detection method for color change in the bromination reactor during the production of fluorobromoethane, step (1) involves constructing a chemical sensing amplification module when the concentration of the target substance is 1×10⁻ 9 g / mL ~1×10⁻ 5 When the concentration range is within g / mL, a fluorescence signal is generated. The intensity of the fluorescence signal has a non-linear monotonic relationship with the concentration of the target substance. The specific implementation process is as follows: A modified fluorescent molecular probe array encapsulated in corrosion-resistant material is arranged inside the reactor. The fluorescent molecular probe array is fixed on the inner wall of the reactor or on the built-in support to ensure that each sensing unit is evenly distributed in the reaction liquid phase region and to avoid response deviation due to local flow field disturbance.

[0034] The modified fluorescent molecular probe uses a benzothiazole derivative as the fluorescent group. Its molecular structure contains a sulfur-nitrogen heterocyclic conjugated system, exhibiting high quantum yield and long fluorescence lifetime under UV excitation, and its electron cloud density is highly sensitive to halogen atoms. To enhance stability in highly corrosive bromination environments, the probe surface is modified by chemical vapor deposition with a perfluoroalkylsilane (such as tridecafluorooctyltrimethoxysilane) to form a hydrophobic-bromine-repellent protective layer with a thickness of 50 nm to 200 nm. This protective layer effectively blocks the oxidative quenching of the fluorescent group by free bromine molecules, while allowing target analytes (trace bromine and fluorinated vinyls) to diffuse to the sensing interface.

[0035] The probe array comprises at least eight spatially distributed sensing units. During synthesis, different substituents (such as methyl, trifluoromethyl, and nitro groups) are introduced into each unit, resulting in varying sensitivities to bromine and fluorinated vinyl groups. This leads to differentiated fluorescence intensity outputs under the same concentration conditions. Bromine acts as a strong fluorescence quencher, reducing the fluorescence intensity of all probe units through a dynamic quenching mechanism, conforming to the Stern-Volmer quenching equation. Fluorinated vinyl groups induce fluorescence enhancement through interaction with the probe groups, with significant differences in enhancement magnitude among probes with different substituents. Based on the differences in response coefficients between the two substances and different probe units, an eight-dimensional response equation system is constructed. Combined with a pre-calibrated single-component response matrix, preliminary signal separation of the two substances can be achieved. Finally, a multi-dimensional response feature vector with a dimension of at least eight is formed. Combined with the nonlinear fitting capability of the backend intelligent algorithm, the cross-interference between the two target substances can be effectively decoupled, avoiding concentration inversion bias caused by the superposition of quenching and enhancement effects.

[0036] Specifically, the fluorescence signal excitation and acquisition system in step (1) includes an ultraviolet LED light source with a center wavelength of 365nm, whose optical power stability is better than ±0.5%, and whose pulse modulation frequency is set to 1kHz to suppress background thermal noise. The excitation light is guided to the inside of the reactor via a quartz optical fiber and irradiates the surface of the probe array. The emitted light is filtered by a bandpass filter with a center wavelength of 450nm and a bandwidth of 20nm, and then received by a high-sensitivity photomultiplier tube (PMT). The dark current of the PMT is less than 1nA, and the gain is adjustable within a range of 10. 4 Up to 10 7 The signal-to-noise ratio is higher than 60dB.

[0037] The signal sampling frequency is no less than 100Hz to ensure dynamic capture of rapid color changes (such as transient bromine release caused by side reactions). Each sensing unit is equipped with independent signal conditioning circuitry, including a preamplifier, a low-pass filter (cutoff frequency 50Hz), and an analog-to-digital converter (ADC, 16-bit resolution), which converts the analog fluorescence intensity signal into a digital signal and transmits it to the central processing unit. The sampling of all sensing units is strictly synchronized, with a timestamp alignment error of less than 1ms, to ensure the temporal consistency of the multidimensional feature vectors.

[0038] In the above method, step (2), constructing the intelligent mapping module, is implemented as follows: The fluorescence signal is acquired and preprocessed, including baseline drift correction, outlier removal, and normalization. Baseline drift correction uses a sliding window mean filtering method with a window length of 10 seconds; outlier removal is based on the 3σ principle, performing linear interpolation to replace data points exceeding the mean ± 3 times the standard deviation; normalization uses minimum-maximum scaling to map each channel signal to the [0,1] interval. The preprocessed data is input into the convolutional neural network and random forest ensemble algorithm model.

[0039] Specifically, step (2) employs a one-dimensional convolutional neural network (1D-CNN), comprising three one-dimensional convolutional layers and two fully connected layers, suitable for extracting local correlation features and temporal variation features of the one-dimensional channel dimension of the fluorescent probe array. The first convolutional layer has an input dimension of 8 (corresponding to the real-time fluorescence intensity sequence of 8 sensing units), 32 convolutional kernels, a kernel size of 3, a stride of 1, and an activation function of ReLU; the one-dimensional convolutional kernels slide along the probe channel dimension to extract response correlation features between different sensing units, effectively capturing the array's differentiated response patterns to the two components; the second convolutional layer has 64 kernels, with the other parameters being the same as above; the third convolutional layer is followed by a global average pooling layer, with an output dimension of 64. Fully connected layer 1 contains 128 neurons, and fully connected layer 2 outputs two concentration values ​​(bromine concentration and fluorinated ethylene concentration). The random forest model consists of 200 decision trees, each with a maximum depth of 15 and a feature sampling ratio of 0.8. The outputs of the two models are fused by weighted average, and the fusion formula is:

[0040]

[0041] This is the predicted concentration value after final fusion. This is the original predicted output of the convolutional neural network. This is the original prediction output of the random forest. and The weights are for the convolutional neural network and the random forest, respectively. The weights are dynamically adjusted based on historical prediction errors. Specifically, after each prediction cycle is completed, the mean squared error (MSE) of the two models on the most recent 100 samples is calculated, and the weights are inversely proportional to the MSE.

[0042] By employing an integrated architecture of one-dimensional convolutional neural networks and random forests, combined with an online incremental learning mechanism, the model can continuously adapt to the nonlinear and time-varying chemical environment in the reaction system, avoiding the problem of static calibration failure. Under long-term operation, the Pearson correlation coefficient between the concentration prediction value and the offline measured value is higher than the set correlation threshold of 0.995, meeting the stringent accuracy requirements of closed-loop control.

[0043] Furthermore, the online incremental learning mechanism in step (2) operates using a sliding time window with a window length of 24 hours, triggering a model fine-tuning every 1000 new samples. The fine-tuning process employs a transfer learning strategy, preserving the weights of the bottom-level feature extraction structure of the convolutional neural network (i.e., the first three convolutional layers) and updating only the parameters of the fully connected layers. For the random forest model, an incremental tree construction algorithm is used, expanding the split nodes only for new samples to avoid retraining the entire tree and thus preventing catastrophic forgetting. The fine-tuning process is completed on edge computing devices, with computation latency controlled within 500ms to ensure real-time performance.

[0044] In the above method, step (3), constructing the electrochemical verification module, is implemented as follows: A bromide ion selective electrode is deployed simultaneously. This electrode is installed near the sampling port at the bottom of the reactor, and continuous monitoring is achieved through a flow cell design. The electrode uses a silver sulfide-silver bromide composite sensitive membrane with a thickness of 10 μm, and its detection limit reaches 10 μm. -9 g / mL, response time less than 5 seconds.

[0045] The electrode signal is corrected by a temperature compensation circuit, which integrates a Pt100 temperature sensor with a sampling frequency of 1Hz. The compensation algorithm is based on the temperature correction term of the Nernst equation, ensuring that the potential measurement error is less than ±0.5mV within the range of 20°C to 80°C. The corrected electrochemical signal and fluorescence signal are aligned at the same timestamp. The alignment mechanism uses hardware-triggered synchronization, driven by the same clock source for sampling by both ADC channels.

[0046] Specifically, in step (3), the electrochemical signal and the fluorescence signal are jointly input into the Extended Kalman Filter (EKF) algorithm for multi-source data fusion and noise suppression. This algorithm linearizes the nonlinear reaction kinetic state equation and observation equation using a Taylor first-order expansion, adapting to the nonlinear and time-varying chemical environment of this system, and finally outputs a cross-validated concentration estimate. (State vector of the Extended Kalman Filter (EKF) algorithm) , This refers to the bromine concentration. The concentration of fluorinated vinylidene is given. The equation of state is constructed based on a nonlinear reaction kinetic model, describing the change in concentration with the reaction progress; the observation equation is composed of both fluorescence and electrochemical signals, representing a nonlinear mapping relationship, as follows:

[0047]

[0048] During the filtering calculation, the state equation and the observation equation are expanded by Taylor first order to achieve linear approximation, which satisfies the application prerequisite of Kalman filtering.

[0049] Observation vector at time It includes two fluorescence-derived concentration estimates and an electrochemical bromide ion concentration; the bromide ion concentration needs to be converted to free elemental bromine using a bromine valence equilibrium model. Concentration: In this bromination reaction system, elemental bromine and bromide ions coexist. The system determines the complexation equilibrium and hydrolysis dissociation equilibrium, with the equilibrium constant changing in real time with temperature and solvent polarity. By online monitoring of the reaction system's redox potential, reaction progress (calculated from the cumulative feed amount and conversion rate model), and temperature parameters, combined with a pre-calibrated equilibrium constant-temperature correlation equation, a nonlinear conversion model is constructed to... Concentration converted to free The concentration is used to achieve cross-validation of the elemental bromine concentration detected by fluorescence with electrochemical signals. For the concentration of fluorinated vinylidene, the system is based on the material conservation relationship of the reaction system, combined with the correspondence between the total bromine concentration and the total amount of fluorinated vinylidene in the feed and the reaction conversion rate. The total amount of bromine obtained by electrochemical detection is used as a constraint condition to indirectly correct the estimated concentration of fluorinated vinylidene, achieving multi-source cross-validation of the two-component concentration. The observation equation is a nonlinear mapping relationship, and the observation matrix... It is a 3×2 matrix whose elements are dynamically adjusted according to the current operating conditions.

[0050] Process noise covariance matrix The reactor stirring rate is adjusted in real time based on temperature fluctuations: the stirring rate is obtained from feedback by the frequency converter; the higher the rate, the better. A larger diagonal element indicates a decrease in concentration fluctuation due to improved mixing uniformity; temperature fluctuation is quantified by the standard deviation, with larger fluctuations indicating greater temperature variation. The larger the value, the better the filter's ability to respond to abrupt changes. The extended Kalman filter is executed every 100ms, outputting a cross-validated concentration estimate with a 95% confidence interval width of less than 5% and a concentration inversion relative error below the 5% error threshold.

[0051] In the above method, step (4), constructing a feedback control module, is implemented as follows: Based on the concentration estimate, a concentration-valve opening controller is established, which acts on the pneumatic regulating valve on the fluorinated ethylene feed pipeline. The controller adopts a proportional-integral-derivative (PID) control structure, and its control parameters are preset values: proportional gain... =2.5, integration time =10s, differential time The valve opening adjustment range is 0% to 100% of its full stroke, with a minimum adjustment step of 0.1%, corresponding to a 12-bit valve positioner resolution. The control objective is to maintain the bromine concentration at (5.0 ± 0.2) × 10⁻⁶. -6 g / mL, fluorinated vinylidene concentration within (1.0±0.05)×10 -5 g / mL.

[0052] Specifically, the concentration mutation threshold in step (4) is set as follows: when the bromine concentration changes by more than 0.5 × 10⁻⁶ within 10 seconds. -6 g / mL, or a change in fluorinated ethylene concentration exceeding 0.1 × 10 g / mL within 10 seconds. -5 At a concentration of g / mL, the system automatically triggers multi-level early warnings. Level 1 is a local audible and visual alarm lasting 30 seconds; Level 2 automatically locks all raw data from the current 10 minutes (including fluorescence signals, electrochemical signals, temperature, pressure, and valve opening) and stores it in read-only memory; Level 3 sends an interrupt command to the central control system via industrial Ethernet, suspending raw material feeding and initiating an inert gas purging procedure. The early warning mechanism has a response delay of less than 1 second, ensuring timely intervention in the early stages of side reactions (such as over-bromination leading to polybrominated compounds).

[0053] Furthermore, the method includes establishing a historical concentration database deployed on a factory server, employing a time-series database architecture (such as InfluxDB), storing no fewer than 100,000 sets of time-series data. Each set of data includes a timestamp, 8-channel fluorescence intensity, electrochemical potential, reaction temperature (accuracy ±0.1℃), system pressure (accuracy ±0.5kPa), and valve opening degree (accuracy ±0.1%), with a sampling interval of 1 second. The database is used for offline training of the initial model (completed before the system's first commissioning) and fault mode backtracking analysis, supporting multi-dimensional queries by time, concentration range, alarm events, and other dimensions.

[0054] The method is integrated into a distributed control system (DCS) and communicates in real time with the reactor PLC (such as Siemens S7-1500) via industrial Ethernet (IEEE 802.3). Data is transmitted using the OPCUA protocol, with a data transmission delay of less than 10ms, ensuring the timeliness and reliability of closed-loop control. All modules are protected by redundant power supplies and watchdog circuits to ensure safe operation, and a single point of failure does not affect the overall functionality.

[0055] To verify the effectiveness of this embodiment, a specific application example was constructed: In a pilot plant with an annual production capacity of 500 tons of fluorobromoethane, the reactor volume was 2000L, the material was Hastelloy C-276, the operating temperature was 60℃, and the pressure was 0.3MPa. After initial feeding, the system detected that the bromine concentration slowly increased to 4.8×10⁻⁶. -6 g / mL, the concentration of fluorinated vinylidene remained stable at 0.95×10 g / mL. -5 g / mL. At the 12th hour, due to catalyst deactivation, the side reaction accelerated, and the bromine concentration surged to 6.2 × 10 g / mL within 8 seconds. -6Upon reaching g / mL, the system immediately triggered a level three warning, reducing the valve opening from 45% to 32% and restoring the concentration to the target range within 20 seconds. Throughout the process, the relative error between the concentration estimate obtained after multi-source fusion and the offline ICP-MS detection results was better than that of single fluorescence sensing or single electrochemical sensing.

[0056] Example 2: Based on Example 1, if it is necessary to further expand the technical solution to cover more complex working conditions, the following alternative structure can be adopted: The modified fluorescent molecular probe in the chemical sensing amplification module adopts a two-photon excited fluorescent probe with an excitation wavelength of 800nm. A femtosecond laser is used as the light source, and the penetration depth can reach 5mm, suitable for high-turbidity reaction solutions. The probe array is increased to 16 units and integrated onto a flexible polyimide substrate using microelectromechanical systems (MEMS) technology, which can conform to the curved inner wall of the reactor. The convolutional neural network in the intelligent mapping module is replaced with a graph convolutional network (GCN), treating the 16 sensing units as graph nodes. The edge weights are determined by both the spatial Euclidean distance and the chemical response similarity, which is calculated using the Pearson correlation coefficient.

[0057] The electrochemical validation module adds a micro-sensor for fluorinated vinyl gas chromatography with a sampling frequency of 1Hz. This sensor, along with the liquid-phase bromine electrode signal, is input to an extended Kalman filter (EKF), expanding the state vector to 3 dimensions (by increasing the concentration of fluorinated vinyl gas). The feedback control module employs a fuzzy PID controller with a rule base containing 15 expert-implemented rules to handle more nonlinear operating conditions. This embodiment further improves the detection limit to 5 × 10⁻⁶. -10 g / mL, suitable for the production of ultra-high purity fluorobromoethane.

Claims

1. A trace chemical detection method for color changes in the bromination reactor during the production of fluorobromoethane, characterized in that, Includes the following steps: A chemical sensing amplification module was constructed, and an array of modified fluorescent molecular probes encapsulated in corrosion-resistant materials was deployed inside the reactor. The fluorescent molecular probes have specific responses to trace amounts of bromine and fluorinated ethylene. When the concentration of the target substance is within the set concentration range, a fluorescent signal is generated. The intensity of the fluorescent signal has a nonlinear monotonic relationship with the concentration of the target substance. An intelligent mapping module is constructed to collect the fluorescence signal, preprocess it, and then input it into a one-dimensional convolutional neural network and random forest ensemble algorithm model. The model parameters are dynamically updated through an online incremental learning mechanism to achieve a high-precision mapping between the fluorescence signal and the concentrations of trace bromine and fluorinated ethylene. An electrochemical verification module was constructed, and a bromide ion selective electrode was deployed simultaneously to detect the bromide ion concentration in the reaction solution in real time. The bromide ion concentration was converted into the free elemental bromide concentration using a bromide valence equilibrium model. The electrochemical signal and the fluorescence signal were input into an extended Kalman filter algorithm for multi-source data fusion and noise suppression, and the cross-validated concentration estimate was output. A feedback control module is constructed, which establishes a concentration-valve opening controller based on the concentration estimate to dynamically adjust the opening of the raw material feed valve to maintain the stability of the reaction system. At the same time, a concentration mutation threshold is set, and an early warning mechanism is triggered when the concentration change rate per unit time exceeds the set limit.

2. The trace chemical detection method for color change in the bromination reactor during the production of fluorobromoethane according to claim 1, characterized in that, The modified fluorescent molecular probe uses a benzothiazole derivative as the fluorescent group, and its surface is modified with perfluoroalkylsilane to enhance its stability in a highly corrosive bromination environment. The probe array contains at least 8 sensing units with different spatial distributions. Each unit has different response sensitivities to the fluorescence quenching effect of bromine and the fluorescence enhancement effect of fluorinated ethylene, forming a multidimensional response feature vector to achieve signal decoupling between the two substances.

3. The trace chemical detection method for color change in the bromination reactor during the production of fluorobromoethane according to claim 2, characterized in that, The one-dimensional convolutional neural network contains multiple one-dimensional convolutional layers and fully connected layers. The input dimension corresponds to the real-time fluorescence intensity of each probe unit, which is used to extract local correlation features of the probe array channel dimension. The random forest model consists of a predetermined number of decision trees. The outputs of the two models are fused by weighted average, and the weights are dynamically adjusted according to historical prediction errors.

4. The trace chemical detection method for color change in the bromination reactor during the production of fluorobromoethane according to claim 3, characterized in that, The convolutional neural network contains multiple convolutional layers and fully connected layers. The input dimension corresponds to the real-time fluorescence intensity of each probe unit. The random forest model consists of a predetermined number of decision trees. The outputs of the two models are fused by weighted average, and the weights are dynamically adjusted according to historical prediction errors.

5. The trace chemical detection method for color change in the bromination reactor during the production of fluorobromoethane according to claim 4, characterized in that, The online incremental learning mechanism operates in a sliding time window manner with a window length of 24 hours. Every 1,000 new samples triggers a model fine-tuning. The fine-tuning process adopts a transfer learning strategy, which preserves the underlying feature extraction structure and only updates the top-level mapping parameters.

6. The trace chemical detection method for color change in the bromination reactor during the production of fluorobromoethane according to claim 5, characterized in that, The bromide ion selective electrode uses a silver sulfide-silver bromide composite sensitive membrane, and its detection limit reaches 10. -9 The response time is less than 5 seconds, and the electrode signal is aligned with the fluorescence signal at the same timestamp after being corrected by the temperature compensation circuit. The bromine valence equilibrium model combines the redox potential, reaction progress and temperature parameters detected in real time in the reaction system, and constructs a nonlinear conversion relationship based on the pre-calibrated bromine dissociation-complexation equilibrium constant.

7. The trace chemical detection method for color change in the bromination reactor during the production of fluorobromoethane according to claim 6, characterized in that, The extended Kalman filter algorithm's state vector includes two variables: bromine concentration and fluorinated ethylene concentration. The state equation is constructed based on a nonlinear reaction kinetic model and linearized through a first-order Taylor expansion. The observation equation is composed of both fluorescence and electrochemical signals, and the process noise covariance matrix is ​​adjusted in real time according to the reactor stirring rate and temperature fluctuations.

8. The trace chemical detection method for color change in the bromination reactor during the production of fluorobromoethane according to claim 7, characterized in that, The concentration-valve opening controller adopts a proportional-integral-derivative control structure, and the valve opening adjustment range is the entire stroke range with a minimum adjustment step of 0.1%.

9. The trace chemical detection method for color change in the bromination reactor during the production of fluorobromoethane according to claim 8, characterized in that, The concentration mutation threshold is set as follows: when the bromine concentration changes by more than 0.5 × 10⁻⁶ within 10 seconds. -6 g / mL, or a change in fluorinated ethylene concentration exceeding 0.1 × 10 g / mL within 10 seconds. -5 When the concentration reaches g / mL, the system automatically triggers multi-level early warnings, including audible and visual alarms, data logging lock, and sending an interrupt command to the central control system.

10. The trace chemical detection method for color change in the bromination reactor during the production of fluorobromoethane according to claim 9, characterized in that, The method also includes establishing a historical concentration database, storing no less than 100,000 sets of time series data. Each set of data includes fluorescence signals, electrochemical signals, temperature, pressure, and valve opening information, which are used for offline training of the initial model and fault mode backtracking analysis.

Images