Multi-source sensing and early warning system for leakage of diazotization and nitrification device

Abstract

The invention discloses a multi-source sensing and early warning system for leakage of a diazotization and nitrification device. The multi-source sensing and early warning system comprises a data monitoring module, a protection module, a data processing module and a grading early warning module, the data monitoring module comprises a physical sensing unit and a visual sensing unit; the visual sensing unit comprises a visual sensor and a color-changing coating sprayed on the outer wall of the device; according to the protection module, the color-changing coating is wrapped with a transparent anti-corrosion film layer, liquid in the controlled device is prevented from overflowing after leaking, the leaked liquid extends on the outer wall of the controlled device, and therefore the color-changing area of the color-changing coating is increased. The data processing module is used for processing data monitored by the physical sensing unit and the visual sensing unit in a deep learning model and judging whether leakage occurs or whether a leakage trend occurs; a grading early warning module which is internally provided with a plurality of threshold values for triggering alarms of different levels; the plurality of thresholds correspond to the judgment result of the data processing module.

Description

Technical Field

[0001] This invention relates to the field of leakage monitoring and early warning technology for nitration units, specifically a multi-source sensing and early warning system for leakage in diazotization nitration units. Background Technology

[0002] In the diazotization and nitration process, highly corrosive media are involved in transportation and reaction. Therefore, media leakage is a core risk factor threatening safe production. Some equipment that transports and reacts with the media is exposed to highly corrosive conditions for extended periods. During long-term operation, it is prone to leakage due to corrosion, stress fatigue, and other reasons, which could trigger a violent reaction and explosion. Therefore, it is necessary to monitor these devices.

[0003] Existing monitoring technologies mainly rely on single sensors for monitoring and judgment; however, the complex workshop environment is easily affected by external interference, making monitoring difficult and inaccurate. In particular, the sensitivity and accuracy of monitoring minute leaks are low, resulting in unsatisfactory monitoring effects. Therefore, this invention provides a multi-source sensing and early warning system for leaks in diazotization and nitration units. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-source sensing and early warning system for leaks in diazotization and nitration units, so as to solve the problem of the single monitoring method for the unit mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A multi-source sensing and early warning system for leaks in a diazotization and nitration unit includes a data monitoring module, a protection module, a data processing module, and a graded early warning module;

[0007] Data monitoring module: includes physical sensing unit and visual sensing unit; physical sensing unit includes temperature sensor, pressure sensor and ultrasonic sensor to monitor the temperature, pressure and wall thickness of the controlled device respectively; visual sensing unit includes visual sensor and color-changing coating sprayed on the outer wall of the device; the color-changing coating changes color when exposed to acid, and the color-changing area of ​​the color-changing coating is recorded by visual sensor.

[0008] Protective module: A transparent, corrosion-resistant film layer with a barrier function is wrapped around the color-changing coating to prevent liquid from overflowing after leakage;

[0009] Data processing module: Uses deep learning models to process the data monitored by the physical sensing unit and the visual sensing unit to determine whether there is leakage or a leakage trend; the deep learning models include LSTM neural network model and convolutional neural network model, which are used to process physical sensing unit data and visual sensing unit data respectively.

[0010] Tiered early warning module: It has multiple thresholds to trigger alarms at different levels; the multiple thresholds correspond to the judgment results of the data processing module; setting threshold A corresponds to the controlled device having a leakage trend and automatically triggers a level 1 alarm; setting threshold B corresponds to the controlled device having a slight leakage and automatically triggers a level 2 alarm; setting threshold C corresponds to the controlled device having a serious leakage and automatically triggers a level 3 alarm.

[0011] Preferably, a Level 1 alarm is triggered when the wall thickness thins to a preset thinning threshold; a Level 2 alarm is triggered when the temperature or pressure of the controlled device exceeds the historical normal range and is within the first percentage threshold, or when the discolored area of ​​the color-changing coating is within the second percentage threshold of the total area; and a Level 3 alarm is triggered when the temperature or pressure of the controlled device exceeds the historical normal range and exceeds the first percentage threshold, or when the discolored area of ​​the color-changing coating exceeds the second percentage threshold of the total area.

[0012] Preferably, when different models trigger alarms of different levels, the higher level shall prevail; if the levels are the same, the judgment result calculated by the LSTM neural network model shall be given priority.

[0013] Preferably, an LSTM neural network model is used to process the physical sensor unit data. The specific steps are as follows:

[0014] S1. Data preprocessing: Temperature sensors, pressure sensors, and ultrasonic sensors collect real-time data from the controlled device, generate time series, and normalize the data.

[0015] S2. Use an LSTM neural network model to process the time-series data of temperature, pressure and wall thickness in the physical sensing unit;

[0016] S3. Compare the LSTM output with the threshold to determine whether an alarm is triggered.

[0017] Preferably, the specific calculation process of the LSTM neural network model is as follows:

[0018] By calculating the forget gate, irrelevant information is discarded;

[0019] By calculating the input gate and candidate memory units, abnormal trends can be captured;

[0020] Accumulate long-term aberration features by calculating cell state updates;

[0021] The LSTM output is obtained by calculating the output gate and the current hidden state.

[0022] Preferably, a convolutional neural network model is used to process the visual sensing unit data to determine whether there is leakage. The specific steps are as follows:

[0023] S1. Image preprocessing: Preprocessing the image by converting it to grayscale and reducing noise to form the input image X;

[0024] S2. Use a convolutional neural network model to analyze the color-changing coating images captured by the visual sensor to identify the color change areas and the degree of leakage;

[0025] S3. Compare the output of the convolutional neural network model with the threshold to determine whether an alarm is triggered.

[0026] Preferably, the specific calculation steps of the convolutional neural network model include S2-1, convolutional layer calculation: applying a convolutional kernel K to the input image X to generate a feature map Y, and outputting the size of the feature map.

[0027] Preferably, it also includes S2-2, which reduces the feature map size and lowers the position sensitivity by calculating the output size through a pooling layer.

[0028] Preferably, it also includes S2-3, calculating through a fully connected layer: flattening the pooled feature map into a one-dimensional vector, mapping it to the classification space through a fully connected layer and outputting the result.

[0029] Preferably, the LSTM output corresponds to threshold A, threshold B, or threshold C; the convolutional neural network model output corresponds to threshold B or threshold C; threshold A automatically triggers a level one alarm, prompting maintenance personnel to check the warning area of ​​the controlled device; threshold B automatically triggers a level two alarm, prompting maintenance personnel to plug leaks in the warning area; threshold C automatically triggers a level three alarm, and the upstream and downstream valves of the controlled device are automatically shut off, and staff are reminded to evacuate through an alarm in the workshop.

[0030] Compared with existing technologies, the beneficial effects of this invention are: improved accuracy in identifying minute leaks. The color-changing effect of the color-changing coating increases its recognizability; simultaneously, in conjunction with the transparent anti-corrosion film, liquid accumulates at the leak point, increasing the color-changing area of ​​the coating. Furthermore, the anti-corrosion film prevents liquid leakage from spilling out and corroding other equipment, thus avoiding damage; it also prevents accidental contact by personnel, reducing safety hazards.

[0031] Multi-source sensing enhances anti-interference capabilities. By combining visual and physical sensing units, the accuracy of assessing leaks in controlled devices can be improved, reducing the false alarm rate.

[0032] Proactive early warning reduces accident losses. Through physical sensing units, changes in pressure, temperature, and wall thickness of the controlled device can be monitored. In conjunction with a graded early warning module, it can promptly detect when the wall thickness of the controlled device becomes thinner and shows a tendency to leak. In the event of a leak, it can automatically trigger corresponding alarms according to the degree of leakage, thereby reducing economic losses. Detailed Implementation

[0033] Example 1:

[0034] This invention provides a technical solution: a multi-source sensing and early warning system for leaks in diazotization and nitration units, comprising a data monitoring module, a protection module, a data processing module, and a graded early warning module;

[0035] The data monitoring module includes a physical sensing unit and a visual sensing unit. The physical sensing unit includes temperature sensors, pressure sensors, and ultrasonic sensors that monitor the temperature, pressure, and wall thickness of the controlled device, respectively. The visual sensing unit includes a visual sensor and a color-changing coating sprayed on the outer wall of the controlled device. The color-changing coating changes color when exposed to acid, and the visual sensor records the area of ​​color change. Sensors are positioned on the outer wall of the controlled device to ensure comprehensive monitoring. Temperature and pressure sensors are used to collect temperature and pressure data inside the controlled device.

[0036] Specifically, the color-changing coating can be made by spraying pH-sensitive polymer microcapsules. The outer shell of the pH-sensitive polymer microcapsules is modified polyurethane, which encapsulates a mixed indicator of bromocresol purple (pH 5.2-6.8) and methyl red (pH 4.4-6.2). When the capsules are exposed to acid (pH<5), the indicator comes into contact with the acid and instantly changes color (response time <0.5 seconds), changing from yellow to purple-red (color difference ΔE>15, visible to the naked eye).

[0037] Protective module: A transparent anti-corrosion film layer with a barrier function is wrapped around the color-changing coating to prevent liquid leakage from the controlled device from overflowing and to allow the leaked liquid to extend on the outer wall of the controlled device, thereby increasing the area of ​​color change of the color-changing coating and improving the monitoring accuracy of the visual sensing unit.

[0038] Specifically, the anti-corrosion membrane can consist of three layers. The inner layer is a hydrophilic modified fluororubber layer, which accelerates the spread of leaking liquid and enlarges the leak warning area to facilitate rapid identification and detection by sensors; the middle layer is expanded PTFE (expanded polytetrafluoroethylene), with axially parallel microgrooves laser-etched on the surface, which guides the leaking liquid to extend axially along the controlled device through capillary effect; the outer layer is perfluoroether rubber, which provides corrosion-resistant sealing.

[0039] Data processing module: This module uses deep learning models to process the data monitored by the physical and visual sensing units to determine whether there is leakage or a leakage trend. The deep learning models include LSTM neural network models and convolutional neural network models, which are used to process data from the physical and visual sensing units, respectively.

[0040] Tiered early warning module: It has multiple thresholds to trigger alarms at different levels; the multiple thresholds correspond to the judgment results of the data processing module; setting threshold A corresponds to the controlled device having a leakage trend and automatically triggers a level 1 alarm; setting threshold B corresponds to the controlled device having a slight leakage and automatically triggers a level 2 alarm; setting threshold C corresponds to the controlled device having a serious leakage and automatically triggers a level 3 alarm.

[0041] When the ultrasonic sensor detects a thinning of the wall thickness (reaching a preset thinning threshold), it indicates a leakage trend in the controlled device, triggering a Level 1 alarm. When the temperature or pressure sensor detects that the temperature or pressure of the controlled device exceeds the historical normal range by less than a first percentage threshold (e.g., 30%), or when the visual sensing unit detects that the discolored area of ​​the color-changing coating is less than a second percentage threshold (e.g., 10%) of the total area of ​​the color-changing coating, it indicates a minor leakage in the controlled device, triggering a Level 2 alarm. When the temperature or pressure sensor detects that the temperature or pressure of the controlled device exceeds the historical normal range by more than a first percentage threshold, or when the visual sensing unit detects that the discolored area of ​​the color-changing coating exceeds a second percentage threshold of the total area of ​​the color-changing coating, it indicates a serious leakage in the controlled device, triggering a Level 3 alarm.

[0042] The controlled device can be a reactor or a pipeline, used to monitor for leaks. Reactors are usually located on the ground, providing convenient conditions for operators to monitor leaks. Pipelines, on the other hand, are long and often arranged in a crisscross pattern in the workshop, or even in the air, making monitoring difficult and leak points hard to find. Therefore, the monitoring scheme of this application is more suitable for monitoring leaks in pipelines than reactors.

[0043] It should be noted that the historical normal range value refers to the baseline range of sensor data (such as the fluctuation range of temperature, pressure, and wall thickness) obtained through long-term monitoring and statistics of the controlled device under leak-free operating conditions. This range needs to be determined through commissioning during the initial deployment of the system, or preset according to the equipment's safety operation standards. The first / second percentage threshold and the thinning threshold are both adjustable parameters, configured according to the actual safety standards of the controlled device.

[0044] When different models (LSTM neural network model and convolutional neural network model) trigger alarms of different levels, the higher level shall prevail; since sudden changes in diazotization process parameters are more dangerous, if the levels are the same, the judgment result of physical sensor data (i.e., LSTM neural network model) shall be given priority.

[0045] An LSTM neural network model is used to process the monitoring data from the physical sensing unit. The LSTM output is compared with a threshold to determine whether an alarm is triggered. If the LSTM output corresponds to threshold A, a Level 1 alarm is automatically triggered, prompting maintenance personnel to check the controlled devices in the warning area. If the LSTM output corresponds to threshold B, a Level 2 alarm is automatically triggered, prompting maintenance personnel to plug leaks in the controlled devices in the warning area. If the LSTM output corresponds to threshold C, a Level 3 alarm is automatically triggered, and the upstream and downstream valves of the controlled device are automatically shut off, with an alarm in the workshop alerting staff to evacuate.

[0046] A convolutional neural network (CNN) model is used to process the monitoring data from the visual sensing unit. The output of the CNN model is compared with a threshold to determine whether an alarm is triggered. If the output of the CNN model corresponds to threshold B, a level two alarm is automatically triggered, prompting maintenance personnel to plug leaks in the controlled devices in the warning area. If the output of the CNN model corresponds to threshold C, a level three alarm is automatically triggered, and the upstream and downstream valves of the controlled devices are automatically shut off. An alarm in the workshop also alerts staff to evacuate.

[0047] Both LSTM and CNN models use the cross-entropy loss function and are trained using the Adam optimizer.

[0048] All thresholds (such as preset thinning thresholds and percentage thresholds) and model parameters (such as dhidden and convolution kernel size K) in this system are configurable variables. In actual deployment, they need to be dynamically set according to the operating conditions of the controlled device, historical data, and safety standards.

[0049] Example 2

[0050] Based on Example 1, the specific process for processing the monitoring data of the physical sensing unit is as follows:

[0051] S1. Data preprocessing: Temperature sensors, pressure sensors, and ultrasonic sensors collect real-time data from the controlled device, generate time series, and normalize the data (specifically using Z-score standardization) to eliminate dimensional differences.

[0052] S2. Use an LSTM neural network model to process the time-series data of temperature, pressure and wall thickness in the physical sensing unit.

[0053] S2-1, Calculation is performed through the gating unit. Formula f is used. t =σ(W f ·[h t-1 ,x t ]+b f ) Calculate the forget gate to discard irrelevant information.

[0054] Where f tσ is the activation vector for the forget gate; σ is the sigmoid activation function with an output range of (0,1); W f b is the weight matrix of the forget gate; f Here is the bias vector for the forget gate; [h] t-1 ,x t [To make h] t-1 With x t pieced together; x t The input vector (normalized data of temperature, pressure, and wall thickness) is defined as dinput for time step t, with an input dimension of dinput (e.g., dinput = 3; model dimension parameters can be determined through training optimization); the hidden layer dimension is dhidden (e.g., dhidden = 64; model dimension parameters can be determined through training optimization); h t-1 This is the hidden state of the previous time step.

[0055] S2-2, Using formula i t =σ(W i ·[h t-1 ,x t ]+b i )and Calculate the input gate and candidate memory units to capture anomalous trends. t W is the activation vector of the input gate; i b is the weight matrix of the input gate; i The bias vector for the input gate; Candidate memory units; Tanh is the hyperbolic tangent activation function with an output range of (-1, 1); W c and b c These are the weight matrix and bias vector of the candidate memory cells, respectively.

[0056] S2-3, Using the formula Calculate cell state updates to accumulate long-term anomalous features. ⊙ represents element-wise multiplication, controlled by forget gates and input gates to integrate historical states with new information; c t The cell state at the current time step; c t-1 This refers to the cell state at the previous time step.

[0057] S2-4, Using formula o t =σ(W o ·[h t-1 ,x t ]+b o ) and h t =o t ⊙tanh(c t Calculate the output gate and the current hidden state. t W is the activation vector of the output gate; o and b oh represents the weight matrix and bias vector of the output gate. t The hidden state at the current time step t represents the probability of leakage risk at each time step.

[0058] S3. Compare the LSTM output with the threshold to determine whether an alarm is triggered.

[0059] All thresholds (such as preset thinning thresholds and percentage thresholds) and model parameters (such as dhidden and convolution kernel size K) in this system are configurable variables. In actual deployment, they need to be dynamically set according to the operating conditions of the controlled device, historical data, and safety standards.

[0060] Example 3

[0061] Based on Example 1, the specific process for processing the monitoring data of the visual sensing unit is as follows:

[0062] S1. Image preprocessing: Preprocessing the image by converting it to grayscale and reducing noise to form the input image X;

[0063] S2. Use a convolutional neural network (CNN) model to analyze the color-changing coating images captured by the visual sensor to identify the color change areas and the degree of leakage.

[0064] S2-1, Convolutional Layer Calculation: Apply a convolution kernel K to the input image X to generate a feature map Y. The dimensions of image X are H×W×C. in C in For the number of input channels, C out H represents the number of output channels, H represents the height, and W represents the width; the size of the convolution kernel K is K×K×C. in ×C out K is the kernel size (usually an odd number, such as 3×3); the formula is used. Perform convolution calculation; Y(i,j) is the value of the generated feature map Y at position (i,j); (i,j) is the position coordinate on the output feature map; W conv (u,v) represents the weights of the convolution kernel; X(i+u,j+v) represents the local region values ​​of the input feature map; b is the bias term; the formula is used. The size of the output feature map; S is the stride, representing the number of pixels the convolution kernel moves in each iteration; P represents the number of zeros padded at the edges of the input feature map; W conv This represents the weight matrix.

[0065] S2-2, Pooling Layer Calculation: Downsampling of the feature map Y, retaining the maximum value in the local region: using the formula... Max pooling is performed, where m and n are used to iterate through all elements within the current pooling window; the formula is used. Perform output dimension calculation.

[0066] S2-3. Fully connected layer calculation: The pooled feature map is flattened into a one-dimensional vector y, and mapped to the classification space through a fully connected layer; the fully connected layer is calculated using the formula y = W·x + b; where x is the flattened input vector and b is the bias term.

[0067] S3. Use a convolutional neural network (CNN) model to process the monitoring data of the visual sensing unit and compare the output of the CNN model with the threshold to determine whether an alarm is triggered.

[0068] All thresholds (such as preset thinning thresholds and percentage thresholds) and model parameters (such as dhidden and convolution kernel size K) in this system are configurable variables. In actual deployment, they need to be dynamically set according to the operating conditions of the controlled device, historical data, and safety standards.

Claims

1. A multi-source sensing and early warning system for leaks in a diazotization and nitration unit, characterized in that: It includes a data monitoring module, a protection module, a data processing module, and a tiered early warning module; Data monitoring module: includes physical sensing unit and visual sensing unit; physical sensing unit includes temperature sensor, pressure sensor and ultrasonic sensor to monitor the temperature, pressure and wall thickness of the controlled device respectively; visual sensing unit includes visual sensor and color-changing coating sprayed on the outer wall of the device; the color-changing coating changes color when exposed to acid, and the color-changing area of ​​the color-changing coating is recorded by visual sensor. Protective module: A transparent, corrosion-resistant film layer with a barrier function is wrapped around the color-changing coating to prevent liquid from overflowing after leakage; Data processing module: Uses deep learning models to process the data monitored by the physical sensing unit and the visual sensing unit to determine whether there is leakage or a leakage trend; the deep learning models include LSTM neural network model and convolutional neural network model, which are used to process physical sensing unit data and visual sensing unit data respectively. Tiered early warning module: It has multiple thresholds to trigger alarms at different levels; Multiple thresholds correspond to the judgment results of the data processing module; setting threshold A corresponds to the controlled device having a leakage trend and automatically triggering a level 1 alarm, setting threshold B corresponds to the controlled device having a slight leakage and automatically triggering a level 2 alarm, and setting threshold C corresponds to the controlled device having a serious leakage and automatically triggering a level 3 alarm.

2. The multi-source sensing and early warning system for leakage in a diazotization nitration unit according to claim 1, characterized in that: When the wall thickness thins to the preset thinning threshold, a level one alarm is triggered; when the temperature or pressure of the controlled device exceeds the historical normal range value within the first percentage threshold, or when the discolored area of ​​the color-changing coating accounts for the second percentage threshold of the total area, a level two alarm is triggered. When the temperature or pressure of the controlled device exceeds the first percentage threshold above the historical normal range, and the discolored area of ​​the color-changing coating exceeds the second percentage threshold of the total area of ​​the color-changing coating, a level three alarm is triggered.

3. The multi-source sensing and early warning system for leakage in a diazotization and nitration unit according to claim 1, characterized in that: When different models trigger alarms of different levels, the higher level shall prevail; if the levels are the same, the judgment result calculated by the LSTM neural network model shall be given priority.

4. A multi-source sensing and early warning system for leakage in a diazotization and nitration unit according to claim 1, characterized in that: The LSTM neural network model is used to process the data from the physical sensing unit. The specific steps are as follows: S1. Data preprocessing: Temperature sensors, pressure sensors, and ultrasonic sensors collect real-time data from the controlled device, generate time series, and normalize the data. S2. Use an LSTM neural network model to process the time-series data of temperature, pressure and wall thickness in the physical sensing unit; S3. Compare the LSTM output with the threshold to determine whether an alarm is triggered.

5. A multi-source sensing and early warning system for leakage in a diazotization nitration unit according to claim 4, characterized in that: The specific calculation process of the LSTM neural network model is as follows: By calculating the forget gate, irrelevant information is discarded; By calculating the input gate and candidate memory units, abnormal trends can be captured; Accumulate long-term aberration features by calculating cell state updates; The LSTM output is obtained by calculating the output gate and the current hidden state.

6. A multi-source sensing and early warning system for leakage in a diazotization and nitration unit according to claim 1, characterized in that: The convolutional neural network model is used to process visual sensor unit data and determine whether there is leakage. The specific steps are as follows: S1. Image preprocessing: Preprocessing the image by converting it to grayscale and reducing noise to form the input image X; S2. Use a convolutional neural network model to analyze the color-changing coating images captured by the visual sensor to identify the color change areas and the degree of leakage; S3. Compare the output of the convolutional neural network model with the threshold to determine whether an alarm is triggered.

7. A multi-source sensing and early warning system for leakage in a diazotization nitration unit according to claim 6, characterized in that: The specific computation steps of the convolutional neural network model include S2-1, convolutional layer computation: applying a convolutional kernel K to the input image X to generate a feature map Y, and outputting the size of the feature map.

8. A multi-source sensing and early warning system for leakage in a diazotization nitration unit according to claim 6, characterized in that: It also includes S2-2, which uses pooling layers to calculate the output size, thereby reducing the feature map size and lowering position sensitivity.

9. A multi-source sensing and early warning system for leakage in a diazotization nitration unit according to claim 6, characterized in that: It also includes S2-3, which is calculated through a fully connected layer: the pooled feature map is flattened into a one-dimensional vector, mapped to the classification space through a fully connected layer, and the result is output.

10. A multi-source sensing and early warning system for leakage in a diazotization nitration unit according to claim 4 or 6, characterized in that: The LSTM output corresponds to threshold A, threshold B, or threshold C; the convolutional neural network model output corresponds to threshold B or threshold C; threshold A automatically triggers a level 1 alarm, prompting maintenance personnel to check the warning area of ​​the controlled device; threshold B automatically triggers a level 2 alarm, prompting maintenance personnel to plug leaks in the warning area; threshold C automatically triggers a level 3 alarm, and the upstream and downstream valves of the controlled device are automatically shut off, and staff are alerted to evacuate the workshop via an alarm.