A resin reactor infrared thermal imaging abnormal hot spot monitoring and alarming system

Abstract

This invention specifically relates to an infrared thermal imaging abnormal hotspot monitoring and alarm system for resin reactors, belonging to the field of industrial process safety monitoring technology. It includes: an infrared thermal image acquisition module; an image preprocessing module; an abnormal hotspot detection and feature extraction module; a graded alarm discrimination module; and an alarm output and linkage control module. This invention sets multi-dimensional abnormality judgment conditions such as absolute temperature threshold, temperature gradient threshold, and temperature difference threshold, and combines pixel-level semantic segmentation of a convolutional neural network with the union fusion of threshold rule detection results. This solves the technical problem of insufficient detection capability of a single threshold judgment method for abnormal hotspots of different causes, achieving the technical effect of improving the detection coverage and segmentation accuracy of abnormal hotspots.

Description

Technical Field

[0001] This invention relates to the field of industrial process safety monitoring technology, and in particular to an infrared thermal imaging abnormal hot spot monitoring and alarm system for resin reactors. Background Technology

[0002] During exothermic reactions such as polymerization and condensation in resin reactors, the temperature distribution on the reactor surface and in key areas such as the agitator shaft seal, inlet and outlet flanges, and jacket cooling water pipe interfaces are important indicators reflecting the reaction status and equipment health. Due to uneven heat dissipation, lining damage, or cooling failure, abnormal hot spots may appear on the reactor surface. If not detected in time, these hot spots could lead to material gelation, reactor wall overheating, or even safety accidents.

[0003] In existing technologies, contact thermocouples are typically used to measure the temperature of key components of a reactor at specific points to monitor the reactor's temperature status. However, contact thermocouple temperature measurement has limitations, including a limited number of measurement points and insufficient spatial coverage. It can only acquire temperature information from a limited number of discrete points, making it difficult to comprehensively capture the spatial temperature distribution on the reactor surface. When abnormal hotspots occur outside the measurement points, existing methods cannot effectively detect them and lack the ability to predict the development trend of abnormal hotspots and a graded response mechanism for abnormal states of different severity, resulting in insufficient coverage and timeliness of monitoring alarms. Summary of the Invention

[0004] The purpose of this invention is to provide an infrared thermal imaging abnormal hot spot monitoring and alarm system for resin reactors in order to solve the above-mentioned problems.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: An infrared thermal imaging abnormal hot spot monitoring and alarm system for a resin reactor includes: The infrared thermal image acquisition module is used to scan the surface of the resin reactor with an infrared thermal imager at a preset acquisition cycle to obtain a sequence of infrared thermal image frames containing information on the spatial distribution of temperature on the surface of the reactor. The image preprocessing module is used to perform radiometric correction and noise filtering on the infrared thermal image frame sequence to generate corrected temperature distribution data; The abnormal hotspot detection and feature extraction module is used to detect abnormal hotspot regions based on the corrected temperature distribution data, using multi-dimensional anomaly judgment conditions including absolute temperature threshold, temperature gradient threshold and temperature difference threshold, and extracting feature parameters including peak temperature, centroid coordinate position, pixel area and heating rate for each detected abnormal hotspot region to generate an abnormal hotspot feature set. The graded alarm discrimination module is used to match the feature parameters of each abnormal hotspot in the abnormal hotspot feature set with the preset graded alarm rules, determine the alarm level corresponding to each abnormal hotspot, including the warning level, alarm level and emergency alarm level, and generate graded alarm signals. The alarm output and linkage control module is used to execute alarm output actions corresponding to the alarm level based on the graded alarm signals, and to send control commands corresponding to the alarm level to the linkage control system.

[0006] Preferably, while acquiring the infrared thermal image frame sequence, a visible light image of the reactor surface is also acquired simultaneously using a visible light camera; based on the pre-calibrated spatial mapping relationship between the infrared thermal imager and the visible light camera, the infrared thermal image and the visible light image are registered to generate fused image data, which simultaneously contains temperature distribution information and structural texture information of the reactor surface; The spatial mapping relationship is obtained through a dual-target calibration method: a calibration plate is placed on the surface of the reactor, and images of the calibration plate are simultaneously acquired by an infrared thermal imager and a visible light camera. The pixel coordinates of the corner points of the calibration plate in the two images are extracted respectively, and the homography matrix between the two images is solved using the corresponding point pairs. During registration, the homography matrix is ​​applied to each pixel coordinate in the infrared thermal image to perform perspective transformation, and its corresponding position in the visible light image coordinate system is obtained. The infrared temperature distribution information and the visible light structure and texture information are aligned and superimposed to generate the fused image data.

[0007] Preferably, the radiation correction process is as follows: based on the preset emissivity parameters of the reactor surface material and the current ambient temperature parameters, the original radiation values ​​of each pixel in the infrared thermal image are compensated and converted to obtain the corrected temperature values; The noise filtering process is as follows: median filtering or Gaussian filtering is used to process the infrared thermal image in the spatial domain, or inter-frame averaging is performed on multiple consecutive thermal images in the time domain.

[0008] Preferably, the multidimensional anomaly determination conditions include: a single point temperature exceeding a preset absolute temperature threshold; a temperature gradient between adjacent regions exceeding a preset gradient threshold; a temperature difference between a local region temperature and the current frame's global average temperature exceeding a preset temperature difference threshold; and when there is a region in the temperature distribution data that meets any one of these conditions, that region is marked as an abnormal hotspot region. The heating rate is calculated as follows: the difference between the peak temperature value of the abnormal hot spot in the current frame and the peak temperature value of the same spatial location in the corresponding frame before the preset time window is divided by the duration of the preset time window to obtain the heating rate value, and the heating rate value is added to the feature parameters of the abnormal hot spot.

[0009] Preferably, when performing the abnormal hotspot region detection, a convolutional neural network is also used to perform semantic segmentation processing on the corrected infrared thermal image; The convolutional neural network employs an encoder-decoder structure. The input is the corrected temperature distribution matrix of a single-frame infrared thermal image, and the output is a pixel-level segmentation mask of the same size as the input image. Each pixel in the segmentation mask is labeled with a category label including normal regions, abnormal hotspot regions, and background regions. The encoder consists of multiple concatenated convolutional blocks, each containing a convolutional layer, a batch normalization layer, and a ReLU activation function. Adjacent convolutional blocks are downsampled using max-pooling layers. The decoder consists of multiple upsampling blocks symmetrical to the encoder. Each upsampling block restores spatial resolution through bilinear interpolation and is then concatenated with the feature map of the corresponding layer of the encoder by channel dimension before being fused through convolutional layers. The output layer is a pixel-wise convolutional layer with the number of channels equal to the number of categories. During training, manually labeled abnormal hotspot segmentation masks are used as supervision labels, and the cross-entropy loss function is employed for optimization. The abnormal hotspot region mask output by the convolutional neural network segmentation is fused with the abnormal hotspot region detected based on the multidimensional anomaly determination conditions, and the fused region is used as the input for feature extraction.

[0010] Preferably, the determination criteria for the graded alarm rule are: The criteria for determining the warning level are: the peak temperature of the abnormal hot spot is within the first temperature range, or the heating rate is within the first rate range. The alarm level is determined by the following conditions: the peak temperature of the abnormal hot spot is in the second temperature range that is higher than the first temperature range, or the heating rate is in the second rate range that is higher than the first rate range, or the area of ​​the abnormal hot spot exceeds the preset area threshold. The criteria for determining the emergency alarm level are: the peak temperature of the abnormal hot spot exceeds the upper limit of the second temperature range, or the peak temperature does not decrease after the alarm level is maintained for a preset duration. The alarm level maintenance time is determined as follows: the timer starts from the moment when the abnormal hot spot first meets the alarm level determination conditions, and in each subsequent frame of detection results, it is continuously verified whether the alarm level determination conditions are still met. If the cumulative duration of continuous satisfaction exceeds the preset duration and the peak temperature does not show a continuous downward trend, an emergency alarm level is triggered.

[0011] Preferably, before matching the abnormal hotspot feature parameters with the graded alarm rules, the persistence of the abnormal hotspot is also verified: when a certain area is detected as an abnormal hotspot in infrared thermal images for a consecutive preset number of frames, and the change trend of the peak temperature does not show a decay and decline feature, the abnormal hotspot is determined as a valid abnormal hotspot and enters the graded alarm discrimination process; for suspected abnormal hotspots that only appear briefly in a single frame or a few frames, they are marked as transient interference and recorded but no alarm is triggered. It also acquires the historical temperature distribution baseline data corresponding to the process formula being executed in the current reactor, calculates the deviation between the temperature distribution of each monitoring area of ​​the current batch and the historical temperature distribution baseline data, and generates an alarm signal at the warning level when the deviation exceeds the preset allowable deviation range, even if the current temperature value has not reached the absolute temperature threshold.

[0012] Preferably, a long short-term memory network is also used to model and predict the peak temperature time series of abnormal hotspots; Before inputting the peak temperature sequence and heating rate sequence into the Long Short-Term Memory (LSTM) network, both are Z-score normalized. The input to the LSM network is a sequence organized in time steps. The input at each time step is a two-dimensional feature vector consisting of the normalized peak temperature value and the corresponding heating rate value. The LSM network consists of an input gate, a forget gate, an output gate, and unit states. After receiving the two-dimensional feature vector at each time step, the forget gate determines the proportion of historical information to be discarded in the unit state, and the input gate determines the amount of new information to be written into the unit state. The unit state is updated after the forgetting and writing operations. The output gate generates the hidden state vector of the current time step based on the updated unit state. The output layer is a linear fully connected layer that maps the hidden state vector output at the last time step to a sequence of predicted peak temperatures within a preset future time step. After inverse normalization, the predicted values ​​are restored to actual temperature values ​​and compared with the warning level temperature threshold. During training, the peak temperature time series of historical batches are used as training samples, and the mean squared error is used as the loss function. When the predicted temperature sequence output by the Long Short-Term Memory Network contains a predicted value that exceeds the warning level temperature threshold, a warning level alarm signal is generated in advance.

[0013] Preferably, the execution of the alarm output action and linkage control action corresponding to the alarm level includes: For alarm signals at the warning level, a warning prompt window will pop up on the monitoring interface, displaying the location, peak temperature and heating rate of the abnormal hotspot, and pushing the warning message to the mobile terminal of the on-duty personnel. For alarm signals at the alarm level, in addition to executing the alarm output action at the early warning level, the on-site audible and visual alarm device is triggered to perform an audible and visual alarm, and a control command to reduce the feeding rate or open the auxiliary cooling circuit is sent to the distributed control system or programmable logic controller. For alarm signals at the emergency alarm level, in addition to executing the alarm output action at the alarm level, a control command to suspend feeding and execute an emergency cooling program is sent to the distributed control system or programmable logic controller. If the temperature of the abnormal hot spot does not show a downward trend after the emergency alarm has lasted for a preset time, an emergency stop command is further sent. While executing the alarm output action and linkage control action, the infrared thermal image frame, visible light image frame, abnormal hot spot characteristic parameters, alarm level, sent linkage control command content and timestamp at the time of alarm triggering are synchronously written into the data management storage system and indexed and stored according to batch and time sequence.

[0014] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention solves the technical problem that a single threshold judgment method is insufficient for detecting abnormal hotspots of different causes by setting multi-dimensional anomaly judgment conditions such as absolute temperature threshold, temperature gradient threshold, and temperature difference threshold, and by combining the pixel-level semantic segmentation of convolutional neural network with the union fusion of threshold rule detection results. This invention achieves the technical effect of improving the detection coverage and segmentation accuracy of abnormal hotspots.

[0015] 2. This invention solves the technical problems of being unable to respond to abnormal states in a graded manner and lacking the ability to predict in advance by dividing the alarm level into three levels: early warning, alarm and emergency alarm, and configuring differentiated linkage control commands. It also combines long short-term memory network to predict and model the peak temperature time series. This achieves the technical effect of being able to take appropriate response measures according to the severity of the abnormality and generate early warning signals in advance before the temperature reaches the threshold. Attached Figure Description

[0016] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0017] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0018] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0019] Example 1

[0020] Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0021] In this embodiment, it includes: Step 1: Acquire infrared thermal imaging data of the reactor surface

[0022] The surface and key parts of the resin reactor are scanned by an infrared thermal imager at a preset acquisition cycle to obtain a sequence of infrared thermal image frames containing information on the spatial distribution of temperature on the reactor surface. Each pixel value in each frame of the infrared thermal image corresponds to the radiation temperature value at the corresponding location on the reactor surface, thus forming a temperature distribution matrix on the reactor surface.

[0023] It should be noted that the acquisition method for the aforementioned infrared thermal image frame sequence can be as follows: An infrared thermal imager can be mounted on a pan-tilt unit, and a data processing terminal can control the pan-tilt unit to perform periodic scans according to a preset inspection path, sequentially acquiring thermal images of each monitoring area of ​​the reactor to obtain temperature distribution data covering the entire surface of the reactor. Alternatively, multiple infrared thermal imagers can be fixedly installed at different locations around the reactor, each responsible for continuous acquisition of its corresponding area, with each imager synchronously or sequentially outputting infrared thermal image frames within its respective field of view.

[0024] In this embodiment, to accurately locate the actual spatial position of abnormal hot spots on the reactor surface in subsequent steps, a visible light image of the reactor surface is simultaneously acquired using a visible light camera. The data processing terminal registers the infrared thermal image and the visible light image based on a pre-defined spatial mapping relationship between the infrared thermal imager and the visible light camera, generating fused image data. The fused image data simultaneously contains temperature distribution information and structural texture information of the reactor surface, enabling subsequently detected abnormal hot spots to be mapped to specific structural parts of the reactor.

[0025] Furthermore, the aforementioned pre-calibrated spatial mapping relationship is obtained through a dual-calibration method: a calibration plate is placed on the surface of the reactor, and images of the calibration plate are simultaneously acquired by an infrared thermal imager and a visible light camera. The pixel coordinates of the corner points of the calibration plate in the two images are extracted respectively. The homography matrix between the two images is solved using the corresponding point pairs, and the homography matrix is ​​stored as a spatial mapping relationship in the data processing terminal. During registration, the homography matrix is ​​applied to the perspective transformation of each pixel coordinate in the infrared thermal image to obtain its corresponding position in the visible light image coordinate system, thereby aligning and superimposing the infrared temperature distribution information with the visible light structure and texture information to generate fused image data.

[0026] Step 2: Preprocess the infrared thermal image to generate corrected temperature distribution data.

[0027] The acquired infrared thermal image frame sequence is subjected to radiometric correction and noise filtering to generate corrected temperature distribution data. Radiometric correction is used to eliminate the influence of environmental radiation reflection, infrared window transmittance attenuation, and other factors on the temperature measurement results, converting pixel radiance values ​​into the true temperature values ​​of the reactor surface. Noise filtering is used to suppress random interference introduced by sensor noise or environmental disturbances in the thermal image, thereby improving the signal-to-noise ratio of the temperature distribution data.

[0028] It should be noted that the above-mentioned radiation correction processing can be based on the preset emissivity parameters of the reactor surface material and the current ambient temperature parameters to compensate and convert the original radiation values ​​of each pixel in the infrared thermal image, thereby obtaining the corrected temperature value. The above-mentioned noise filtering processing can be performed on the infrared thermal image in the spatial domain using median filtering or Gaussian filtering, or it can be performed on multiple consecutive frames of thermal images in the temporal domain using inter-frame averaging to suppress transient noise.

[0029] Step 3: Detect abnormal hotspots based on the corrected temperature distribution data and extract abnormal hotspot features.

[0030] The corrected temperature distribution data is analyzed pixel-by-pixel or region-by-region to detect hotspot regions that meet the anomaly judgment criteria. Feature parameters are extracted from each detected abnormal hotspot region to generate an abnormal hotspot feature set.

[0031] The above-mentioned abnormality determination conditions include: the temperature at a single point exceeds a preset absolute temperature threshold. The temperature gradient between adjacent regions exceeds a preset gradient threshold. ; and the temperature difference between a local area and the global average temperature exceeds a preset temperature difference threshold. .in, The upper limit of temperature is a pre-set value based on the process safety requirements of the reactor, in °C. This represents the maximum allowable rate of temperature change between adjacent pixels, expressed in °C / pixel. This represents the maximum permissible deviation of the local temperature from the global average temperature of the current frame, expressed in °C. When any region in the temperature distribution data satisfies any of the above conditions, that region is marked as an abnormal hotspot.

[0032] For each marked anomalous hotspot region, the following feature parameters are extracted: peak temperature of the anomalous hotspot region. (Unit: °C) Centroid coordinates of anomalous hotspot regions in infrared thermal images Pixel area of ​​abnormal hot spots The system also includes the feature vector of the abnormal hotspot, which is determined by mapping the centroid coordinates to the actual spatial coordinates on the reactor surface.

[0033] It should be noted that the above-mentioned methods for detecting abnormal hot spots can be either by comparing the corrected temperature distribution data with a preset static threshold pixel by pixel, or by using an adaptive threshold method, which dynamically calculates the threshold based on the statistical characteristics (such as mean and standard deviation) of the current frame temperature distribution data, and identifies areas that deviate from the mean by more than a preset multiple of the standard deviation as abnormal hot spots.

[0034] In this embodiment of the application, in order to evaluate the development trend of abnormal hotspots, in addition to extracting static feature parameters, the heating rate of the abnormal hotspots is calculated based on the peak temperature sequence of the same abnormal hotspot region in multiple consecutive frames of infrared thermal images. Specifically, the heating rate The calculation formula is: ; in, This represents the peak temperature value (in °C) of the abnormal hotspots in the current frame. For time window The peak temperature value (in °C) at the same spatial location in the previous frame. The duration of the preset time window (in seconds). The time corresponding to the current frame. The unit is ℃ / s. The heating rate value... Add to the feature vector of the abnormal hotspot.

[0035] Furthermore, to accurately segment and locate anomalous hotspot regions, a convolutional neural network (CNN) is used to perform semantic segmentation processing on the corrected infrared thermal image. The CNN employs an encoder-decoder structure. The input to the CNN is the temperature distribution matrix of the corrected single-frame infrared thermal image, where each pixel value represents the corrected temperature value at its corresponding location. The output of the CNN is a pixel-level segmentation mask of the same size as the input image. Each pixel in the mask has a category label indicating the region type it belongs to, including normal regions, anomalous hotspot regions, and background regions. The encoder extracts the spatial temperature features of the thermal image layer by layer, and the decoder restores the spatial resolution layer by layer and generates pixel-level classification results. The output layer is a pixel-by-pixel classification layer, outputting the probability distribution of each pixel belonging to one of the three regions. The category with the highest probability is taken as the predicted label for that pixel, thus obtaining the segmentation mask for the region type of each pixel. During training, manually labeled anomalous hotspot segmentation masks are used as supervision labels. The cross-entropy loss function is used to calculate the pixel-by-pixel classification error between the predicted mask and the labeled mask, and the Adam optimization algorithm is used to update the network parameters. The abnormal hotspot region mask output by the convolutional neural network segmentation is fused with the abnormal hotspot region detected based on the threshold rule to improve the detection recall of abnormal hotspot regions. The fused region is used as the input for subsequent feature extraction.

[0036] Furthermore, the specific implementation of the encoder-decoder structure is as follows: the encoder consists of multiple convolutional blocks connected in series. Each convolutional block contains a convolutional layer, a batch normalization layer, and a ReLU activation function. Adjacent convolutional blocks are downsampled through a max pooling layer to gradually reduce the spatial resolution of the input thermal image while increasing the number of feature channels, thereby extracting the local temperature distribution features of the thermal image at different scales. The decoder consists of multiple upsampling blocks symmetrical to the encoder. Each upsampling block restores the spatial resolution of the feature map to the previous scale through bilinear interpolation, and is then concatenated with the feature map of the corresponding layer of the encoder by channel dimension before being fused by a convolutional layer to preserve the detailed spatial information during the encoding process. The final output layer is a pixel-wise convolutional layer with the number of channels equal to the number of categories (normal region, abnormal hotspot region, and background region), outputting the predicted probability distribution of the three regions at each pixel position.

[0037] Step 4: Based on the abnormal hotspot feature set, perform early warning discrimination and generate hierarchical alarm signals.

[0038] The feature parameters of each abnormal hotspot in the abnormal hotspot feature set are matched with the preset hierarchical alarm rules to determine the alarm level corresponding to each abnormal hotspot and generate hierarchical alarm signals.

[0039] The above-mentioned tiered alarm rules divide alarm levels into three levels: warning level, alarm level, and emergency alarm level. The criterion for determining the warning level is the peak temperature of the abnormal hotspot. Within the first temperature range Internal, or heating rate In the first rate range Within this range, it indicates that the hotspot is in the early stages of abnormal development. Among them, This is the lower limit of the first temperature range. This is the upper limit value of the first temperature range. This is the lower bound of the first speed range. This is the upper limit of the first rate range. The judgment condition corresponding to the alarm level is: the peak temperature of the abnormal hotspot. The second temperature range is higher than the first temperature range. Internal, or heating rate The second rate range is higher than the first rate range. Inside, or abnormal hotspot area Exceeding the preset area threshold This indicates that the anomaly of the hotspot has reached a level requiring intervention. Among them, This is the upper limit value of the second temperature range. This is the upper limit of the second rate range. This is a preset area threshold. The criterion for determining the emergency alarm level is the peak temperature of the abnormal hotspot. Exceeding the upper limit of the second temperature range Or the alarm level is maintained for a longer than preset duration. The temperature did not decrease afterward, indicating that the hotspot posed an immediate safety risk. Among other things, The duration for which the preset alarm level is maintained.

[0040] Furthermore, the method for determining the alarm level maintenance time in the above-mentioned graded alarm rules is as follows: the timer starts from the moment when the abnormal hotspot first meets the alarm level determination conditions, and in the detection results of each subsequent frame, it is continuously verified whether the abnormal hotspot still meets the alarm level determination conditions. If the cumulative duration of continuously meeting the alarm level determination conditions exceeds the preset duration... And the peak temperature of the abnormal hotspot during the period If no continuous downward trend is observed, it is determined that the alarm level maintenance time has exceeded [a certain limit]. This triggers the emergency alarm level determination criteria. Among them, The unit is s, and its specific value is pre-configured according to the process formula and reaction stage performed by the reactor.

[0041] It should be noted that the boundary values ​​of each temperature range, the boundary values ​​of each rate range, and the area threshold in the above-mentioned graded alarm rules are... and duration The specific values ​​can be pre-configured according to the different process formulations and reaction stages performed by the resin reactor. For example, during the peak stage of the exothermic polymerization reaction, the threshold values ​​can be set relatively loosely to avoid false alarms caused by normal exothermic reactions; while during the cooling and heat preservation stage after the reaction, the threshold values ​​can be set more strictly to capture minute temperature changes that may indicate equipment abnormalities.

[0042] In this embodiment, to reduce false alarms caused by environmental disturbances or transient operating condition changes, the persistence of abnormal hotspots is verified before matching the characteristic parameters of abnormal hotspots with the graded alarm rules. Specifically, when a certain area is detected as an abnormal hotspot in infrared thermal images for a consecutive preset number of frames, and the peak temperature of the abnormal hotspot reaches a certain value, the verification is performed. Only when the trend of change does not show a decay or decline characteristic will the abnormal hotspot be judged as a valid abnormal hotspot and enter the hierarchical alarm discrimination process. For suspected abnormal hotspots that only appear briefly in a single frame or a few frames, the suspected abnormal hotspots are marked as transient interference and recorded, but no alarm is triggered.

[0043] In this embodiment, to further improve the accuracy of anomaly detection, historical operating data is introduced for comparative analysis in addition to rule-based hierarchical alarm detection. The data processing terminal acquires the historical temperature distribution baseline data corresponding to the process formula being executed by the current reactor. The historical temperature distribution baseline data is the temperature distribution statistics of normally operating batches under the same process formula and the same reaction stage. The deviation between the temperature distribution of each monitoring area of ​​the current batch and the historical temperature distribution baseline data is calculated. When the deviation exceeds the preset allowable deviation range, even if the current temperature value has not yet reached the absolute temperature threshold, an anomaly is detected. It also generates early warning alarm signals to enable early prediction of potential anomalies.

[0044] Furthermore, to predict the future temperature trends of anomalous hotspots and generate early warning signals before their temperatures reach alarm thresholds, a Long Short-Term Memory (LSTM) network is used to model and predict the peak temperature time series of the anomalous hotspots. Before inputting the peak temperature series and the heating rate series into the LSTM network, both are Z-score normalized to eliminate the impact of the dimensional difference between temperature (in °C) and heating rate (in °C / s) on network training. The input to the LSTM network is the peak temperature series of the anomalous hotspots organized in time steps, with the normalized peak temperature value at each time step as the input. and the corresponding heating rate value The resulting two-dimensional feature vector; the output layer is a linear fully connected layer that maps the hidden state vector output by the last time step of the Long Short-Term Memory network to a sequence of predicted peak temperatures within a preset time step. The number of output nodes equals the preset number of prediction time steps. The output predicted values ​​are numerical values ​​in the standardized domain, which are then de-standardized to restore the actual temperature values ​​(in °C) before being compared with the warning level temperature threshold. A comparison was made. During training, peak temperature time series of the same monitoring area from historical batches were used as training samples, mean squared error was used as the loss function, and the Adam optimization algorithm was used to update the network parameters. When the predicted temperature series output by the Long Short-Term Memory network contains temperatures exceeding the warning level threshold... When the predicted value is obtained, an early warning level alarm signal is generated in advance and the subsequent graded alarm discrimination process is initiated.

[0045] Furthermore, the specific implementation of the Long Short-Term Memory (LSTM) network for modeling and predicting peak temperature time series is as follows: The LSM network consists of an input gate, a forget gate, an output gate, and unit states. After receiving the two-dimensional feature vector at the current time step, the forget gate determines the proportion of historical information to be discarded in the unit state based on the combination of the current input and the hidden state of the previous time step. The input gate determines the amount of new information to be written into the unit state in the current input. The unit state is updated after the forgetting and writing operations. The output gate generates the hidden state vector of the current time step based on the updated unit state. After traversing all time steps of the input sequence, the hidden state vector output at the last time step is taken and input into a fully connected linear layer. The fully connected linear layer maps the hidden state vector into an output vector with a length equal to the number of prediction time steps through matrix multiplication. The output vector is the peak temperature prediction value sequence in the normalized domain.

[0046] Step 5: Output alarm information based on the graded alarm signals and send control commands to the linkage control system.

[0047] Based on the generated graded alarm signals, execute alarm output actions and linkage control actions corresponding to the alarm level.

[0048] For alarm signals at the warning level, the data processing terminal will display a warning prompt window on the monitoring interface, showing the location of the abnormal hotspot. Peak temperature and heating rate The system also sends alert messages to the mobile devices of on-duty personnel.

[0049] For alarm signals at the alarm level, the data processing terminal, in addition to executing the alarm output action at the early warning level, triggers the on-site audible and visual alarm device to perform an audible and visual alarm, and sends control commands to the distributed control system or programmable logic controller system to reduce the feeding rate or open the auxiliary cooling circuit.

[0050] For alarm signals at the emergency alarm level, the data processing terminal, in addition to executing the alarm output action, sends control commands to the distributed control system or programmable logic controller system to suspend feeding and execute an emergency cooling procedure. If the abnormal hot spot temperature does not show a decreasing trend after the emergency alarm has lasted for a preset time, an emergency stop command is further sent to the distributed control system or programmable logic controller system.

[0051] It should be noted that the specific content of the control commands sent to the distributed control system or programmable logic controller system can be configured differently based on the structural location label of the abnormal hotspot. For example, when the abnormal hotspot is located in the jacket cooling water pipe interface area, the linkage control command is to increase the cooling water flow in that area; when the abnormal hotspot is located in the agitator shaft seal area, the linkage control command is to reduce the agitator speed and activate the shaft seal cooling.

[0052] In this embodiment, to support post-event traceability and process optimization analysis, while executing the alarm output and linkage control, the following data is also synchronously written to the data management and storage system: the infrared thermal image frame at the alarm trigger time, the corresponding visible light image frame, abnormal hotspot characteristic parameters, alarm level, the content of the sent linkage control command, and the timestamp. The data management and storage system indexes and stores the above data according to batches and time series for subsequent historical playback queries and temperature trend curve analysis.

[0053] In this embodiment, based on step 5, periodic statistical analysis is performed on the historically stored alarm records and temperature trend data to generate temperature trend reports and alarm frequency statistics reports for each monitoring area. By analyzing the distribution characteristics of alarm frequency in the time and spatial dimensions, process engineers are assisted in identifying weak areas and high-risk operating conditions on the reactor surface, thereby providing data support for adjusting process parameters and developing equipment maintenance plans.

[0054] Example 2

[0055] This embodiment uses an infrared thermal imager to collect non-contact temperature distribution data on the surface of the resin reactor, obtaining temperature distribution data covering the entire surface of the reactor. This overcomes the limitations of traditional contact thermocouples, which have limited measuring points and insufficient spatial coverage, enabling the acquisition of complete spatial temperature distribution information on the reactor surface. Furthermore, by performing radiometric correction and noise filtering preprocessing on the infrared thermal images, the accuracy of the temperature measurement data is improved, providing a reliable data foundation for subsequent abnormal hotspot detection.

[0056] During the abnormal hotspot detection phase, by setting absolute temperature thresholds... Temperature gradient threshold and temperature difference threshold Among the multi-dimensional abnormal determination conditions, abnormal hot spots caused by different reasons such as local uneven heat dissipation, lining damage or cooling failure can be detected, rather than relying only on a single temperature threshold for judgment. At the same time, by extracting the peak temperature , position , area and heating rate and other multi-dimensional characteristic parameters, sufficient criterion information is provided for subsequent classification alarm discrimination. Further, a convolutional neural network is used to perform pixel-level semantic segmentation on the infrared thermal image and fuse it with the detection result of the threshold rule, which improves the detection accuracy and recall rate of the abnormal hot spot area, especially having stronger segmentation ability for hot spot areas with blurred boundaries or irregular shapes.

[0057] In the early warning discrimination stage, by dividing the alarm level into three levels: early warning, alarm and emergency alarm, and setting corresponding multi-dimensional determination conditions for each level, abnormal states with different severity levels can be distinguished and processed. Further, a long short-term memory network is used to predict and model the time series of the peak temperature of the abnormal hot spot , which can identify the rising trend in advance and generate an early warning signal before the temperature reaches the alarm threshold, improving the forward-looking nature of the alarm. Correspondingly, in the alarm output and linkage control stage, different alarm levels trigger differentiated control instructions, from prompting an early warning to reducing the feed rate until an emergency stop, forming a gradually progressive response strategy. Therefore, this implementation method can take corresponding response measures according to the actual severity of the abnormal hot spot, avoiding over-intervention or insufficient response that may be caused by adopting a unified response method for all abnormal states.

[0058] In addition, through the verification of the persistence of the abnormal hot spot and the deviation analysis based on the historical temperature distribution baseline data, while reducing false alarms caused by transient interference, potential abnormalities that have not reached the absolute temperature threshold but have deviated from the normal working condition can be predicted in advance, thereby improving the reliability and timeliness of the monitoring alarm.

[0059] The No. 5 resin reactor (R-505) in the polyester resin workshop of a chemical plant is currently undergoing an unsaturated polyester resin condensation reaction process (batch number B-20XX0317-05). The reactor has a rated volume of 8 cubic meters, a jacketed cooling water system on its outer wall, a stirring shaft seal at the top, and inlet / outlet flanges and cooling water pipe interfaces on the side of the reactor. The reaction is currently in the peak stage of exothermic condensation, with an internal temperature of approximately 195℃, and the jacketed cooling water is circulating normally. The monitoring system consists of one infrared thermal imager (mounted on a pan-tilt unit, resolution 384×288 pixels) and one visible light camera. The data processing terminal is connected to the workshop's DCS via an industrial Ethernet connection, with a data acquisition cycle set to 5 seconds per frame. During a certain monitoring period, when the system scanned the surface of the reactor, it detected a suspected abnormal hot spot in the cooling water pipe interface area, triggering a complete detection-early warning-linkage process.

[0060] With a data acquisition cycle of 5 seconds per frame, the infrared thermal imager continuously scans the surface of the R-505 vessel, outputting a sequence of infrared thermal image frames with a resolution of 384×288 pixels. Simultaneously, a visible light camera acquires images of the vessel surface. The data processing terminal calls a pre-calibrated homography matrix H to perform perspective transformation on the pixel coordinates of each frame of the infrared thermal image, obtaining their corresponding positions in the visible light image coordinate system. This achieves pixel-by-pixel alignment of the infrared temperature distribution with the visible light structural texture, generating fused image data. During calibration, a 9×6 corner calibration plate is placed on the vessel surface, and both images are acquired simultaneously, extracting the corner pixel coordinates. The resulting homography matrix is ​​stored in the terminal configuration file.

[0061] In the fused image data, the infrared channel carries the radiation temperature value of each pixel, while the visible light channel provides the structural texture of the vessel body (such as the outline of the cooling water pipe interface and the flange edge). The superposition of the two allows subsequent detected abnormal hot spots to be accurately mapped to specific structural parts of the vessel body. Taking frame 47 (corresponding to 14:23:15 on March 17, 20XX) as an example, the cooling water pipe interface area in this frame of the fused image is clearly identifiable in the visible light texture layer, providing a structural reference for subsequent localization.

[0062] Table 1 Basic Information of Infrared Thermal Image Frames and Fuded Images

[0063] After acquiring the 47th frame of the infrared thermal image, the data processing terminal sequentially performs radiometric correction and noise filtering. Radiometric correction compensates for the original radiometric values ​​of each pixel based on the material of the outer wall of the vessel (carbon steel with paint, emissivity ε=0.92) and the current ambient temperature (38℃, read in real-time by a field temperature sensor). Taking a pixel in the cooling water pipe interface area as an example, the apparent temperature corresponding to the original radiometric value is 152.3℃. After compensation for emissivity and ambient radiation reflection, the corrected temperature calculation formula is: ; Substituting the specific values ​​(radiation temperature calculated in Kelvin and then converted back to Celsius), the corrected temperature of this pixel is 148.7℃, which is a correction of 3.6℃ compared to the apparent temperature.

[0064] Noise removal employed a 3×3 median filter to process single-frame images in the spatial domain, while inter-frame averaging was performed on five consecutive frames (frames 43 to 47) to suppress transient interference introduced by vessel vibration and sensor noise. After processing, the global average temperature was 112.4℃, and the local average temperature in the cooling water pipe interface area was 146.2℃. The signal-to-noise ratio was significantly improved, and the temperature distribution data became smoother.

[0065] Table 2 Comparison of radiation correction and noise filtering before and after frame 47

[0066] Based on the corrected temperature distribution data, the data processing terminal performs multi-dimensional anomaly detection on frame 47. The preset detection thresholds for the current process stage (condensation exothermic peak) are: absolute temperature threshold 145℃, gradient threshold 8℃ / pixel, and temperature difference threshold 30℃.

[0067] Verify each item in the cooling water pipe interface area (the connected region with pixel coordinates (271, 184) as its centroid): The peak temperature of 157.3℃ is greater than the absolute temperature threshold of 145℃, thus satisfying the absolute temperature threshold condition. The temperature gradient between this region and the adjacent normal jacket region is 11.2℃ / pixel, which is greater than the gradient threshold of 8℃ / pixel, thus satisfying the gradient threshold condition. The difference between the local peak temperature and the global average temperature is: ; If all three conditions are met, the area will be marked as an abnormal hotspot.

[0068] A convolutional neural network (encoder-decoder structure, with 4 convolutional blocks in the encoder and 4 upsampling blocks in the decoder) performs pixel-level semantic segmentation on the corrected temperature matrix of frame 47, outputting three types of masks (normal region, abnormal hotspot region, and background region). The cooling water pipe interface region is marked as an abnormal hotspot region in the CNN output mask, with a pixel area of ​​312 pixels. The CNN segmentation result is fused with the threshold rule detection result, and the resulting abnormal hotspot region has a pixel area of ​​318 pixels (including some blurred boundary regions identified by the CNN).

[0069] The heating rate was calculated based on the peak temperature sequence of the same spatial location across 5 consecutive frames (frames 43 to 47, with a time window of 20 seconds): ; After mapping the centroid pixel coordinates (271, 184) to the actual spatial coordinates of the vessel body using a homography matrix, the corresponding structural part label is determined as "cooling water pipe interface area (southwest side, 1.2 meters from the bottom of the vessel)".

[0070] Table 3. Set of characteristic parameters for abnormal hotspots (Frame 47)

[0071] In terms of continuous verification, the system checked five consecutive frames from frame 43 to 47. This area was detected as an abnormal hot spot in each frame, and the peak temperature sequence (148.1→150.2→152.7→155.1→157.3℃) showed a continuous upward trend without any decay or decline characteristics. It was determined to be a valid abnormal hot spot and entered the graded alarm discrimination process.

[0072] Regarding historical baseline comparison, the data processing terminal retrieved historical normal batch temperature baseline data for the same process formulation (unsaturated polyester resin condensation) and the same reaction stage (peak exothermic period). The historical baseline average for the cooling water pipe interface area was 121.3℃, and the current deviation is: ; If the temperature exceeds the preset allowable deviation range (±20℃), the warning condition has been triggered even without relying on the absolute threshold.

[0073] For LSTM prediction, the peak temperature sequence and heating rate sequence of frames 43 to 47 are used as input. After Z-score normalization (mean 152.7℃, standard deviation 3.5℃), they are input into the LSTM network and output the predicted temperature sequence for the next 6 time steps (30 seconds). After inverse normalization, the predicted values ​​are: 159.6, 161.9, 164.2, 166.4, 168.1, and 169.7℃. The temperature exceeds the warning level threshold of 155℃ in the first prediction step, and a warning signal is generated in advance.

[0074] The preset graded alarm thresholds for the current process stage are: first temperature range [155℃, 165℃), second temperature range [165℃, 175℃), first rate range [0.3℃ / s, 0.6℃ / s), second rate range [0.6℃ / s, 1.0℃ / s), area threshold 500 pixels, and duration 120 seconds.

[0075] The feature parameters are matched with the classification rules: the peak temperature of 157.3℃ belongs to the first temperature range [155℃, 165℃), meeting the temperature condition for the warning level; the heating rate of 0.46℃ / s belongs to the first rate range [0.3℃ / s, 0.6℃ / s), meeting the rate condition for the warning level; the hotspot area of ​​318 pixels is less than the area threshold of 500 pixels, failing to meet the area condition. Based on the comprehensive judgment: both the temperature and rate conditions meet the warning level, and a warning level alarm signal is generated.

[0076] Table 4. Matching results of graded alarm thresholds and current characteristic parameters

[0077] The data processing terminal performs the following actions based on the warning level: At the interface layer, the monitoring interface pops up an early warning window, showing the location of the abnormal hot spot (centroid coordinates (271, 184), corresponding structural part: southwest side of the cooling water pipe interface, 1.2 meters from the bottom of the vessel), peak temperature 157.3℃, heating rate 0.46℃ / s, and pushes an early warning message to the on-duty engineer's mobile terminal (employee number OP-07).

[0078] At the linkage control layer, the on-site audible and visual alarm device (R-505 reactor side audible and visual column) is triggered to start the audible and visual alarm, and at the same time, a differentiated control command is sent to the DCS: since the abnormal hot spot is located in the "cooling water pipe interface area", the linkage control command is "increase the flow rate of the southwest cooling water pipe, and increase the target flow rate from the current 8.5m³ / h to 14.0m³ / h", instead of the general command to reduce the feed rate.

[0079] At the data storage layer, the system synchronously writes the following data into the data management storage system: the 47th frame infrared thermal image, the corresponding visible light image, the set of abnormal hot spot characteristic parameters, the alarm level (early warning level), the content of the sent DCS command (increase cooling water flow) and the timestamp (March 17, 20XX, 14:23:15).

[0080] Table 5 Alarm Output and Linkage Control Execution Record

[0081] The entire data flow process demonstrates a clear logical progression between steps: Step 1, the raw infrared thermal image frame sequence (384×288 pixels, including radiance), undergoes radiometric correction (emissivity 0.92, ambient temperature 38℃) and median filtering and inter-frame averaging in Step 2, transforming it into corrected temperature distribution data with a higher signal-to-noise ratio. The peak temperature in the cooling water pipe interface area is corrected from the apparent value of 152.3℃ to the corrected value of 148.7℃. Step 3, based on the corrected data, accurately marks this area as an abnormal hotspot through the union fusion of three-dimensional threshold rules and CNN semantic segmentation, improving... The peak temperature of 157.3℃, hotspot area of ​​318 pixels, and heating rate of 0.46℃ / s are extracted to form a complete feature vector. Step 4 matches the feature vector with the graded alarm rules, combining LSTM prediction (predicting a maximum temperature of 169.7℃) and historical baseline deviation (+36℃) for dual verification. Both the peak temperature and heating rate fall within the warning level range, ultimately determining it to be at the warning level. Step 5 outputs differentiated DCS linkage commands (increasing cooling water flow) based on the warning level and structural location label (cooling water pipe interface area), and archives all process data. The output data from each step serves as direct input for the next step, ensuring data consistency and traceability throughout the entire data flow from the original radiation signal to the final control command.

[0082] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0083] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0084] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0085] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A resin reactor infrared thermal imaging abnormal hot spot monitoring alarm system, characterized in that, include: The infrared thermal image acquisition module is used to scan the surface of the resin reactor with an infrared thermal imager at a preset acquisition cycle to obtain a sequence of infrared thermal image frames containing information on the spatial distribution of temperature on the surface of the reactor. The image preprocessing module is used to perform radiometric correction and noise filtering on the infrared thermal image frame sequence to generate corrected temperature distribution data; The abnormal hotspot detection and feature extraction module is used to detect abnormal hotspot regions based on the corrected temperature distribution data, using multi-dimensional anomaly judgment conditions including absolute temperature threshold, temperature gradient threshold and temperature difference threshold, and extracting feature parameters including peak temperature, centroid coordinate position, pixel area and heating rate for each detected abnormal hotspot region to generate an abnormal hotspot feature set. The graded alarm discrimination module is used to match the feature parameters of each abnormal hotspot in the abnormal hotspot feature set with the preset graded alarm rules, determine the alarm level corresponding to each abnormal hotspot, including the warning level, alarm level and emergency alarm level, and generate graded alarm signals. The alarm output and linkage control module is used to execute alarm output actions corresponding to the alarm level based on the graded alarm signals, and to send control commands corresponding to the alarm level to the linkage control system.

2. The infrared thermal imaging abnormal hot spot monitoring and alarm system for a resin reactor according to claim 1, characterized in that, While acquiring the infrared thermal image frame sequence, a visible light image of the reactor surface is also acquired simultaneously by a visible light camera; based on the pre-calibrated spatial mapping relationship between the infrared thermal imager and the visible light camera, the infrared thermal image and the visible light image are registered to generate fused image data, which simultaneously contains temperature distribution information and structural texture information of the reactor surface; The spatial mapping relationship is obtained through a dual-target calibration method: a calibration plate is placed on the surface of the reactor, and images of the calibration plate are simultaneously acquired by an infrared thermal imager and a visible light camera. The pixel coordinates of the corner points of the calibration plate in the two images are extracted respectively, and the homography matrix between the two images is solved using the corresponding point pairs. During registration, the homography matrix is ​​applied to each pixel coordinate in the infrared thermal image to perform perspective transformation, and its corresponding position in the visible light image coordinate system is obtained. The infrared temperature distribution information and the visible light structure and texture information are aligned and superimposed to generate the fused image data.

3. The infrared thermal imaging abnormal hot spot monitoring and alarm system for a resin reactor according to claim 1, characterized in that, The radiation correction process is as follows: based on the preset emissivity parameters of the reactor surface material and the current ambient temperature parameters, the original radiation values ​​of each pixel in the infrared thermal image are compensated and converted to obtain the corrected temperature values. The noise filtering process is as follows: median filtering or Gaussian filtering is used to process the infrared thermal image in the spatial domain, or inter-frame averaging is performed on multiple consecutive thermal images in the time domain.

4. The infrared thermal imaging abnormal hot spot monitoring and alarm system for a resin reactor according to claim 1, characterized in that, The multidimensional anomaly detection criteria include: a single point temperature exceeding a preset absolute temperature threshold; a temperature gradient between adjacent regions exceeding a preset gradient threshold; a temperature difference between a local region temperature and the current frame's global average temperature exceeding a preset temperature difference threshold; and when there is a region in the temperature distribution data that meets any one of these criteria, that region is marked as an abnormal hotspot region. The heating rate is calculated as follows: the difference between the peak temperature value of the abnormal hot spot in the current frame and the peak temperature value of the same spatial location in the corresponding frame before the preset time window is divided by the duration of the preset time window to obtain the heating rate value, and the heating rate value is added to the feature parameters of the abnormal hot spot.

5. The infrared thermal imaging abnormal hot spot monitoring and alarm system for a resin reactor according to claim 1, characterized in that, When detecting the abnormal hotspot areas, a convolutional neural network is also used to perform semantic segmentation processing on the corrected infrared thermal image; The convolutional neural network employs an encoder-decoder structure. The input is the corrected temperature distribution matrix of a single-frame infrared thermal image, and the output is a pixel-level segmentation mask of the same size as the input image. Each pixel in the segmentation mask is labeled with a category label including normal regions, abnormal hotspot regions, and background regions. The encoder consists of multiple concatenated convolutional blocks, each containing a convolutional layer, a batch normalization layer, and a ReLU activation function. Adjacent convolutional blocks are downsampled using a max-pooling layer. The decoder consists of multiple upsampling blocks symmetrical to the encoder. Each upsampling block restores spatial resolution through bilinear interpolation and is then concatenated with the feature map of the corresponding layer of the encoder by channel dimension before being fused through a convolutional layer. The output layer is a pixel-wise convolutional layer with the number of channels equal to the number of categories. During training, manually labeled abnormal hotspot segmentation masks are used as supervision labels, and the cross-entropy loss function is used for optimization. The abnormal hotspot region mask output by the convolutional neural network segmentation is fused with the abnormal hotspot region detected based on the multidimensional anomaly determination conditions, and the fused region is used as the input for feature extraction.

6. The infrared thermal imaging abnormal hot spot monitoring and alarm system for a resin reactor according to claim 1, characterized in that, The criteria for determining the tiered alarm rules are as follows: The criteria for determining the warning level are: the peak temperature of the abnormal hot spot is within the first temperature range, or the heating rate is within the first rate range. The alarm level is determined by the following conditions: the peak temperature of the abnormal hot spot is in the second temperature range that is higher than the first temperature range, or the heating rate is in the second rate range that is higher than the first rate range, or the area of ​​the abnormal hot spot exceeds the preset area threshold. The criteria for determining the emergency alarm level are: the peak temperature of the abnormal hot spot exceeds the upper limit of the second temperature range, or the peak temperature does not decrease after the alarm level is maintained for a preset duration. The alarm level maintenance time is determined as follows: the timer starts from the moment when the abnormal hot spot first meets the alarm level determination conditions, and in each subsequent frame of detection results, it is continuously verified whether the alarm level determination conditions are still met. If the cumulative duration of continuous satisfaction exceeds the preset duration and the peak temperature does not show a continuous downward trend, an emergency alarm level is triggered.

7. The infrared thermal imaging abnormal hot spot monitoring and alarm system for a resin reactor according to claim 1, characterized in that, Before matching the abnormal hotspot feature parameters with the graded alarm rules, the persistence of the abnormal hotspot is also verified: when a certain area is detected as an abnormal hotspot in infrared thermal images for a consecutive preset number of frames, and the trend of the peak temperature does not show a decay and decline feature, the abnormal hotspot is determined to be a valid abnormal hotspot and enters the graded alarm discrimination process; for suspected abnormal hotspots that only appear briefly in a single frame or a few frames, they are marked as transient interference and recorded but no alarm is triggered. It also acquires the historical temperature distribution baseline data corresponding to the process formula being executed in the current reactor, calculates the deviation between the temperature distribution of each monitoring area of ​​the current batch and the historical temperature distribution baseline data, and generates an alarm signal at the warning level when the deviation exceeds the preset allowable deviation range, even if the current temperature value has not reached the absolute temperature threshold.

8. The infrared thermal imaging abnormal hot spot monitoring and alarm system for a resin reactor according to claim 1, characterized in that, Long Short-Term Memory (LSTM) networks were also used to model and predict the peak temperature time series of anomalous hotspots. Before inputting the peak temperature sequence and heating rate sequence into the Long Short-Term Memory (LSTM) network, both are Z-score normalized. The input to the LSM network is a sequence organized in time steps. The input at each time step is a two-dimensional feature vector consisting of the normalized peak temperature value and the corresponding heating rate value. The LSM network consists of an input gate, a forget gate, an output gate, and unit states. After receiving the two-dimensional feature vector at the current time step, the forget gate determines the proportion of historical information to be discarded in the unit state, and the input gate determines the amount of new information to be written into the unit state in the current input. The unit state is updated after the forgetting and writing operations. The output gate generates the hidden state vector of the current time step based on the updated unit state. The output layer is a linear fully connected layer that maps the hidden state vector output at the last time step to a sequence of predicted peak temperatures within a preset future time step. After inverse normalization, the predicted values ​​are restored to actual temperature values ​​and compared with the warning level temperature threshold. During training, the peak temperature time series of historical batches were used as training samples, and the mean squared error was used as the loss function. When the predicted temperature sequence output by the Long Short-Term Memory Network contains a predicted value that exceeds the warning level temperature threshold, a warning level alarm signal is generated in advance.

9. The infrared thermal imaging abnormal hot spot monitoring and alarm system for a resin reactor according to claim 1, characterized in that, The alarm output actions and linkage control actions corresponding to the alarm level include: For alarm signals at the warning level, a warning prompt window will pop up on the monitoring interface, displaying the location, peak temperature and heating rate of the abnormal hotspot, and pushing the warning message to the mobile terminal of the on-duty personnel. For alarm signals at the alarm level, in addition to executing the alarm output action at the early warning level, the on-site audible and visual alarm device is triggered to perform an audible and visual alarm, and a control command to reduce the feeding rate or open the auxiliary cooling circuit is sent to the distributed control system or programmable logic controller. For alarm signals at the emergency alarm level, in addition to executing the alarm output action at the alarm level, a control command to suspend feeding and execute an emergency cooling program is sent to the distributed control system or programmable logic controller. If the temperature of the abnormal hot spot does not show a downward trend after the emergency alarm has lasted for a preset time, an emergency stop command is further sent. While executing the alarm output action and linkage control action, the infrared thermal image frame, visible light image frame, abnormal hot spot characteristic parameters, alarm level, sent linkage control command content and timestamp at the time of alarm triggering are synchronously written into the data management storage system and indexed and stored according to batch and time sequence.

Images