Grate cooler abnormal working condition diagnosis method and system based on AI image recognition

Abstract

The invention discloses a grate cooler abnormal working condition diagnosis method and system based on AI image recognition, and relates to the technical field of intelligent diagnosis of cement production equipment. Carrying out frame cutting every 100ms, carrying out 5 * 5 Gaussian filtering denoising, and correcting a gray value by PT100 temperature data; training a convolutional neural network model by 50000 + frame samples; the preprocessed image is input into the model, features are extracted and compared with a threshold value, and an identification result and 0-1 confidence coefficient are output; a result is pushed to an APC system through the Ethernet, and a DCS is reversely controlled according to a Modbus protocol to trigger a sound-light alarm; a new data iteration model is used every 7 days, a temperature compensation mechanism is set, and stable operation of equipment is guaranteed. The grate cooler abnormity diagnosis method has the advantages that the grate cooler abnormity diagnosis precision, timeliness, adaptability and management and control capability are obviously broken through, the adaptability is enhanced through closed-loop optimization and high-temperature-resistant design, and the production benefit and the intelligent level are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent diagnostic technology for cement production equipment, and in particular to a method and system for diagnosing abnormal operating conditions of grate coolers based on AI image recognition. Background Technology

[0002] The grate cooler is the core equipment of the new dry-process cement clinker calcination system, undertaking key functions such as clinker cooling, system heat recovery, material conveying, and crushing. Its operating status directly determines the efficiency, energy consumption, and clinker quality of cement production. During the cement production process, the grate cooler is prone to abnormal operating conditions such as clinker flow fluctuations, air temperature and pressure imbalances, and uneven ventilation of the grate plates, resulting in a "snowman-like" accumulation. In this condition, clinker accumulates and clumps in specific areas of the grate cooler, forming irregular accumulations resembling "snowmen." This anomaly directly blocks the grate ventilation channel, preventing the clinker from cooling in time. This manifests as a drop in secondary air temperature and a sudden rise in tertiary air temperature, which in turn disrupts the ventilation and heat exchange balance within the kiln, worsening the calcination conditions. The clinker temperature exiting the grate cooler rises to over 150°C, not only reducing quality indicators such as clinker strength but also exacerbating the erosion of components such as the grate and kiln inlet guards. In severe cases, it can cause malfunctions such as positive pressure at the kiln head and material leakage at the kiln tail, forcing the production line to reduce its load or even shut down. The economic loss caused by a single malfunction can reach hundreds of thousands of yuan.

[0003] Traditional diagnostic methods for abnormal operating conditions in grate coolers rely on manual observation of television images by central control operators, a method with significant drawbacks. Operators must rely on years of accumulated experience to visually identify changes in the shape and color of the clinker to determine if a "snowman" formation has occurred. However, the high-temperature environment inside the grate cooler causes image distortion and blurring of details. Furthermore, manual observation has a 0.5-2 minute recognition lag. By the time an anomaly is detected, the snowman has already reached a considerable size. Activating air cannons or mechanical cleaning devices at this point not only increases the difficulty of handling the problem but also results in prolonged thermal efficiency loss. In addition, the accuracy of manual judgment is greatly affected by the operator's condition and experience, easily leading to misjudgments or missed diagnoses. Misjudgments result in frequent activation of air cannons, wasting compressed air and impacting the grate; missed diagnoses allow the anomaly to continue to expand, triggering subsequent cascading failures. Simultaneously, operators must continuously monitor multiple devices, leading to high labor intensity and visual fatigue over long periods, further reducing diagnostic reliability.

[0004] Existing technologies for diagnosing grate cooler anomalies attempt to incorporate sensors or simple image recognition solutions, but significant shortcomings remain. Some technologies rely on temperature and pressure sensors to collect data and indirectly infer anomalies through process parameters, failing to directly identify clinker accumulation morphology and prone to misdiagnosis due to parameter cross-interference. Some simple image recognition technologies employ traditional machine learning algorithms, with sample libraries covering only a few operating conditions. They fail to consider the impact of the complex environment inside the grate cooler (80℃-150℃ high temperature, dust, vibration, etc.) on image quality, lacking temperature compensation and image enhancement design. At high temperatures, image edges become blurred and grayscale drift occurs, resulting in an accuracy rate below 85%. Furthermore, the lack of an iterative optimization mechanism after model training makes it unable to adapt to changes in clinker composition and production load. Simultaneously, many existing technologies lack deep integration with the cement production line's APC intelligent optimization control system and DCS system. Identification results require manual secondary transmission to trigger control actions, further extending response time and failing to meet the need for rapid handling of grate cooler anomalies. As the cement industry transforms towards intelligent and efficient operation, traditional diagnostic methods and existing technologies are no longer adequate to meet the high-precision, real-time diagnostic needs of large grate coolers. There is an urgent need for an intelligent diagnostic technology that balances environmental adaptability, diagnostic accuracy, and linkage control capabilities. Summary of the Invention

[0005] The present invention proposes an AI image recognition-based method and system for diagnosing abnormal operating conditions of a grate cooler, in order to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for diagnosing abnormal operating conditions of a grate cooler based on AI image recognition includes the following steps:

[0008] Image acquisition steps: Use a wide-angle camera to capture real-time images of the inside of the grate cooler from all angles, and collect video stream data of clinker morphology, distribution and internal environment. It is adapted to the internal environment of the grate cooler of 80℃-150℃ and has a continuous running time of not less than 8000 hours.

[0009] Data preprocessing steps: The acquired video stream is captured as a frame every 100ms interval, and noise is removed by a 5×5 Gaussian filter kernel. Combined with real-time temperature data collected by the PT100 temperature sensor, the grayscale values ​​of the image pixels are linearly corrected.

[0010] AI model training steps: Construct an abnormal working condition recognition model based on a convolutional neural network, input a database of 50,000+ frames of image samples, and train it using the batch gradient descent method;

[0011] Real-time diagnostic steps: Input the preprocessed image into the trained model, extract the clinker accumulation area, edge contour shape and gray-scale distribution features in the image, compare them with the snowman feature threshold preset by the model, and output the recognition result and the confidence level in the 0-1 interval.

[0012] Linkage control steps: The identification result is pushed to the APC intelligent optimization control system via Ethernet. The system writes the value back to the DCS system according to the Modbus protocol. The DCS system triggers an audible and visual alarm and completes the central control abnormality pop-up prompt within 1 second.

[0013] Closed-loop optimization steps: Collect the recognition accuracy data marked by operators through the central control interaction unit, combine it with newly collected working condition images to supplement the sample library, and iterate and update the model parameters every 7 days;

[0014] High-temperature adaptation steps: A temperature compensation mechanism is introduced to correct image features. For every 10°C increase in temperature, the image edge detection threshold is reduced by 5%. At the same time, the fan speed of the camera heat dissipation system is optimized. When the temperature is ≥120°C, the fan speed is increased to 3500r / min to maintain the temperature of the internal components of the camera ≤60°C.

[0015] Furthermore, it also includes an image feature enhancement step, using the following formula to calculate the overall confidence level of abnormal operating conditions: ,in The overall confidence level for abnormal operating conditions. For morphological feature weights, The weights are the grayscale distribution feature weights. Associating process parameters with weights, For morphological feature matching degree, The grayscale distribution feature matching degree, The correlation between image data and process parameters is determined; when the confidence level is ≥0.8, it is judged as a high-confidence anomaly, directly triggering the DCS alarm and operation suggestion push in the linkage control step; when the confidence level is 0.6-0.8, a manual review pop-up window is triggered on the central control display screen, and the operator needs to complete the review and confirmation within 30 seconds; when the confidence level is <0.6, it is judged as a normal working condition, and the system records the current image data to the sample library.

[0016] Furthermore, it also includes an abnormal operating condition classification process, with the classification criteria calculated using the following formula: ,in This represents the severity value of the abnormality. The overall confidence level for abnormal operating conditions. This is the historical correlation coefficient. This refers to the frequency of similar abnormalities occurring within the past 30 days; based on... Values ​​are differentiated. When the value is <0.8, only the DCS system's audible and visual alarm is triggered; production parameters are not affected. When the value is ≤0.8, the DCS system's audible and visual alarm is triggered. <1.5 hours triggers the activation of the air cannon device; press and hold during air cannon firing. Value linear adjustment, Spray for 5 seconds when =0.8. Spray for 15 seconds when =1.5. When the value is ≥1.5, an emergency operation is triggered to reduce the kiln feed rate by 10%, and the abnormal handling process is simultaneously pushed to the central control display screen.

[0017] Furthermore, the AI ​​model training process employs a transfer learning strategy, fine-tuning the pre-trained ResNet50 model by freezing the parameters of the first 10 convolutional layers and only fine-tuning the parameters of the last 15 layers. The fully connected layers use the ReLU activation function, and the output layer uses the Softmax function to achieve work condition classification. During model training, five-fold cross-validation is used, dividing the sample library into training and validation sets in an 8:2 ratio. After each round of training, the accuracy, recall, and F1 score of the validation set are calculated. Training stops when the accuracy of the validation set is ≥95% and the F1 score is ≥0.94 for 10 consecutive rounds. The trained model is stored in ONNX format.

[0018] Furthermore, in the data preprocessing step, temperature correction adopts a linear correction algorithm. The correction formula refers to the fitting relationship between real-time temperature and image grayscale value. For every 10°C increase in temperature, the image brightness gain is linearly increased by 5%. Image frame capture adopts a hardware triggering method, and the frame capture interval is controlled by the camera's built-in timer. The frame capture time deviation is ≤1ms.

[0019] Furthermore, a model performance evaluation system is established in the closed-loop optimization step, using three core indicators: accuracy, recall, and F1 score. Accuracy = (Number of correctly identified anomalies / Total number of identified anomalies) × 100%; Recall = (Number of correctly identified anomalies / Actual number of anomalies) × 100%; F1 score = 2 × Accuracy × Recall / (Accuracy + Recall). When any indicator falls below the preset threshold in three consecutive evaluations, an emergency optimization process is triggered, and the sample library is expanded and the model parameters are updated within 24 hours. After the update, the performance evaluation is re-executed until all indicators recover to above the threshold.

[0020] Furthermore, in the linkage control step, after receiving the identification results, the APC intelligent optimization control system automatically extracts the secondary air temperature, tertiary air temperature, and clinker flow rate data from the past hour, and calculates operating condition adjustment suggestions through a data fusion algorithm. After the parameter suggestions are pushed to the DCS system, the system displays them in priority order, with priority based on the severity of the anomaly. Sure, The higher the value, the higher the priority. After the operator selects the execution suggestion, the system automatically records the operation time, adjusted parameters, and subsequent changes in operating conditions.

[0021] Furthermore, the system for diagnosing abnormal operating conditions of a grate cooler based on AI image recognition includes the following modules:

[0022] Image acquisition module: It consists of a wide-angle camera, a 304 stainless steel shell, an axial fan cooling system and a PT100 temperature sensor. The camera lens is made of special high-temperature resistant glass with an anti-fog coating. The temperature sensor has a measurement range of 0℃-200℃. The module transmits video stream data through an Ethernet interface.

[0023] Data preprocessing module: Equipped with an ARM Cortex-A9 embedded processor with an operating frequency of 1.2GHz, it integrates image frame capture, Gaussian filtering, adaptive histogram equalization and temperature correction algorithms, supports data interaction with the AI ​​diagnostic model module via PCIe bus, and stores preprocessed images in JPEG format;

[0024] AI diagnostic model module: It adopts NVIDIA Jetson TX2 GPU acceleration computing unit with a computing performance of 1.3 TFLOPS. It has a built-in trained convolutional neural network model and supports manual parameter updates via USB interface or automatic updates at preset cycles. It outputs abnormal working condition identification results and confidence level within 200ms after receiving preprocessed images.

[0025] Linkage control module: includes APC intelligent optimization control unit and DCS system interface unit. The APC unit adopts industrial-grade server, and the DCS interface unit supports RS485 bus and ModbusRTU protocol to realize real-time transmission of recognition results and output of audible and visual alarm signals.

[0026] Closed-loop optimization module: It consists of a 1TB SSD data storage unit and a model update unit. The data storage unit records the recognition results, operator feedback data and working condition images for the past year. The model update unit supports automatic monthly updates or manual updates via the central control unit.

[0027] High-temperature adaptation module: includes an axial fan cooling system and a temperature compensation unit. The cooling system fan speed ranges from 1500 to 3500 r / min and is equipped with a 2mm thick aluminum heat-conducting sheet. The temperature compensation unit corrects image grayscale deviations under high-temperature environments through software algorithms.

[0028] Operation interaction module: It consists of a 19-inch display screen and a feedback input unit. The display screen supports the visualization of abnormal results, operation suggestions and model performance indicators. The feedback input unit supports the recognition of the correctness of touch annotations. The annotation data is transmitted to the closed-loop optimization module in real time.

[0029] Furthermore, the camera mounting bracket of the image acquisition module is made of 304 stainless steel and is equipped with a nitrile rubber shockproof and buffer structure; the camera lens has a focal length of 8mm, a depth of field of 0.5m-5m, and can identify clinker particles with a diameter of 5mm or larger inside the grate cooler. The lens has an IP68 protection rating and a lens cleaning cycle of ≥30 days.

[0030] Furthermore, the linkage control module also includes an emergency handling unit, which uses a Siemens S7-1200 PLC as its control core; when the identification result... When the value is ≥1.5, the emergency response unit drives the electric actuator to adjust the opening of the grate cooler feed regulating valve. The valve opening is adjusted according to... The value decreases linearly. When the value is 1.5, the opening decreases by 30%. When the opening is 2.0, the opening is reduced by 50%, and a steam flow adjustment command is sent to the waste heat power generation system, reducing the steam flow by 15%. The execution time, parameter change range, and operating condition feedback data of all emergency operations are automatically stored in the SSD data storage unit of the closed-loop optimization module.

[0031] Compared with existing technologies, the beneficial effects of this invention are:

[0032] In terms of diagnostic accuracy and real-time performance, the convolutional neural network model constructed in this invention incorporates multi-dimensional image features and process parameter correlation analysis. Through training with a large-scale labeled sample library, combined with noise removal, contrast enhancement, and temperature correction in image preprocessing, it effectively solves the image distortion problem caused by high temperature inside the grate cooler. The recognition results are more in line with the actual working conditions. Compared with the experience-based reliance of traditional manual diagnosis and the single feature analysis of existing image recognition, the diagnostic accuracy is greatly improved. Moreover, it completes image frame capture and analysis every 100ms, which is 0.5-2 minutes slower than manual recognition. It can realize the instant capture of abnormal working conditions and reserve sufficient time for subsequent handling.

[0033] In terms of anomaly handling and coordinated control, the anomaly classification mechanism established in this invention is deeply integrated with the APC and DCS systems. Differentiated handling strategies can be implemented based on the severity of the anomaly: minor anomalies only trigger alarm prompts, avoiding unnecessary production intervention; moderate anomalies automatically initiate pre-processing measures to prevent escalation; and severe anomalies quickly intervene to adjust production parameters, reducing failure losses. This coordinated mode significantly improves response efficiency compared to existing technologies that require manual transmission of identification results. Simultaneously, the emergency handling unit can automatically execute key control actions without manual intervention, further shortening the handling cycle and ensuring the stability of the firing system.

[0034] Regarding continuous optimization and environmental adaptability, the closed-loop optimization mechanism of this invention can periodically update model parameters by combining operator feedback and new operating condition data, so that the diagnostic capability gradually increases with the use time, avoiding the problem of decreased adaptability caused by the solidification of existing technology models. At the same time, the high-temperature resistant design of the image acquisition module and the temperature compensation and heat dissipation optimization of the high-temperature adaptation module can stably adapt to the high temperature and vibration environment inside the grate cooler, and the equipment has high reliability for continuous operation. Compared with existing image acquisition equipment, which is prone to damage due to high temperature and image shift due to vibration, the durability and stability are greatly improved.

[0035] In terms of ease of operation and production efficiency, the visualization and feedback functions of the operation interaction module reduce the labor intensity of operators who need to constantly monitor the screen. The long-term storage of fault data facilitates subsequent traceability and analysis, providing data support for production process optimization. From a long-term operation perspective, this invention, through accurate diagnosis and timely handling, can reduce energy waste and equipment wear caused by abnormal operating conditions, improve clinker quality stability and production continuity. Compared with the frequent abnormal shutdowns and load fluctuations under traditional diagnostic methods, the production efficiency and economic benefits of enterprises are significantly improved. At the same time, it promotes the improvement of the intelligent level of cement production lines and provides a reliable technical reference for the intelligent diagnosis of similar equipment in the industry. Attached Figure Description

[0036] Figure 1 This is a schematic block diagram of the abnormal operating condition diagnosis method for grate coolers based on AI image recognition proposed in this invention;

[0037] Figure 2 This is a schematic block diagram of the abnormal operating condition diagnosis system for a grate cooler based on AI image recognition proposed in this invention;

[0038] Figure 3 A bar chart comparing the accuracy of abnormal operating condition identification at different temperatures;

[0039] Figure 4 This is a biaxial line graph showing the relationship between grate cooler operating time and the occurrence rate of snowman-building anomalies. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0042] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0043] Reference Figures 1 to 4 A method for diagnosing abnormal operating conditions of a grate cooler based on AI image recognition includes the following steps:

[0044] Image acquisition steps: A wide-angle high-definition camera is used to capture real-time images of the inside of the grate cooler from all angles. The camera is equipped with a 304 stainless steel high-temperature resistant shell and an axial fan cooling system with thermally conductive silicone pads. The shooting frame rate is 25fps and the resolution is 1920×1080. The horizontal field of view of the camera is 120° and the vertical field of view is 60°. The camera collects video stream data of clinker morphology, distribution and internal environment. It is suitable for the internal environment of the grate cooler of 80℃-150℃ and has a continuous running time of not less than 8000 hours.

[0045] Data preprocessing steps: The acquired video stream is segmented into frames every 100ms. High-frequency noise is removed using a 5×5 Gaussian filter kernel. An adaptive histogram equalization algorithm is used to enhance image contrast. Combined with real-time temperature data collected by a PT100 temperature sensor, the grayscale values ​​of the image pixels are linearly corrected. The grayscale deviation of the corrected image is ≤3%.

[0046] AI model training steps: Construct an abnormal working condition recognition model based on convolutional neural networks. The model includes an input layer, 5 convolutional layers, 3 pooling layers, 2 fully connected layers, and an output layer. The input includes a sample library of labeled images of snowmen and normal working conditions. The sample library covers more than 50,000 frames of images in the temperature range of 20℃-180℃ and the clinker flow rate of 100t / h-200t / h. The model is trained using batch gradient descent with a batch size of 64. The learning rate is dynamically adjusted in the range of 0.001-0.01 using a cosine annealing strategy. After 500 iterations, the accuracy of the validation set is ≥95%.

[0047] Real-time diagnostic steps: Input the preprocessed image into the trained model, extract the clinker accumulation area, edge contour shape features and gray-scale distribution features in the image, compare them with the snowman feature threshold preset by the model, and output the snowman abnormal working condition identification result and the confidence level in the 0-1 interval. The confidence level calculation incorporates the correlation weight between real-time temperature and clinker flow rate.

[0048] Linkage control steps: The identification result is pushed to the APC intelligent optimization control system via Ethernet at a rate of 100Mbps. The system writes the value back to the DCS system according to the Modbus protocol. The DCS system triggers an audible and visual alarm with an alarm sound intensity of 85dB and a light flashing frequency of 3Hz. Within 1 second after the alarm signal is triggered, an abnormal pop-up window is displayed on the central control screen.

[0049] Closed-loop optimization steps: Collect the recognition accuracy data marked by operators through the central control interaction unit, and supplement the sample library with newly collected working condition images. Iterate and update the model parameters every 7 days. After the update, the recognition accuracy of the model on the same test set is improved by ≥2% compared with that before the update.

[0050] High-temperature adaptation steps: A temperature compensation mechanism is introduced to correct image features. For every 10°C increase in temperature, the image edge detection threshold is reduced by 5%. At the same time, the fan speed of the camera heat dissipation system is optimized. When the temperature is ≥120°C, the fan speed is increased to 3500r / min to maintain the temperature of the internal components of the camera ≤60°C.

[0051] This invention also includes an image feature enhancement step, which improves recognition accuracy through multi-dimensional feature fusion, and calculates the comprehensive confidence level of abnormal working conditions using the following formula: in The overall confidence level for abnormal operating conditions ranges from 0 to 1. The morphological feature weight is 0.5. This represents the grayscale distribution feature weight, with a value of 0.3. The process parameter is assigned a weight and its value is 0.2. The morphological feature matching degree is defined as the overlap rate between the clinker accumulation area in the recognition image and the standard shape of a snowman. The grayscale distribution feature matching degree is defined as the similarity between the grayscale histogram of the identified image and the typical grayscale histogram of the snowman. The correlation between image data and process parameters is defined as the correlation coefficient between real-time clinker flow rate, air temperature data, and image features. This confidence level is used to determine the reliability of abnormal operating conditions. A confidence level ≥ 0.8 indicates a highly reliable anomaly, directly triggering a DCS alarm and operation suggestion push in the linkage control steps. A confidence level between 0.6 and 0.8 triggers a manual review pop-up on the central control display screen, requiring operators to complete the review and confirmation within 30 seconds. A confidence level < 0.6 indicates a normal operating condition, and the system records the current image data to the sample database.

[0052] This invention also includes an abnormal operating condition classification step, which combines real-time identification results with historical abnormal data from the past 30 days to classify the severity of the abnormality. The classification standard is calculated using the following formula: in This represents the severity of the anomaly, ranging from 0 to 2. The overall confidence level for abnormal operating conditions. The historical correlation coefficient is 0.8, which corresponds to an average processing time of 15 minutes for similar anomalies over the past 30 days. The frequency of similar anomalies within the past 30 days is calculated by taking the number of times snowmen were built in the same grate cooler area within the past 30 days; based on... Values ​​are differentiated. When the value is <0.8, only the DCS system's audible and visual alarm is triggered; production parameters are not affected. When the value is ≤0.8, the DCS system's audible and visual alarm is triggered. <1.5 hours triggers the activation of the air cannon device; press and hold during air cannon firing. Value linear adjustment, Spray for 5 seconds when =0.8. Spray for 15 seconds when =1.5. When the value is ≥1.5, an emergency operation is triggered to reduce the kiln feed rate by 10%, and the abnormal handling process is simultaneously pushed to the central control display screen.

[0053] In this invention, the AI ​​model training step employs a transfer learning strategy, fine-tuning a pre-trained ResNet50 model by freezing the parameters of the first 10 convolutional layers and only fine-tuning the parameters of the last 15 layers (including 3 fully connected layers). The fully connected layers use the ReLU activation function, and the output layer uses the Softmax function to achieve work condition classification. During model training, five-fold cross-validation is used, dividing the sample library into training and validation sets in an 8:2 ratio. After each round of training, the accuracy, recall, and F1 score of the validation set are calculated. Training stops when the accuracy of the validation set is ≥95% and the F1 score is ≥0.94 for 10 consecutive rounds. The trained model is stored in ONNX format, occupying ≤8GB of storage space and loading time ≤2 seconds.

[0054] In this invention, the temperature correction in the data preprocessing step adopts a linear correction algorithm. The correction formula refers to the fitting relationship between real-time temperature and image grayscale value. For every 10°C increase in temperature, the image brightness gain is linearly increased by 5%, and the brightness gain adjustment range is 0-30%. Image frame capture adopts a hardware triggering method, and the frame capture interval is precisely controlled by the camera's built-in timer. The frame capture time deviation is ≤1ms, avoiding time drift caused by software frame capture. After Gaussian filtering, the image noise variance is ≤5. After adaptive histogram equalization processing, the image contrast is improved by 20%-40%, ensuring that the clinker edge contour is clearly distinguishable.

[0055] In this invention, a model performance evaluation system is established in the closed-loop optimization step, using three core indicators: accuracy, recall, and F1 score. Accuracy = (Number of correctly identified anomalies / Total number of identified anomalies) × 100%; Recall = (Number of correctly identified anomalies / Actual number of anomalies) × 100%; F1 score = 2 × Accuracy × Recall / (Accuracy + Recall). When any indicator falls below a preset threshold (accuracy 92%, recall 90%, F1 score 0.91) in three consecutive evaluations, an emergency optimization process is triggered. Within 24 hours, the sample library is expanded and the model parameters are updated. After the update, the performance evaluation is re-executed until all indicators recover to above the threshold.

[0056] In this invention, after receiving the identification results in the linkage control step, the APC intelligent optimization control system automatically extracts the secondary air temperature, tertiary air temperature, and clinker flow rate data from the past hour. The secondary air temperature is collected at 10-second intervals, the tertiary air temperature at 15-second intervals, and the clinker flow rate at 5-second intervals. The system then calculates operational adjustment suggestions using a data fusion algorithm. These suggestions include specific parameters such as the fan speed adjustment range (±5%-15%) and the air cannon activation interval (10-30 seconds). After the parameter suggestions are pushed to the DCS system, the system displays them in priority order, with priority determined by the severity of the anomaly. Sure, The higher the value, the higher the priority. After the operator selects the execution suggestion, the system automatically records the operation time, adjusted parameters and subsequent changes in operating conditions. The recorded data is kept for ≥1 year.

[0057] This invention includes the following modules:

[0058] Image acquisition module: It consists of a wide-angle high-definition camera, a 304 stainless steel high-temperature resistant shell, an axial fan cooling system and a PT100 temperature sensor. The camera lens is made of special high-temperature resistant glass with an anti-fog coating and a light transmittance of ≥95%. The temperature sensor has a measurement range of 0℃-200℃ and an accuracy of ±1℃. The module transmits video stream data through an Ethernet interface with a transmission rate of ≥100Mbps.

[0059] Data preprocessing module: Equipped with an ARM Cortex-A9 embedded processor with an operating frequency of 1.2GHz, it integrates image frame capture, Gaussian filtering, adaptive histogram equalization and temperature correction algorithms, with a processing latency of ≤150ms. It supports data interaction with the AI ​​diagnostic model module via the PCIe bus. The preprocessed image is stored in JPEG format with a compression ratio of 10:1.

[0060] AI diagnostic model module: It adopts NVIDIA Jetson TX2 GPU acceleration computing unit with a computing performance of 1.3 TFLOPS. It has a built-in trained convolutional neural network model with a model storage capacity of ≥10GB. It supports manual parameter updates via USB interface or automatic updates at preset cycles. It outputs abnormal working condition identification results and confidence level within 200ms after receiving preprocessed images.

[0061] Linkage control module: includes APC intelligent optimization control unit and DCS system interface unit. APC unit adopts industrial-grade server, CPU model Intel Xeon E3-1230, memory 16GB. DCS interface unit supports RS485 bus and Modbus RTU protocol to realize real-time transmission of recognition results and output of audible and visual alarm signals.

[0062] Closed-loop optimization module: It consists of a 1TB SSD data storage unit and a model update unit. The data storage unit records the recognition results, operator feedback data and working condition images for the past year. The model update unit supports automatic monthly updates or manual updates via the central control unit. The update process does not interrupt the diagnostic process.

[0063] High-temperature adaptation module: includes an axial fan cooling system and a temperature compensation unit. The cooling system fan speed is adjustable from 1500-3500r / min and is equipped with a 2mm thick aluminum heat-conducting sheet with a thermal conductivity ≥200W / (m·K). The temperature compensation unit corrects image grayscale deviation under high-temperature environments through software algorithms.

[0064] Operation interaction module: It consists of a 19-inch touch central control display screen and a feedback input unit. The display screen has a resolution of 1920×1080 and supports the visualization of abnormal results, operation suggestions and model performance indicators. The feedback input unit supports the correctness of touch annotation recognition and the annotation data is transmitted to the closed-loop optimization module in real time.

[0065] In this invention, the camera mounting bracket of the image acquisition module is made of 304 stainless steel. The bracket is equipped with a 5mm thick nitrile rubber shockproof and buffer structure with a rubber hardness of 60 Shore A. When the grate cooler vibrates at 50Hz-100Hz, the lens displacement is ≤0.1mm. The camera lens has a focal length of 8mm and a depth of field of 0.5m-5m, which can clearly identify clinker particles with a diameter of 5mm or larger inside the grate cooler. The lens has an IP68 protection rating, which can resist the corrosion of dust and high-temperature steam inside the grate cooler. The lens cleaning cycle is ≥30 days.

[0066] In this invention, the linkage control module also includes an emergency handling unit, which uses a Siemens S7-1200 PLC as the control core, with a response time ≤500ms; when the identification result is a severe anomaly ( When the value is ≥1.5, the emergency response unit drives the electric actuator to adjust the opening of the grate cooler feed regulating valve. The valve opening is adjusted according to... The value decreases linearly. When the value is 1.5, the opening decreases by 30%. =2.0 When the opening is reduced by 50%, a steam flow adjustment command is sent to the waste heat power generation system, and the steam flow is reduced by 15%. The execution time, parameter change range and operating condition feedback data of all emergency operations are automatically stored in the SSD data storage unit of the closed-loop optimization module. The stored data is retained for ≥2 years and supports reading data through the I2C bus for fault tracing analysis.

[0067] The following two examples further illustrate specific embodiments of the present invention:

[0068] Example 1: Diagnosis of Abnormal Operating Conditions of the Grate Cooler in a 7500t / d New Dry Process Cement Production Line

[0069] This embodiment is applied to the No. 3 7500t / d clinker production line of Huzhou Huaikan Southern Cement Co., Ltd. This production line adopts the "second generation" new dry process technology and is equipped with a fourth generation grate cooler with an effective cooling area of ​​120㎡. The kiln tail is equipped with a double-series six-stage low-pressure loss cyclone preheater and a TDF decomposition furnace. The daily clinker flow rate is 150t / h-180t / h, and the internal temperature of the grate cooler is stable at 90℃-140℃. Historically, the snowman-like failure occurred an average of 3-4 times per month. The average response time of traditional manual diagnosis was 1.2 minutes, and it took more than 30 minutes to restore production after the failure was dealt with.

[0070] I. System Deployment

[0071] Image acquisition module deployment: Three wide-angle high-definition cameras are installed at the head (feed end), middle (main cooling zone), and tail (discharge end) of the grate cooler, forming a comprehensive, blind-spot-free monitoring system. The camera lenses use special high-temperature resistant glass with an anti-fog coating, an 8mm focal length, a horizontal field of view of 120°, a vertical field of view of 60°, and an IP68 protection rating. The high-temperature resistant housing is made of 304 stainless steel, 3mm thick, and includes two axial flow fans with adjustable speeds from 1500-3500 r / min. Each camera is equipped with a 2mm thick aluminum heat-conducting fin with a thermal conductivity of 200W / (m·K). Each camera is paired with a PT100 temperature sensor with a measurement range of 0℃-200℃ and an accuracy of ±1℃. The sensor and camera are fixed to the same bracket, which uses a nitrile rubber shockproof structure with a hardness of 60 Shore A. The installation height is 3m above the grate to prevent clinker splashing and impact.

[0072] Data preprocessing module deployment: The module is integrated into the industrial control cabinet in the production line's central control room. It is equipped with an ARM Cortex-A9 embedded processor with a computing frequency of 1.2GHz and a cache capacity of 256MB. It transmits data via Ethernet at a transmission rate of 100Mbps, receives video stream data from three cameras, and has a built-in 5×5 Gaussian filter core with a sigma value set to 0.8, an adaptive histogram equalization algorithm, a cliplimit set to 2.0, and a temperature correction program. The processing latency is stable at 120ms-140ms. The preprocessed image is stored in JPEG format with a compression ratio set to 10:1 and transmitted to the AI ​​diagnostic model module via the PCIe bus.

[0073] AI diagnostic model module deployment: Utilizing GPU-accelerated computing units with a computing performance of 1.3 TFLOPS and 8GB of memory, it incorporates a ResNet50-based convolutional neural network model, stored in ONNX format, occupying 7.5GB of space, with a loading time of 1.8 seconds. The model sample library contains 50,000 labeled images, including 30,000 frames of normal operating conditions, covering a temperature range of 20℃-180℃ and a clinker flow rate range of 100t / h-200t / h, and 20,000 frames of snowman-like anomalies, divided into mild, moderate, and severe anomalies in a 3:4:3 ratio. During training, the parameters of the first 10 convolutional layers are frozen, and the parameters of the remaining 15 layers are fine-tuned. The batch size is 64, and the learning rate is reduced from 0.01 to 0.001 using a cosine annealing strategy. After 500 iterations, the validation set accuracy is 96.5%, recall is 95.8%, and F1 score is 0.961.

[0074] Deployment of the linkage control and closed-loop optimization modules: The APC intelligent optimization control unit of the linkage control module adopts an industrial-grade server with a CPU speed of 1.5GHz and a memory capacity of 16GB. The DCS interface unit supports RS485 bus and Modbus RTU protocol, seamlessly connecting with the existing DCS system on the production line. The DCS system control cycle is 100ms. The closed-loop optimization module is equipped with a 1TB SSD data storage unit with a write speed of 500MB / s. The model update unit is set to automatically update every Monday at 2:00 AM and supports manual emergency updates triggered via the central control touch screen. The operation interaction module is a 19-inch touch screen with a resolution of 1920×1080 and a response time of 50ms. It is installed on the central control console and displays images from three cameras, anomaly recognition results, and operation suggestions in real time.

[0075] II. Diagnostic Method Implementation Steps

[0076] Image acquisition steps: Three cameras simultaneously capture images at a frame rate of 25fps, with each camera capturing one frame every 100ms. Temperature sensors simultaneously collect ambient temperatures, for example, the head camera collects temperatures of 125℃, the middle camera 110℃, and the tail camera 95℃. The video stream and temperature data are transmitted in real time to the data preprocessing module via Ethernet. The camera's cooling fan automatically adjusts its speed according to the temperature: 1500r / min when the temperature is <100℃, 2500r / min when the temperature is 100℃-130℃, and 3500r / min when the temperature is >130℃, ensuring that the temperature of the internal components of the camera remains stable at 55℃-60℃.

[0077] Data preprocessing steps: First, a 5×5 Gaussian filter is applied to the captured image to remove high-frequency noise caused by clinker dust. The noise variance of the filtered image is ≤4. Then, adaptive histogram equalization is used to improve contrast, expanding the grayscale range of the processed image from 60-180 to 40-200. Finally, grayscale correction is performed using temperature sensor data: the head camera temperature is 125℃, which is 100℃ higher than the 25℃ reference temperature, resulting in a 50% increase in brightness gain. Based on the standard of 5% increase for every 10℃ increase, the average grayscale value of the head image is adjusted from 120 to 180 after correction to ensure consistency with the grayscale values ​​of the middle and tail images.

[0078] AI model training and real-time diagnosis steps: The preprocessed image is input into the GPU acceleration unit. The model extracts the clinker accumulation area, edge contour irregularity, and gray-level distribution features from the image. Specifically, the head image has an accumulation area of ​​0.25㎡, an edge contour irregularity of 0.72, and a gray-level histogram peak value of 180. The morphological feature matching degree is calculated. =0.85, meaning the overlap rate between the clinker accumulation area in the identified image and the standard shape of a snowman is 85%, and the grayscale distribution feature matching degree is 0.85. =0.9, meaning the similarity between the gray-level histogram of the identified image and the typical gray-level histogram of a snowman is 90%, and the correlation of process parameters. =0.8, which is the correlation coefficient between the real-time clinker flow rate of 170t / h, the secondary air temperature of 310℃, and the image features. Substituting this into the formula for the comprehensive confidence level of abnormal operating conditions: Combined with the frequency of snowman building incidents in the area over the past 30 days =2, meaning there were two instances of snowman-like formations in the head area within the past 30 days. Substituting this into the formula for the severity of the anomaly: The result was determined to be a severe anomaly. The model output the recognition result, which included a confidence score of 0.855 and a severity score of 2.223, taking 180ms.

[0079] Linkage control and closed-loop optimization steps: The identification results are pushed to the APC system. The system extracts secondary air temperature, tertiary air temperature, and clinker flow rate data for the past hour. The average secondary air temperature is 310℃, the average tertiary air temperature is 860℃, and the average clinker flow rate is 170t / h. Using a data fusion algorithm, the system calculates adjustment suggestions: increase fan speed by 15%, activate air cannon for 15 seconds, and reduce feed regulating valve opening by 30%. This is suggested to trigger an audible and visual alarm via the DCS system, with an alarm sound intensity of 85dB and a light flashing frequency of 3Hz, simultaneously triggering a pop-up window on the central control screen. The emergency handling unit drives the electric actuator to complete the valve opening adjustment within 450ms, simultaneously sending a command to the waste heat power generation system to reduce steam flow by 15%. After the operator confirms the abnormal handling, they mark the identification as correct via the touchscreen. This data includes the image, identification result, and operation record, and is stored in the SSD unit. It is added to the sample library 7 days later during model updates. After the update, the model's accuracy on the same test set improves to 97.2%.

[0080] III. Implementation Results Verification

[0081] Table 1 Comparison of the effects of different diagnostic methods in Example 1

[0082]

[0083] Table 1 clearly demonstrates the advantages of this invention in a 7500t / d production line. Traditional manual diagnosis relies on experience, with an accuracy rate of less than 80% and a response time as long as 72 seconds, resulting in 65.2% of anomalies not being handled in a timely manner, a monthly equipment wear rate of 1.8%, and a monthly downtime of 8.5 hours. Although existing image recognition has shortened the response time to 500ms, its accuracy rate is only 84.3% because it does not consider high-temperature image correction and multi-feature fusion, and the anomaly handling rate and equipment wear rate are still unsatisfactory. This invention, through high-temperature adaptability design, multi-dimensional feature fusion, and GPU acceleration, improves the accuracy rate to 96.5%, shortens the response time to 180ms, achieves an anomaly handling rate of 98.3%, reduces the monthly equipment wear rate to 0.5%, and reduces the monthly downtime to only 0.8 hours, significantly improving production continuity and economy, while reducing the labor intensity of operators, meeting the needs of intelligent production lines.

[0084] Example 2: Diagnosis of Abnormal Operating Conditions of the Grate Cooler in a 5000t / d New Dry Process Cement Production Line

[0085] This embodiment is applied to a 5000t / d clinker production line of a building materials group, equipped with a third-generation grate cooler with an effective cooling area of ​​80㎡. The kiln tail is equipped with a five-stage cyclone preheater and a decomposition furnace. The daily clinker flow rate is 100t / h-130t / h, and the internal temperature of the grate cooler is 80℃-120℃. Traditional manual diagnosis has a 20% missed detection rate due to the blind spot in the tail area monitoring. The accuracy of the existing simple image recognition system drops to below 75% when the temperature is >110℃.

[0086] I. System Deployment

[0087] Image acquisition module deployment: One wide-angle high-definition camera is installed at the head and one at the tail of the grate cooler. The head camera covers the feeding and main cooling areas, while the tail camera covers the discharging and transition areas, avoiding redundant monitoring of the middle area. The camera lens has a focal length of 6mm, a horizontal field of view of 110°, a vertical field of view of 55°, a high-temperature resistant shell thickness of 2.5mm, and a built-in axial flow fan with a speed of 1500-3000r / min. The temperature sensor and camera are integrated, with a measurement range of 0℃-180℃ and an accuracy of ±1℃. The shock-absorbing structure of the bracket is made of silicone material with a hardness of 55 Shore A, and the installation height is 2.5m from the grate plate.

[0088] Data preprocessing module deployment: Equipped with an ARM Cortex-A7 embedded processor, with an operating frequency of 1.0GHz, a cache capacity of 128MB, a built-in 5×5 Gaussian filter core, a sigma value set to 0.7, an adaptive histogram equalization algorithm, a cliplimit set to 1.8, and temperature correction performed according to the standard of increasing brightness gain by 4.5% for every 10℃. The processing latency is 130ms-150ms. Images are transmitted to the AI ​​module via Ethernet at a transmission rate of 100Mbps.

[0089] AI diagnostic model module deployment: GPU acceleration unit computing performance of 1.0 TFLOPS, model sample library of 40,000 labeled images, including 25,000 frames under normal operating conditions and 15,000 frames under abnormal operating conditions, covering the temperature range of 20℃-160℃ and the clinker flow rate range of 80t / h-150t / h, based on ResNet50 fine-tuning iteration of 300 rounds, validation set accuracy of 95.8%, recall of 94.9%, F1 score of 0.953, and model loading time of 2.0 seconds.

[0090] Deployment of linkage control and closed-loop optimization modules: The APC unit adopts an industrial-grade server with a CPU speed of 1.2GHz and a memory capacity of 8GB. The DCS interface unit supports Ethernet and Modbus TCP protocols. The closed-loop optimization module has an SSD storage capacity of 512GB, and the model is automatically updated every Tuesday at 3:00 AM. The operation interaction module is a 17-inch touch screen with a resolution of 1920×1080.

[0091] II. Diagnostic Method Implementation Steps

[0092] Image acquisition steps: Two cameras capture images at a frame rate of 25fps, with a frame taken every 100ms. The rear camera captures the temperature at 115℃, and the fan speed is automatically adjusted to 3000r / min. Temperature sensor data and images are synchronously transmitted to the preprocessing module.

[0093] Data preprocessing steps: The tail image is first subjected to Gaussian filtering, reducing the noise variance to 3.8; then, adaptive histogram equalization is used to improve contrast, expanding the grayscale range to 50-190; finally, grayscale correction is performed in combination with temperature data: because the temperature of 115℃ is 90℃ higher than that of 25℃, the brightness gain is increased by 40.5%, and the grayscale mean of the corrected image is adjusted from 110 to 154, eliminating the grayscale darkening caused by high temperature.

[0094] AI model training and real-time diagnosis steps: The preprocessed image is input into the GPU acceleration unit. The model extracts the clinker accumulation area (0.18㎡), edge irregularity (0.68), and grayscale distribution matching degree (0.85) of the tail image, indicating an 85% similarity between the grayscale histogram of the identified image and the typical grayscale histogram of a snowman. The correlation degree of the process parameters (T=0.75) is calculated, which is the correlation coefficient between the clinker flow rate (120t / h), the secondary air temperature (290℃), and the image features. The comprehensive confidence level of the abnormal working condition is then calculated. Combined with the frequency of snowman building in the tail area over the past 30 days =1, substituting into the formula for the severity of the anomaly: If the value is in the range of 0.8-1.5, it is judged as a moderate anomaly. The model outputs the recognition result, which takes 190ms.

[0095] Linkage control and closed-loop optimization steps: The APC system generates operation suggestions: increase fan speed by 10%, activate air cannon for 10 seconds, trigger audible and visual alarms in the DCS system, and after the operator executes the suggestions, the correct identification is marked, and the data is stored in the SSD unit. Seven days later, the model is updated with 800 new images, and the accuracy rate is improved to 96.3%. Under a high temperature environment of 115℃, the recognition accuracy rate is improved from 92.1% before the update to 95.5%.

[0096] III. Implementation Results Verification

[0097] Table 2 Comparison of the effects of different diagnostic methods in Example 2

[0098]

[0099] Table 2 data verifies the adaptability and advantages of this invention in a 5000t / d production line. Traditional manual diagnosis suffers from a 20.1% false negative rate due to blind spots in tail-end monitoring, and an accuracy rate of only 70.2% under high-temperature conditions, requiring 25.3 minutes for a single anomaly handling. Existing image recognition does not solve the problem of image distortion at high temperatures, with a high-temperature accuracy rate of 74.8% and a false negative rate of 12.3%. This invention, through targeted tail-end control, high-temperature brightness correction, and multi-feature fusion, achieves an overall accuracy rate of 95.8%, a high-temperature accuracy rate of 95.5%, a false negative rate of 1.2%, and a single anomaly handling time of 8.6 minutes, effectively solving the blind spots and high-temperature adaptability problems of traditional diagnosis. Simultaneously, the system deployment cost is lower than existing technologies, reducing the need for one camera, adapting to the budget requirements of small and medium-sized production lines, balancing economy and practicality, further demonstrating the wide adaptability of this invention.

[0100] The two embodiments cover large and medium-sized cement production lines respectively. By refining the equipment deployment, step parameters and formula applications, the stability and effectiveness of the invention under different working conditions are fully verified. All technical details do not exceed the scope of the invention and supplement the installation location, algorithm parameters and other contents not detailed in the invention, to ensure that the technical solution can be implemented and reproduced. At the same time, the significant advantages of the invention in terms of accuracy, response time and anomaly handling are intuitively reflected by the tabular data, which meets the needs of intelligent transformation in the cement industry.

[0101] Reference Figure 3This figure visually illustrates the accuracy differences of three diagnostic methods at different temperatures, highlighting the high-temperature adaptability advantage of this invention. As the temperature rises from 50℃ to 170℃, traditional manual diagnosis suffers from image grayscale distortion and increased human experience-based judgment errors due to high temperatures, resulting in an accuracy rate drop from 82% to 65%. Existing image recognition systems do not optimize preprocessing for high temperatures, leading to a significant decrease in feature extraction accuracy as temperature increases, with the accuracy rate dropping from 88% to 68%. This invention, equipped with a temperature compensation mechanism and a high-temperature resistant image acquisition design, achieves an accuracy rate drop of only 1-2 percentage points for every 30℃ increase in temperature, maintaining 92% accuracy even at 170℃, far exceeding traditional methods and existing technologies. This advantage ensures that the grate cooler maintains high diagnostic accuracy within the common operating temperature range of 80℃-150℃, avoiding misdiagnosis and missed diagnosis in high-temperature environments, and providing a reliable basis for subsequent anomaly handling.

[0102] Reference Figure 4 This figure verifies the practical application value of the invention from a long-term operational perspective, demonstrating the dual benefits of "reduced incidence and shorter handling time." Under traditional diagnostic methods, the occurrence rate of snowman-like anomalies remained stable at 3.5-3.9 times / month. Due to a 0.5-2 minute lag in manual observation, some early anomalies were not addressed in time and developed into overt faults. Furthermore, handling required manual confirmation and manual operation of air cannons, with an average time of 26.3-28.5 minutes per instance. The present invention, through its closed-loop optimization mechanism, continuously improves identification accuracy, enabling the capture of early, minute accumulations. The anomaly occurrence rate decreased from 1.2 times / month to 0.6 times / month, a cumulative reduction of 17 anomalies over 5 months. Simultaneously, relying on the linkage control with the DCS system, no manual intervention is required in core operations, reducing handling time from 9.2 minutes / event to 7.9 minutes / event. Long-term operational data shows that the present invention not only reduces the frequency of anomalies but also shortens the time required for each handling, reducing thermal efficiency loss and equipment wear in the grate cooler caused by snowman-like accumulation, thus ensuring the continuous and stable operation of the cement production line.

[0103] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for diagnosing abnormal operating conditions of a grate cooler based on AI image recognition, characterized in that, Includes the following steps: Image acquisition steps: Use a wide-angle camera to capture real-time images of the inside of the grate cooler from all angles, and collect video stream data of clinker morphology, distribution and internal environment. It is adapted to the internal environment of the grate cooler of 80℃-150℃ and has a continuous running time of not less than 8000 hours. Data preprocessing steps: The acquired video stream is captured as a frame every 100ms interval, and noise is removed by a 5×5 Gaussian filter kernel. Combined with real-time temperature data collected by the PT100 temperature sensor, the grayscale values ​​of the image pixels are linearly corrected. AI model training steps: Construct an abnormal working condition recognition model based on a convolutional neural network, input an image sample library, and train it using the batch gradient descent method; Real-time diagnostic steps: Input the preprocessed image into the trained model, extract the clinker accumulation area, edge contour shape and gray-scale distribution features in the image, compare them with the snowman feature threshold preset by the model, and output the recognition result and the confidence level in the 0-1 interval. Linkage control steps: The identification result is pushed to the APC intelligent optimization control system via Ethernet. The system writes the value back to the DCS system according to the Modbus protocol. The DCS system triggers an audible and visual alarm and completes the central control abnormality pop-up prompt within 1 second. Closed-loop optimization steps: Collect the recognition accuracy data marked by operators through the central control interaction unit, combine it with newly collected working condition images to supplement the sample library, and iterate and update the model parameters every 7 days; High-temperature adaptation steps: A temperature compensation mechanism is introduced to correct image features. For every 10°C increase in temperature, the image edge detection threshold is reduced by 5%. At the same time, the fan speed of the camera heat dissipation system is optimized. When the temperature is ≥120°C, the fan speed is increased to 3500r / min to maintain the temperature of the internal components of the camera ≤60°C.

2. The method for diagnosing abnormal operating conditions of a grate cooler based on AI image recognition according to claim 1, characterized in that, It also includes an image feature enhancement step, using the following formula to calculate the overall confidence level of abnormal operating conditions: ,in The overall confidence level for abnormal operating conditions. For morphological feature weights, The weights are the grayscale distribution feature weights. Associating process parameters with weights, For morphological feature matching degree, The grayscale distribution feature matching degree, The correlation between image data and process parameters; When the confidence level is ≥0.8, it is judged as a high-confidence anomaly, and the DCS alarm and operation suggestion push in the linkage control steps are directly triggered. When the confidence level is 0.6-0.8, a manual review pop-up window is triggered on the central control display screen. The operator must complete the review and confirmation within 30 seconds. When the confidence level is <0.6, it is judged as a normal working condition, and the system records the current image data to the sample library.

3. The method for diagnosing abnormal operating conditions of a grate cooler based on AI image recognition according to claim 1, characterized in that, It also includes an abnormal operating condition classification process, and the classification standard is calculated using the following formula: ,in This represents the severity value of the abnormality. The overall confidence level for abnormal operating conditions. Historical correlation coefficient This refers to the frequency of similar abnormalities occurring within the past 30 days. in accordance with Values ​​are differentiated. When the value is <0.8, only the DCS system's audible and visual alarm is triggered; production parameters are not affected. When the value is ≤0.8, the DCS system's audible and visual alarm is triggered. <1.5 hours triggers the activation of the air cannon device; press and hold during air cannon firing. Value linear adjustment, Spray for 5 seconds when =0.

8. Spray for 15 seconds when =1.

5. When the value is ≥1.5, an emergency operation is triggered to reduce the kiln feed rate by 10%, and the abnormal handling process is simultaneously pushed to the central control display screen.

4. The method for diagnosing abnormal operating conditions of a grate cooler based on AI image recognition according to claim 1, characterized in that, The AI ​​model training process employs a transfer learning strategy, fine-tuning a pre-trained ResNet50 model by freezing the parameters of the first 10 convolutional layers and fine-tuning only the parameters of the last 15 layers. The fully connected layers use the ReLU activation function, and the output layer uses the Softmax function for work condition classification. Five-fold cross-validation is used during model training, dividing the sample library into training and validation sets in an 8:2 ratio. After each training round, the accuracy, recall, and F1 score of the validation set are calculated. Training stops when the validation set achieves an accuracy ≥ 95% and an F1 score ≥ 0.94 for 10 consecutive rounds. The trained model is stored in ONNX format.

5. The method for diagnosing abnormal operating conditions of a grate cooler based on AI image recognition according to claim 1, characterized in that, In the data preprocessing step, temperature correction adopts a linear correction algorithm. The correction formula refers to the fitting relationship between real-time temperature and image grayscale value. For every 10°C increase in temperature, the image brightness gain is linearly increased by 5%. Image frame capture adopts a hardware triggering method, and the frame capture interval is controlled by the camera's built-in timer. The frame capture time deviation is ≤1ms.

6. The method for diagnosing abnormal operating conditions of a grate cooler based on AI image recognition according to claim 1, characterized in that, In the closed-loop optimization step, a model performance evaluation system is established, using three core indicators: accuracy, recall, and F1 score. Accuracy = (Number of correctly identified anomalies / Total number of identified anomalies) × 100%, Recall = (Number of correctly identified anomalies / Actual number of anomalies) × 100%, F1 score = 2 × accuracy × recall / (accuracy + recall). If any indicator falls below the preset threshold in three consecutive evaluations, an emergency optimization process is triggered. Within 24 hours, the sample library is expanded and the model parameters are updated. After the update, the performance evaluation is re-executed until all indicators recover to above the threshold.

7. The method for diagnosing abnormal operating conditions of a grate cooler based on AI image recognition according to claim 1, characterized in that, In the linkage control process, after receiving the identification results, the APC intelligent optimization control system automatically extracts the secondary air temperature, tertiary air temperature, and clinker flow rate data from the past hour, and calculates operating condition adjustment suggestions through a data fusion algorithm. After the parameter suggestions are pushed to the DCS system, the system displays them in order of priority, with priority based on the severity of the anomaly. Sure, The higher the value, the higher the priority. After the operator selects the execution suggestion, the system automatically records the operation time, adjusted parameters, and subsequent changes in operating conditions.

8. A system for diagnosing abnormal operating conditions of a grate cooler based on AI image recognition as described in any one of claims 1-7, characterized in that, Includes the following modules: Image acquisition module: It consists of a wide-angle camera, a 304 stainless steel shell, an axial fan cooling system and a PT100 temperature sensor. The camera lens is made of special high-temperature resistant glass with an anti-fog coating. The temperature sensor has a measurement range of 0℃-200℃. The module transmits video stream data through an Ethernet interface. Data preprocessing module: Equipped with an ARM Cortex-A9 embedded processor with an operating frequency of 1.2GHz, it integrates image frame capture, Gaussian filtering, adaptive histogram equalization and temperature correction algorithms, supports data interaction with the AI ​​diagnostic model module via PCIe bus, and stores preprocessed images in JPEG format; AI diagnostic model module: It adopts NVIDIA Jetson TX2 GPU acceleration computing unit with a computing performance of 1.3 TFLOPS. It has a built-in trained convolutional neural network model and supports manual parameter updates via USB interface or automatic updates at preset cycles. It outputs abnormal working condition identification results and confidence level within 200ms after receiving preprocessed images. Linkage control module: includes APC intelligent optimization control unit and DCS system interface unit. The APC unit adopts industrial-grade server, and the DCS interface unit supports RS485 bus and ModbusRTU protocol to realize real-time transmission of recognition results and output of audible and visual alarm signals. Closed-loop optimization module: It consists of a 1TB SSD data storage unit and a model update unit. The data storage unit records the recognition results, operator feedback data and working condition images for the past year. The model update unit supports automatic monthly updates or manual updates via the central control unit. High-temperature adaptation module: includes an axial fan cooling system and a temperature compensation unit. The cooling system fan speed ranges from 1500 to 3500 r / min and is equipped with a 2mm thick aluminum heat-conducting sheet. The temperature compensation unit corrects image grayscale deviations under high-temperature environments through software algorithms. Operation interaction module: It consists of a display screen and a feedback input unit. The display screen supports the visualization of abnormal results, operation suggestions and model performance indicators. The feedback input unit supports the recognition of the correctness of touch annotation. The annotation data is transmitted to the closed-loop optimization module in real time.

9. The system for diagnosing abnormal operating conditions of a grate cooler based on AI image recognition according to claim 8, characterized in that, The camera mounting bracket of the image acquisition module is made of 304 stainless steel and is equipped with a nitrile rubber shockproof and buffer structure; the camera lens has a focal length of 8mm, a depth of field of 0.5m-5m, and can identify clinker particles with a diameter of 5mm or larger inside the grate cooler. The lens has an IP68 protection rating and a lens cleaning cycle of ≥30 days.

10. The system for diagnosing abnormal operating conditions of a grate cooler based on AI image recognition according to claim 8, characterized in that, The linkage control module also includes an emergency handling unit, which uses a Siemens S7-1200 PLC as its control core; when the identification result... When the value is ≥1.5, the emergency response unit drives the electric actuator to adjust the opening of the grate cooler feed regulating valve. The valve opening is adjusted according to... The value decreases linearly. When the value is 1.5, the opening decreases by 30%. When the opening is 2.0, the opening is reduced by 50%, and a steam flow adjustment command is sent to the waste heat power generation system, reducing the steam flow by 15%. The execution time, parameter change range, and operating condition feedback data of all emergency operations are automatically stored in the SSD data storage unit of the closed-loop optimization module.

Images