A hot-dip galvanized steel strip coating control method based on deep learning

Abstract

The application relates to the technical field of industrial control systems and artificial intelligence, and particularly discloses a control method for a hot galvanized steel strip coating based on deep learning. The method is characterized in that a high-sensitivity microphone array is arranged in a zinc pot area, and acoustic signals, strip surface images and process sensor data are synchronously collected; after preprocessing, cross-modal attention fusion network is used to dynamically associate three-mode information to generate fusion perception features; a deep time series prediction model is used to deduce future coating thickness distribution and identify abnormalities; and coordinated adjustment instructions for air knife pressure, angle and strip speed are generated to implement closed-loop control. The application realizes reliable perception and forward-looking regulation of the coating state, improves control robustness and product quality, reduces dependence on expensive thickness measuring equipment, and is suitable for harsh industrial scenes.

Description

Technical Field

[0001] This invention belongs to the field of industrial control systems and artificial intelligence, specifically relating to a control method for hot-dip galvanized steel strip coating based on deep learning. Background Technology

[0002] With the intelligent transformation of the steel manufacturing industry, hot-dip galvanizing, as a core process for improving the corrosion resistance and surface quality of strip steel, significantly impacts product performance and resource utilization due to the precision of its process control. Modern hot-dip galvanizing production lines typically integrate high-precision automated control systems designed to ensure the uniformity and stability of the zinc layer thickness through real-time adjustment of key process parameters. In a highly automated industrial environment, building a robust monitoring system has become a crucial technical support for improving production line efficiency and ensuring product consistency.

[0003] The air knife control system in the zinc pot area is a key piece of equipment determining the coating weight. Its working principle involves using high-pressure gas to remove excess zinc from the strip surface. This technology requires the control system to be able to frequently sense and respond to the air knife's operating status, strip surface characteristics, and environmental feedback. Conventional solutions primarily rely on visual sensors or laser thickness gauges to capture the production line's operating status and use control algorithms to adjust production parameters to achieve the preset coating accuracy target.

[0004] The zinc pot area of ​​a hot-dip galvanizing production line is subjected to an extreme industrial environment characterized by high temperature, high humidity, and high dust levels. This makes traditional optical imaging equipment and precision laser measuring instruments highly susceptible to vapor obstruction and ambient light interference, leading to sensor failure or a significant decline in measurement accuracy. Because key process variables are unmeasurable in the blind zone environment, existing systems often struggle to detect potential problems such as zinc dross clogging the air knife or excessive localized thickness of the strip, resulting in lag in control actions.

[0005] Traditional models lack the ability to analyze complex interference signals in the production site, making it difficult to correlate multi-source heterogeneous information and compensate for nonlinear states. As a result, the perception accuracy and response speed of the system under extreme working conditions cannot meet the needs of high-quality production. Therefore, a control method for hot-dip galvanized steel strip coating based on deep learning is desired. Summary of the Invention

[0006] The purpose of this invention is to provide a control method for hot-dip galvanized steel strip coating based on deep learning, which can solve the problems in the background art mentioned above.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for controlling the coating of hot-dip galvanized strip steel based on deep learning, comprising the following specific steps: Step 1: Deploy a high-sensitivity microphone array in the zinc pot area of ​​the hot-dip galvanizing production line to collect acoustic signals generated during the process of air knife blowing zinc liquid in real time, and simultaneously acquire strip surface images and process sensor data; Step 2: Preprocess the acoustic signal, including noise reduction, framing and spectral feature extraction, to generate an acoustic feature vector characterizing the working state of the air knife; Step 3: Enhance and segment the image of the strip surface to extract visual feature maps that reflect the uniformity of the coating; Step 4: Input the acoustic feature vector, visual feature map and process sensor data into the cross-modal attention fusion network, establish dynamic correlation weights between the acoustic, visual and sensing modalities through a multi-head cross-attention mechanism, and generate fused perception features; Step 5: Based on the fused sensing features, use the deep temporal prediction model to deduce the coating thickness distribution trend in the next few time steps, and identify whether there is zinc dross clogging the air knife or local excessive thickness anomaly. Step 6: Based on the coating thickness distribution trend and anomaly identification results, generate coordinated adjustment commands for air knife pressure, angle, and strip running speed, and send them to the actuator for closed-loop control.

[0008] Preferably, the high-sensitivity microphone array in step 1 consists of no fewer than 8 omnidirectional industrial-grade microphones arranged in a ring within a predetermined distance around the air knife outlet, with a sampling frequency of no less than 48 kHz, and has a high-temperature and high-humidity resistant encapsulation structure to ensure continuous and stable acquisition of acoustic signals in the extreme environment of the zinc pot area.

[0009] Preferably, in step 2, the noise reduction process adopts an adaptive filtering algorithm based on wavelet threshold, the frame window length is set to a fixed number of milliseconds, the frame shift ratio is 50%, and the spectral feature extraction includes Mel frequency cepstral coefficients, short-time energy and zero-crossing rate, which together constitute an acoustic feature vector with fixed dimensions.

[0010] Preferably, in step 3, the image enhancement adopts an illumination correction method based on Retinex theory, combined with nonlocal mean denoising, and the segmentation process uses a semantic segmentation neural network to perform pixel-level classification on the strip surface, distinguishing normal coating areas, zinc particle accumulation areas and exposed iron defect areas, and outputting a visual feature map that is spatially aligned with the original image.

[0011] Preferably, in step 4, the cross-modal attention fusion network includes three independent encoders that process acoustic, visual, and sensor data respectively. Then, the attention weights between any two modalities are calculated through a cross-modal interaction layer, and finally, a unified high-dimensional fusion perception feature is generated by weighted fusion. This feature retains the key discrimination information of each modality and suppresses environmental interference noise.

[0012] Preferably, the deep time-series prediction model in step 5 adopts a gated cyclic unit architecture. Its input is a fusion sensing feature sequence of multiple consecutive time steps, and its output is a predicted value of the coating thickness distribution along the strip width direction in the future several time steps. The anomaly detection module compares the deviation of the predicted value with the historical normal mode. When the deviation exceeds a preset threshold, it is determined that there is a risk of zinc slag blockage or local excessive thickness.

[0013] Preferably, the generation of the coordinated adjustment command in step 6 follows the principle of multi-objective optimization. Under the premise of ensuring that the coating weight meets the process requirements, the air knife pressure is adjusted first to quickly suppress local overthickness, while the air knife tilt angle is finely adjusted to compensate for edge effects. When necessary, the strip running speed is reduced in conjunction to extend the response window. All adjustment actions are transmitted to the actuator in real time through the industrial fieldbus.

[0014] Preferably, the process sensor data includes strip temperature, tension, running speed, and air knife nitrogen flow rate. After standardization, these data are used as the third modal input of the fusion network to constrain the physical rationality of acoustic and visual features and improve the generalization ability of fusion perception under non-steady-state conditions.

[0015] Preferably, the cross-modal attention fusion network is optimized end-to-end using an labeled industrial field dataset during the training phase. The labeled information includes manually verified measured values ​​of coating thickness and fault event records. The loss function comprehensively considers prediction error and anomaly detection accuracy to ensure that the model has high robustness in real production line environments.

[0016] Preferably, the deep time series prediction model supports online incremental learning. When the system detects a new abnormal pattern and it is confirmed by a human, the new sample can be added to the training set and the model can be fine-tuned, so that the control system has the ability to continuously evolve and adapt to equipment aging and process drift during long-term operation of the production line.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention avoids the technical problem of optical sensors being prone to failure in high temperature, high humidity and high dust environments by introducing a high-sensitivity microphone array to collect acoustic signals in the air knife area, and realizes indirect but reliable perception of key process variables.

[0018] 2. This invention utilizes a cross-modal attention mechanism to deeply integrate three types of heterogeneous information: acoustic, visual, and process sensing, thereby improving the system's state recognition capability within the measurement blind zone and overcoming the risk of misjudgment caused by the one-sidedness of single-modal information.

[0019] 3. The depth time-series prediction model constructed in this invention can predict the evolution trend of coating thickness in advance, enabling the control system to shift from passive response to active intervention and shorten the response delay to potential problems such as zinc slag blockage or excessive local thickness.

[0020] 4. The collaborative adjustment strategy of this invention comprehensively considers the coupling relationship of multiple actuators, optimizing energy consumption and equipment lifespan while ensuring product quality. The overall solution does not rely on expensive and fragile laser thickness measurement equipment, reducing system maintenance costs and deployment complexity. It is suitable for various harsh industrial scenarios, providing a highly robust and adaptable technical path for the intelligent upgrade of hot-dip galvanizing production lines. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture according to the present invention; Figure 2 This is a schematic diagram illustrating the core principle framework of the cross-modal attention fusion network according to the present invention; Figure 3 This is a flowchart illustrating the logical process of acoustic feature extraction and visual feature extraction according to the present invention. Figure 4 This is a flowchart illustrating the logical process of coating thickness distribution deduction and anomaly identification based on the depth time-series prediction model according to the present invention. Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between the perception fusion, decision prediction and execution mechanisms in this invention. Detailed Implementation

[0022] Example 1: Please refer to the appendix Figure 1 To be continued Figure 5 To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0023] In the production process of hot-dip galvanized steel strip, precise control of the coating thickness is a key indicator for measuring product quality. This embodiment proposes a deep learning-based control method for the coating of hot-dip galvanized steel strip. By integrating acoustic sensing, computer vision, and multimodal data fusion technologies, it achieves real-time monitoring and closed-loop adjustment of coating quality under complex working conditions.

[0024] In the above method, step 1 involves deploying a high-sensitivity microphone array in the zinc pot area of ​​the hot-dip galvanizing production line to collect acoustic signals generated during the process of air knife blowing of zinc liquid in real time, and simultaneously acquiring images of the strip surface and process sensor data.

[0025] The high-sensitivity microphone array in step 1 consists of no fewer than eight omnidirectional industrial-grade microphones. These microphones are arranged in a ring within a predetermined distance around the air knife outlet, forming a spatial sound field capture network. To cope with the high temperature, high humidity, and high dust environment of the zinc pot area, each microphone is equipped with a high-temperature and high-humidity resistant encapsulation structure. Its shell is made of corrosion-resistant special alloy material and has a microporous sound-permeable diaphragm to prevent zinc vapor and dust from entering the pickup head. The sampling frequency is set to no less than 48 kHz to ensure that the ultrasonic frequency characteristics and fine turbulence noise generated by the high-pressure gas jet at the air knife nozzle can be captured.

[0026] While acquiring acoustic signals, the system obtains images of the strip surface through an industrial camera deployed directly above the strip. This camera has a high frame rate acquisition capability, and its shutter speed is triggered synchronously with the strip's running speed. Process sensor data is retrieved in real time through the production line's existing industrial control bus, and the data items involved include the strip's real-time temperature, the tension on the strip, the strip's linear velocity, and the instantaneous flow rate of nitrogen in the air knife.

[0027] All these data are aligned using a unified timestamp to form a synchronous data stream containing multi-dimensional information of sound, light, and electricity. Before entering the subsequent processing logic, the process sensor data needs to be standardized by subtracting the mean and dividing by the standard deviation to transform it into a numerical sequence with unified dimensions. This sequence serves as the third modal input to the fusion network, constraining the rationality of acoustic and visual features at the physical logic level.

[0028] In the above method, step 2 involves preprocessing the acoustic signal, including noise reduction, framing, and spectral feature extraction, to generate an acoustic feature vector characterizing the working state of the air knife.

[0029] The noise reduction process in step 2 employs an adaptive filtering algorithm based on wavelet thresholding. This algorithm performs multi-scale wavelet decomposition on the acquired raw acoustic signal, breaking it down into approximate and detail components at different frequency ranges. The system calculates an adaptive threshold based on the noise distribution characteristics of each component, suppressing or eliminating noise components smaller than this threshold. Finally, a wavelet reconstruction algorithm is used to synthesize the denoised, clean acoustic signal.

[0030] The denoised signal enters the framing stage, with the framing window length set to a fixed range of 25 to 40 milliseconds and the frame shift ratio between adjacent frames set to 50%. This overlapping sampling ensures the continuity of the signal in the time domain and avoids spectral leakage. When extracting spectral features, the system not only calculates the conventional Mel frequency cepstral coefficients but also simultaneously extracts parameters such as short-time energy, zero-crossing rate, and spectral centroid. The Mel frequency cepstral coefficient extraction process includes pre-emphasis, framing and windowing, fast Fourier transform, Mel filter bank filtering, and discrete cosine transform. Finally, these parameters are concatenated into a high-dimensional vector with fixed dimensions, i.e., the acoustic feature vector. This vector can sensitively capture subtle changes in the sound field caused by air knife pressure fluctuations, partial nozzle blockage, or strip vibration.

[0031] In the above method, step 3 involves enhancing and segmenting the image of the strip surface to extract visual feature maps that reflect the uniformity of the coating.

[0032] The image enhancement in step 3 employs an illumination correction method based on Retinex theory. This method assumes that an image is composed of both illuminance and reflectance components. The illuminance component is estimated by performing multi-scale Gaussian blurring on the image, and then the illuminance component is extracted from the original image in the logarithmic domain to obtain the reflectance component, which reflects the true surface morphology of the strip steel. The system combines a nonlocal mean denoising algorithm to smooth thermal noise generated by the high-temperature environment using redundant information in the image.

[0033] Based on the enhanced image, the system utilizes a semantic segmentation neural network for pixel-level classification. This network employs a deep convolutional encoder-decoder structure, extracting texture and edge features of the strip surface through multi-level convolutional layers. The segmentation process aims to divide the image into three main regions: a normal coating area, a zinc particle accumulation area caused by uneven air knife spraying, and an exposed iron defect area caused by poor wettability or mechanical abrasion.

[0034] The output of the semantic segmentation neural network is a visual feature map perfectly aligned with the spatial scale of the original image, where the value of each pixel represents the probability that the location belongs to a category. This visual feature map provides an intuitive spatial reference for evaluating the lateral uniformity of the coating thickness distribution.

[0035] In the above method, step 4 involves inputting the acoustic feature vector, visual feature map, and process sensor data into a cross-modal attention fusion network. A dynamic correlation weight between the acoustic, visual, and sensing modalities is established through a multi-head cross-attention mechanism to generate fused perception features.

[0036] The cross-modal attention fusion network in step 4 comprises three independent parallel encoders, which perform deep feature abstraction for the acoustic modality, visual modality, and sensor data modality, respectively. The acoustic encoder uses a one-dimensional convolutional network, the visual encoder uses a residual convolutional architecture, and the sensor encoder uses a simple fully connected stacked structure. After feature extraction, the system introduces a cross-attention mechanism. Using acoustic features as the query vector and visual features and sensor data as the key and value vectors, the correlation matrix among the three is calculated. During the calculation, the query vector and the key vector are multiplied by a dot product, and the result is scaled and then transformed into a weight distribution using a normalized exponential function.

[0037] These weights represent the degree to which different modal information contributes to judging the coating quality at the current moment. For example, when on-site smoke and dust blur visual features, the attention mechanism automatically reduces the weight of visual features while increasing the weight of acoustic and sensor features, achieving cross-modal information complementarity. Finally, the features of each modality are weighted and summed according to the calculated weights to generate a unified high-dimensional fusion perception feature.

[0038] This high-dimensional fusion sensing feature can retain key discriminative information in each modality and suppress environmental interference noise that may exist in a single sensor. The cross-modal attention fusion network is optimized end-to-end using a labeled industrial field dataset during the training phase. The loss function is designed to comprehensively consider the mean square error between predicted and measured values, as well as the cross-entropy loss for anomaly classification, ensuring the model has extremely high robustness in real production line environments.

[0039] In the above method, step 5 involves using a deep temporal prediction model based on the fused sensing features to predict the coating thickness distribution trend over several future time steps and identify whether there is zinc slag clogging the air knife or local excessive thickness anomalies.

[0040] The deep time-series prediction model in step 5 employs a gated recurrent unit architecture. This architecture controls the forgetting and transmission of information through reset and update gates, enabling it to handle industrial sequence data with long-distance dependencies. The model's input is a fused sensing feature sequence from multiple consecutive historical time steps. Through nonlinear mapping in the hidden layer, the output is a predicted value of the coating thickness distribution along the strip width direction within several future time steps.

[0041] The system integrates an anomaly detection module. This module continuously compares the current predicted trend value with pre-stored characteristic benchmarks from historical normal operating modes. When the predicted coating thickness deviates continuously from the standard range at a certain local location, and the deviation exceeds a preset dynamic threshold, the system determines that there is a potential risk of zinc slag clogging the air knife nozzle or excessively thick local coating. This time-series-based identification method can issue early warnings through subtle acoustic and air pressure fluctuations before the anomaly fully manifests on the surface of the finished strip steel, thus shortening the system's response delay.

[0042] In the above method, step 6 generates coordinated adjustment commands for air knife pressure, angle, and strip running speed based on the coating thickness distribution trend and anomaly identification results, and sends them to the actuator to implement closed-loop control.

[0043] The generation of coordinated adjustment instructions in step 6 follows a multi-objective optimization principle. When calculating control increments, the primary objective is to ensure that the coating weight per unit area of ​​the strip meets the customer's required process range; the secondary objective is to optimize nitrogen consumption. When the system detects localized excessive thickness anomalies, the controller prioritizes calculating the compensation amount for the air knife pressure, increasing the blowing intensity in the corresponding area to suppress zinc melt accumulation. The system fine-tunes the overall tilt angle of the air knife, utilizing the shear force of the airflow to compensate for excessive edge thickness caused by hydrodynamic effects at the strip edges.

[0044] If the predicted thickness deviation is large and air knife adjustment alone cannot restore it to normal within a controllable time, the control command will trigger the frequency converter to reduce the strip's running speed. This increases the adjustment redundancy and provides a wider response window by extending the air knife's blowing time at the position. All adjustment commands are transmitted in real-time to the air knife regulating valve, angle actuator, and main drive motor via industrial Ethernet or fieldbus. The entire control logic forms a complete data closed loop, achieving intelligent and precise control of the hot-dip galvanizing process through a continuous "sensing-prediction-decision-execution" cycle.

[0045] The deep time-series prediction model supports online incremental learning. When the production line encounters a new and unprecedented abnormal condition, and this condition has been manually verified and labeled, the system adds the sample to the incremental training set, triggering fine-tuning of the model's local parameters. This mechanism enables the control system to continuously evolve, adapting to wear and tear on mechanical components, equipment aging, and drift in process formulations caused by long-term operation of the production line, ensuring the system's control accuracy throughout its entire lifecycle.

[0046] Example 2: Based on Example 1, this example further refines the interactive collaboration logic of acoustic signal processing and visual feature extraction to cope with more extreme low visibility conditions.

[0047] In the above method, the high-sensitivity microphone array described in step 1 employs shock-absorbing connectors for physical installation. Each microphone's pickup head is encased in a high-temperature resistant silicone pad to isolate the mechanical vibration of the air knife holder itself from interference with the air acoustic signals. To ensure data synchronization, the system uses an industrial controller with multi-channel synchronous sampling capabilities. Its internal clock deviation is controlled at the microsecond level, ensuring absolute alignment of the bus sampling of the 8 acoustic signals, 1 image trigger signal, and the process sensor on the time axis.

[0048] To address the acoustic feature extraction in step 2, this embodiment adds a spectral subtraction noise reduction step. During the no-load phase before the air knife is ignited, the system automatically records the power spectrum distribution of the background ambient noise. In the formal production phase, this background noise component is subtracted from the power spectrum of the noisy signal. When the denoised signal undergoes frame processing, a Hamming window is used to reduce frequency component leakage at signal cutoff points. In addition to the Mel-frequency cepstral coefficients, the frequency centroid and frequency divergence are specifically extracted from the spectral feature vector to characterize the stability of the high-pressure airflow at the air knife nozzle. A higher frequency centroid generally indicates a faster nozzle airflow velocity and a stronger corresponding scouring force.

[0049] Regarding the image processing in step 3, when the dust concentration in the production line environment exceeds a preset optical clarity threshold, the system will automatically activate the image contrast stretching algorithm. This algorithm maps relatively concentrated low-contrast pixels to the full-range grayscale space by statistically analyzing the grayscale histogram of the image, thereby enhancing the visual contrast between zinc droplets, zinc slag, and the strip steel substrate.

[0050] In the semantic segmentation stage, this embodiment employs a neural network with a multi-scale pooling module, which can capture defect features on the strip surface under different receptive fields. For the segmented "zinc particle accumulation area," the system calculates its centroid coordinates and equivalent area, and transforms these spatial parameters into part of the visual feature vector.

[0051] In the cross-modal attention fusion process of step 4, this embodiment defines in detail the implementation path of the multi-head cross-attention mechanism. The system projects the feature matrix output by the acoustic encoder into multiple different subspaces, each representing an "attention head". Within each head, the system calculates the attention weights of acoustic features to visual features, as well as the correlation strength of acoustic features to process sensor data. Through this parallel computation, the model can simultaneously capture the nonlinear relationships between different modalities from multiple dimensions. For example, the first head may focus on capturing the coupling between "airflow sound frequency and strip speed", while the second head focuses on the correlation between "visual image brightness and air knife pressure". The outputs of all heads are finally stitched together and passed through a linear projection layer to restore a unified fused perceptual feature.

[0052] In step 5, the GRU network of the deep time-series prediction model is configured as a multi-layer stacked structure. The bottom layer network is responsible for capturing rapid fluctuations on short time scales, such as instantaneous jumps in air knife pressure; the top layer network is responsible for learning long-term process evolution, such as slow drifts in strip temperature. To improve the accuracy of anomaly identification, the anomaly detection module introduces a pattern matching algorithm based on dynamic time warping. This algorithm calculates the minimum alignment cost between the current predicted sequence and the standard process path. When the alignment cost exceeds a first preset threshold, the system triggers a level one warning; when the alignment cost exceeds a more stringent second preset threshold, it is directly determined to be a severe blockage fault.

[0053] In the coordinated adjustment of step 6, the control algorithm incorporates a predictive model control strategy. This strategy considers not only the current error value but also the cumulative error over multiple future prediction steps. When generating instructions, the system constrains the amplitude of the actuator's movements to prevent production line oscillations caused by excessive adjustments. The adjustment of the air knife pressure is quantified as the change in the opening of the pressure valve. By consulting a preset pressure-flow characteristic table, the required increase in blowing force is converted into the actual valve control current. Simultaneously, the fine-tuning instructions for the strip speed take into account the motor's acceleration and deceleration curves to ensure that the strip tension remains constant during speed adjustments, avoiding the risk of strip deviation or breakage due to sudden speed changes.

[0054] Example 3: This example focuses on how to use the method of the present invention to perform fine compensation of process parameters under the hot-dip galvanizing condition of high carbon steel, and details the adaptive online correction mechanism of the system.

[0055] In the above method, the surface roughness parameter of the strip steel before it enters the zinc pot is specifically included in the process sensor data collected in step 1. This parameter is obtained by a laser scanner in the upstream process and is input into the fusion network as a static feature. Because the surface activity of high-carbon steel differs from that of conventional low-carbon steel, its adsorption capacity for zinc liquid is different. Therefore, the acoustic signal exhibits unique spectral envelope characteristics when the strip steel is blown by an air knife.

[0056] For step 2, a real-time spectrum equalization process is added to the acoustic preprocessing. The system automatically adjusts the weights of different frequency bands based on the material type of the strip. For high-carbon steel, because the liquid zinc layer on its surface is prone to generating tiny high-frequency splashes under the action of the air knife, the system will strengthen the monitoring of signals in the 15 kHz to 22 kHz high-frequency band.

[0057] In step 3, the system introduces dynamic masking technology for image segmentation of the strip surface. Since strong specular reflection interference often occurs at the edges of the strip during hot-dip galvanizing, the system automatically generates an edge masking area in the image based on the real-time strip width measured by the process sensor. Pixels within the masking area are processed using a specialized edge detection operator to extract visual features reflecting edge effects. This solves the problem of edge misjudgment caused by light and shadow interference in the visual feature image.

[0058] In step 4, a modal confidence estimator is added within the cross-modal attention fusion network. This estimator calculates the data quality for each modality in real time. If a sensor in the microphone array fails, causing abnormal fluctuations in the variance of the acoustic features, the confidence estimator immediately reduces the weight of that channel during fusion and sends a maintenance prompt to the operation backend. This mechanism ensures that even in the event of hardware failure, the system can still maintain basic control logic using residual sensor information, thus improving the fault tolerance of the entire system.

[0059] In the time-series prediction stage of step 5, this embodiment uses a Long Short-Term Memory network with an attention mechanism instead of the basic GRU network. The output of the prediction model includes not only the future thickness distribution but also a confidence interval. When the predicted value deviates from the target but the confidence interval is very wide, it indicates that the system is currently in an uncertain transitional state. In this case, the controller will adopt a more conservative adjustment strategy to avoid erroneous operations.

[0060] For step 6, a nonlinear compensation step is added to the collaborative adjustment command generation algorithm. The relationship between air knife pressure and final coating thickness is not completely linear, but rather exhibits a negatively correlated exponential relationship. The system uses a built-in nonlinear function to convert the required thickness adjustment into a precise pressure adjustment. When adjusting the angle, the system calculates the change in the vertical distance between the air knife injection point and the strip surface, and adjusts the height of the lifting platform accordingly to maintain the optimal geometric relationship between the air knife and the strip.

[0061] This embodiment details the implementation process of online incremental learning. When the system detects a new anomaly and it is manually verified, the backend server automatically extracts full-modal data within 10 minutes before and after the anomaly occurred. This data is automatically truncated, labeled, and expanded to form a micro-training package. The system uses a lightweight gradient descent algorithm to iterate the model in the background, and the updated weight parameters are sent to the front-end controller via hot-loading technology. The entire fine-tuning process is completed without interrupting production line operation, achieving a smooth upgrade of the control logic. This continuous evolution capability enables the present invention to cope with complex control requirements under different steel grades, different seasonal temperature and humidity environments, and different air knife conditions.

[0062] Example 4: This example describes the specific performance and implementation details of the present invention in dealing with extreme steam interference in the zinc pot area, focusing on the compensation mechanism of acoustic features for visual blind spots.

[0063] In the above method, the industrial camera deployed in step 1 is equipped with a specialized air curtain dust cover. By continuously introducing clean compressed air into the cover, a positive pressure protective air curtain is formed in front of the lens, reducing the condensation of zinc vapor on the lens. However, when the production line is engaged in high-speed continuous production, intense evaporation of molten zinc occurs around the air knife, creating localized blind spots. At this time, the quality of the synchronously acquired strip surface image will be significantly reduced.

[0064] In the parallel processing of steps 2 and 3, the system determines the reliability of visual information by evaluating the entropy value of the image feature map. If the image entropy value drops below a preset sharpness threshold, it indicates that the visual modality has partially failed. At this point, the cross-modal attention fusion network in step 4 plays a crucial role. The cross-attention matrix within the network detects the precipitous drop in visual modality weights, and as compensation, the system automatically extracts low-frequency energy features related to "jetting resistance" from the acoustic feature vector.

[0065] After extensive data training, the system established an acoustic-thickness mapping model, indicating a strong correlation between the sound spectrum at the air knife nozzle and the thickness of the zinc liquid film under varying air knife pressure and strip speed. When vision is unavailable, the system primarily relies on the power spectral density distribution of the acoustic signal to infer the thickness distribution across the strip width. For example, if the amplitude of the acoustic signal suddenly increases in a certain frequency band, and this change does not match the fluctuation of the air knife pressure, the system will combine this with the strip tension fluctuations in the sensor data to determine that local zinc liquid backflow or accumulation has occurred.

[0066] In step 5, the depth time-series prediction model receives this acoustically-driven fusion feature. Because acoustic signals respond almost instantaneously to changes in physical state, the prediction model can capture process disturbances at the millisecond level. Compared to traditional hysteresis control relying solely on thickness gauges, this acoustically-based prediction can detect thickness anomalies approximately 10 to 15 seconds in advance.

[0067] In the coordinated adjustment of step 6, the system precisely controls the segmented regulating valves of the air knife based on the abnormal locations detected acoustically. The air knife contains multiple independent chambers; by adjusting the air intake ratio of each chamber, differentiated control of the blowing force at different transverse positions of the strip can be achieved. The control commands generated by the system are refined to the duty cycle of the solenoid valve in each chamber. By changing the duty cycle, the instantaneous flow rate is adjusted to precisely smooth out locally excessively thick zinc layers. To prevent single-mode drift caused by acoustic sensing, the system periodically uses the calibration values ​​of an infrared thickness gauge to perform benchmark adjustments on the prediction model, ensuring the long-term stability of the closed-loop control.

[0068] Example 5: This example illustrates the operating strategy of the present invention in a specific energy-saving mode, which aims to reduce nitrogen consumption through refined acoustic sensing while ensuring coating quality.

[0069] In the above method, the data collected in step 1 includes the gas supply pressure of the nitrogen main pipe for the air knife and the ambient atmospheric pressure. In energy-saving mode, the system not only focuses on achieving the required thickness but also on maximizing the blowing efficiency.

[0070] In step 2, the extraction of spectral features of the acoustic signal includes an indicator called "harmonic distortion". When the air knife nozzle is at its optimal working gap and the airflow is stable, the harmonic structure of the acoustic signal is clear and regular; when the airflow becomes turbulent or its efficiency decreases, the harmonic distortion increases.

[0071] In step 3, the visual feature map focuses on extracting the morphology of the overflowing zinc liquid at the edge of the strip. If the amount of overflow at the edge is extremely small, it indicates that the current air knife pressure may be redundant and there is room for further reduction.

[0072] During the fusion process in step 4, the system uses attention weights to identify the feature combination most sensitive to "thickness redundancy". The optimal pressure balance point under the current operating conditions is determined by cross-validating the strip speed in the sensing mode with the blowing intensity in the acoustic mode.

[0073] In step 5, the time-series prediction model not only predicts the thickness but also the energy consumption trend under the current adjustment path. If the prediction results show that after reducing the air knife pressure by 2%, the coating thickness can still remain near the upper limit of the process tolerance range within the next 30 seconds, the system will determine that the adjustment scheme is feasible.

[0074] In step 6, the coordinated adjustment command generator outputs a series of stepped decompression commands. After each decompression, the system monitors the airflow status in real time via acoustic signals to determine if it is stable. If the acoustic characteristics show that the harmonic distortion remains within the normal range, it indicates that the decompression has not affected the jetting quality. This micro-adjustment based on acoustic feedback allows the system to operate close to physical limits.

[0075] To compensate for the risk of edge thickening that may result from decompression, the system will fine-tune the deflection angle of the air knife, allowing the airflow energy to be more concentrated on the strip area. The control process in the entire energy-saving mode is dynamic; the system will automatically find a new equilibrium point based on changes in the strip material and linear velocity, achieving intensive utilization of nitrogen resources while ensuring quality.

[0076] Through the detailed description of the above embodiments, this invention demonstrates its powerful sensing capabilities and decision-making accuracy in complex hot-dip galvanizing industrial environments. By deeply fusing heterogeneous data, the system overcomes the limitations of single sensors, elevating traditional experience-based control to intelligent predictive control based on deep learning and multimodal fusion, thereby improving the uniformity of the strip coating and the operational stability of the production line.

[0077] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for controlling the coating of hot-dip galvanized steel strip based on deep learning, characterized in that, Includes the following steps: Step 1: Deploy a microphone array in the zinc pot area of ​​the hot-dip galvanizing production line to collect acoustic signals generated during the air knife blowing of zinc liquid in real time, and simultaneously acquire images of the strip surface and process sensor data; Step 2: Preprocess the acoustic signal, including noise reduction, framing and spectral feature extraction, to generate an acoustic feature vector characterizing the working state of the air knife; Step 3: Enhance and segment the image of the strip surface to extract visual feature maps that reflect the uniformity of the coating; Step 4: Input the acoustic feature vector, visual feature map and process sensor data into the cross-modal attention fusion network, establish dynamic correlation weights among the acoustic, visual and sensing modalities through a multi-head cross-attention mechanism, and generate fused perception features; Step 5: Based on the fused sensing features, use the deep temporal prediction model to deduce the coating thickness distribution trend within future time steps and identify anomalies; Step 6: Based on the coating thickness distribution trend and anomaly identification results, generate coordinated adjustment commands for air knife pressure, angle, and strip running speed, and send them to the actuator for closed-loop control.

2. The method for controlling the coating of hot-dip galvanized strip steel based on deep learning according to claim 1, characterized in that: The microphone array deployed in step 1 consists of multiple omnidirectional industrial-grade microphones. The microphones are arranged in a ring within a predetermined distance around the air knife outlet. Each microphone is wrapped in a high-temperature resistant silicone pad. The microphone's encapsulation structure uses a corrosion-resistant alloy material shell and is equipped with a microporous sound-permeable membrane. The process sensor data includes strip temperature, strip tension, strip linear speed, and air knife nitrogen flow rate; Before being input into the cross-modal attention fusion network, the process sensor data needs to undergo standardization processing. The standardization processing calculates the difference between the observed value and the mean of the process sensor data, divides the difference by the standard deviation, and transforms the data of different dimensions into a unified numerical sequence, which serves as the input modality of the fusion network. This is used to constrain the acoustic feature vector and the visual feature map at the physical logic level.

3. The method for controlling the coating of hot-dip galvanized strip steel based on deep learning according to claim 1, characterized in that: The preprocessing of the acoustic signal in step 2 specifically includes: using an adaptive filtering algorithm based on wavelet threshold to perform multi-scale wavelet decomposition on the original acoustic signal, decomposing it into approximate components and detail components in different frequency ranges, calculating an adaptive threshold based on the distribution characteristics of noise in each component, suppressing noise components smaller than the adaptive threshold, and performing wavelet reconstruction to synthesize a clean acoustic signal. The pure acoustic signal is divided into segments according to a fixed time window length, and an overlapping area is set between adjacent frames; Spectral features are extracted from the segmented signal frames, including Mel frequency cepstral coefficients, short-time energy, zero-crossing rate, spectral centroid, frequency centroid, and frequency divergence. The process of extracting Mel frequency cepstral coefficients includes performing pre-emphasis, windowing, fast Fourier transform, Mel filter bank filtering, and discrete cosine transform on the signal frame, and concatenating all the extracted feature parameters into an acoustic feature vector with fixed dimensions.

4. The method for controlling the coating of hot-dip galvanized steel strip based on deep learning according to claim 1, characterized in that: The process of extracting visual feature maps in step 3 specifically includes: processing the strip surface image using an illumination correction method, estimating the illuminance component by performing multi-scale Gaussian blurring on the original image, and stripping the illuminance component from the original image in the logarithmic domain to obtain the reflection component, and smoothing the thermal noise caused by the environment by combining a nonlocal mean denoising algorithm. The enhanced image is classified at the pixel level using a semantic segmentation neural network. The semantic segmentation neural network adopts a deep convolutional encoding to decoding structure and extracts the texture and edge features of the strip surface through multi-level convolutional layers, dividing the image into normal coating area, zinc particle accumulation area and exposed iron defect area. The output of the semantic segmentation neural network is a visual feature map aligned with the spatial scale of the original image, where the value of each pixel represents the probability that the location belongs to a region category. When the concentration of smoke and dust in the environment exceeds the preset optical clarity threshold, low-contrast pixels are mapped to the full-range grayscale space by statistically analyzing the image grayscale histogram to enhance visual contrast.

5. The method for controlling the coating of hot-dip galvanized strip steel based on deep learning according to claim 1, characterized in that: The process of feature fusion performed by the cross-modal attention fusion network in step 4 specifically includes: using an acoustic encoder, a visual encoder, and a sensor encoder to perform deep feature abstraction on the acoustic modality, visual modality, and sensor data modality, respectively. The acoustic encoder adopts a one-dimensional convolutional network, the visual encoder adopts a residual convolutional architecture, and the sensor encoder adopts a fully connected stacked structure. After feature extraction is completed, a cross-attention mechanism is introduced, using acoustic features as query vectors, visual features and sensor data as key vectors and value vectors, and calculating the correlation matrix among the three. The calculation process includes performing a dot product operation on the query vector and the key vector, scaling the dot product result, and converting it into a weight distribution using a normalized exponential function. The weight distribution represents the degree of contribution of different modal information to judging the coating quality. The features of each modality are weighted and summed according to the weight distribution to generate a unified high-dimensional fusion perception feature; The cross-modal attention fusion network has a modality confidence evaluator that calculates the data quality of each modality in real time. When the sensor shows abnormal fluctuations, the weight of that modality is reduced during fusion.

6. The method for controlling the coating of hot-dip galvanized strip steel based on deep learning according to claim 1, characterized in that: The process of inferring trends in the deep time series prediction model in step 5 specifically includes: using a gated cyclic unit architecture to process sequence data with long-distance dependencies, wherein the gated cyclic unit architecture controls the forgetting and transmission of information by resetting and updating the gates; The fused sensing feature sequence of multiple consecutive historical time steps is input into the deep time series prediction model, and nonlinear mapping is performed in the hidden layer to output the predicted value of the coating thickness distribution along the strip width direction in the future multiple time steps. The anomaly detection module continuously compares the current predicted trend value with the feature benchmark of the pre-stored historical normal working mode to calculate the minimum alignment cost between the current predicted sequence and the standard process path. When the predicted coating thickness deviates from the standard range at a local location and the minimum alignment cost is greater than the preset dynamic threshold, it is determined that there is an abnormal potential for zinc slag clogging the air knife nozzle or excessive local coating thickness. The output of the deep time series prediction model also includes a confidence interval, which indicates the degree of certainty of the current prediction result.

7. The method for controlling the coating of hot-dip galvanized strip steel based on deep learning according to claim 1, characterized in that: The process of generating the coordinated adjustment command in step 6 specifically includes: following the principle of multi-objective optimization, optimizing energy consumption while ensuring that the coating weight per unit area of ​​the strip steel meets the process requirements; When the identification result is a local overthickness anomaly, the compensation amount of the air knife pressure is calculated first to increase the blowing intensity of the corresponding area; Fine-tune the overall tilt angle of the air knife to compensate for the excessive thickness of the strip edge using airflow shearing force; If the predicted thickness deviation exceeds the preset range, the frequency converter will reduce the running speed of the strip and increase the adjustment redundancy by extending the blowing time of the air knife at the position. When generating instructions, the action range of the actuator is constrained. The required increase in blowing force is converted into the change in the opening of the pressure valve by looking up the preset pressure and flow characteristic table. The strip speed is adjusted according to the acceleration and deceleration curve of the motor to maintain constant strip tension. The coordinated adjustment commands are transmitted in real time to the air knife regulating valve, angle actuator, and main drive motor via an industrial fieldbus, forming a data closed loop.

8. The method for controlling the coating of hot-dip galvanized strip steel based on deep learning according to claim 1, characterized in that: The control method also includes an online incremental learning process. When the system determines that a new abnormal working condition has occurred and confirms it, it automatically extracts full-modal data within a predetermined time period before and after the occurrence of the abnormality, and truncates, labels and expands the data to form a training package. A lightweight gradient descent algorithm is used to iterate the cross-modal attention fusion network and the deep temporal prediction model in the background, updating the model's weight parameters. The updated weight parameters are sent to the front-end controller via hot loading, completing the upgrade of the control logic without interrupting the production line operation, in order to adapt to the wear and tear of mechanical parts, equipment aging, and process formula drift caused by long-term operation of the production line.

9. The method for controlling the coating of hot-dip galvanized strip steel based on deep learning according to claim 1, characterized in that: The control method also includes a compensation mechanism for visual blind spots, which determines the reliability of visual information by evaluating the entropy value of the visual feature map during the processing. When the image entropy value is lower than the preset sharpness threshold and the visual modality is partially ineffective, the cross-modal attention fusion network automatically reduces the weight of the visual modality and extracts the low-frequency energy features related to the jet resistance from the acoustic feature vector. Using a pre-trained acoustic-thickness mapping model, the thickness distribution of the strip width is deduced based on the sound spectrum morphology at the air knife nozzle under air knife pressure and strip speed. If the amplitude of the acoustic signal increases in the frequency band and does not match the fluctuation of the air knife pressure, combined with the strip tension fluctuation in the sensor data, it is determined that local zinc liquid backflow or accumulation has occurred, and the segmented regulating valve of the air knife is adjusted accordingly.

10. The method for controlling the coating of hot-dip galvanized strip steel based on deep learning according to claim 1, characterized in that: The control method also includes an energy-saving control mode, in which harmonic distortion index is extracted from the acoustic signal in step 2 to characterize airflow stability; In step 3, the morphology of the overflowing zinc liquid at the edge of the strip is extracted from the visual feature map; In step 4, attention weights are used to identify feature combinations that are sensitive to thickness redundancy, and cross-validation is performed by combining strip speed and blowing intensity to determine the pressure balance point under the current working condition. In step 5, the depth time-series prediction model is used to predict the thickness evolution trend after the air knife pressure is reduced. If the prediction result shows that the coating thickness is still within the process tolerance range, a step-type pressure reduction command is output in step 6. After each decompression, the adjustment effect is fed back by monitoring whether the harmonic distortion of the acoustic signal remains within the normal range, and the air knife deflection angle is finely adjusted during decompression to compensate for the risk of edge thickening.

Images