Grate cooler snowman identification method and system based on Fourier transform residual neural network

Abstract

The invention relates to a grate cooler snowman identification method and system based on a Fourier transform residual neural network. The method comprises the following steps: acquiring an internal image of the grate cooler, and converting the internal image of the grate cooler into a grayscale image; fourier transform is carried out on the grayscale image to obtain a first transformed image; performing fast Fourier transform translation on the first transformed image to obtain a second transformed image; keeping components with the frequency greater than the target cut-off frequency in the second transformation image to obtain a frequency domain image; performing inverse Fourier transform on the frequency domain image to obtain a third transform image; inputting the third transformation image into a residual neural network to output a grate cooler snowman identification result; the residual neural network comprises a plurality of residual processing modules, each residual processing module comprises a plurality of residual blocks connected in sequence, the first residual block in each residual processing module is provided with a convolutional layer with the step length of 2, and jump connection is arranged in each residual block. And accurate identification of the snowman form can be realized.

Description

Technical Field

[0001] Several embodiments of this specification relate to the field of snowman recognition technology for grate coolers, specifically to a snowman recognition method and system for grate coolers based on Fourier transform residual neural networks. Background Technology

[0002] In cement production, the grate cooler is one of the key pieces of equipment, and its operating status directly affects the cooling efficiency and quality of clinker. During the operation of the grate cooler, due to factors such as the moisture content, particle size, and shape of the material, it is easy for material to accumulate on the surface of the grate, a phenomenon known as the "snowman phenomenon." The "snowman phenomenon" in the grate cooler increases the ventilation resistance, reduces the cooling effect, and in severe cases, can even cause equipment shutdown, resulting in significant production losses.

[0003] Traditional methods for identifying "snowmen" in grate coolers mainly rely on manual observation, which suffers from low efficiency, high latency, and inability to quantify. While existing deep learning technologies such as convolutional neural networks and residual neural networks have achieved significant results in image recognition, their application in grate cooler video analysis for "snowman" identification is still insufficient. Summary of the Invention

[0004] This specification provides a method and system for snowman recognition in a grate cooler based on Fourier transform residual neural network. It fully combines Fourier transform and residual neural network and applies it to the "snowman" recognition scenario in a grate cooler to achieve accurate recognition of the "snowman" in the grate cooler.

[0005] The technical solution is as follows: This specification provides an embodiment of a snowman recognition method for a grate cooler based on a Fourier transform residual neural network, including: Acquire images of the interior of the grate cooler and convert them into grayscale images; Perform a Fourier transform on the grayscale image to obtain the first transformed image; The first transformed image is translated by performing a Fast Fourier Transform to obtain the second transformed image; The components in the second transformed image with frequencies greater than the target cutoff frequency are retained to obtain the frequency domain image; Perform an inverse Fourier transform on the frequency domain image to obtain the third transformed image; The third transformed image is input into the residual neural network to output the snowman recognition result of the grate cooler; The residual neural network includes an input layer, multiple residual processing modules, a global average pooling layer, and a fully connected layer connected in sequence. Each residual processing module includes multiple residual blocks connected in sequence. The first residual block in each residual processing module is equipped with a convolutional layer with a stride of 2, and each residual block is equipped with a skip connection.

[0006] As a preferred embodiment, converting the image inside the grate cooler into a grayscale image includes: The image inside the grate cooler is converted to grayscale using a weighted average method to obtain the first processed image; The region of interest is cropped from the first processed image to obtain the second processed image; The second image is scaled to obtain a grayscale image.

[0007] As a preferred embodiment, the target cutoff frequency is determined based on the spectrum corresponding to the second transformed image.

[0008] As a preferred approach, the number of residual blocks in the multiple residual processing modules connected sequentially from front to back increases first and then decreases.

[0009] As a preferred embodiment, each residual block includes a consistent number of convolutional layers, batch normalization layers, and ReLU activation function layers. Each convolutional layer is followed by a batch normalization layer, and each batch normalization layer is followed by a ReLU activation function layer. The last batch normalization layer is connected to the ReLU activation function layer after a skip connection with the input of the residual block.

[0010] Secondly, the embodiments of this specification provide a snowman recognition system for a grate cooler based on a Fourier transform residual neural network, including a first processing module, a second processing module, a third processing module, a fourth processing module, a fifth processing module, and an output module; The first processing module acquires an image of the inside of the grate cooler and converts the image of the inside of the grate cooler into a grayscale image; The second processing module performs a Fourier transform on the grayscale image to obtain a first transformed image; The third processing module performs a fast Fourier transform on the first transformed image to obtain the second transformed image; The fourth processing module retains the components in the second transformed image whose frequency is greater than the target cutoff frequency to obtain a frequency domain image. The fifth processing module performs an inverse Fourier transform on the frequency domain image to obtain the third transformed image; The output module inputs the second transformed image into the residual neural network to output the snowman recognition result of the grate cooler; The residual neural network in the output module includes an input layer, multiple residual processing modules, a global average pooling layer, and a fully connected layer connected in sequence. Each residual processing module includes multiple residual blocks connected in sequence. The first residual block in each residual processing module is equipped with a convolutional layer with a stride of 2, and each residual block is equipped with a skip connection.

[0011] As a preferred embodiment, the first processing module converts the image inside the grate cooler into a grayscale image, including: The image inside the grate cooler is converted to grayscale using a weighted average method to obtain the first processed image; The region of interest is cropped from the first processed image to obtain the second processed image; The second image is scaled to obtain a grayscale image.

[0012] As a preferred option, a target cutoff frequency acquisition module is also included; The target cutoff frequency acquisition module determines the target cutoff frequency based on the spectrum corresponding to the second transformed image.

[0013] As a preferred embodiment, the number of residual blocks in the multiple residual processing modules connected sequentially from front to back in the residual neural network of the output module first increases and then decreases.

[0014] As a preferred embodiment, each residual block in the residual neural network of the output module includes a consistent number of convolutional layers, batch normalization layers, and ReLU activation function layers. Each convolutional layer is followed by a batch normalization layer, and each batch normalization layer is followed by a ReLU activation function layer. The last batch normalization layer is connected to the ReLU activation function layer after a skip connection with the input of the residual block.

[0015] Thirdly, embodiments of this specification provide an electronic device, including a processor and a memory; the processor is connected to the memory; the memory is used to store executable program code; the processor reads the executable program code stored in the memory to run a program corresponding to the executable program code, so as to perform the steps described in the first aspect of the above embodiments.

[0016] Fourthly, embodiments of this specification provide a computer storage medium storing a plurality of instructions adapted for loading by a processor and executing the steps described in the first aspect of the above embodiments.

[0017] The beneficial effects of the technical solutions provided in some embodiments of this specification include at least the following: By preprocessing the grate cooler image using Fourier transform to extract high-frequency features, and then using a residual neural network for deep learning training, accurate recognition of the snowman shape and automatic alarm can be achieved, thereby improving recognition efficiency and accuracy and reducing labor costs. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a snowman recognition method for a grate cooler based on a Fourier transform residual neural network, as provided in the embodiments of this specification. Figure 2 This is a schematic diagram of the residual neural network structure involved in the embodiments of this specification; Figure 3 This is a schematic diagram of the residual block involved in the embodiments of this specification; Figure 4 This is a schematic diagram of a snowman recognition system for a grate cooler based on a Fourier transform residual neural network, provided in the embodiments of this specification. Figure 5 This is a schematic diagram of the structure of an electronic device provided in the embodiments of this specification. Detailed Implementation

[0020] The technical solutions in the embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings.

[0021] The terms "first," "second," "third," etc., in the description, claims, and accompanying drawings are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0022] The following description provides examples and does not limit the scope, applicability, or examples set forth in the claims. Changes may be made to the function and arrangement of the described elements without departing from the scope of this specification. Various processes or components may be appropriately omitted, substituted, or added to the examples. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Furthermore, features described with respect to some examples may be combined into other examples.

[0023] Reference Figure 1 As shown, Figure 1A flowchart illustrating a load sharing method for multiple thermoelectric converter units provided in one embodiment of this specification may include at least: Step 102: Obtain an image of the inside of the grate cooler and convert it into a grayscale image; Step 104: Perform a Fourier transform on the grayscale image to obtain the first transformed image; Step 106: Perform a Fast Fourier Transform on the first transformed image to obtain the second transformed image; Step 108: Retain the components in the second transformed image whose frequencies are greater than the target cutoff frequency to obtain a frequency domain image; Step 110: Perform an inverse Fourier transform on the frequency domain image to obtain the third transformed image; Step 112: Input the third transformed image into the residual neural network to output the snowman recognition result of the grate cooler; The residual neural network includes an input layer, multiple residual processing modules, a global average pooling layer, and a fully connected layer connected in sequence. Each residual processing module includes multiple residual blocks connected in sequence. The first residual block in each residual processing module is equipped with a convolutional layer with a stride of 2, and each residual block is equipped with a skip connection.

[0024] The residual neural network structure can be referenced from Figure 2 As shown in the figure, the example of setting up four residual processing modules is illustrated.

[0025] Currently, there are methods for recognizing and alerting on the shape of snowmen in grate coolers using pure spatial domain deep learning: for example, directly training deep learning models such as CNNs on the original images. Although CNNs can learn features, for targets like grate cooler snowmen that may have specific textures and edge features (such as stacked edges and irregular surfaces), direct learning may require deeper networks and more data, and may be quite sensitive to changes in lighting, background interference, and changes in snowman shape (size, shape, density).

[0026] Currently, there are also methods for recognizing and alerting on the snowman-like shape of grate coolers using traditional image processing and simple classification: for example, relying on traditional image processing methods (such as edge detection and threshold segmentation) to extract features, and then using traditional machine learning methods such as SVM and decision trees for classification. However, these methods often have complex feature engineering, weak generalization ability, and are difficult to cope with complex and ever-changing real-world working conditions, resulting in limited accuracy and robustness.

[0027] Snowmen, often appearing as a mass of material, require precise edge and texture information for accurate recognition. Fourier transform converts the image to the frequency domain, separating high-frequency and low-frequency components. By selectively preserving high-frequency components, we can actively enhance the edge and texture features the snowman may possess, while suppressing low-frequency interference from background, lighting, and other factors. This is crucial for improving the ability of subsequent models to capture key features.

[0028] The complex internal environment of a grate chiller and the diverse shapes of snowmen necessitate a model with strong feature extraction capabilities, often requiring deep networks. However, deep networks are prone to vanishing / exploding gradients and degradation. Residual Neural Networks (ResNet) effectively address these issues by introducing skip connections, making it possible to train very deep networks. Deep networks can learn hierarchical features from low-level (edges) to high-level (overall snowman shape and complex relationships with the background), which is crucial for accurately distinguishing different snowman shapes and differentiating snowmen from the background / normal materials.

[0029] Furthermore, by setting up multiple residual processing modules connected sequentially in the residual neural network, this staged processing can be seen as a process of gradually filtering out irrelevant background information and highlighting the target (the snowman). In the early stages, textures and edges in the background may be captured, but as the network deepens, through downsampling and higher-level feature learning, the model can better focus its attention on those cross-scale features that constitute the key structure of the "snowman," thereby suppressing background interference.

[0030] Although it's a staged process, the skip connections within the residual blocks still allow information to flow between blocks of different depths. This means that even in later stages, the model can indirectly utilize low-level features extracted in earlier stages, which is especially important for processing snowmen that are partially occluded or blend into the background.

[0031] Targeted feature extraction: Fourier transform preprocessing actively extracts and enhances high-frequency features such as edges and textures most relevant to snowman recognition, suppressing interference and allowing ResNet to focus more on learning effective representations of the snowman, thereby improving recognition accuracy. Simultaneously, it reduces sensitivity to changes in lighting and background within the grate cooler, enhancing robustness.

[0032] Powerful deep feature learning capability: ResNet can use these preprocessed features to learn deeper and more abstract snowman representations, thereby distinguishing snowmen of different shapes and understanding their relationship with complex backgrounds.

[0033] Improved recognition efficiency: Although a Fourier transform preprocessing step was added, the enhanced features allowed for the subsequent use of a relatively shallow ResNet or a reduction in the amount of training data, thus improving the overall training efficiency and subsequent recognition efficiency.

[0034] Because snowmen in grate cooling machines come in various shapes and sizes and are located in complex industrial environments, ResNet can effectively capture the key features of the snowmen, including subtle textures and overall structure, through multi-stage downsampling and feature extraction, thus achieving high-precision recognition.

[0035] In some embodiments of this specification, converting the internal image of the grate cooler into a grayscale image includes: The image inside the grate cooler is converted to grayscale using a weighted average method to obtain the first processed image; The region of interest is cropped from the first processed image to obtain the second processed image; The second image is scaled to obtain a grayscale image.

[0036] In the process of converting the grayscale of the internal image of the grate cooler using the weighted average method, the following formula may be used, but is not limited to:

[0037] Where Gray represents the grayscale value, and R, G, and B are the pixel values ​​of the red, green, and blue channels in the image, respectively.

[0038] The first image is cropped to a region of interest. The cropped region needs to be specifically analyzed to ensure that the area where the grate cooler snowman frequently appears can cover the key features of the grate cooler snowman phenomenon.

[0039] The image scaling process for the second image is specifically for facilitating subsequent Fourier transform at a uniform size and for appropriately enlarging and cropping the image.

[0040] Understandably, the Fourier transform is based on the fact that any function can be represented as a sum of sine and cosine functions. For a grate cooler fire image, this means the image can be decomposed into sine waves of different frequencies, each with its specific amplitude and phase. For an image, a two-dimensional FFT is required because the image is a two-dimensional signal.

[0041] In step 104 above, a Fast Fourier Transform (FFT) is applied to the grayscale image to transform it from the spatial domain to the frequency domain. Using the FFT algorithm can significantly improve computational efficiency. A frequency domain image is generated, where each point represents the amplitude and phase information of the corresponding frequency in the image.

[0042] Understandably, in the first transformed image, the high-frequency components typically contain edge and detail information. This can be achieved by setting a high-pass filter.

[0043] Design and apply high-pass filters to filter out low-frequency components and retain high-frequency components. High-pass filter design can be based on ideal filters, Butterworth filters, or Gaussian filters, etc.

[0044] In some embodiments of this specification, the target cutoff frequency of the high-pass filter can also be set based on the image size of the second transformed image, for example, set to about one-tenth of the image size. For example, if the input image size is 224*224, the target cutoff frequency can be, but is not limited to, 20.

[0045] In some embodiments of this specification, the target cutoff frequency of the high-pass filter can also be determined based on the spectrum corresponding to the second transformed image.

[0046] Understandably, a spectrogram visually presents the intensity distribution of each frequency component in an image. By observing the spectrogram, the intensity characteristics of different frequency components can be clearly identified, thus providing an intuitive basis for determining the cutoff frequency. Therefore, the frequency energy distribution in the spectrogram can be observed, and the boundaries of the energy concentration regions can be found as a reference for setting the target cutoff frequency. This can be done, but is not limited to, using the following methods: Multiple extension rays are uniformly extended outward from the center of the spectrum graph corresponding to the second transformed image, and the amplitude value corresponding to each point on each ray is obtained. Based on the amplitude values ​​of each point on each ray, the cutoff radius of each ray is determined. The target cutoff frequency is obtained based on the cutoff radius corresponding to each ray.

[0047] Specifically, the following methods can be used: Starting from the center of the spectrogram, draw rays along multiple angles (0°, 10°, 20°, ..., 350°); For each ray, starting from the center of the spectrum graph and moving outwards along the ray, record the radius and amplitude value of each point; For each ray, determine a cutoff radius (e.g., calculate the point on each ray where the amplitude value decreases by the largest gradient, and determine the cutoff radius based on the radius of this point). Take the average (or median) cutoff radius of all rays as the global cutoff radius (Note: i.e., the target cutoff frequency).

[0048] Furthermore, an inverse Fourier transform (IFFT) can be applied to the processed frequency domain image to convert it back to the spatial domain. IFFT can recover the spatial information of the image.

[0049] Through the above steps, high-frequency features in the grate cooler image can be effectively extracted, providing higher-quality input data for the subsequent residual neural network, thereby improving the recognition efficiency and accuracy of targets with specific textures and edge features (such as stacked edges and irregular surfaces), like the Yeti.

[0050] In some embodiments of this specification, the number of residual blocks in a plurality of residual processing modules connected sequentially from front to back first increases and then decreases.

[0051] By setting the number of residual blocks in this way, increasing the number of blocks in the intermediate layer (key feature extraction stage) allows for more thorough learning and combination of complex features; appropriately reducing the number of blocks in the early and late stages saves computational resources and reduces the risk of overfitting; and allocating computational resources more effectively to the most important feature learning stage can achieve better performance at a lower cost.

[0052] Furthermore, in some embodiments of this specification, each residual block includes a consistent number of convolutional layers, multiple batch normalization layers, and multiple ReLU activation function layers. Each convolutional layer is followed by a batch normalization layer, and each batch normalization layer is followed by a ReLU activation function layer. The last batch normalization layer is connected to the ReLU activation function layer after a skip connection with the input of the residual block.

[0053] The structure of the residual block can be referenced. Figure 3 As shown in the figure (Note: The figure uses 2 convolutional layers, 2 batch normalization layers, and 2 ReLU activation function layers as an example for illustration).

[0054] To better acquire images of the grate cooler's interior, high-temperature, high-definition industrial cameras need to be installed at key locations within the cooler to ensure clear, real-time image capture. A stable and easily maintained transmission interface is also required to transmit the captured images reliably and in real-time. Based on the cooler's structure and common areas, the camera's installation location must be precisely determined to ensure complete coverage without blind spots. Appropriate brackets and mounting methods should be selected to ensure the camera remains stable under high temperature and vibration conditions. After installation, the camera's focus and viewing angle should be adjusted to ensure clear images and full coverage of the required monitoring area.

[0055] Further configure image parameters: Set the camera's image parameters, including resolution, frame rate, exposure time, etc., to adapt to the light and temperature conditions inside the grate cooler.

[0056] Further network configuration is required: Configure the camera's network settings, including IP address, subnet mask, gateway, etc., to ensure that image data can be transmitted to the algorithm server in real time.

[0057] The following is a detailed explanation of how to train ResNet-50: ResNet-50 is a specific implementation of ResNet that contains 50 weight layers. The network structure can be summarized as follows: Initial layer: a 7x7 convolutional layer with a stride of 2, followed by a max pooling layer with a stride of 2.

[0058] Four residual stages: Each stage contains multiple residual blocks, the number of which can be, but is not limited to, 3, 4, 6, and 3 respectively. The first residual block of each stage uses a convolution with a stride of 2 to reduce the size of the feature map.

[0059] Global average pooling: reduces the spatial dimension of the feature map to 1x1.

[0060] Fully connected layer: A 2-way (snowman, no snowman) fully connected layer for classification tasks.

[0061] I. Dataset Preparation A large number of grate cooler images, including images in normal and snowman states, were collected and labeled as the training dataset. Dataset preparation is a crucial step in the training process of the residual neural network in this invention, directly affecting the model training effect and the final recognition performance. The following are the detailed dataset preparation steps: 1.1 Data Collection On-site data acquisition: High-definition cameras are used to acquire real-time image data at the grate cooler's operating site. The cameras are ensured to operate stably under various working conditions, capturing images of the grate cooler in both normal and abnormal states.

[0062] Data diversity: Collect images of grate coolers under different time periods and operating conditions to ensure that the dataset contains diverse samples that reflect the operating status of grate coolers under different conditions.

[0063] Data Scale: Collect enough image data to meet the training needs of deep learning models. Generally speaking, the larger the dataset, the better the model training effect, but the quality and diversity of the data should also be considered.

[0064] 1.2 Data Labeling Labeling Standards: Develop a detailed set of image labeling standards, clearly defining what constitutes a "normal state" and a "snowman state," and how to distinguish between the two. This may require collaboration with grate cooler operators and maintenance personnel to ensure labeling accuracy.

[0065] Manual annotation: Professionals annotate the collected images to determine whether each image contains the Yeti phenomenon. This step is labor-intensive and requires a large amount of manual work.

[0066] Annotation tools: Develop or utilize existing image annotation tools to assist annotation work and improve annotation efficiency and accuracy. These tools can be graphical or programmable to suit different annotation needs. For example, the labelimg tool is commonly used for annotation.

[0067] 1.3 Data Cleaning Quality control: The collected images are inspected for quality, and low-quality images caused by camera malfunction, poor lighting conditions, etc. are removed.

[0068] Standardize the format: Convert all images to a uniform format and size to facilitate subsequent processing and training. This may involve operations such as image cropping and scaling.

[0069] 1.4 Data Augmentation Augmentation techniques: Apply data augmentation techniques, such as rotation, flipping, scaling, and color adjustment, to simulate different shooting angles and lighting conditions, increasing the diversity of the dataset.

[0070] Synthetic data: When necessary, more training samples can be generated through image synthesis techniques, especially when the number of samples for certain categories is insufficient.

[0071] 1.5 Data Partitioning Training, validation, and test sets: The labeled dataset is divided into training, validation, and test sets. A common ratio is 70% training set, 15% validation set, and 15% test set.

[0072] Random partitioning: Ensure that the dataset is partitioned randomly to avoid any potential biases affecting the model training and evaluation results.

[0073] 1.6 Data Storage Organizational structure: Create a good file organization structure on the server or cloud storage to facilitate data management and access.

[0074] Backup: Back up your dataset regularly to prevent data loss or corruption.

[0075] By following the steps above, we can prepare a high-quality, diverse, and accurately labeled grate cooler image dataset, laying a solid foundation for training an efficient residual neural network model.

[0076] II. Online Training 2.1 Setting training parameters Choosing a loss function: Select an appropriate loss function based on the task type (e.g., classification or regression), such as cross-entropy loss or mean squared error loss. Since our snowman recognition task is a classification task, we choose the cross-entropy loss function.

[0077] Optimizer Selection: Choose a suitable optimizer, such as SGD (Stochastic Gradient Descent) or Adam (Adaptive Moment Estimation), to update the network weights. For Snowman Recognition, we use DGD (Stochastic Gradient Descent) as our optimizer during training.

[0078] Setting the learning rate: Determine the initial learning rate and the learning rate adjustment strategy, such as learning rate decay or learning rate warm-up. The initial learning rate for snowman recognition is set to lr=0.001.

[0079] Set batch size: This determines the number of samples in each training batch. Batch size affects the stability of model training and memory consumption. Set Batch_size to 32.

[0080] Setting the training cycle: Determine the maximum number of training cycles (epochs), which is the number of times the entire training set is traversed. Here, for Snowman Recognition, we set epochs=100.

[0081] 2.2 Data Loading and Preprocessing Data loading: The PyTorch framework's data loader is used to load training data from storage media and perform real-time preprocessing.

[0082] Data preprocessing includes operations such as data augmentation and normalization to improve the model's generalization ability and accelerate the training process.

[0083] 2.3 Forward Propagation Input data: Input the preprocessed training data into the residual neural network.

[0084] Output calculation: The network is propagated forward through its hierarchical structure, and the output of each layer is calculated until the final prediction result is obtained.

[0085] 2.4 Calculate the loss Loss calculation: Based on the prediction results and the true labels, the loss value is calculated using a loss function. This value represents the accuracy of the model's prediction.

[0086] 2.5 Backpropagation Gradient calculation: The gradient of the loss function with respect to the network parameters is calculated using the backpropagation algorithm.

[0087] Weight update: Update the weights and biases in the network based on the calculated gradients and the optimizer's policy.

[0088] 2.6 Iterative Training Periodic iteration: Repeat the forward propagation, loss calculation, back propagation and weight update steps until the preset training period is reached or the performance no longer improves.

[0089] Performance monitoring: After each training cycle, evaluate the model's performance on the validation set and monitor whether the model is overfitting.

[0090] 2.7 Model Saving and Early Stopping Model saving: At the end of each period, if the performance on the validation set has improved, the current best model is saved.

[0091] Early stopping strategy: If the performance on the validation set does not improve over several consecutive periods, an early stopping mechanism is triggered to avoid overfitting.

[0092] 2.8 Hyperparameter Adjustment Performance evaluation: Based on the performance feedback on the validation set, adjust hyperparameters such as learning rate, batch size, and network structure to achieve better training results.

[0093] 2.9 Model Fine-tuning Fine-tuning strategies: In the later stages of training, model fine-tuning strategies, such as learning rate decay and weight decay, can be used to refine weight adjustments and improve model performance.

[0094] 2.10 Training completed Final evaluation: After training is complete, the model is evaluated using a test set to determine its generalization ability.

[0095] Model Deployment: The trained model is deployed into the actual grate cooler snowman recognition system for real-time monitoring and recognition.

[0096] Through the above steps, the residual neural network can learn the feature representations of the normal state and the snowman state in the grate cooler image, thereby achieving efficient and accurate snowman recognition.

[0097] III. Model Validation Evaluate the model performance on the validation dataset and adjust the network parameters to achieve the best recognition results.

[0098] The following are the steps in the model validation phase: 3.1 Validation Set Preparation Separate the validation set from the dataset and perform the same preprocessing as the training set.

[0099] 3.2 Performance Evaluation The validation set is used to evaluate the model and calculate key metrics such as accuracy and recall.

[0100] 3.3 Error Analysis Analyze situations where the model makes prediction errors to identify its shortcomings.

[0101] 3.4 Model Adjustment Adjust the model parameters and structure based on the evaluation results, and make fine-tuning adjustments.

[0102] 3.5 Selecting the best model Compare different model versions and select the model that performs best on the validation set.

[0103] 3.6 Final Confirmation Verify model performance on the final validation set and prepare for model deployment.

[0104] IV. Online Snowman Recognition for Grate Coolers After training the model, we will deploy the trained ResNet50 model online for real-time prediction and output the results (0 represents normal, 1 represents a snowman). The detailed steps are as follows: 4.1 Model Conversion and Optimization Model conversion: Convert the trained ResNet50 model into a deployment-friendly format, such as ONNX, so that it can run on different platforms.

[0105] 4.2 Deployment Environment Preparation Hardware selection: Choose the appropriate hardware platform based on actual needs, such as GPU servers or edge computing devices.

[0106] Software configuration: Install the necessary software environment, including deep learning frameworks (such as TensorFlow or PyTorch), dependency libraries, etc.

[0107] 4.3 Model Loading and Service-Oriented Architecture Model loading: Loads the transformed model file in the deployment environment.

[0108] Service encapsulation: Encapsulate the model as a service, enabling it to receive input and return prediction results. This can be achieved using model inference service frameworks such as Triton.

[0109] 4.4 Interface Development API Development: Develop API interfaces that allow the front-end or other systems to interact with the model via HTTP requests.

[0110] Parameter configuration: Configure the input and output parameters of the API interface to ensure consistency with the input and output of the model.

[0111] 4.5 Real-time data stream access Data acquisition: Real-time acquisition of image data from the grate cooler, which can be obtained through cameras or other sensors.

[0112] Data preprocessing: Perform necessary preprocessing on the acquired image data, such as resizing and normalization, to match the model input.

[0113] 4.6 Model Reasoning Request sending: Send the preprocessed data to the model service via API.

[0114] Model prediction: After receiving a request, the model service performs inference calculations and outputs the prediction results.

[0115] 4.7 Results Processing and Feedback Result analysis: The predicted results output by the model are analyzed into specific category labels (0 or 1).

[0116] Feedback output: The prediction results are fed back to the control system or display interface for subsequent processing or monitoring.

[0117] 4.8 Performance Monitoring and Optimization Performance monitoring: Monitor the online performance of the model, including metrics such as response time and accuracy.

[0118] Model optimization: Based on performance monitoring results, the model is optimized and adjusted to improve the accuracy and efficiency of predictions.

[0119] By following the steps above, the trained ResNet50 model can be deployed online for real-time recognition of snowmen and output prediction results.

[0120] V. Automatic Voice Alarm of DCS System The trained ResNet50 model is deployed online for real-time prediction, and the results are output (0 represents normal, 1 represents a snowman). The detailed steps are as follows: 5.1 Write the algorithm output to Kafka Kafka configuration: Configure the Kafka server and create a Topic to receive model prediction results.

[0121] Results push: Push the prediction results (0 or 1) of the ResNet50 model to the specified topic in Kafka in real time.

[0122] 5.2 APC software retrieves Kafka data Data access: The APC (Advanced Process Control) software subscribes to topics that output the model results through a Kafka client and obtains data in real time.

[0123] Data processing: The APC software parses the received data and prepares to write it into the DCS system.

[0124] 5.3 Writing data to DCS DCS Interface: Utilizes the open interface of the DCS system to write the data parsed by the APC software into the DCS system.

[0125] Data storage: In a DCS system, the received prediction results are stored in the corresponding data points or database.

[0126] 5.4 DCS Development of Voice Alarm Function Alarm logic configuration: Configure alarm logic in the DCS system so that a voice alarm is triggered when the model prediction result is 1 (snowman state).

[0127] Voice file preparation: Prepare or record voice files for different alarm levels and alarm messages.

[0128] Alarm point configuration: In the DCS system, all alarm information is configured as LD digital points, and alarm attributes are configured for each digital point.

[0129] 5.5 Alarm Control Room Voice broadcast: When the DCS system triggers an alarm, a voice broadcast will be made through the loudspeaker system in the central control room to remind the operators.

[0130] Alarm confirmation: Operators confirm alarms in the central control room through the DCS interface, and the system records the alarm event and the operator's reaction time.

[0131] By following the above steps, the entire process from model prediction to automatic voice alarm of the DCS system can be realized, ensuring that the status of the grate cooler can be monitored in real time and notified to the operators in a timely manner, thereby improving production safety and efficiency.

[0132] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0133] Please refer to the following. Figure 4 , Figure 4 This specification shows a schematic diagram of a snowman recognition system for a grate cooler based on a Fourier transform residual neural network, which may include at least a first processing module, a second processing module, a third processing module, a fourth processing module, a fifth processing module, and an output module. The first processing module acquires an image of the inside of the grate cooler and converts the image of the inside of the grate cooler into a grayscale image; The second processing module performs a Fourier transform on the grayscale image to obtain a first transformed image; The third processing module performs a fast Fourier transform on the first transformed image to obtain the second transformed image; The fourth processing module retains the components in the second transformed image whose frequency is greater than the target cutoff frequency to obtain a frequency domain image. The fifth processing module performs an inverse Fourier transform on the frequency domain image to obtain the third transformed image; The output module inputs the second transformed image into the residual neural network to output the snowman recognition result of the grate cooler; The residual neural network in the output module includes an input layer, multiple residual processing modules, a global average pooling layer, and a fully connected layer connected in sequence. Each residual processing module includes multiple residual blocks connected in sequence. The first residual block in each residual processing module is equipped with a convolutional layer with a stride of 2, and each residual block is equipped with a skip connection.

[0134] In some embodiments of this specification, the first processing module converts the image inside the grate cooler into a grayscale image, including: The image inside the grate cooler is converted to grayscale using a weighted average method to obtain the first processed image; The region of interest is cropped from the first processed image to obtain the second processed image; The second image is scaled to obtain a grayscale image.

[0135] In some embodiments of this specification, the snowman recognition system further includes a target cutoff frequency acquisition module; The target cutoff frequency acquisition module determines the target cutoff frequency based on the spectrum corresponding to the second transformed image.

[0136] In some embodiments of this specification, the number of residual blocks in the multiple residual processing modules connected sequentially from front to back in the residual neural network of the output module first increases and then decreases.

[0137] In some embodiments of this specification, each residual block in the residual neural network of the output module includes a consistent number of convolutional layers, batch normalization layers, and ReLU activation function layers. Each convolutional layer is followed by a batch normalization layer, and each batch normalization layer is followed by a ReLU activation function layer. The last batch normalization layer is connected to the ReLU activation function layer after a skip connection with the input of the residual block.

[0138] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, the embodiment of the grate cooler snowman recognition system is relatively simple in description because it is fundamentally similar to the embodiment of the grate cooler snowman recognition method; relevant parts can be referred to in the description of the embodiment of the grate cooler snowman recognition method.

[0139] Please see Figure 5 The diagram shown is a structural schematic of an electronic device provided in an embodiment of this specification.

[0140] like Figure 5 As shown, the electronic device 500 may include at least one processor 501, at least one network interface 504, a user interface 503, a memory 505, and at least one communication bus 502.

[0141] The communication bus 502 can be used to realize the connection and communication of the above components.

[0142] The user interface 503 may include buttons, and the optional user interface may also include a standard wired interface or a wireless interface.

[0143] The network interface 504 may include, but is not limited to, Bluetooth modules, NFC modules, Wi-Fi modules, etc.

[0144] The processor 501 may include one or more processing cores. The processor 501 connects to various parts within the electronic device 500 using various interfaces and lines. It executes various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 505, and by calling data stored in the memory 505. Optionally, the processor 501 may be implemented using at least one hardware form of DSP, FPGA, or PLC. The processor 501 may integrate one or more of the following: CPU, GPU, and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed on the screen; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 501 and may be implemented as a separate chip.

[0145] The memory 505 may include RAM or ROM. Optionally, the memory 505 may include a non-transitory computer-readable medium. The memory 505 may be used to store instructions, programs, code, code sets, or instruction sets. The memory 505 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for at least one function (such as touch function, sound playback function, image playback function, etc.), instructions for implementing the above-described method embodiments, etc.; the data storage area may store data involved in the above-described method embodiments, etc. Optionally, the memory 505 may also be at least one storage device located remotely from the aforementioned processor 501. As a computer storage medium, the memory 505 may include an operating system, a network communication module, a user interface module, and a grate cooler snowman recognition application. The processor 501 may be used to call the grate cooler snowman recognition program stored in the memory 505 and execute the grate cooler snowman recognition steps mentioned in the foregoing embodiments.

[0146] This specification also provides a computer-readable storage medium storing instructions that, when executed on a computer or processor, cause the computer or processor to perform one or more steps in the above-described embodiments of the snowman identification method for the grate cooler. If the constituent modules of the above-described electronic device are implemented as software functional units and sold or used as independent products, they can be stored in the computer-readable storage medium.

[0147] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this specification are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in or transmitted through a computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., Digital Versatile Discs (DVDs)), or semiconductor media (e.g., Solid State Disks (SSDs)).

[0148] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks. Unless otherwise specified, the technical features of this embodiment and its implementation can be combined arbitrarily.

[0149] The embodiments described above are merely preferred embodiments of this specification and are not intended to limit the scope of this specification. Any modifications and improvements made by those skilled in the art to the technical solutions of this specification without departing from the spirit of this specification should fall within the protection scope defined by the claims of this specification.

Claims

1. A snowman recognition method for a grate cooler based on Fourier transform residual neural network, characterized in that, include: Acquire images of the interior of the grate cooler and convert them into grayscale images; Perform a Fourier transform on the grayscale image to obtain the first transformed image; The first transformed image is translated by performing a Fast Fourier Transform to obtain the second transformed image; The components in the second transformed image with frequencies greater than the target cutoff frequency are retained to obtain the frequency domain image; Perform an inverse Fourier transform on the frequency domain image to obtain the third transformed image; The third transformed image is input into the residual neural network to output the snowman recognition result of the grate cooler; The residual neural network includes an input layer, multiple residual processing modules, a global average pooling layer, and a fully connected layer connected in sequence. Each residual processing module includes multiple residual blocks connected in sequence. The first residual block in each residual processing module is equipped with a convolutional layer with a stride of 2, and each residual block is equipped with a skip connection.

2. The snowman recognition method for a grate cooler based on Fourier transform residual neural network according to claim 1, characterized in that, The process of converting the internal image of the grate cooler into a grayscale image includes: The image inside the grate cooler is converted to grayscale using a weighted average method to obtain the first processed image; The region of interest is cropped from the first processed image to obtain the second processed image; The second image is scaled to obtain a grayscale image.

3. The snowman recognition method for a grate cooler based on Fourier transform residual neural network according to claim 1, characterized in that, The target cutoff frequency is determined based on the spectrum corresponding to the second transformed image.

4. The snowman recognition method for a grate cooler based on Fourier transform residual neural network according to claim 1, characterized in that, The number of residual blocks set in the multiple residual processing modules connected sequentially from front to back first increases and then decreases.

5. The snowman recognition method for a grate cooler based on Fourier transform residual neural network according to claim 1, characterized in that, Each residual block includes a consistent number of convolutional layers, batch normalization layers, and ReLU activation function layers. Each convolutional layer is followed by a batch normalization layer, and each batch normalization layer is followed by a ReLU activation function layer. The last batch normalization layer is connected to the ReLU activation function layer after a skip connection with the input of the residual block.

6. A snowman recognition system for a grate cooler based on Fourier transform residual neural network, characterized in that, It includes a first processing module, a second processing module, a third processing module, a fourth processing module, a fifth processing module, and an output module; The first processing module acquires an image of the inside of the grate cooler and converts the image of the inside of the grate cooler into a grayscale image; The second processing module performs a Fourier transform on the grayscale image to obtain a first transformed image; The third processing module performs a fast Fourier transform on the first transformed image to obtain the second transformed image; The fourth processing module retains the components in the second transformed image whose frequency is greater than the target cutoff frequency to obtain a frequency domain image. The fifth processing module performs an inverse Fourier transform on the frequency domain image to obtain the third transformed image; The output module inputs the second transformed image into the residual neural network to output the snowman recognition result of the grate cooler; The residual neural network in the output module includes an input layer, multiple residual processing modules, a global average pooling layer, and a fully connected layer connected in sequence. Each residual processing module includes multiple residual blocks connected in sequence. The first residual block in each residual processing module is equipped with a convolutional layer with a stride of 2, and each residual block is equipped with a skip connection.

7. The snowman recognition system for a grate cooler based on Fourier transform residual neural network according to claim 6, characterized in that, The first processing module converts the image inside the grate cooler into a grayscale image, including: The image inside the grate cooler is converted to grayscale using a weighted average method to obtain the first processed image; The region of interest is cropped from the first processed image to obtain the second processed image; The second image is scaled to obtain a grayscale image.

8. The snowman recognition system for a grate cooler based on Fourier transform residual neural network according to claim 6, characterized in that, It also includes a target cutoff frequency acquisition module; The target cutoff frequency acquisition module determines the target cutoff frequency based on the spectrum corresponding to the second transformed image.

9. The snowman recognition system for a grate cooler based on Fourier transform residual neural network according to claim 6, characterized in that, In the residual neural network of the output module, the number of residual blocks in the multiple residual processing modules connected sequentially from front to back first increases and then decreases.

10. The snowman recognition system for a grate cooler based on Fourier transform residual neural network according to claim 6, characterized in that, In the residual neural network of the output module, each residual block includes a consistent number of convolutional layers, batch normalization layers, and ReLU activation function layers. Each convolutional layer is followed by a batch normalization layer, and each batch normalization layer is followed by a ReLU activation function layer. The last batch normalization layer is connected to the ReLU activation function layer after a skip connection with the input of the residual block.

Images