Visual inspection method for monitoring feeding port of conveyor based on AI algorithm

Abstract

The invention relates to the technical field of industrial production, and discloses a visual inspection method for monitoring a feeding port of a conveyor based on an AI algorithm, which comprises the following steps: acquiring original image data of the feeding port by using an image acquisition device mounted at the feeding port of the conveyor, and constructing an image data set for subsequent analysis based on the original image data; and performing preprocessing operation on the image data set, and removing noise and irrelevant interference information in the image through preprocessing to obtain preprocessed image data. Original image data are acquired and the image data set is constructed by using the image acquisition device mounted at the feed port of the conveyor, so that a comprehensive data basis is provided for subsequent analysis, and the problems that limited sample data are mostly adopted in traditional conveyor feed port monitoring methods, and the data are not comprehensive, so that the monitoring accuracy is poor are solved. Therefore, the problem of inaccurate judgment of the state of the feeding port is solved.

Description

Technical Field

[0001] This invention relates to the field of industrial production technology, specifically to a visual inspection method for monitoring the feed port of a conveyor based on AI algorithms. Background Technology

[0002] In industrial production, conveyors are key equipment for material handling, and monitoring the status of the conveyor feed inlet is crucial for ensuring the smooth operation of the entire conveying process. Traditional methods for monitoring conveyor feed inlets mostly rely on manual periodic inspections or simple sensor threshold judgments to monitor material status and equipment operation. However, these methods often only obtain limited information, making it difficult to make comprehensive and accurate judgments about the feed inlet, which can easily lead to incomplete data and inaccurate assessments of the feed inlet status. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a visual detection method for monitoring conveyor feed ports based on AI algorithms. This method improves upon the problem that traditional conveyor feed port monitoring methods mostly rely on limited sample data, which leads to inaccurate judgments of the feed port status due to incomplete data.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a visual inspection method for monitoring the feed port of a conveyor based on an AI algorithm, comprising the following steps: S1. Obtain raw image data of the feed inlet using an image acquisition device installed at the feed inlet of the conveyor, and construct an image dataset for subsequent analysis based on the raw image data; S2. Perform preprocessing on the image dataset to remove noise and irrelevant interference information from the image, and obtain preprocessed image data; S3. Extract key feature information representing the material state and feed port equipment characteristics from the preprocessed image data; S4. Input the key feature information into the pre-trained AI algorithm-based detection model. The AI ​​algorithm-based detection model analyzes and outputs the material status and equipment operation status results at the feed port based on the input key feature information. S5. Based on the output material status and equipment operating status results, compare them with the preset normal status reference standard to determine whether there is an abnormal situation. If an abnormal situation is determined, trigger the alarm mechanism to issue an alarm signal.

[0005] By adopting the above technical solution, the image acquisition device installed at the feed port of the conveyor is used to acquire raw image data and construct an image dataset, thereby providing a comprehensive data foundation for subsequent analysis. This improves the problem that traditional conveyor feed port monitoring methods mostly use limited sample data, which leads to inaccurate judgment of the feed port status due to incomplete data.

[0006] Preferably, step S1 includes: using a camera with appropriate resolution and frame rate to continuously collect images of the feed port of the conveyor at a set time interval, integrating the collected multiple frames of images to form the original image data, and then constructing an image dataset containing multiple frames of images at different times to comprehensively reflect the real-time status and changes of the feed port.

[0007] Preferably, step S2 includes: sequentially converting each frame of the image dataset to grayscale, transforming the color image into a grayscale image, using the formula: ,in Represents grayscale value, These represent the red, green, and blue channel values ​​of the pixels in the color image, respectively. This formula converts the image into a single-channel grayscale image for easier subsequent processing. A filtering algorithm is then used to filter the grayscale image, removing salt-and-pepper noise and Gaussian noise interference. The filtering algorithm can be either median filtering or mean filtering to obtain the preprocessed image data.

[0008] Preferably, step S2 further includes: performing data augmentation operations on the image dataset, the data augmentation operations including rotating, flipping, scaling, adding random noise and then restoring the images to expand the size of the dataset.

[0009] Preferably, step S3 includes: using a feature extraction method based on a convolutional neural network, taking the preprocessed image data as input, performing feature extraction and feature dimensionality reduction operations on the image sequentially through convolutional layers and pooling layers, the parameters of the convolutional kernels in the convolutional layers are obtained through training, and the pooling layers use max pooling or average pooling to reduce the size of the feature map, and finally outputting a feature vector containing key feature information of material shape, texture and feed port equipment outline.

[0010] Preferably, step S4 includes: the detection model based on the AI ​​algorithm adopts a multilayer perceptron structure, the input layer receives the key feature information, after nonlinear transformation of the hidden layer, the output layer outputs the classification results of the material state and equipment operating state at the feed port, and the connection weights between each layer are continuously adjusted and optimized through the backpropagation algorithm during the training process so that the model can accurately analyze and judge based on the input key feature information.

[0011] Preferably, step S4 further includes: the pre-trained detection model based on AI algorithm is trained in the following way: image data of the conveyor feed port under various working conditions are collected, and the accurate material status and equipment operating status corresponding to each image are manually labeled to form a training sample set; The training sample set is divided into a training set and a validation set according to a certain ratio. The initial detection model is trained using the training set. The difference between the model's prediction results and the labeled true results is calculated using the cross-entropy loss function, as shown in the formula: ,in This represents the cross-entropy loss value. The number of training samples. For the number of categories, For the sample Category The true label, For the sample Predicted as category The probability of; The model parameters are adjusted by backpropagation based on the difference calculated by the loss function using the gradient descent algorithm. During the training process, the model is validated using a validation set. When the loss function value on the validation set no longer decreases significantly or the accuracy reaches the preset requirement, training is stopped, and a pre-trained detection model based on the AI ​​algorithm is obtained.

[0012] Preferably, step S4 further includes: performing normalization processing, using the normalization formula as follows: ,in It is a feature value in the original key feature information. This is the mean of the feature across all samples. The standard deviation of this feature across all samples. These are the normalized eigenvalues. Normalization ensures that the eigenvalues ​​are within a suitable range.

[0013] Preferably, step S5 includes: using a large amount of image data of the conveyor feed inlet under normal operating conditions and corresponding manually labeled accurate status information, statistically analyzing and obtaining normal status reference standards for various material status parameters and equipment operating status indicators; comparing the output material status and equipment operating status results with the normal status reference standards one by one; determining an abnormal situation when material status parameters exceed the corresponding normal range or equipment operating status indicators do not meet normal requirements; triggering an audible and visual alarm device to issue a visible and audible alarm signal, and simultaneously transmitting detailed information about the abnormal situation to a remote monitoring terminal via a communication module, so that staff can be informed in a timely manner and take appropriate measures.

[0014] A visual inspection system for monitoring the feed inlet of a conveyor based on an AI algorithm, used in the aforementioned visual inspection method for monitoring the feed inlet of a conveyor based on an AI algorithm, includes the following modules: Image acquisition module: It is installed on the upper side of the conveyor and captures images of the conveyor feed port and materials under the control of the controller to obtain images of the material status at the conveyor feed port; Dataset building module: used to collect images of conveyor belts with normal quantities of material, some material, a small amount of material, almost no material, and no material at all, and to use labeling software to label the material status, generating a dataset containing labeling information for different material statuses; Network model processing module: used for feature fusion of effective feature layers in feature pyramid structure, processing images acquired by image acquisition module, and detecting the status of materials on conveyor; Objective function calculation module: used to optimize the detection accuracy of the network model; Model training and deployment module: This module is used to deploy the best weight file and network model obtained after training on specific hardware devices, so as to realize the application of the model in real-world scenarios and to verify the model performance on mobile devices. The detection result output module is used to output the results after processing by the network model, determine the status of the conveyor feed port, and generate alarm signals or send control commands to relevant equipment when an abnormal status is detected according to preset rules.

[0015] This invention provides a visual inspection method for monitoring the feed inlet of a conveyor based on an AI algorithm. It has the following advantages: 1. This invention utilizes an image acquisition device installed at the feed inlet of a conveyor to acquire raw image data and construct an image dataset, thereby providing a comprehensive data foundation for subsequent analysis. This improves upon the problem that traditional conveyor feed inlet monitoring methods mostly use limited sample data, which leads to inaccurate judgment of the feed inlet status due to incomplete data.

[0016] 2. This invention improves the image data quality by preprocessing the image dataset to remove noise and irrelevant interference information. This addresses the problem that traditional image data, when used directly for analysis, is prone to errors in subsequent feature extraction and state judgment due to noise and other interference factors.

[0017] 3. This invention uses an AI-based detection model to analyze the extracted key feature information and output the results of material status and equipment operating status, thereby achieving accurate status judgment. This improves the problem that traditional conveyor feed port monitoring methods mostly rely on manual experience or simple rule judgments. Due to the lack of intelligent analysis methods, the judgment of feed port status under complex working conditions is inaccurate and untimely.

[0018] 4. This invention determines abnormal situations by comparing the material status and equipment operating status results output by the model with preset normal status reference standards, and triggers an alarm mechanism to issue alarm signals and transmit abnormal information, thereby ensuring that abnormal situations can be known and handled in a timely manner. This improves the problem in traditional conveyor feed port monitoring where it is difficult to notify relevant personnel in a timely manner after an abnormality is detected, and the untimely transmission of information causes the abnormal situation to not be resolved quickly, affecting the normal operation of production. Attached Figure Description

[0019] Figure 1 This is a flowchart of a visual inspection method for monitoring the feed port of a conveyor based on AI algorithms, as proposed in this invention. Figure 2 This is an architecture diagram of a vision inspection system for monitoring the feed port of a conveyor based on AI algorithms, as proposed in this invention. Detailed Implementation

[0020] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] Please see the appendix Figure 1 This invention provides a visual inspection method for monitoring the feed port of a conveyor based on an AI algorithm, comprising the following steps: S1. Obtain raw image data of the feed inlet using an image acquisition device installed at the feed inlet of the conveyor, and construct an image dataset for subsequent analysis based on the raw image data; S2. Perform preprocessing on the image dataset to remove noise and irrelevant interference information from the image, and obtain preprocessed image data; S3. Extract key feature information representing the material state and feed port equipment characteristics from the preprocessed image data; S4. Input the key feature information into the pre-trained AI algorithm-based detection model. The AI ​​algorithm-based detection model analyzes and outputs the material status and equipment operation status results at the feed port based on the input key feature information. S5. Based on the output material status and equipment operating status results, compare them with the preset normal status reference standard to determine whether there is an abnormal situation. If an abnormal situation is determined, trigger the alarm mechanism to issue an alarm signal.

[0022] Specifically, an image acquisition device is installed at the feed inlet of the conveyor to acquire raw image data and construct an image dataset. This dataset comprehensively and intuitively records the actual state of the feed inlet at different times. Using a camera with appropriate resolution and frame rate, images are continuously acquired at set time intervals. Multiple frames are integrated to form a dataset that covers various possible material distributions, equipment configurations, and other conditions. This provides a rich and comprehensive data foundation for subsequent accurate analysis of the feed inlet's status, ensuring that subsequent detection and analysis are based on sufficient sample information and avoiding inaccurate judgments of the feed inlet's status due to missing or incomplete data.

[0023] Preprocessing image datasets offers several positive benefits. First, grayscale conversion transforms color images into grayscale images. Using a specific grayscale value calculation formula, multi-channel color pixels are converted into single-channel grayscale representations, simplifying the image data structure, reducing the amount and complexity of subsequent processing, and highlighting key information such as texture and contours, thus facilitating subsequent feature extraction. Next, filtering algorithms remove interference such as salt-and-pepper noise and Gaussian noise, effectively improving image quality and preventing noise-induced deviations in feature extraction. This ensures that the extracted features reflect the true state of the materials and equipment, thereby guaranteeing the accuracy of the entire visual inspection method. Analysis based on the processed image data more accurately reflects the actual situation at the feed port.

[0024] By employing a feature extraction method based on convolutional neural networks to extract key feature information from preprocessed image data, deep and representative features can be uncovered. Through convolutional layers utilizing kernel parameters learned during training, key details such as the shape and texture of materials and the outline of the feed inlet can be automatically captured in the image. These features are crucial indicators of the material status at the feed inlet and the operational status of the equipment. Pooling layers employ appropriate pooling methods to reduce the feature map size, minimizing data volume while preserving the most critical feature information. This further improves the efficiency of subsequent analysis and processing, making the extracted feature vectors more targeted and effective, thus enhancing the detection model's ability to accurately determine the feed inlet status based on these features.

[0025] The extracted key feature information is input into a pre-trained AI-based detection model, which utilizes its multilayer perceptron structure to achieve effective state analysis. After receiving the key feature information, the input layer performs nonlinear transformations in the hidden layer to model complex feature relationships and uncover deep-seated correlations between features. This results in accurate classification results of the material state and equipment operating status at the feed port being output at the output layer. Furthermore, the connection weights between layers are continuously adjusted and optimized through backpropagation during training, enabling the model to continuously self-adjust and learn based on the input feature information to adapt to the characteristics of the feed port state under different operating conditions. This improves the accuracy of judging the material state and equipment operating status, achieving precise monitoring of the overall operation of the feed port.

[0026] Based on a large amount of image data of the conveyor feed inlet under normal operating conditions and corresponding manually labeled accurate status information, statistical analysis is performed to derive a reference standard for normal conditions. The material status and equipment operating status results output by the detection model are then compared with this standard to reliably determine abnormal situations. When material status parameters exceed the corresponding normal range or equipment operating status indicators fail to meet normal requirements, the system can accurately identify abnormalities at the feed inlet, such as material blockage, insufficient material, or equipment malfunction. Triggering an audible and visual alarm device emits a visible and audible alarm signal, promptly alerting on-site personnel to the abnormal condition at the feed inlet. Simultaneously, a communication module transmits detailed information about the abnormal situation to a remote monitoring terminal, allowing personnel to fully understand the situation and quickly take appropriate measures. This helps reduce production delays and equipment damage caused by untimely handling of feed inlet abnormalities, ensuring the stable operation of the conveyor system.

[0027] Step S1 includes: using a camera with appropriate resolution and frame rate to continuously collect images of the conveyor feed port at set time intervals, integrating the collected multiple frames of images to form raw image data, and then constructing an image dataset containing multiple frames of images at different times to comprehensively reflect the real-time status and changes of the feed port.

[0028] Specifically, an appropriate resolution ensures that the acquired images clearly present the detailed features of the material at the feed inlet and the relevant shape of the equipment, avoiding the loss of key information due to too low a resolution, or the excessive data volume due to too high a resolution, which increases the burden on subsequent processing. An appropriate frame rate ensures that enough image frames are acquired per unit time, which can fully capture the dynamic changes of the material at the feed inlet without generating too much redundant data due to an excessively high frame rate.

[0029] By continuously acquiring images at set time intervals, the acquired images can be continuously and evenly distributed in the time dimension, thus comprehensively covering the state of the feed port at different times.

[0030] The collected images are integrated to form raw image data, and then an image dataset containing multiple frames of images at different times is constructed. The significance of this is that by integrating images from multiple times, various possible material distributions, equipment forms, material flow states, etc. can be covered, providing rich and comprehensive data samples for subsequent steps. This ensures that subsequent operations such as image preprocessing, feature extraction, and model-based state analysis have sufficient basis, thereby ensuring that the analysis results based on this data can reflect the actual real-time status and changes of the feed port as accurately as possible, avoiding deviations in the judgment of the feed port status due to missing or incomplete data.

[0031] Step S2 includes: sequentially converting each frame of the image dataset to grayscale, transforming the color image into a grayscale image, using the following formula: ,in Represents grayscale value, These represent the red, green, and blue channel values ​​of the pixels in the color image, respectively. This formula converts the image into a single-channel grayscale image for easier subsequent processing. A filtering algorithm is then used to filter the grayscale image, removing salt-and-pepper noise and Gaussian noise interference. Median filtering or mean filtering algorithms can be used to obtain the preprocessed image data.

[0032] Specifically, in the grayscale processing section, each frame of the image dataset is sequentially converted to grayscale. Based on the corresponding grayscale value calculation formula, the color image is transformed into a single-channel grayscale image. This operation simplifies the image data structure and reduces the data dimensionality and volume in subsequent processing. A color image contains red, green, and blue channel information. After conversion to grayscale, only the grayscale values ​​that reflect the overall brightness and contour features of the image are retained. This allows subsequent processing to focus on more critical and representative image features, avoiding the increased processing complexity and computational load caused by multi-channel data complexity. It also facilitates the efficient execution of subsequent operations such as filtering and feature extraction, laying a simplified and effective data foundation for the entire visual inspection process.

[0033] In the filtering section, filtering algorithms are used to filter the grayscale image to remove common noise interference such as salt-and-pepper noise and Gaussian noise. Salt-and-pepper noise manifests as randomly occurring black and white pixels in the image, while Gaussian noise causes pixel values ​​to exhibit random variations following a Gaussian distribution. This noise can interfere with the accurate representation of material and equipment features in the image, leading to deviations in subsequent feature extraction and status analysis. Median filtering effectively removes salt-and-pepper noise by replacing the center pixel value with the median value within the pixel's neighborhood; mean filtering reduces the impact of Gaussian noise by replacing the center pixel with the average value calculated within the pixel's neighborhood. These filtering operations improve image quality, ensuring that the extracted image features are based on accurate and clean image information. This, in turn, guarantees the accuracy of the judgments made on the conveyor feed port status based on the image data, enabling the entire visual inspection method to function more reliably.

[0034] Step S2 also includes performing data augmentation operations on the image dataset, including but not limited to rotating, flipping, scaling, adding random noise and then restoring the images to expand the size of the dataset.

[0035] Specifically, in terms of expanding the dataset size, data augmentation operations such as rotation, flipping, scaling, and adding random noise before restoration can effectively expand the dataset size. The original image dataset may be relatively limited in number, but these operations can generate more images of different shapes based on existing images. For example, rotation can simulate images taken from different angles, flipping can produce new images that are symmetrical to the original horizontally or vertically, scaling can reflect the effects of images taken from different distances, and adding random noise before restoration can enrich the sample performance of images under different noise environments. In this way, the number of samples in the dataset increases, providing the model with richer data for training, avoiding insufficient model learning due to insufficient data, and thus laying the foundation for training a more accurate and robust model.

[0036] In terms of improving model generalization ability, data augmentation helps to enhance this ability. In real-world applications, the environment at the conveyor feed inlet and the state of the materials are complex and diverse. The model needs to be able to accurately judge various situations that have not been fully encountered in the training set. The diverse image samples generated through data augmentation allow the model to encounter image features from different angles, shapes, and with varying noise interference during training. This enables it to learn more general and universal feature patterns, rather than being limited to the specific situations reflected in the original dataset. As a result, when the model is applied to actual conveyor feed inlet monitoring, it can more accurately extract features and perform analysis and judgment on image inputs under various actual working conditions, reducing judgment errors caused by changes in working conditions and improving the applicability and reliability of the AI-based visual detection method for conveyor feed inlet monitoring in different real-world environments.

[0037] Step S3 includes: using a feature extraction method based on convolutional neural networks, taking preprocessed image data as input, performing feature extraction and feature dimensionality reduction operations on the image sequentially through convolutional layers and pooling layers. The parameters of the convolutional kernels in the convolutional layers are learned through training, and the pooling layers use max pooling or average pooling to reduce the size of the feature map. Finally, the output is a feature vector containing key feature information of material shape, texture, and feed port equipment outline.

[0038] Specifically, using preprocessed image data as input ensures the quality of the input data and avoids interference factors such as noise in the image affecting the accuracy of feature extraction. Based on convolutional neural networks, features are extracted using convolutional layers. The parameters of the convolutional kernels are learned through training, enabling the convolutional layers to automatically capture feature information at different levels and dimensions in the image. Especially for shape features reflecting the material state, texture features, and contour features reflecting the condition of the feeding port equipment, key features can be adaptively extracted based on the data characteristics without the need for manually designing complex feature extraction rules, greatly improving the efficiency and accuracy of feature extraction.

[0039] Next, feature dimensionality reduction is performed by using max pooling or average pooling to reduce the feature map size. Max pooling selects the maximum value within a local region as the representative, preserving the most prominent features in the image, such as prominent texture features at material edges. Average pooling, on the other hand, summarizes the features of a local region by calculating the average value, effectively summarizing the overall features. This feature dimensionality reduction operation reduces the amount of data while highlighting key features, avoiding problems such as excessive computational resource consumption and slow processing speed caused by excessive data volume in subsequent processing. It also retains key information sufficient to distinguish different material states and equipment operating states, enabling the final output feature vector containing key feature information such as material shape, texture, and feed port equipment outline to provide accurate and efficient input for subsequent AI-based detection models. This helps to accurately determine the material state and equipment operating state at the conveyor feed port, improving the reliability and effectiveness of the entire visual inspection method.

[0040] Step S4 includes: The detection model based on AI algorithm adopts a multilayer perceptron structure. The input layer receives key feature information, which is then transformed by the nonlinear transformation of the hidden layer. The output layer outputs the classification results of the material status and equipment operation status at the feed port. The connection weights between each layer are continuously adjusted and optimized through the backpropagation algorithm during the training process so that the model can accurately analyze and judge based on the input key feature information.

[0041] Specifically, regarding the multilayer perceptron structure, a multilayer perceptron structure is adopted. Its input layer receives key feature information extracted from the previous step, enabling the model to directly acquire key feature data related to the material and equipment at the conveyor feed port, laying the foundation for subsequent analysis and judgment. The hidden layer in the middle performs in-depth processing of the input feature information through nonlinear transformation. Since there are often complex nonlinear relationships between the material state at the feed port and the equipment operating state in reality, the nonlinear transformation capability of the hidden layer can effectively model these complex relationships, uncovering deep-seated correlations and patterns between feature information. This allows the model to better understand and analyze the meaning implied by these features in a way that closely reflects the actual situation.

[0042] In terms of output classification results, the output layer outputs the classification results of the material status and equipment operation status of the feed port, realizing the transformation from input feature information to specific and identifiable status results. It can intuitively present the current material quantity at the feed port, the material accumulation pattern, and whether the equipment has operating statuses such as belt misalignment or feed port blockage. This provides a clear basis for further judgment on whether there are any abnormalities and meets the core requirement of effectively monitoring the feed port of the conveyor.

[0043] Regarding the optimization of connection weights, the connection weights between each layer are continuously adjusted and optimized through the backpropagation algorithm during training, a crucial mechanism. During training, based on the given accurately labeled training samples, the model uses the backpropagation algorithm to adjust the connection weights of each layer according to the difference between the output results and the actual labels. This allows the model to continuously learn and adapt to the feature performance under different working conditions, continuously optimizing its ability to analyze and judge input feature information. This improves the accuracy of judging the material status at the feed port and the operating status of the equipment, ensuring that the entire AI-based visual detection method for monitoring the conveyor feed port can function reliably and accurately, avoiding misjudgments of the actual situation at the feed port due to inaccurate model judgments.

[0044] Step S4 also includes: The pre-trained AI algorithm-based detection model is trained in the following way: image data of the conveyor feed port under various working conditions are collected, and the accurate material status and equipment operating status corresponding to each image are manually labeled to form a training sample set; The training sample set is divided into a training set and a validation set according to a certain ratio. The initial detection model is trained using the training set. The difference between the model's prediction results and the labeled true results is calculated using the cross-entropy loss function, as shown in the formula: ,in The number of training samples. For the number of categories, For the sample Category The true label, For the sample Predicted as category The probability of; The model parameters are adjusted by backpropagation based on the difference calculated by the loss function using the gradient descent algorithm. During the training process, the model is validated using a validation set. When the loss function value on the validation set no longer decreases significantly or the accuracy reaches the preset requirement, training is stopped, and a pre-trained detection model based on the AI ​​algorithm is obtained.

[0045] Specifically, regarding the construction of the training sample set, image data of the conveyor feed port under various operating conditions were collected, and the accurate material state and equipment operating state corresponding to each image were manually labeled to form the training sample set. This approach can comprehensively cover various situations that may occur in actual applications. The image data under different operating conditions includes various situations such as different material quantities, different material accumulation shapes, and different equipment operating states. Through accurate manual labeling, clear learning objectives and judgment criteria are provided for the model, enabling the model to learn based on rich and realistic samples that reflect the actual situation. This avoids the situation where the learned features and judgment rules cannot adapt to complex and diverse actual operating conditions due to single or one-sided samples. Thus, a comprehensive and reliable data foundation is provided for training a model that can accurately cope with various conveyor feed port states.

[0046] Regarding the division of training and validation sets, the training sample set is divided into training and validation sets according to a certain ratio. This division method helps to scientifically train and evaluate the model. The initial detection model is trained using the training set, allowing the model to learn how to judge the material status and equipment operating status from the input feature information based on the data. The validation set, independent of the training process, is used to evaluate the model's performance at different stages during training. By performing training and validation tasks separately on different data subsets, overfitting of the model to the training set data can be effectively prevented. This ensures that the model adapts to the feature patterns of the training data while also possessing the ability to accurately judge new data not used in training, guaranteeing good generalization performance and enabling it to function stably and accurately in actual conveyor feed port monitoring scenarios.

[0047] Regarding the use of the cross-entropy loss function, it is employed to calculate the difference between the model's predicted results and the labeled true results. Mathematically, this function effectively measures the degree of inconsistency between the model's predicted output and the true labels. For classification tasks, a smaller cross-entropy loss function value indicates that the model's predicted results are closer to the actual situation. Using this as a basis to measure model performance provides a clear direction for model training optimization, guiding the model to continuously adjust parameters during training, learning towards making the predicted results more closely match the true labels, thereby improving the model's accuracy in classifying material states and equipment operating states.

[0048] Regarding parameter adjustment based on gradient descent and backpropagation, the model's parameters are adjusted via backpropagation using the gradient descent algorithm based on the differences calculated from the loss function. This mechanism enables automatic model optimization. The gradient descent algorithm determines the parameter update direction that reduces the loss function value based on the gradient information of the loss function under the current parameters. The backpropagation algorithm efficiently transmits the gradient information of the loss function with respect to the output layer to each layer from back to front, thereby accurately calculating the gradient corresponding to the parameters of each layer for updating the parameters of each layer. Through this iterative update process, the model can continuously self-optimize, gradually improving its ability to analyze and judge input features, enabling it to output more accurate material status and equipment operating status results when faced with different conveyor feed port image feature inputs.

[0049] Regarding the use of a validation set to stop training, the model is validated using a validation set during training. Training is stopped when the loss function value on the validation set no longer decreases significantly or the accuracy reaches a preset requirement. This strategy ensures the rationality and effectiveness of model training. Continuous training may lead to overfitting of the training data, resulting in poor performance on new, unseen data. By observing changes in the validation set metrics and stopping training in a timely manner, it ensures that the model has learned enough effective features and patterns while maintaining good generalization ability. The resulting pre-trained AI-based detection model can then stably and accurately analyze and output reasonable material and equipment operating status results based on the input key feature information when used for visual inspection of conveyor feed ports, ensuring the reliability and effectiveness of the entire visual inspection method.

[0050] Step S4 also includes: normalization processing, using the following normalization formula: ,in It is a feature value in the original key feature information. This is the mean of the feature across all samples. The standard deviation of this feature across all samples. These are the normalized feature values. Normalization ensures that the feature values ​​are within a suitable range, which helps improve the efficiency and accuracy of model training and detection.

[0051] Specifically, regarding data range adjustment, a given normalization formula is used to process the original key feature information, mapping each feature value to a suitable range. In actual conveyor feed port monitoring scenarios, the value ranges of different key feature information may vary significantly, with some feature values ​​having large ranges and others small ranges. Through normalization, based on the mean and standard deviation of the feature across all samples, the original feature values ​​are transformed into normalized feature values, ensuring that different features are within a relatively uniform and standardized range. This avoids numerical instability issues that may arise during subsequent model training and detection due to excessively large differences in feature value ranges. For example, some excessively large or small feature values ​​may dominate gradient updates during calculation, affecting the overall learning and judgment performance of the model.

[0052] In terms of improving model training efficiency, normalization helps to enhance training efficiency. During model training, calculations based on normalized feature values ​​within an appropriate range allow optimization algorithms to operate more smoothly and efficiently. For example, when calculating gradients to update model parameters, if feature values ​​are not normalized, features of different magnitudes can lead to vastly different gradient changes, making it difficult to control the step size of parameter updates during training, potentially resulting in slow convergence or even failure to converge. Normalized feature values, on the other hand, ensure a more balanced impact of each feature on the gradient during training, helping the optimization algorithm update parameters with a more reasonable step size, accelerating model convergence, reducing training time and computational resources, and improving overall training efficiency.

[0053] In terms of improving model detection accuracy, normalization also plays a positive role. During the detection phase, the model needs to accurately determine the material state and equipment operating status at the conveyor feed port based on the input feature information. Normalized feature values ​​allow the model to better learn the inherent relationships and patterns between different features, avoiding the masking of true correlations between features due to inconsistent feature value ranges. Thus, when the model analyzes and judges based on the processed features, it can more accurately capture subtle differences in features under different states, thereby improving the accuracy of judging the material state and equipment operating status, and ensuring that the entire AI-based visual detection method for conveyor feed port monitoring outputs reliable detection results.

[0054] Step S5 includes: First, using a large amount of image data from the conveyor feed inlet under normal operating conditions and corresponding manually labeled accurate status information, statistical analysis is performed to derive normal status reference standards for various material status parameters and equipment operating status indicators. Then, the output material status and equipment operating status results are compared one-to-one with the normal status reference standards. If a material status parameter exceeds the corresponding normal range or an equipment operating status indicator fails to meet normal requirements, an abnormal situation is determined. Finally, an audible and visual alarm device is triggered to issue a visible and audible alarm signal, and detailed information about the abnormal situation is transmitted to the remote monitoring terminal via the communication module so that staff can be promptly informed and take appropriate measures.

[0055] Specifically, in determining the reference standards for normal operation, a large amount of image data of the conveyor feed inlet under normal operating conditions, along with corresponding manually labeled accurate status information, is used to statistically analyze and derive the reference standards for various parameters of the material status and various indicators of the equipment operation status. This process provides a reliable and objective reference for subsequent anomaly identification. Determining these standards based on sufficient normal operating data comprehensively covers the reasonable range of material quantity and form under normal operating conditions, as well as the reasonable range of equipment operating indicators. This ensures that subsequent comparative judgments have accurate and realistic references, avoiding misjudgments that may be caused by subjectively set standards, and guaranteeing the scientific rigor and accuracy of the entire judgment process.

[0056] The anomaly detection section compares the output material status and equipment operating status results with normal status reference standards. When material status parameters exceed the corresponding normal range or equipment operating status indicators fail to meet normal requirements, an anomaly is identified. This rigorous comparison mechanism accurately identifies deviations from normal conditions during actual feed port operation. Whether it's excessive or insufficient material, abnormal material accumulation patterns, or equipment issues such as belt misalignment or feed port blockage, these problems can be effectively detected by this comparison mechanism. This ensures accurate capture of abnormal feed port operation, meeting the core requirement of a monitoring system for timely problem detection.

[0057] In the alarm and information transmission section, the triggered audible and visual alarm device emits a visible and audible alarm signal, immediately alerting nearby staff to any abnormalities at the feed inlet. This ensures staff are quickly aware of the anomaly and prevents further escalation due to delayed detection. Simultaneously, detailed information about the anomaly is transmitted to a remote monitoring terminal via a communication module. This allows off-site staff to gain a comprehensive and timely understanding of the anomaly, facilitating coordinated action and remote command of on-site personnel for troubleshooting and equipment parameter adjustments. Overall, this enhances the timeliness and effectiveness of responding to anomalies, ensuring the conveyor system quickly returns to normal operation and minimizing potential production interruptions and equipment damage caused by feed inlet anomalies. This fully leverages the effectiveness of the visual inspection method in monitoring the conveyor feed inlet and ensuring its stable operation.

[0058] Please see the appendix Figure 2 A visual inspection system for monitoring conveyor feed inlets based on AI algorithms, comprising the following modules: Image acquisition module: It is installed on the upper side of the conveyor and captures images of the conveyor feed port and materials under the control of the controller to obtain images of the material status at the conveyor feed port; Dataset building module: used to collect images of conveyor belts with normal quantities of material, some material, a small amount of material, almost no material, and no material at all, and to use labeling software to label the material status, generating a dataset containing labeling information for different material statuses; Network model processing module: used for feature fusion of effective feature layers in feature pyramid structure, processing images acquired by image acquisition module, and detecting the status of materials on conveyor; Objective function calculation module: used to optimize the detection accuracy of the network model; Model training and deployment module: This module is used to deploy the best weight file and network model obtained after training on specific hardware devices, so as to realize the application of the model in real-world scenarios and to verify the model performance on mobile devices. The detection result output module is used to output the results after processing by the network model, determine the status of the conveyor feed port, and generate alarm signals or send control commands to relevant equipment when an abnormal status is detected according to preset rules.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A visual inspection method for monitoring the feed inlet of a conveyor based on AI algorithms, characterized in that, Includes the following steps: S1. Obtain raw image data of the feed inlet using an image acquisition device installed at the feed inlet of the conveyor, and construct an image dataset for subsequent analysis based on the raw image data; S2. Perform preprocessing on the image dataset to remove noise and irrelevant interference information from the image, and obtain preprocessed image data; S3. Extract key feature information representing the material state and feed port equipment characteristics from the preprocessed image data; S4. Input the key feature information into the pre-trained AI algorithm-based detection model. The AI ​​algorithm-based detection model analyzes and outputs the material status and equipment operation status results at the feed port based on the input key feature information. S5. Based on the output material status and equipment operating status results, compare them with the preset normal status reference standard to determine whether there is an abnormal situation. If an abnormal situation is determined, trigger the alarm mechanism to issue an alarm signal.

2. The visual inspection method for monitoring the feed port of a conveyor based on an AI algorithm as described in claim 1, characterized in that, Step S1 includes: using a camera with appropriate resolution and frame rate to continuously collect images of the feed port of the conveyor at set time intervals, integrating the collected multiple frames of images to form the original image data, and then constructing an image dataset containing multiple frames of images at different times to comprehensively reflect the real-time status and changes of the feed port.

3. The visual inspection method for monitoring the feed port of a conveyor based on an AI algorithm according to claim 1, characterized in that, Step S2 includes: sequentially converting each frame of the image dataset to grayscale, transforming the color image into a grayscale image, using the following formula: ,in Represents grayscale value, These represent the red, green, and blue channel values ​​of the pixels in the color image, respectively. This formula converts the image into a single-channel grayscale image for easier subsequent processing. A filtering algorithm is then used to filter the grayscale image, removing salt-and-pepper noise and Gaussian noise interference. The filtering algorithm can be either median filtering or mean filtering to obtain the preprocessed image data.

4. The visual inspection method for monitoring the feed port of a conveyor based on an AI algorithm according to claim 1, characterized in that, The S2 step further includes: performing data augmentation operations on the image dataset, including rotating, flipping, scaling, adding random noise and then restoring the images to expand the size of the dataset.

5. The visual inspection method for monitoring the feed port of a conveyor based on an AI algorithm according to claim 1, characterized in that, The S3 step includes: using a feature extraction method based on a convolutional neural network, taking the preprocessed image data as input, performing feature extraction and feature dimensionality reduction operations on the image sequentially through convolutional layers and pooling layers, the parameters of the convolutional kernels in the convolutional layers are obtained through training, and the pooling layers use max pooling or average pooling to reduce the size of the feature map, and finally outputting a feature vector containing key feature information of material shape, texture and feed port equipment outline.

6. The visual inspection method for monitoring the feed port of a conveyor based on an AI algorithm according to claim 1, characterized in that, The S4 step includes: the detection model based on the AI ​​algorithm adopts a multilayer perceptron structure. The input layer receives the key feature information, and after nonlinear transformation of the hidden layer, the output layer outputs the classification results of the material state and equipment operation state at the feed port. The connection weights between each layer are continuously adjusted and optimized through the backpropagation algorithm during the training process so that the model can accurately analyze and judge based on the input key feature information.

7. The visual inspection method for monitoring the feed port of a conveyor based on an AI algorithm according to claim 1, characterized in that, The S4 step also includes: the pre-trained detection model based on AI algorithm is trained in the following way: image data of the conveyor feed port under various working conditions are collected, and the accurate material status and equipment operating status corresponding to each image are manually labeled to form a training sample set; The training sample set is divided into a training set and a validation set according to a certain ratio. The initial detection model is trained using the training set. The difference between the model's prediction results and the labeled true results is calculated using the cross-entropy loss function, as shown in the formula: ,in This represents the cross-entropy loss value. The number of training samples. For the number of categories, For the sample Category The true label, For the sample Predicted as category The probability of; The model parameters are adjusted by backpropagation based on the difference calculated by the loss function using the gradient descent algorithm. During the training process, the model is validated using a validation set. When the loss function value on the validation set no longer decreases significantly or the accuracy reaches the preset requirement, training is stopped, and a pre-trained detection model based on the AI ​​algorithm is obtained.

8. The visual inspection method for monitoring the feed port of a conveyor based on an AI algorithm according to claim 1, characterized in that, Step S4 further includes: normalization processing, using the following normalization formula: ,in It is a feature value in the original key feature information. This is the mean of the feature across all samples. The standard deviation of this feature across all samples. These are the normalized eigenvalues. Normalization ensures that the eigenvalues ​​are within a suitable range.

9. The visual inspection method for monitoring the feed port of a conveyor based on an AI algorithm according to claim 1, characterized in that, Step S5 includes: Firstly, by statistically analyzing a large amount of image data of the conveyor feed inlet under normal operating conditions and corresponding manually labeled accurate status information, to derive normal status reference standards for various material status parameters and equipment operating status indicators; secondly, comparing the output material status and equipment operating status results with the normal status reference standards one by one; thirdly, determining an abnormal situation when material status parameters exceed the corresponding normal range or equipment operating status indicators do not meet normal requirements; fourthly, triggering an audible and visual alarm device to issue a visible and audible alarm signal, and simultaneously transmitting detailed information about the abnormal situation to a remote monitoring terminal via a communication module, so that staff can be promptly informed and take appropriate measures.

10. A visual inspection system for monitoring the feed inlet of a conveyor based on AI algorithms, characterized in that, A visual inspection method for monitoring the feed port of a conveyor based on an AI algorithm, as described in any one of claims 1-9, comprises the following modules: Image acquisition module: It is installed on the upper side of the conveyor and captures images of the conveyor feed port and materials under the control of the controller to obtain images of the material status at the conveyor feed port; Dataset building module: used to collect images of conveyor belts with normal quantities of material, some material, a small amount of material, almost no material, and no material at all, and to use labeling software to label the material status, generating a dataset containing labeling information for different material statuses; Network model processing module: used for feature fusion of effective feature layers in feature pyramid structure, processing images acquired by image acquisition module, and detecting the status of materials on conveyor; Objective function calculation module: used to optimize the detection accuracy of the network model; Model training and deployment module: This module is used to deploy the best weight file and network model obtained after training on specific hardware devices, so as to realize the application of the model in real-world scenarios and to verify the model performance on mobile devices. The detection result output module is used to output the results after processing by the network model, determine the status of the conveyor feed port, and generate alarm signals or send control commands to relevant equipment when an abnormal status is detected according to preset rules.

Images