A method and system for monitoring the humidity of cardboard boxes

Abstract

The present application relates to a method and system for monitoring the humidity of a cardboard box, belonging to the technical field of cardboard box management. The method includes obtaining color information on the surface of the cardboard box based on an RGB camera; obtaining temperature distribution information on the surface of the cardboard box based on an infrared thermal imaging device; obtaining structural change information on the surface of the cardboard box based on a 3D scanner; inputting the color information, temperature distribution information, and structural change information into a multi-modal fusion network for feature-level fusion to obtain a cardboard box humidity feature representation; performing decision fusion based on a temperature change algorithm and the cardboard box humidity feature representation to obtain a cardboard box humidity estimated value; determining whether the cardboard box humidity estimated value is a preset value; if so, generating a corresponding control action.

Description

Technical Field

[0001] This application relates to the technical field of carton management, and particularly to a method and system for monitoring the humidity of cartons. Background Art

[0002] With the rapid development of the logistics industry, the requirements for protecting packaging materials have also been increasing. After the cartons are produced, they are usually stored in a warehouse. During storage, the cartons are prone to moisture absorption, resulting in a decrease in strength, which is not conducive to packing goods.

[0003] Therefore, how to automatically monitor the humidity of cartons and take actions is a technical problem that those skilled in the art need to overcome. Summary of the Invention

[0004] To at least partially solve the above technical problems, this application provides a method and system for monitoring the humidity of cartons.

[0005] In a first aspect, a method for monitoring the humidity of cartons provided by this application adopts the following technical solutions.

[0006] Obtain the color information of the carton surface based on an RGB camera;

[0007] Obtain the temperature distribution information of the carton surface based on an infrared thermal imaging device;

[0008] Obtain the structural change information of the carton surface based on a 3D scanner;

[0009] Input the color information, temperature distribution information, and structural change information into a multi-modal fusion network for feature-level fusion to obtain a carton humidity feature representation;

[0010] Perform decision fusion according to the temperature change algorithm and the carton humidity feature representation to obtain a carton humidity estimated value;

[0011] Judge whether the carton humidity estimated value is a preset value; if so, generate a corresponding control action.

[0012] Optionally, obtaining the color information of the carton surface based on an RGB camera includes:

[0013] Control the adjustable light source to work for illumination;

[0014] Feed back the light source brightness information through an ambient light sensor;

[0015] Based on the light source brightness information, control the brightness of the adjustable light source through a PID algorithm to adjust it until the re-obtained feedback light source brightness information reaches a preset range;

[0016] Obtain the carton surface image through an RGB camera;

[0017] Perform image quality enhancement processing on the collected carton surface image; the image quality enhancement processing includes contrast enhancement and sharpening;

[0018] Separate the carton surface image from the background through image segmentation technology; extract the color information of the carton surface from the separated carton surface image to obtain the color information of the carton surface.

[0019] Optionally, based on the light source brightness information, control the brightness of the adjustable light source through the PID algorithm to adjust until the newly obtained feedback light source brightness information reaches the preset range, including:

[0020] Obtain the current ambient brightness based on the light source brightness information;

[0021] Subtract the current ambient brightness from the target brightness to obtain the brightness error;

[0022] Calculate the proportional term, integral term, and differential term in the PID algorithm to obtain the output value;

[0023] Control the brightness of the adjustable light source for adjustment according to the output value.

[0024] Optionally, input the color information, temperature distribution information, and structural change information into a multi-modal fusion network for feature-level fusion to obtain the carton humidity feature representation, including:

[0025] Perform preprocessing on the color information, temperature distribution information, and structural change information, and the preprocessing includes normalization, denoising, and feature point extraction;

[0026] Input the preprocessed information into each branch of the multi-modal fusion network;

[0027] Perform cross-modal feature interaction learning inside the network to generate richer feature representations;

[0028] Perform feature-level fusion at the output layer of the network to generate the carton humidity feature representation;

[0029] Among them, performing cross-modal feature interaction learning inside the network to generate richer feature representations includes:

[0030] Obtain data samples of different modalities from multiple data sources;

[0031] Use the feature extractors designed for each modality to extract the preliminary feature vectors;

[0032] Input the preliminary feature vectors into the cross-modal fusion module, and the cross-modal fusion module realizes feature transformation and enhancement between modalities through the preset mapping relationship;

[0033] In the feature interaction stage, the attention mechanism is used to dynamically select the features with the strongest correlation between different modalities for fusion, so as to generate a richer feature representation;

[0034] Optimize the model parameters through the backpropagation algorithm;

[0035] After the training is completed, use the generated feature representation to make predictions or decisions on the newly input data.

[0036] Optionally, perform decision fusion based on the temperature change algorithm and the cardboard box humidity feature representation to obtain the cardboard box humidity estimation value, including:

[0037] Analyze the cardboard box humidity feature representation to obtain a preliminary humidity estimation value;

[0038] Combine the temperature distribution information and use the temperature change algorithm to correct the preliminary humidity estimation value;

[0039] Compare the corrected humidity estimation value with the historical data for trend analysis;

[0040] Perform decision fusion according to the trend analysis result to obtain the final cardboard box humidity estimation value;

[0041] Among them, the trend analysis includes:

[0042] Establish a time series of humidity estimation values;

[0043] Calculate the moving average value of the time series;

[0044] Determine the change trend of the moving average value;

[0045] Adjust the humidity estimation value according to the change trend.

[0046] Optionally, determine whether the cardboard box humidity estimation value is a preset value; if so, generate corresponding control actions, including:

[0047] Set multiple humidity thresholds to distinguish the risk levels of different humidity levels;

[0048] When the cardboard box humidity estimation value exceeds the first-level threshold, record and mark the cardboard box;

[0049] When the cardboard box humidity estimation value exceeds the second-level threshold, send a warning notice to the operator and start the drying equipment;

[0050] When the cardboard box humidity estimation value exceeds the third-level threshold, immediately start the drying equipment and send out an alarm message.

[0051] Optionally, perform cross-modal feature interaction learning within the network to generate a richer feature representation, including:

[0052] Obtain data samples of different modalities from multiple data sources;

[0053] Extract preliminary feature vectors using feature extractors designed for each modality;

[0054] Input the preliminary feature vectors into a cross-modal fusion module, and within the cross-modal fusion module, implement feature transformation and enhancement between modalities through a preset mapping relationship;

[0055] In the feature interaction stage, use the attention mechanism to dynamically select the features with the strongest correlation between different modalities for fusion, thereby generating a richer feature representation.

[0056] Optionally, the cross-modal fusion module specifically includes:

[0057] Input the preliminary feature vectors of different modalities into their respective sub-networks for preliminary feature transformation; the sub-network includes a fully connected layer and a convolutional layer;

[0058] Convert the preliminary feature vectors of different modalities into a unified spatial representation through a mapping relationship;

[0059] Input the transformed feature vectors into the gated recurrent unit of the cross-modal fusion module for interactive learning of features;

[0060] Inside the gated recurrent unit, implement feature transformation and enhancement between modalities through a multi-layer neural network;

[0061] Add residual connections after each layer to alleviate the problem of gradient vanishing; and perform feature fusion after the last layer to generate a comprehensive feature representation.

[0062] Optionally, the attention mechanism includes:

[0063] Calculate the similarity matrix between features of different modalities;

[0064] Apply the softmax function to the similarity matrix to obtain the attention weight matrix;

[0065] Perform weighted summation on the features of different modalities according to the attention weight matrix to obtain a comprehensive feature representation;

[0066] Pass the comprehensive feature representation to the next layer for further feature enhancement.

[0067] In a second aspect, a carton humidity monitoring method provided by the present application adopts the following technical solution.

[0068] A carton humidity monitoring system, characterized in that it includes:

[0069] A first processing module for: obtaining color information on the surface of the carton based on an RGB camera;

[0070] A second processing module, configured to: obtain temperature distribution information on the surface of the carton based on an infrared thermal imaging device;

[0071] A third processing module, configured to: obtain structural change information on the surface of the carton based on a 3D scanner;

[0072] A fourth processing module, configured to: input the color information, temperature distribution information, and structural change information into a multi-modal fusion network for feature-level fusion to obtain a carton humidity feature representation;

[0073] A fifth processing module, configured to: perform decision fusion according to a temperature change algorithm and the carton humidity feature representation to obtain a carton humidity estimation value;

[0074] A sixth processing module, configured to: perform decision fusion according to a temperature change algorithm and the carton humidity feature representation to obtain a carton humidity estimation value. Description of the Drawings

[0075] Figure 1 is a flowchart of a method for monitoring the humidity of a carton according to an embodiment of the present application;

[0076] Figure 2 is a system block diagram of a system for monitoring the humidity of a carton according to an embodiment of the present application;

[0077] In the figure, 201 is the first processing module; 202 is the second processing module; 203 is the third processing module; 204 is the fourth processing module; 205 is the fifth processing module; 206 is the sixth processing module. Detailed Embodiments

[0078] The following further describes the present application in conjunction with the attached Figure 1-2 drawings and specific embodiments:

[0079] An embodiment of the present application discloses a method for monitoring the humidity of a carton. Referring to Figure 1 , as an implementation manner of a method for monitoring the humidity of a carton, a method for monitoring the humidity of a carton includes the following steps:

[0080] Step 101: Obtain color information on the surface of the carton based on an RGB camera;

[0081] Step 102: Obtain temperature distribution information on the surface of the carton based on an infrared thermal imaging device;

[0082] Step 103: Obtain structural change information on the surface of the carton based on a 3D scanner;

[0083] Step 104: Input the color information, temperature distribution information, and structural change information into a multi-modal fusion network for feature-level fusion to obtain a carton humidity feature representation;

[0084] Step 105: Perform decision fusion based on the temperature change algorithm and the cardboard box humidity feature representation to obtain an estimated value of the cardboard box humidity;

[0085] Step 106: Determine whether the estimated value of the cardboard box humidity is a preset value; if so, generate a corresponding control action.

[0086] Specifically, information such as the color, temperature distribution, and structural changes of the cardboard box are obtained through an RGB camera, an infrared thermal imaging device, and a 3D scanner respectively, comprehensively perceiving the state of the cardboard box from multiple angles.

[0087] Input the above different types of sensor data into a multi-modal fusion network. At this stage, different data sets are processed and integrated together to form a more complete and detailed description of the cardboard box state - that is, the cardboard box humidity feature representation. This process utilizes the complementarity between different sensors, enabling information that may be missed by a single sensor to be supplemented by the data of other sensors.

[0088] Perform decision fusion based on the temperature change algorithm combined with the previously obtained cardboard box humidity feature representation. Compare the calculated humidity estimated value with a preset standard or safety threshold. If the humidity exceeds the allowed range, trigger corresponding control actions, such as alarming or starting a drying device, thereby protecting the items inside the cardboard box from the influence of too high or too low humidity.

[0089] As a specific implementation of a cardboard box humidity monitoring method, based on an RGB camera to obtain the color information on the surface of the cardboard box, including:

[0090] Control the adjustable light source to work for illumination;

[0091] Feedback the light source brightness information through an ambient light sensor;

[0092] Based on the light source brightness information, control the brightness of the adjustable light source through a PID algorithm to adjust it until the re-obtained feedback light source brightness information reaches a preset range;

[0093] Obtain an image of the surface of the cardboard box through an RGB camera;

[0094] Perform image quality enhancement processing on the collected image of the surface of the cardboard box; the image quality enhancement processing includes contrast enhancement and sharpening;

[0095] Separate the image of the surface of the cardboard box from the background through image segmentation technology; extract the color information from the image of the surface of the cardboard box separated from the background to obtain the color information on the surface of the cardboard box.

[0096] Specifically, an adjustable light source is used to illuminate the cardboard box, and the ambient light sensor is used to monitor the light intensity in real time to ensure consistent lighting conditions in any environment, thereby improving the quality of subsequent image processing. The light source brightness information is dynamically adjusted through the PID (Proportional-Integral-Derivative) algorithm to stabilize it within an ideal range, which can eliminate image quality fluctuations caused by changes in external light. An RGB camera is used to capture images of the cardboard box surface. The acquired images are processed through a series of image processing techniques, such as contrast enhancement and sharpening. Image quality enhancement can significantly improve the clarity of the images, making the color information more reliable. Image segmentation technology is used to separate the cardboard box surface from the background, so that the focus can be on the cardboard box itself without being disturbed by the environment. Once the cardboard box surface is successfully segmented, the color information can be extracted from it.

[0097] As a specific implementation of a cardboard box humidity monitoring method, based on the light source brightness information, the brightness of the adjustable light source is controlled through the PID algorithm to be adjusted until the re-obtained feedback light source brightness information reaches a preset range, including:

[0098] Obtain the current ambient brightness based on the light source brightness information;

[0099] Subtract the current ambient brightness from the target brightness to obtain a brightness error;

[0100] Calculate the proportional term, integral term, and derivative term in the PID algorithm to obtain an output value;

[0101] Control the brightness of the adjustable light source to be adjusted according to the output value.

[0102] Specifically, the brightness information of the current environment is obtained through the ambient light sensor; the target brightness is a preset ideal brightness value, representing the best lighting conditions that the system hopes to achieve. By subtracting the current ambient brightness from the target brightness, a brightness error value can be obtained. This error value is the basis for subsequent PID control. The output value calculated by the PID algorithm is used to adjust the brightness of the adjustable light source to make the actual brightness as close as possible to the target brightness. Through continuous feedback and adjustment, the system can maintain a stable lighting level and ensure the stability of image acquisition.

[0103] As a specific implementation of a cardboard box humidity monitoring method, input the color information, temperature distribution information, and structural change information into a multi-modal fusion network for feature-level fusion to obtain a cardboard box humidity feature representation, including:

[0104] Preprocess the color information, temperature distribution information, and structural change information. The preprocessing includes normalization, denoising, and feature point extraction;

[0105] Input the preprocessed information into each branch of the multi-modal fusion network;

[0106] Conduct cross-modal feature interaction learning inside the network to generate richer feature representations;

[0107] Perform feature-level fusion at the output layer of the network to generate the carton humidity feature representation;

[0108] Among them, conducting cross-modal feature interaction learning inside the network to generate richer feature representations includes:

[0109] Obtain data samples of different modalities from multiple data sources;

[0110] Use feature extractors designed for each modality to extract preliminary feature vectors;

[0111] Input the preliminary feature vectors into the cross-modal fusion module, and inside the cross-modal fusion module, implement feature transformation and enhancement between modalities through a preset mapping relationship;

[0112] In the feature interaction stage, use the attention mechanism to dynamically select the features with the strongest correlation between different modalities for fusion, so as to generate richer feature representations;

[0113] Optimize the model parameters through the backpropagation algorithm;

[0114] After training is completed, use the generated feature representations to make predictions or decisions on newly input data.

[0115] Specifically, preprocessing is performed on color information, temperature distribution information, and structural change information, including normalization, denoising, and feature point extraction. The purpose of preprocessing is to ensure the quality of the input data, making them more suitable for subsequent feature extraction and fusion. Different types of information are fed into network branches specifically designed to process data of corresponding modalities, ensuring that the information of each modality can be effectively extracted and processed. At the output layer of the network, the feature information of all modalities is integrated together to form the final representation of the humidity of the cardboard box. This process further enhances the expressive power of the features and the generalization ability of the model. The model parameters are continuously optimized through the backpropagation algorithm until the model can achieve the expected performance on the training set. After training, the model can be used to make predictions or decisions on new input data, such as judging the humidity state of the cardboard box. By combining information from multiple modalities such as color, temperature distribution, and structural change, the system can capture the features of the humidity of the cardboard box from more perspectives, improving the comprehensiveness and accuracy of monitoring. The cross-modal fusion module allows the features between different modalities to be converted and enhanced with each other, while the attention mechanism helps the system focus on the most relevant feature combinations, thus improving the quality of the feature representation. Since the multi-modal fusion network can process various types of input data and enhance the feature representation through feature interaction learning, the model has better generalization ability when facing data in different scenarios.

[0116] As one of the implementation manners of a cardboard box humidity monitoring method, a cardboard box humidity estimation value is obtained through decision fusion according to a temperature change algorithm and the cardboard box humidity feature representation, including:

[0117] Analyze the cardboard box humidity feature representation to obtain a preliminary humidity estimation value;

[0118] Combine the temperature distribution information and use the temperature change algorithm to correct the preliminary humidity estimation value;

[0119] Compare the corrected humidity estimation value with historical data for trend analysis;

[0120] Perform decision fusion according to the trend analysis result to obtain the final cardboard box humidity estimation value;

[0121] Among them, the trend analysis includes:

[0122] Establish a time series of humidity estimation values;

[0123] Calculate the moving average value of the time series;

[0124] Determine the change trend of the moving average value;

[0125] Adjust the humidity estimation value according to the change trend.

[0126] Specifically, first, based on the cardboard box humidity feature representation (features obtained through multimodal fusion of color information, temperature distribution information, structural change information, etc. as described above), the system calculates a preliminary humidity estimate value. Considering the influence of temperature change on the cardboard box humidity, the preliminary humidity estimate value is corrected through a temperature change algorithm. This step utilizes temperature as an external variable to calibrate the humidity estimate, increasing the accuracy of the estimate value. By comprehensively considering the preliminary humidity estimate value, the estimate value corrected by temperature change, and the result of trend analysis, decision fusion is performed to obtain a more accurate and reliable cardboard box humidity estimate value. By combining the preliminary estimate value with the correction of the temperature change algorithm, the system not only relies on a single data source but also comprehensively considers the influence of temperature on humidity, thereby improving the accuracy of humidity estimation. By comprehensively considering the preliminary humidity estimate value, the corrected humidity estimate value, and the result of trend analysis, the finally obtained humidity estimate value not only reflects the current situation but also considers historical trends, which can better meet the humidity monitoring requirements in complex environments.

[0127] As one implementation manner of a cardboard box humidity monitoring method, it is judged whether the cardboard box humidity estimate value is a preset value; if so, corresponding control actions are generated, including:

[0128] Set multiple humidity thresholds to distinguish the risk levels of different humidity levels;

[0129] When the cardboard box humidity estimate value exceeds the first-level threshold, record and mark the cardboard box;

[0130] When the cardboard box humidity estimate value exceeds the second-level threshold, send a warning notice to the operator and start the drying equipment;

[0131] When the cardboard box humidity estimate value exceeds the third-level threshold, immediately start the drying equipment and send an alarm message.

[0132] Specifically, three risk levels are defined according to different humidity levels, and each level corresponds to a humidity threshold, namely, the first-level threshold, the second-level threshold, and the third-level threshold. These thresholds are set to distinguish the potential risk levels faced by the cartons under different humidity levels. Exceeding the first-level threshold: At this time, the humidity has exceeded the normal range, but it is not yet severe enough to cause serious quality problems. The system will record and mark the carton for subsequent inspection. Exceeding the second-level threshold: The humidity further increases to a level that may affect the product quality. The system will not only record and mark the carton but also send a warning notice to the operator and start the drying equipment to reduce the humidity. Exceeding the third-level threshold: This is the most serious situation, with extremely high humidity, which may cause damage to the carton or other adverse consequences. The system will immediately start the drying equipment and send an alarm message to remind the relevant personnel to handle it urgently. By setting multiple humidity thresholds, the system can take corresponding early warning measures according to different humidity levels. This means that even minor humidity changes can be detected in a timely manner, and appropriate actions can be taken according to their severity, thus preventing small problems from evolving into big problems.

[0133] As one implementation of a method for monitoring the humidity of cartons, cross-modal feature interaction learning is carried out within the network to generate richer feature representations, including:

[0134] Obtain data samples of different modalities from multiple data sources;

[0135] Use feature extractors designed for each modality to extract preliminary feature vectors;

[0136] Input the preliminary feature vectors into a cross-modal fusion module, and within the cross-modal fusion module, feature transformation and enhancement between modalities are achieved through a preset mapping relationship;

[0137] In the feature interaction stage, the attention mechanism is used to dynamically select the features with the strongest correlation between different modalities for fusion, thereby generating richer feature representations.

[0138] Specifically, due to the adoption of the cross-modal feature interaction learning method, the system can obtain information from multiple data sources and convert this information into useful feature vectors through a feature extractor. By introducing a cross-modal fusion module and using an attention mechanism to dynamically select features, the model can exhibit better adaptability and stability when facing different types of input data. Even if the data quality of a certain modality is poor or missing, the data of other modalities can still provide supplementary information to help the model make accurate judgments. In the application of carton humidity monitoring, the comprehensive feature representation obtained by cross-modal feature interaction learning can more accurately reflect the actual situation. For example, by analyzing the interactions between temperature, humidity, and other environmental factors, the system can more precisely evaluate the humidity state of the carton and thus make more reasonable humidity control decisions. The application of the attention mechanism means that the system can more efficiently allocate computing resources when processing multi-modal data. By only focusing on the most informative parts, unnecessary computational overhead can be reduced while ensuring that important information is not overlooked.

[0139] As one of the implementation manners of a carton humidity monitoring method, the cross-modal fusion module specifically includes:

[0140] Input the preliminary feature vectors of different modalities into their respective sub-networks for preliminary feature transformation; the sub-network includes a fully connected layer and a convolutional layer;

[0141] Convert the preliminary feature vectors of different modalities into a unified spatial representation through a mapping relationship;

[0142] Input the transformed feature vectors into the gated recurrent unit of the cross-modal fusion module for feature interaction learning;

[0143] Inside the gated recurrent unit, implement feature transformation and enhancement between modalities through a multi-layer neural network;

[0144] Add a residual connection after each layer to alleviate the problem of gradient vanishing; and perform feature fusion after the last layer to generate a comprehensive feature representation.

[0145] As one of the implementation manners of a carton humidity monitoring method, the attention mechanism includes:

[0146] Calculate the similarity matrix between different modality features;

[0147] Apply the softmax function to the similarity matrix to obtain the attention weight matrix;

[0148] Perform weighted summation on the features of different modalities according to the attention weight matrix to obtain a comprehensive feature representation;

[0149] Pass the comprehensive feature representation to the next layer for further feature enhancement.

[0150] Specifically, the initial feature vectors of different modalities are first input into their respective sub-networks, which contain fully connected layers and convolutional layers. This can perform preliminary processing on the original feature vectors and extract more meaningful feature representations. The fully connected layers help capture the global relationships between features, while the convolutional layers can extract local features. Through the mapping relationship, the feature vectors of different modalities are transformed into the same feature space, enabling features from different sources to be directly compared and fused, enhancing the consistency and comparability of the features. The transformed feature vectors are fed into a gated recurrent unit (GRU), which is a variant of the recurrent neural network (RNN) and is particularly suitable for processing sequential data. The GRU controls the flow of information through a gating mechanism, allowing the preservation of long-term dependencies, thereby effectively performing interactive learning of features. Adding residual connections after each layer helps alleviate the vanishing gradient problem and ensures the training effect of deep networks. After learning through multiple layers of neural networks, feature fusion is performed at the last layer to generate a comprehensive feature representation. This representation contains cross-modal information and more comprehensively reflects the state of the humidity of the cardboard box. By combining the cross-modal fusion module and the attention mechanism, this method for monitoring the humidity of the cardboard box not only improves the quality of the feature representation but also enhances the overall performance of the model.

[0151] This application also provides a cardboard box humidity monitoring system, including:

[0152] A first processing module 201, configured to: obtain color information on the surface of the cardboard box based on an RGB camera;

[0153] A second processing module 202, configured to: obtain temperature distribution information on the surface of the cardboard box based on an infrared thermal imaging device;

[0154] A third processing module 203, configured to: obtain structural change information on the surface of the cardboard box based on a 3D scanner;

[0155] A fourth processing module 204, configured to: input the color information, temperature distribution information, and structural change information into a multi-modal fusion network for feature-level fusion to obtain a cardboard box humidity feature representation;

[0156] A fifth processing module 205, configured to: perform decision fusion based on a temperature change algorithm and the cardboard box humidity feature representation to obtain a cardboard box humidity estimated value;

[0157] A sixth processing module 206, configured to: perform decision fusion based on a temperature change algorithm and the cardboard box humidity feature representation to obtain a cardboard box humidity estimated value.

[0158] It should be noted that the above embodiments are only used to illustrate the present application and do not limit the technical solutions described in the present application. Although the present specification has described the present application in detail with reference to the above embodiments, those of ordinary skill in the art should understand that those skilled in the art can still modify the present application or make equivalent substitutions, and all technical solutions and improvements that do not depart from the spirit and scope of the present application should be covered within the scope of the claims of the present application.

Claims

1. A method for monitoring carton humidity, characterized in that: include: Obtain the color information of the carton surface based on the RGB camera; Obtain the temperature distribution information on the carton surface based on infrared thermal imaging equipment; Obtain structural change information on the carton surface based on a 3D scanner; Inputting the color information, temperature distribution information and structural change information into a multimodal fusion network for feature-level fusion to obtain a feature representation of carton humidity; A decision fusion is performed based on the temperature change algorithm and the carton humidity feature representation to obtain a carton humidity estimation value; Determine whether the estimated value of the carton humidity is a preset value; if so, generate a corresponding control action; The color information, temperature distribution information and structural change information are input into a multimodal fusion network for feature-level fusion to obtain a carton humidity feature representation, including: Preprocessing the color information, temperature distribution information and structural change information, wherein the preprocessing includes normalization, denoising and feature point extraction; Input the preprocessed information into each branch of the multimodal fusion network; Interactive learning of cross-modal features within the network to generate richer feature representations; Perform feature-level fusion at the output layer of the network to generate a feature representation of carton humidity; Among them, interactive learning of cross-modal features is performed within the network to generate richer feature representations, including: Obtain data samples of different modalities from multiple data sources; Preliminary feature vectors were extracted using a feature extractor designed for each modality; The preliminary feature vector is input into a cross-modal fusion module, and the cross-modal fusion module realizes feature conversion and enhancement between modalities through a preset mapping relationship; In the feature interaction stage, the attention mechanism is used to dynamically select the most correlated features between different modalities for fusion, thereby generating a richer feature representation; Optimize model parameters through back-propagation algorithm; After training is complete, the generated feature representation is used to make predictions or decisions on new input data; According to the temperature change algorithm and the carton humidity feature representation, decision fusion is performed to obtain the carton humidity estimation value, including: Analyzing the carton humidity characteristic representation to obtain a preliminary humidity estimation value; Combined with the temperature distribution information, the temperature change algorithm is used to correct the initial humidity estimate; Compare the revised humidity estimates with historical data for trend analysis; Decision fusion is performed based on the trend analysis results to obtain the final estimated value of carton humidity; The trend analysis includes: Build a time series of humidity estimates; Calculate the moving average of a time series; Determine the trend of the moving average; Adjust humidity estimates based on changing trends.

2. A carton humidity monitoring method according to claim 1, characterized in that: Obtain the color information of the carton surface based on the RGB camera, including: Control adjustable light source to work for lighting; Feedback of light source brightness information through ambient light sensors; Based on the light source brightness information, the brightness of the adjustable light source is controlled by a PID algorithm to adjust until the feedback light source brightness information obtained again reaches a preset range; Acquire the carton surface image through the RGB camera; Performing image quality enhancement processing on the collected carton surface image; the image quality enhancement processing includes contrast enhancement and sharpening; The carton surface image is separated from the background by image segmentation technology; the color information of the carton surface image separated from the background is extracted to obtain the color information of the carton surface.

3. A carton humidity monitoring method according to claim 2, characterized in that: Based on the light source brightness information, the brightness of the adjustable light source is controlled by a PID algorithm to adjust until the retrieved feedback light source brightness information reaches a preset range, including: Obtaining the current environment brightness based on the light source brightness information; Subtract the current environment brightness from the target brightness to obtain the brightness error; Calculate the proportional term, integral term and differential term in the PID algorithm to obtain the output value; The brightness of the adjustable light source is controlled and adjusted according to the output value.

4. A method for monitoring carton humidity according to claim 3, characterized in that: Determine whether the estimated value of the carton humidity is a preset value; if so, generate a corresponding control action, including: Set multi-level humidity thresholds to distinguish risk levels at different humidity levels; When the estimated moisture content of a carton exceeds the first-level threshold, the carton is recorded and marked; When the estimated carton moisture exceeds the secondary threshold, a warning notification is sent to the operator and the drying equipment is started; When the estimated carton humidity exceeds the third-level threshold, the drying equipment is immediately started and an alarm message is issued.

5. A method for monitoring carton humidity according to claim 4, characterized in that: Interactive learning of cross-modal features is performed within the network to generate richer feature representations, including: Obtain data samples of different modalities from multiple data sources; Preliminary feature vectors were extracted using a feature extractor designed for each modality; The preliminary feature vector is input into a cross-modal fusion module, and the cross-modal fusion module realizes feature conversion and enhancement between modalities through a preset mapping relationship; In the feature interaction stage, the attention mechanism is used to dynamically select the most correlated features between different modalities for fusion, thereby generating richer feature representations.

6. A method for monitoring carton humidity according to claim 5, characterized in that: The cross-modal fusion module specifically includes: Inputting preliminary feature vectors of different modalities into respective sub-networks for preliminary feature conversion; the sub-networks include fully connected layers and convolutional layers; The preliminary feature vectors of different modalities are converted into a unified spatial representation through mapping relationships; The converted feature vector is input into the gated recurrent unit of the cross-modal fusion module for interactive feature learning; Inside the gated recurrent unit, feature conversion and enhancement between modalities are achieved through a multi-layer neural network; Residual connections are added after each layer to alleviate the gradient vanishing problem, and feature fusion is performed after the last layer to generate a comprehensive feature representation.

7. A method for monitoring carton humidity according to claim 6, characterized in that: Attention mechanism, including: Calculate the similarity matrix between different modal features; Apply the softmax function to the similarity matrix to obtain the attention weight matrix; The features of different modalities are weighted and summed according to the attention weight matrix to obtain a comprehensive feature representation; The comprehensive feature representation is passed to the next layer for further feature enhancement.

8. A carton humidity monitoring system, characterized in that: include: The first processing module is used to: obtain color information of the carton surface based on an RGB camera; The second processing module is used to obtain the temperature distribution information of the carton surface based on the infrared thermal imaging device; The third processing module is used to obtain structural change information on the surface of the carton based on a 3D scanner; The fourth processing module is used to: input the color information, temperature distribution information and structure change information into a multimodal fusion network for feature-level fusion to obtain a carton humidity feature representation; A fifth processing module is used to: perform decision fusion according to the temperature change algorithm and the carton humidity feature representation to obtain a carton humidity estimation value; A sixth processing module is used to: perform decision fusion according to the temperature change algorithm and the carton humidity feature representation to obtain a carton humidity estimation value; The color information, temperature distribution information and structural change information are input into a multimodal fusion network for feature-level fusion to obtain a carton humidity feature representation, including: Preprocessing the color information, temperature distribution information and structural change information, wherein the preprocessing includes normalization, denoising and feature point extraction; Input the preprocessed information into each branch of the multimodal fusion network; Interactive learning of cross-modal features within the network to generate richer feature representations; Perform feature-level fusion at the output layer of the network to generate a feature representation of carton humidity; Among them, interactive learning of cross-modal features is performed within the network to generate richer feature representations, including: Obtain data samples of different modalities from multiple data sources; Preliminary feature vectors were extracted using a feature extractor designed for each modality; The preliminary feature vector is input into a cross-modal fusion module, and the cross-modal fusion module realizes feature conversion and enhancement between modalities through a preset mapping relationship; In the feature interaction stage, the attention mechanism is used to dynamically select the most correlated features between different modalities for fusion, thereby generating a richer feature representation; Optimize model parameters through back-propagation algorithm; After training is complete, the generated feature representation is used to make predictions or decisions on new input data; According to the temperature change algorithm and the carton humidity feature representation, decision fusion is performed to obtain the carton humidity estimation value, including: Analyzing the carton humidity characteristic representation to obtain a preliminary humidity estimation value; Combined with the temperature distribution information, the temperature change algorithm is used to correct the initial humidity estimate; Compare the revised humidity estimates with historical data for trend analysis; Decision fusion is performed based on the trend analysis results to obtain the final estimated value of carton humidity; The trend analysis includes: Build a time series of humidity estimates; Calculate the moving average of a time series; Determine the trend of the moving average; Adjust humidity estimates based on changing trends.

Images