Cottonseed mildew probability prediction method and system based on multi-input deep learning model

Abstract

The invention discloses a cottonseed mildew probability prediction method based on a multi-input deep learning model. The method comprises the following steps: acquiring a high-resolution image and near infrared spectrum data of a cottonseed and microbial activity, temperature and humidity and conductivity data in a cottonseed storage environment; preprocessing the high-resolution image, the near infrared spectrum, the microbial activity, the temperature and humidity and the conductivity data; and based on the preprocessed data, utilizing a multi-input deep learning model to predict the cottonseed mildew probability. According to the cottonseed mildew probability prediction method, high-resolution images and near infrared spectrum data of cottonseeds and multiple data sources such as microbial activity, temperature and humidity and conductivity in a storage environment are fused, and feature extraction and fusion are performed by using a specially designed deep learning architecture, so that accurate prediction of the cottonseed mildew probability is realized; and effective guarantee is provided for cotton seed storage management and quality control.

Description

Technical Field

[0001] The present invention relates to the field of cottonseed mildew probability prediction, and more specifically to a cottonseed mildew probability prediction method and system based on a multi-input deep learning model. Background Art

[0002] Cottonseed mildew refers to the process of cottonseed spoilage caused by microbial infection during storage. Mildew not only reduces the quality of cottonseed, but also produces harmful substances, such as aflatoxin and other fungal toxins, which seriously affect the safety and economic value of cottonseed.

[0003] At present, the traditional cottonseed mildew detection methods mainly include the following:

[0004] Manual visual inspection: Experienced experts use their naked eyes to observe the mildew spots, color changes and other signs on the surface of cotton seeds to judge the mildew situation. This method is simple and easy to operate, but it is highly subjective and easily affected by personal experience and working conditions.

[0005] Physical and chemical testing: including measuring the moisture content, acid value and other indicators of cotton seeds. Although these indicators can provide certain reference information, they cannot directly reflect the degree of mildew and require a long time to test.

[0006] Microbiological culture: The possibility of mildew is determined by culturing cottonseed samples and observing the growth of specific molds. This method is time-consuming, usually taking several days or even weeks to get results, and the operation is complicated.

[0007] Single sensor monitoring: Use a single sensor (such as temperature and humidity sensors) to monitor changes in the storage environment. Although this monitoring method can monitor environmental parameters in real time, for complex mold mechanisms, a single environmental parameter is not enough to fully assess the risk of mold.

[0008] The above cottonseed mildew detection method has the following defects:

[0009] Traditional manual visual inspection and microbial culture methods require a lot of manpower and material resources, have a long detection cycle, and cannot meet the needs of large-scale real-time detection. Manual visual inspection is highly subjective and easily affected by individual differences; physical and chemical testing can only indirectly reflect the degree of mold and cannot accurately predict the occurrence of mold.

[0010] At the same time, a single sensor can only reflect some environmental factors, and fails to comprehensively consider the characteristics of cottonseed itself (such as changes in chemical composition) and the combined effects of environmental factors. It is difficult to adapt to different storage conditions and environmental changes, and has limited adaptability to complex and changeable actual application environments.

[0011] In addition, due to the lack of an effective early warning system, mold can often only be discovered after it occurs, resulting in significant economic losses when remedial measures are taken. High-level data such as high-resolution images, near-infrared spectra, and long-term trend analysis of environmental parameters are not fully utilized, resulting in the potential value of information not being fully explored.

[0012] Therefore, how to design a cottonseed mildew probability prediction method based on a multi-input deep learning model to overcome the defects of traditional cottonseed mildew detection methods such as low efficiency, poor accuracy, and weak adaptability, provide early warning for cottonseed storage, and effectively reduce the economic losses caused by mildew is an urgent problem that technical personnel in this field need to solve. Summary of the invention

[0013] In view of this, the present invention integrates high-resolution images of cotton seeds, near-infrared spectral data, and multiple data sources such as microbial activity, temperature and humidity, and conductivity in the storage environment, and uses a specially designed deep learning architecture for feature extraction and fusion. This achieves accurate prediction of the probability of cottonseed mildew, providing effective protection for cottonseed storage management and quality control.

[0014] In order to achieve the above object, the present invention adopts the following technical solution:

[0015] In a first aspect, the present invention provides a method for predicting the probability of cottonseed mildew based on a multi-input deep learning model, comprising the following steps:

[0016] S1. Obtain high-resolution images, near-infrared spectral data, and microbial activity, temperature, humidity, and conductivity data in the cottonseed storage environment;

[0017] S2, preprocessing the high-resolution image, near-infrared spectrum, microbial activity, temperature and humidity, and conductivity data;

[0018] S3. Based on the preprocessed data, a multi-input deep learning model is used to predict the probability of cottonseed mold.

[0019] Wherein, in said S2, the preprocessing includes:

[0020] Enhance and segment high-resolution image data;

[0021] Smoothing of near infrared spectral data;

[0022] The microbial activity, temperature, humidity and conductivity data were filtered and smoothed.

[0023] Furthermore, in S3, the multi-input deep learning model includes:

[0024] Image branch, uses pre-trained convolutional neural network to extract image features;

[0025] Spectral branch, uses multiple fully connected layers to extract spectral features;

[0026] Time series branch, using long short-term memory network LSTM to extract microbial activity, temperature and humidity, and conductivity characteristics;

[0027] The fusion layer fuses the output vectors of the image branch, spectrum branch, and time series branch through a fully connected layer to obtain a fused feature vector;

[0028] The output layer uses the Sigmoid function to output the mold probability based on the fused feature vector.

[0029] Furthermore, in the image branch, the convolutional neural network includes a plurality of convolutional layers and a global average pooling layer;

[0030] The shallow convolution layer in the convolution layer is used to capture low-level features, and the deep convolution layer in the convolution layer is used to extract high-level features;

[0031] The global average pooling layer is used to convert the feature map output by the convolutional layer into a vector of fixed length.

[0032] Furthermore, in the spectral branch, the number of nodes of the first fully connected layer is set to match the number of wavelengths of the near-infrared spectral data, and the number of nodes of the subsequent fully connected layers is reduced based on a preset ratio.

[0033] Furthermore, in the time series branch, the hidden state of the last time step of the long short-term memory network LSTM is used as the final representation of each time series.

[0034] Furthermore, in the fusion layer, the fusion feature vector is obtained as:

[0035] V combined =[V image ; V spectrum ; V time ]

[0036] Among them, [·;·;·] represents the concatenation operation of feature vectors, V combined represents the comprehensive feature vector, V image represents the output vector of the image branch, V spectrum represents the output vector of the spectral branch, V time The output vector representing the time series branch;

[0037] The comprehensive feature vector V combined Mapped to a lower dimensional feature space and output a fused feature vector.

[0038] Furthermore, in the output layer, the Sigmoid function is used to output the mold probability as:

[0039]

[0040] Among them, z represents the fused feature vector.

[0041] Furthermore, the mold probability is compared with a preset threshold value, and if the mold probability is greater than the preset threshold value, an alarm signal is generated.

[0042] In a second aspect, the present invention provides a cottonseed mildew probability prediction system based on a multi-input deep learning model, comprising:

[0043] Data acquisition module: used to obtain high-resolution images of cotton seeds, near-infrared spectral data, and microbial activity, temperature, humidity, and conductivity data in the cotton seed storage environment;

[0044] Preprocessing module: used for preprocessing the high-resolution image, near-infrared spectrum, microbial activity, temperature and humidity, and conductivity data;

[0045] Mildew probability prediction module: used to predict the probability of cottonseed mildew based on preprocessed data using a multi-input deep learning model.

[0046] It can be seen from the above technical solution that compared with the prior art, the technical solution of the present invention has the following advantages:

[0047] Beneficial effects:

[0048] 1. This method combines high-resolution images of cotton seeds, near-infrared spectral data, and multiple data sources such as microbial activity, temperature, humidity, and conductivity in the storage environment. These multimodal data can comprehensively reflect the state of cotton seeds and changes in their storage environment, thereby improving the accuracy and reliability of mold prediction.

[0049] 2. In the preprocessing stage, the image data is enhanced and segmented to enhance the model's robustness to changes in different lighting conditions and angles. The near-infrared spectral data is smoothed to reduce the impact of noise and improve data quality. The microbial activity, temperature, humidity and conductivity data are filtered and smoothed to further improve the quality and availability of time series data.

[0050] 3. Use multi-input deep learning models for targeted feature extraction. The image branch uses a pre-trained convolutional neural network (CNN), which can effectively extract low-level features (such as edges, textures) and high-level features (such as shapes, patterns) in images. The spectral branch uses multiple fully connected layers to extract features related to changes in cottonseed composition from near-infrared spectral data. The time series branch uses a long short-term memory network (LSTM) to capture trends and patterns in microbial activity, temperature, humidity, and conductivity over time. The output vectors of the image branch, spectral branch, and time series branch are fused through a fusion layer to obtain a comprehensive feature vector, which enables the model to comprehensively consider all modal information, thereby improving prediction accuracy.

[0051] 4. Using the Sigmoid function to output the probability of mildew, we can directly get a value between 0 and 1, which intuitively indicates the possibility of cottonseed mildew. By comparing with the preset threshold and generating an alarm signal, users can take timely measures to prevent the spread of mildew and reduce losses. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0053] Figure 1 A flow chart of a cottonseed mildew probability prediction method based on a multi-input deep learning model provided in an embodiment of the present invention;

[0054] Figure 2 A framework diagram of a cottonseed mildew probability prediction method based on a multi-input deep learning model provided in an embodiment of the present invention;

[0055] Figure 3 Framework diagram of a cottonseed mildew probability prediction system based on a multi-input deep learning model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] Embodiment 1;

[0058] like Figure 1 and Figure 2 As shown, this embodiment provides a cottonseed mildew probability prediction method based on a multi-input deep learning model, comprising the following steps:

[0059] S1. Obtain high-resolution images, near-infrared spectral data, and microbial activity, temperature, humidity, and conductivity data in the cottonseed storage environment;

[0060] S2, preprocessing the high-resolution image, near-infrared spectrum, microbial activity, temperature and humidity, and conductivity data;

[0061] S3. Based on the preprocessed data, a multi-input deep learning model is used to predict the probability of cottonseed mold.

[0062] This cottonseed mold probability prediction method integrates multiple data sources such as high-resolution images of cottonseeds, near-infrared spectral data, microbial activity, temperature and humidity, and conductivity in the storage environment, and uses a specially designed deep learning architecture for feature extraction and fusion. It achieves accurate prediction of the cottonseed mold probability and provides effective protection for cottonseed storage management and quality control.

[0063] The above steps and related technical features are further described in detail below;

[0064] In this embodiment S1, high-resolution images of cotton seeds, near-infrared spectral data, and microbial activity, temperature, humidity, and conductivity data in the cotton seed storage environment are obtained, specifically including:

[0065] Choose a high-resolution camera, such as an industrial camera, to ensure that the image has enough detail to observe subtle changes in the cottonseed surface. Use a uniform light source, such as an LED ring light, to ensure consistent lighting conditions and avoid shadows and reflections that affect image quality. And establish a stable acquisition platform to ensure that the position and angle of the cottonseed are consistent each time you shoot, so as to facilitate subsequent image processing. Save the image in a suitable data format (such as TIFF or JPEG) to ensure that the image quality is not affected by compression.

[0066] Use a near infrared spectrometer and set the appropriate spectral range (e.g., 700-2500nm) to collect near infrared spectral data of cotton seeds. Record the spectral data of each sample and make appropriate annotations to facilitate subsequent data analysis.

[0067] Gas sensors are used to monitor carbon dioxide emissions in the cottonseed storage environment to infer the level of microbial activity. Carbon dioxide emissions are recorded regularly, along with other parameters that may affect microbial activity, such as oxygen concentration.

[0068] Deploy temperature and humidity sensors at key locations in the cottonseed storage area to ensure coverage of all areas that may affect the occurrence of mildew. Continuously record temperature and humidity data to ensure data accuracy and completeness.

[0069] Using a conductivity sensor, ensure the cottonseed is in the proper state (i.e. dry or wet) so that conductivity can be accurately measured.

[0070] In this embodiment S2, the pretreatment includes:

[0071] Enhance and segment high-resolution image data, including: adjusting the image according to the overall brightness of the image to make it more suitable for subsequent feature extraction. And combining histogram equalization or contrast-limited adaptive histogram equalization methods to increase the contrast of the image and make the features more obvious.

[0072] Image segmentation techniques (such as color, texture or shape-based segmentation algorithms) are used to remove the background and retain only the cottonseed area, ensuring that each cottonseed can be analyzed individually.

[0073] Finally, all images are resized to the same size and the pixel values ​​of the images are normalized to the range [0, 1] to reduce the impact of the numerical range on model training.

[0074] The near-infrared spectral data are smoothed, including: using a moving average method or a Savitzky-Golay filter to smooth the near-infrared spectral data. The Savitzky-Golay filter is a polynomial fitting smoothing method suitable for removing random noise in the spectrum while keeping the basic shape of the data unchanged.

[0075] Filter and smooth the microbial activity, temperature, humidity and conductivity data, including:

[0076] Use a low-pass filter (such as a Butterworth filter) to remove high-frequency noise and retain useful information. If you know that there is important information in a specific frequency range, you can choose a bandpass filter to extract the information at these frequencies.

[0077] Use an exponentially weighted moving average (EWMA) to smooth time series data and reduce the impact of short-term fluctuations.

[0078] Through the above detailed preprocessing steps, we can ensure that the data quality is optimal, thereby improving the performance and prediction accuracy of subsequent deep learning models.

[0079] In this embodiment S3, the multi-input deep learning model includes:

[0080] The image branch uses a pre-trained convolutional neural network to extract image features. In the image branch, the convolutional neural network includes multiple convolutional layers and a global average pooling layer.

[0081] The shallow convolution layer in the convolution layer is used to capture low-level features, and the deep convolution layer in the convolution layer is used to extract high-level features;

[0082] The global average pooling layer is used to convert the feature map output by the convolutional layer into a vector of fixed length.

[0083] In the spectral branch, multiple fully connected layers are used to extract spectral features. In the spectral branch, the number of nodes in the first fully connected layer is set to match the number of wavelengths of the near-infrared spectral data, and the number of nodes in the subsequent fully connected layers is reduced based on a preset ratio.

[0084] In the time series branch, the long short-term memory network (LSTM) is used to extract the characteristics of microbial activity, temperature and humidity, and conductivity. In the time series branch, the hidden state of the last time step of the long short-term memory network (LSTM) is used as the final representation of each time series.

[0085] The fusion layer fuses the output vectors of the image branch, the spectrum branch, and the time series branch through the fully connected layer to obtain a fused feature vector; in the fusion layer, the fused feature vector is expressed as:

[0086] V combined =[V image ; V spectrum ; V time ]

[0087] Among them, [·;·;·] represents the concatenation operation of feature vectors, V combined represents the comprehensive feature vector, V image represents the output vector of the image branch, V spectrum represents the output vector of the spectral branch, V time The output vector representing the time series branch;

[0088] The comprehensive feature vector V combined Mapped to a lower dimensional feature space and output a fused feature vector.

[0089] The output layer uses the Sigmoid function to output the mold probability based on the fused feature vector. In the output layer, the Sigmoid function is used to output the mold probability as:

[0090]

[0091] Among them, z represents the fused feature vector.

[0092] The multi-input deep learning model in this embodiment can comprehensively capture the various characteristics of cottonseed mildew by integrating image, spectrum and time series data. The image branch uses a pre-trained convolutional neural network to efficiently extract the visual features of the cottonseed surface; the spectrum branch uses multiple fully connected layers to deeply explore the changes in chemical composition in the near-infrared spectrum; the time series branch uses a long short-term memory network (LSTM) to capture the changing trends of environmental factors over time. These multimodal features are integrated through the fusion layer to improve the accuracy of mildew prediction, and finally the Sigmoid function is used to output the mildew probability, providing strong technical support for cottonseed storage management and quality control.

[0093] In this embodiment, the method further includes comparing the mold probability with a preset threshold value, and generating an alarm signal if the mold probability is greater than the preset threshold value.

[0094] The preset threshold is combined with a dynamic threshold adjustment strategy, including: determining the basic threshold for different seasons. Continuously collecting ambient temperature and humidity data. Adjusting the threshold according to the current season and environmental conditions. Triggering an alarm when the probability of mold exceeds the adjusted threshold. Feedback to the threshold adjustment algorithm based on the actual mold situation, and continuously optimizing the threshold adjustment mechanism.

[0095] For example, assume that the initial threshold is 0.5. The summer threshold is adjusted to 0.45, and the winter threshold is adjusted to 0.55. The threshold is further adjusted based on the real-time monitored environmental conditions. For example, if the humidity is continuously above a certain threshold, the threshold is further reduced to 0.4.

[0096] Embodiment 2;

[0097] In this embodiment, the prediction of cottonseed mildew probability using a multi-input deep learning model is further described in detail:

[0098] In multi-input deep learning models,

[0099] Image branch, using pre-trained convolutional neural network to extract image features and obtain feature vector V image , V image = [0.2, 0.1, 0.15, 0.05, 0.5];

[0100] Spectral branch, using multiple fully connected layers to extract spectral features and obtain feature vector V spectrum , V spectrum = [0.05, 0.1, 0.1, 0.05, 0.7];

[0101] Time series branch, using long short-term memory network LSTM to extract microbial activity, temperature and humidity, and conductivity features to obtain feature vector V time , V time They include: Vmicrobial 、V tem-hum and V conductivity , V microbial =[0.1, 0.1, 0.1, 0.1, 0.6], V tem-hum =[0.1, 0.1, 0.1, 0.1, 0.6], V conductivity = [0.1, 0.1, 0.1, 0.1, 0.6];

[0102] The output vectors V of the image branch, spectral branch, and time series branch are transformed into image 、V spectrum 、V microbial 、V tem-hum and V conductivity Fusion is performed to obtain the fused feature vector V combined , specifically including:

[0103] First, concatenate the above five vectors into a vector of length 25;

[0104] Then, the fused feature vector z=W is calculated fution ·V combined +b fution ;

[0105] In this implementation, the weight matrix W_{fusion} of the fusion layer is a 25x1 matrix, specifically [0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.35, 0.40, 0.45, 0.50, 0.55, 0.60, 0.65, 0.70, 0.75, 0.80, 0.85, 0.90, 0.95, 1.00, 1.05, 1.10, 1.15, 1.20, 1.25]; at the same time, the bias term b_{fusion} = -2.95.

[0106] Calculate each item separately: (0.2×0.05)=0.01, (0.1×0.10)=0.01, (0.15×0.15)=0.0225, (0.05×0.20)=0.01, (0.5×0.25)=0.125, (0.05×0.30)=0.015, (0.1×0.35)=0.035, (0.1×0.40)=0.04, (0.05×0.45)=0.0225, (0.7×0.50)=0.35, (0.1×0.55)=0.055, (0.1×0.60)=0.06, (0.1×0.65)=0.065, (0.1×0.70)=0.07, (0.6×0.75)=0.45, (0.1×0.80)=0.08, (0.1×0.85)=0.085, (0.1×0.90)=0.09, (0.1×0.95)=0.095, (0.6×1.00)=0.6, (0.1×1.05)=0.105, (0.1×1.10)=0.11, (0.1×1.15)=0.115, (0.1×1.20)=0.12, (0.6×1.25)=0.75;

[0107] Adding the bias term of -2.95:

[0108] z=0.01+0.01+0.0225+0.01+0.125+0.015+0.035+0.04+0.0225+0.35+0.055+0.06+0.065+0. 07+0.45+0.08+0.085+0.09+0.095+0.6+0.105+0.11+0.115+0.12+0.75-2.95=4.45-2.95=1.5

[0109] Furthermore, P(mold) is obtained based on the formula:

[0110]

[0111] That is, the probability of mildew is 0.82.

[0112] Embodiment 3;

[0113] like Figure 3 As shown, this embodiment provides a cottonseed mildew probability prediction system based on a multi-input deep learning model, including:

[0114] Data acquisition module: used to obtain high-resolution images of cotton seeds, near-infrared spectral data, and microbial activity, temperature, humidity, and conductivity data in the cotton seed storage environment;

[0115] Preprocessing module: used for preprocessing the high-resolution image, near-infrared spectrum, microbial activity, temperature and humidity, and conductivity data;

[0116] Mildew probability prediction module: used to predict the probability of cottonseed mildew based on preprocessed data using a multi-input deep learning model.

[0117] In the above embodiment, a cottonseed mildew probability prediction method and system based on a multi-input deep learning model is proposed, by collecting and preprocessing high-resolution images of cottonseeds, near-infrared spectral data, and multivariate data such as microbial activity, temperature, humidity, and conductivity in the storage environment, and using a multi-input deep learning model including image branches, spectral branches, and time series branches to extract features of different modal data, and then integrating these features through the fusion layer to obtain the cottonseed mildew probability prediction result through the output layer. The system can effectively monitor the risk of cottonseed mildew and issue an alarm in time, thereby helping to improve the safety of cottonseed storage and the efficiency of quality management.

[0118] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0119] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A cottonseed mildew probability prediction method based on a multi-input deep learning model, characterized in that: The following steps are involved: S1. Obtain high-resolution images, near-infrared spectral data, and microbial activity, temperature, humidity, and conductivity data in the cottonseed storage environment; S2, preprocessing the high-resolution image, near-infrared spectrum, microbial activity, temperature and humidity, and conductivity data; S3. Based on the preprocessed data, a multi-input deep learning model is used to predict the probability of cottonseed mold.

2. The method for predicting cottonseed mildew probability based on a multi-input deep learning model according to claim 1, characterized in that: In S2, the preprocessing includes: Enhance and segment high-resolution image data; Smoothing of near infrared spectral data; The microbial activity, temperature, humidity and conductivity data were filtered and smoothed.

3. The method for predicting cottonseed mildew probability based on a multi-input deep learning model according to claim 1, characterized in that: In S3, the multi-input deep learning model includes: Image branch, uses pre-trained convolutional neural network to extract image features; Spectral branch, uses multiple fully connected layers to extract spectral features; Time series branch, using long short-term memory network LSTM to extract microbial activity, temperature and humidity, and conductivity characteristics; The fusion layer fuses the output vectors of the image branch, spectrum branch, and time series branch through a fully connected layer to obtain a fused feature vector; The output layer uses the Sigmoid function to output the mold probability based on the fused feature vector.

4. The method for predicting cottonseed mildew probability based on a multi-input deep learning model according to claim 3, characterized in that: In the image branch, the convolutional neural network includes multiple convolutional layers and a global average pooling layer; The shallow convolution layer in the convolution layer is used to capture low-level features, and the deep convolution layer in the convolution layer is used to extract high-level features; The global average pooling layer is used to convert the feature map output by the convolutional layer into a vector of fixed length.

5. The method for predicting cottonseed mildew probability based on a multi-input deep learning model according to claim 3, characterized in that: In the spectral branch, the number of nodes of the first fully connected layer is set to match the number of wavelengths of the near-infrared spectral data, and the number of nodes of the subsequent fully connected layers is reduced based on a preset ratio.

6. The method for predicting cottonseed mildew probability based on a multi-input deep learning model according to claim 3, characterized in that: In the time series branch, the hidden state of the last time step of the long short-term memory network LSTM is used as the final representation of each time series.

7. The method for predicting cottonseed mildew probability based on a multi-input deep learning model according to claim 1, characterized in that: In the fusion layer, the fusion feature vector is obtained as: V combined =[V image ;V spectrum ;V time ] Among them, [·;·;·] represents the concatenation operation of feature vectors, V combined represents the comprehensive feature vector, V image represents the output vector of the image branch, V spectrum represents the output vector of the spectral branch, V time The output vector representing the time series branch; The comprehensive feature vector V combined Mapped to a lower dimensional feature space and output a fused feature vector.

8. The method for predicting cottonseed mildew probability based on a multi-input deep learning model according to claim 7, characterized in that: In the output layer, the Sigmoid function is used to output the mold probability, which is expressed as: Among them, z represents the fused feature vector.

9. The method for predicting cottonseed mildew probability based on a multi-input deep learning model according to claim 7, characterized in that: The mold probability is compared with a preset threshold, and if the mold probability is greater than the preset threshold, an alarm signal is generated.

10. A cottonseed mildew probability prediction system based on a multi-input deep learning model, characterized in that: include: Data acquisition module: used to obtain high-resolution images of cotton seeds, near-infrared spectral data, and microbial activity, temperature, humidity, and conductivity data in the cotton seed storage environment; Preprocessing module: used for preprocessing the high-resolution image, near-infrared spectrum, microbial activity, temperature and humidity, and conductivity data; Mildew probability prediction module: used to predict the probability of cottonseed mildew based on preprocessed data using a multi-input deep learning model.

Images