Liquid bacterial fermentation monitoring method based on multispectral diffuse reflectance and edge computing

Abstract

The present invention provides a liquid culture fermentation monitoring method based on multispectral diffuse reflectance and edge computing, comprising the following steps: Step 1: Acquire characteristic spectral information of key biological parameters of the liquid culture fermentation process; Step 2: Develop a multispectral diffuse reflectance imaging system based on the characteristic spectral band information of the fermentation process; Step 3: Construct a lightweight target detection deep learning model that can integrate multispectral data processing to facilitate embedded machine vision deployment; Step 4: Use a cascade calibration method to enable the visual model to identify and locate the fermentation window position, and then perform key target detection and analysis within the region of interest; Step 5: Deploy the above-mentioned deep learning visual model on an intelligent edge computing device to process the fermentation process atlas information in real time. The application of this technical solution can realize the application of embedded machine vision in the fermentation industry site.

Description

Technical Field

[0001] The present invention relates to the technical field of in-situ real-time monitoring of fermentation engineering, and in particular to a liquid strain fermentation monitoring method based on multispectral diffuse reflectance and edge computing. Background Art

[0002] Liquid cultures are essentially fermentation processes, the goal of which is to obtain a large quantity of target bacterial strains with high purity and good vitality in the shortest possible time. Compared to traditional solid cultures, liquid cultures have the advantages of shorter production cycles, lower costs, consistent bacterial viability, and ease of industrialization. However, the liquid culture fermentation process places high demands on monitoring technology, ensuring that the fermentation process is free of contaminants, that the target bacterial strain grows consistently, and that the optimal fermentation endpoint is determined. To this end, in-situ, real-time, non-contact monitoring of the fermentation process has become necessary to achieve stable and high-quality output. However, the fermentation process of liquid cultures is typically a dynamic growth and evolution process, involving the fission growth of the bacterial strain, nutrient consumption in the fermentation broth, and accumulation of fermentation products. Although online monitoring of some physical parameters (such as temperature and air pressure) and chemical parameters (such as pH and dissolved oxygen concentration) is now possible, real-time monitoring of key biological parameters of the fermentation process (such as target bacterial biomass and solution substrate and product concentrations) is not ideal.

[0003] There are three main methods for monitoring biological parameters during liquid culture fermentation: 1. Manual sensory evaluation; 2. Offline sampling and physical and chemical analysis; and 3. Online invasive testing. Manual sensory evaluation relies on extensive experience, is subjective, and cannot generate objective and stable indicators. Offline sampling and monitoring can be relatively objective and accurate, but has the disadvantages of time lag and consumes chemical materials and manpower. Online invasive sensors, on the other hand, increase the size of the fermentation equipment and pose potential contamination risks to the liquid culture fermentation process. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a liquid strain fermentation monitoring method based on multispectral diffuse reflectance and edge computing, so as to realize the application of embedded machine vision in industrial sites.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a liquid strain fermentation monitoring method based on multispectral diffuse reflectance and edge computing, comprising the following steps:

[0006] Step 1: Obtain characteristic spectral information of key biological parameters during the liquid bacterial fermentation process;

[0007] Step 2: Develop a multispectral diffuse reflectance imaging system based on the acquired characteristic spectral band information of the fermentation process;

[0008] Step 3: Build a lightweight deep learning vision model for target detection that can integrate multispectral data processing and make it suitable for embedded machine vision applications;

[0009] Step 4: Using a cascade calibration method, the visual model can identify and locate the fermentation window position, and then perform key target detection and analysis within the region of interest;

[0010] Step 5: Deploy the above deep learning vision model on the edge computing device to monitor the fermentation process in real time.

[0011] In a preferred embodiment, step 1 specifically includes: using a hyperspectral spectrometer covering ultraviolet, visible, and near-infrared to collect spectral information of key biological parameters at different concentrations, collect spectral information of culture solutions at different concentrations, collect spectral information of fermentation solutions at different growth stages, and collect spectral information of fermentation solutions under healthy and polluted conditions; and based on the above spectral information, calculate the spectral index to obtain characteristic bands of interest that can robustly characterize the corresponding parameters.

[0012] In a preferred embodiment, step 2 is performed by using a filter wheel camera. The filter wheel is located in front of the sensor or lens, and the filter is replaced by rotating the filter wheel to capture a multi-channel spectral image. The spectral reflectance of each pixel is then estimated from the multi-spectral image. The advantage of the filter wheel camera is the full spatial resolution of each band, and the filters can be customized and replaced according to different spectral requirements.

[0013] In a preferred embodiment, step 3 requires building a lightweight deep learning model for target detection that can integrate multispectral data processing to facilitate the application of embedded machine vision.

[0014] The lightweight improvement of step 3 is specifically as follows: lightweight model structure and customization of main functions; lightweight is achieved by changing the backbone network, replacing the optimization algorithm and post-processing algorithm; function customization is achieved by changing the attention mechanism and cascade classifier; the specific implementation methods of lightweight and customized improvements are: changing the backbone network, the backbone network of YOLOv5 is CSPDarknet53 or ResNet, in order to reduce the amount of calculation, a lightweight ShuffleNet network is used instead; changing the attention mechanism so that the model can pay more attention to the key target area, thereby improving the accuracy of the model; improving the optimization algorithm to Lion; improving the loss function, since in liquid strain monitoring, according to specific needs, the classification of the target is more critical, the recognition accuracy can be improved by increasing the weight of the classification loss; using the Soft-NMS (non-maximum suppression) algorithm to improve the detection rate of the model.

[0015] The step 3 of fusion multispectral data processing is as follows: multispectral information fusion adopts a mid-term feature fusion strategy, that is, different modal data are first converted into high-dimensional feature expressions, and then fused in the middle layer of the model; the common mode and differential mode features are extracted by using the CAMFF algorithm, and the main feature extraction network of LSYOLO Nano is copied into two feature extraction networks for F X Band area and F Y The feature extraction of the band area image is performed; the fusion component is then used to fuse the three layers of multi-scale feature signals respectively. The fused three-layer feature output is sent to the Neck layer for a series of convolution, upsampling, downsampling and concat splicing processes to obtain the final lightweight LSYOLO Nano model that can fuse multispectral data.

[0016] In a preferred embodiment, the step 4 is specifically as follows: through the cascade calibration method, the model can identify and locate the window position, segment the area of ​​interest, and then perform key target detection and analysis in the area of ​​interest; the cascade calibration can be implemented in two stages, the first stage window area locking and the second stage target detection; in the first stage window area locking stage, firstly, combined with the characteristics of the fermentation window having a clear outline and a shape of a perfect circle of uniform size, the center red dot positioning and fixed circle size calibration method are used to anchor the fermentation window and perform circle correction, and then the circle diameter of the same specified size is used for masking and segmentation; in the second stage key target detection stage, the image data is obtained after the mask segmentation of the first stage, and the target bacteria and fermentation liquid in the fermentation tank are detected, and threshold segmentation and morphological segmentation can be used for processing to obtain the atlas information of the key target.

[0017] In a preferred embodiment, step 5 specifically includes: the edge computing device is selected to use an edge computing device with an independent graphics card to perform real-time processing of multispectral visual data at the edge node end, and solve the time lag and data security problems by performing inference and prediction directly at a location close to the data source;

[0018] Furthermore, the edge computing device is connected to other sensors through the GPIO interface expanded by the carrier board to measure environmental parameters related to the fermentation process; after data normalization, the environmental parameters are used as auxiliary features of the optical map information to form synchronization indicators, which are jointly input into the deep learning model for training to increase the detection accuracy and robustness of the model.

[0019] Compared with the existing technology, the present invention has the following beneficial effects: using a multispectral industrial camera covering characteristic bands to collect fermentation process map information, using a lightweight deep learning model that integrates multispectral image processing and target detection, and using intelligent edge computing devices to deploy the above-mentioned software and hardware at the front end, to achieve non-invasive in-situ real-time monitoring of the fermentation process, regress the fermentation growth curve, evaluate the fermentation situation, and determine the optimal fermentation endpoint. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Attachment Figure 1 This is a flow chart of a method for in-situ real-time monitoring of a liquid bacterial fermentation process based on multispectral diffuse reflectance imaging according to a preferred embodiment of the present invention;

[0021] Attachment Figure 2 This is a diagram showing the overall structure of a prototype edge computing device for a liquid bacterial fermentation process based on multispectral diffuse reflectance imaging according to a preferred embodiment of the present invention;

[0022] Attachment Figure 3 Schematic diagram of a prototype of an in-situ real-time monitoring device for a liquid bacterial fermentation process based on multispectral diffuse reflectance imaging according to a preferred embodiment of the present invention;

[0023] Attachment Figure 4 This is the differential enhancement module structure of the preferred embodiment of the present invention;

[0024] Attachment Figure 5 This is the common mode selection module structure of the preferred embodiment of the present invention;

[0025] Attachment Figure 6 This is a network architecture diagram of LSYOLO Nano according to a preferred embodiment of the present invention;

[0026] Attachment Figure 7 This is a schematic diagram of the IO port expansion of the edge computing device Jetson Nano in a preferred embodiment of the present invention;

[0027] Attachment Figure 8 This is the effect of target detection and segmentation in the first stage of the preferred embodiment of the present invention;

[0028] Attachment Figure 9 This is the effect of the two-stage target detection and data analysis of the preferred embodiment of the present invention;

[0029] Attachment Figure 3 Chinese: 1. Fermentation tank 2. Multispectral diffuse reflectance imaging equipment 3. Jatso n Nano 4, fermentation window 5, monitoring host 6, remote workstation computer. DETAILED DESCRIPTION

[0030] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0031] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meaning as commonly understood by those skilled in the art to which the present application belongs.

[0032] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form, and it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or their combinations.

[0033] refer to Figures 1 to 9 The liquid bacterial fermentation monitoring method based on diffuse reflection and edge computing includes the following steps:

[0034] Step 1: Obtain characteristic spectral information of key biological parameters during the liquid culture fermentation process. Use an ASD hyperspectrometer and a SiderWinder near-infrared spectrometer to extract key biological parameters. Specifically, obtain full-wavelength hyperspectral information of fermentation broths from normally growing liquid cultures at different growth stages and extract characteristic bands; obtain full-wavelength hyperspectral information of fermentation broths from liquid cultures contaminated with different degrees of foreign bacteria and obtain characteristic bands; obtain spectral information of bacterial clusters with different degrees of viability and obtain characteristic bands; obtain spectral information of fermentation broths with varying fermentation substrate concentration gradients at different stages and obtain characteristic bands; obtain spectral information of fermentation broths with varying fermentation product concentration gradients at different stages and obtain characteristic bands; obtain spectral information of key biological indicators at different concentrations and obtain characteristic bands; further, calculate spectral indices based on the characteristic bands to characterize key parameter information.

[0035] Step 2: Based on the above characteristic spectral indices, develop a multispectral diffuse reflectance imaging system. The main implementation schemes of the multispectral diffuse reflectance imaging system are: 1. Use a filter wheel camera to capture multi-channel spectral images by rotating the filter in the filter wheel in front of the sensor or lens, and then estimate the spectral reflectance of each pixel from the multispectral image; 2. Use multiple independent cameras and point the cameras equipped with different filters to the same target for shooting. The disadvantage is that there will be non-negligible optical parallax from cameras from different directions, and it is impossible to perfectly fuse the spectral information of specific pixels in images in different directions. Other methods include: push-broom cameras for multispectral imaging (line scanning), line scanning cameras with multi-line sensors, multispectral cameras with beam splitter elements, etc. Taking into account the requirements of cost, application goals, imaging speed and applicable environment, fermentation field monitoring mainly adopts the first method to develop a multispectral diffuse reflectance imaging system.

[0036] In this specific example, due to the factory environment's requirements for camera stability, a multispectral solution was implemented using a filter wheel camera. The filter wheel used in this prototype supports six wavelength bands: 420nm, 550nm, 710nm, 820nm, 1100nm, and 1300nm. The filters can be customized and replaced according to different application requirements. The light source uses an integrated LED light source that emits corresponding wavelength bands.

[0037] Step 2: In order to make the multispectral diffuse reflectance imaging system suitable for the fermentation window (circular window, generally 10 cm in diameter) and observe and analyze the fermentation situation, the present invention first uses a cascade calibration method to enable the visual model to identify and locate the window position, that is, find the area of ​​interest, and then perform key target detection and analysis in the area of ​​interest; specifically, the first-level window area locking and second-level semantic segmentation are used to realize data processing. In the first-level window area locking stage, first, combined with the characteristics of the fermentation window with clear outline and the shape of a perfect circle of uniform size, the center red dot positioning and fixed circle size calibration method are used to anchor the fermentation window and perform circular correction, and then the circle diameter of the same specified size is used for masking and segmentation; in the second-level key target detection stage, the image data is obtained after the mask segmentation of the first stage, and the target bacteria and fermentation liquid in the fermenter are detected. Threshold segmentation and morphological segmentation can be used for processing to obtain image information of key targets.

[0038] Furthermore, the threshold segmentation method described above adopts an adaptive threshold segmentation method, which is a type of local threshold segmentation. Unlike global thresholding, which uses a single threshold for the entire matrix, each position of the input matrix has a corresponding threshold. Assuming that the input image is I, the steps of the adaptive threshold segmentation algorithm are as follows:

[0039] Thresh=(1-ratio)*f smooth (I)

[0040]

[0041] Among them, f smooth (I) indicates smoothing of the image, which can replace mean smoothing, Gaussian smoothing, and median smoothing;

[0042] Furthermore, the morphological processing employs a closing operation to address noise and PV frame adhesion issues present in threshold segmented images through processing methods such as corrosion and dilation. Corrosion can eliminate the boundaries of the target object, resulting in an area one pixel smaller than the original target. This can separate previously adhered shapes, but reduces the size of the original shape. Dilation, on the other hand, has the opposite effect, fusing two adhered shapes. Selecting a circular structural element for the closing operation in multispectral images of liquid bacterial strains can fill small holes and cracks in the image without changing the overall position and shape of the shape, thereby improving the detection rate of bacterial clusters.

[0043] Step 3: Build an improved deep learning model LSYOLO Nano that integrates multispectral data processing to obtain a machine vision model suitable for fermentation process detection.

[0044] Furthermore, the improved LSYOLO Nano includes: using a lightweight ShuffleNet backbone network; changing the attention mechanism and replacing the optimizer Lion; optimizing the loss function; and adopting the Soft-NMS post-processing algorithm;

[0045] LSYOLO Nano uses Lion as its optimization algorithm. Optimizers play a fundamental role in training neural networks. Lion, proposed by Google and UCLA, uses programmatic search to discover optimization algorithms, addressing the challenges of searching in infinite and sparse spaces. The update process of the Lion optimizer is compared to the commonly used AdamW as follows:

[0046]

[0047] As can be seen above, Lion has fewer parameters than AdamW and eliminates square root and division operations, thus saving more video memory and processors. Furthermore, Lion has strong generalization capabilities, transcending limitations of architectures, datasets, and tasks.

[0048] Furthermore, the optimization process of the loss function used by LSYOLO Nano is as follows:

[0049] L CIoU =1-IoU+R CIoU

[0050]

[0051]

[0052] Due to L CIoU When the length and width of the gradient are between 0 and 1, the value is usually very small, so it will be replaced by 1 when detecting small targets in the fermentation process.

[0053] Furthermore, in the LSYOLO Nano model, an attention mechanism can be introduced to enhance the model's focus on the target. The SENet (Squeeze-and-Excitation Network) model strengthens important feature channels by learning the weights of each channel, thereby improving the accuracy of the model.

[0054] Furthermore, the LSYOLO Nano improves the post-processing algorithm of deep learning, which is responsible for screening and correcting the results output by the model. It uses the Soft-NMS (non-maximum suppression) algorithm to screen overlapping targets, and effectively improves the recall and detection rates of the model by reducing the confidence of the candidate boxes instead of directly deleting them. At the same time, some correction strategies are introduced.

[0055] Furthermore, the multispectral data fusion adopts mid-term feature fusion, that is, different modal data are first converted into high-dimensional feature expressions and then fused in the middle layer of the model. The CMAFF algorithm is used to fuse spectral images of different range bands. The fusion process is as follows: first, the multispectral data of 6 bands are divided into 2-way 3-channel image data, namely, pseudo-visible band X spectrum and pseudo-invisible band Y spectrum, and then differential mode and common mode feature extraction is performed;

[0056] Furthermore, the above differential mode is obtained by subtracting the two modes to obtain the differential feature map F D , then the differential enhancement module infers the feature attention map of the channel dimension based on the differential modal feature map, and multiplies the attention map with each input feature map to obtain a more refined feature map. Then, the input feature map and the refined feature map are added to obtain the enhanced feature map of each modality. Finally, F X and F Y The enhanced feature maps of the modalities are added to obtain the output of the differential enhancement module, and the structure is as follows: Figure 4 The process can be described as follows:

[0057]

[0058] s2=f GMP (F D )=max(F D (i,j))

[0059] z1=f sc (s1)

[0060] z2=f sc (s2)

[0061] M DM =σ(z1+z2)

[0062]

[0063] Among them, f GAP represents global average pooling, f GMP represents the global maximum pooling, f sc Two layers of 1*1 convolutional network, It means that the new feature map is obtained by multiplying the elements of different feature maps, and σ represents the sigmoid function.

[0064] Furthermore, the common mode selection module is based on the SoftMax attention mechanism, which can adaptively select bimodal channel features according to the common modal features and recombine them into a new feature. First, for a given size F X Convolutional features and F Y Convolution feature map, directly add it to get the common mode feature map F C Then calculate F X Band feature attention map and F Y The characteristic attention map of the band is multiplied by the input of its corresponding mode. Finally, the results obtained in the previous step are added to get the output of the common mode selection module. The structure is as follows Figure 5 The process can be described as follows:

[0065]

[0066]

[0067]

[0068]

[0069]

[0070] Furthermore, in order to extract the features of multispectral images, the backbone feature extraction network of LSYOLO Nano is copied into two feature extraction networks for F X Band area and F Y Feature extraction of band region images. The fusion component then fuses the three layers of multi-scale feature signals. The resulting fused three-layer feature output is sent to the Neck layer for a series of convolution, upsampling, downsampling, and concat processing, resulting in the final lightweight LSYOLO Nano network structure capable of fusing multispectral data.

[0071] Step three: Research and development of edge computing equipment, integration of map information acquisition and control, image preprocessing and deep learning models, and deployment of intelligent edge computing devices on terminals. The intelligent edge computing embedded device selected for this preferred example is NVDIA's JetsonNano 4GB, which uses a quad-core 64-bit ARM central processing unit (CPU) and a 128-core integrated NVDIA graphics card (GPU), which can provide 472GFLOPS of computing performance. It is equipped with the Ubuntu operating system and a rich CUDA toolkit, and is fully compatible with mainstream open source machine learning frameworks such as Pytorch and TensorFlow, and computer vision processing tool libraries such as OpenCV and ROS, making it easy to deploy artificial intelligence-based machine learning reasoning work on terminals. The expansion and deployment of intelligent edge computing devices are shown in the attached figure. Figure 7 shown.

[0072] Step 4: Expand the IO port of the edge device to realize the perception and processing of one-dimensional sequence information such as key physical parameters and chemical parameters, integrate the two-dimensional map machine vision data, conduct comprehensive analysis, and realize more comprehensive and accurate fermentation process monitoring. For example, it can judge the concentration of liquid bacteria per unit area, the purity of liquid bacteria fermentation liquid, the morphology of bacteria (spherical and flocculent), and the vitality of liquid bacteria. Figure 8 , as shown in 9.

[0073] Furthermore, the fusion process of one-dimensional environmental parameters and two-dimensional map information primarily involves serializing and normalizing the different environmental parameter data, using them as auxiliary features of the map information, and inputting them into the model for joint training. In a multi-sensor system, the environmental information provided by each information source has a certain degree of uncertainty. The fusion of this uncertain information is actually an uncertain reasoning process. Multimodal sensor data fusion can achieve ideal reasoning results by distributing and improving different weights on each parameter.

[0074] In summary, the present invention realizes non-invasive in-situ real-time monitoring of liquid bacterial fermentation process through multispectral diffuse reflectance imaging and deep learning model deployed in intelligent edge computing. Among them, the multispectral industrial camera covering characteristic bands selects pseudo-visible F according to the characteristic wavelength of target bacterial cluster and fermentation liquid. X and the invisible F Y A total of 6-band spectra were developed to obtain spectral information of key biological parameters in the fermentation process; the improved target detection algorithm LSYOLO Nano is a lightweight deep learning model that can be deployed on intelligent edge computing devices to perform machine vision tasks such as image processing, target detection and segmentation in real time, thereby conducting real-time monitoring and analysis of key parameters of the fermentation process.

Claims

1. A liquid bacterial fermentation monitoring method based on multispectral diffuse reflectance and edge computing, characterized in that: The following steps are involved: Step 1: Obtain characteristic spectral information of key biological parameters during the liquid bacterial fermentation process; Step 2: Develop a multispectral diffuse reflectance imaging system based on the acquired characteristic spectral band information of the fermentation process; Step 3: Build a lightweight deep learning vision model for target detection called LSYOLONano that can integrate multispectral data processing and make it suitable for embedded machine vision applications. Step 4: Using a cascade calibration method, the visual model can identify and locate the fermentation window position, and then perform key target detection and analysis within the region of interest; Step 5: Deploy the deep learning vision model on the edge computing device to monitor the fermentation process in real time; The lightweight improvements in step 3 specifically include: lightweighting the model structure and customizing the main functions; lightweighting is achieved by changing the backbone network, replacing the optimization algorithm and post-processing algorithm; function customization is achieved by changing the attention mechanism and cascade calibration method; the specific implementation methods of lightweight and customized improvements are: 1) changing the backbone network. The backbone network of YOLOv5 is CSPDarknet53 or ResNet. In order to reduce the amount of computation, a lightweight ShuffleNet network is used instead; 2) changing the optimization algorithm to Lion; 3) optimizing the loss function; 4) using the Soft-NMS non-maximum suppression algorithm to improve the detection rate of the model; The step 3 of fusion multispectral data processing is as follows: multispectral information fusion adopts a mid-term feature fusion strategy, that is, different modal data are first converted into high-dimensional feature expressions, and then fused in the middle layer of the model; the common mode and differential mode features are extracted by using the CAMFF algorithm, and the main feature extraction network of LSYOLO Nano is copied into two feature extraction networks for F X Band area and F Y The feature extraction of the band region image is then performed. The fusion component is then used to fuse the three layers of multi-scale feature signals. The fused three-layer feature output is sent to the Neck layer for a series of convolution, upsampling, downsampling, and concat splicing processes to obtain the final lightweight LSYOLO Nano model that can fuse multispectral data. The cascade calibration method of step 4 is specifically as follows: first, the model is made identifiable and located at the window position, the region of interest is segmented, and then key target detection and analysis are performed within the region of interest.

2. The liquid strain fermentation monitoring method based on multispectral diffuse reflectance and edge computing according to claim 1 is characterized in that: Step 1 specifically includes: using a hyperspectral spectrometer covering ultraviolet, visible, and near-infrared to collect spectral information of key biological parameters at different concentrations, spectral information of culture solutions at different concentrations, spectral information of fermentation solutions at different growth stages, and spectral information of fermentation solutions under healthy and polluted conditions; and calculating the spectral index based on the above spectral information to obtain characteristic bands of interest that can robustly characterize the corresponding parameters.

3. The liquid strain fermentation monitoring method based on multispectral diffuse reflectance and edge computing according to claim 1 is characterized in that: The step 2 is to use a filter wheel camera, where the filter wheel is located in front of the sensor or lens. The filter is replaced by rotating the filter wheel to capture a multi-channel spectral image, and then the spectral reflectance of each pixel is estimated from the multi-spectral image.

4. The liquid strain fermentation monitoring method based on multispectral diffuse reflectance and edge computing according to claim 1 is characterized in that: Specifically, step 5 includes: the edge computing device is selected to use an edge computing device with an independent graphics card to perform real-time processing of multispectral visual data at the edge node end, so that the data can be directly inferred and predicted close to the data source, solving time lag and data security issues; The edge computing device is connected to other sensors through the GPIO interface extended by the carrier board to measure parameters related to the fermentation process environment; After data normalization, the environmental parameters are used as auxiliary features of the optical map information to form synchronization indicators, which are jointly input into the deep learning vision model for training to increase the detection accuracy and robustness of the model.

Images