Method and system for detecting freshness of aquatic products

Abstract

The application discloses an aquatic product freshness detection method, which comprises the following steps: 1) constructing a fluorescence sensing array label based on CDs with AIE characteristics; 2) collecting fluorescence sensing array label images and corresponding freshness data; 3) establishing a deep convolutional neural network model, and training the deep convolutional neural network model based on the historical collected fluorescence sensing array label images and corresponding freshness, to obtain a trained deep convolutional neural network model; and 4) inputting the real-time collected fluorescence sensing array label images into the trained deep convolutional neural network model, and outputting aquatic product freshness classification results. The application constructs a detection system combining the fluorescence sensing array label based on CDs with AIE characteristics and a light-weight deep convolutional neural network, can realize visual real-time detection of aquatic product freshness, has the advantages of simple operation, no damage to aquatic products, rapid detection and the like, and is suitable for large-scale application.

Description

Technical Field

[0001] This invention relates to the field of non-destructive testing technology for food, specifically to a method and system for detecting the freshness of aquatic products that combines an AIE fluorescence sensor array with a deep convolutional neural network. Background Technology

[0002] Aquatic products are a general term for aquatic animal and plant products and their processed products produced by marine and freshwater fisheries. They are generally popular with consumers due to their tender texture, delicious taste, and rich nutrition. However, during transportation, storage, and sales, aquatic products are susceptible to protein decomposition and fat oxidation caused by endogenous enzymes, microorganisms, and other factors, which may also lead to the release of unpleasant gases, resulting in a decline in quality and freshness, or even spoilage. Food safety issues caused by aquatic product spoilage cannot be ignored and can involve public safety. Therefore, accurately monitoring the freshness of aquatic products in the supply chain is an essential link in ensuring their safety.

[0003] Currently, methods for detecting the freshness of aquatic products mainly include traditional physicochemical analysis methods (analytical indicators include TVB-N, pH value, K value, and thiobarbituric acid value, etc.) and high-performance liquid chromatography (HPLC). However, while these methods can meet certain detection needs, they can cause serious sample damage during the testing process and cannot meet the requirements for in-situ, real-time, and visual feedback of aquatic product freshness during storage, transportation, and sales. In addition, some on-site detection methods (such as machine vision and handheld Raman spectroscopy) are used for on-site detection of aquatic product freshness, but these methods all suffer from problems such as expensive testing instruments, poor equipment applicability, and low detection efficiency, and cannot fully meet the requirements of practical applications.

[0004] Therefore, developing a simple, rapid, in-situ, and visualized method and system for detecting the freshness of aquatic products is of great significance.

[0005] The above statements are merely background information related to the present invention and do not necessarily constitute prior art. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for detecting the freshness of aquatic products.

[0007] The technical solution adopted in this invention is:

[0008] A method for detecting the freshness of aquatic products includes the following steps:

[0009] 1) A fluorescent sensing array tag is obtained by arranging various carbon quantum dots with aggregation-induced emission properties in an array and fixing them on the substrate surface;

[0010] 2) Place the aquatic product samples into a preservation box and fix the fluorescent sensor array tag to the inside of the box lid. After sealing the preservation box, store it at low temperature. Then, collect images of the fluorescent sensor array tag at regular intervals and simultaneously measure the TVB-N (total volatile basic nitrogen) value of the aquatic product samples. Determine the freshness of the aquatic product samples based on the TVB-N value.

[0011] 3) Establish a deep convolutional neural network model, and train the deep convolutional neural network model based on historically acquired fluorescent sensor array tag images and their corresponding freshness to obtain the trained deep convolutional neural network model.

[0012] 4) Input the real-time acquired fluorescent sensor array tag images into the trained deep convolutional neural network model to output the aquatic product freshness classification results.

[0013] Preferably, the fluorescent sensing array tag in step 1) contains nine carbon quantum dots with aggregation-induced emission properties, arranged in a 3×3 array.

[0014] Preferably, the nine carbon quantum dots with aggregation-induced emission properties are prepared by a method comprising the following steps:

[0015] Preparation of CD1: Pyromellitic acid (PDMA) and o-phenylenediamine (o-PDA) were dissolved in an aqueous ethanol solution and subjected to a microwave reaction. The mass ratio of pyromellitic acid to o-phenylenediamine was 1:0.30-0.75. The products were then separated to obtain carbon quantum dots CD1 with aggregation-induced emission properties.

[0016] Preparation of CD2: 2,2'-dithiodibenzoic acid (DTSA) and ammonium sulfate were dissolved in glacial acetic acid and subjected to a solvothermal reaction. The mass ratio of 2,2'-dithiodibenzoic acid to ammonium sulfate was 1:0.12 to 1.50. The products were then separated to obtain carbon quantum dots CD2 with aggregation-induced emission properties.

[0017] Preparation of CD3: 2,2'-dithiodibenzoic acid, urea and copper salt were dissolved in glacial acetic acid and subjected to a solvothermal reaction. The mass ratio of 2,2'-dithiodibenzoic acid, urea and copper salt was 1:0.15~2.50:0.004~0.125. The products were then separated to obtain carbon quantum dots CD3 with aggregation-induced emission properties.

[0018] Preparation of CD4: 2,2'-dithiodibenzoic acid and urea were dissolved in glacial acetic acid and subjected to a solvothermal reaction. The mass ratio of 2,2'-dithiodibenzoic acid to urea was 1:0.25-1.00. The products were then separated to obtain carbon quantum dots CD4 with aggregation-induced emission properties.

[0019] Preparation of CD5: 2,2'-dithiodibenzoic acid and urea were dissolved in glacial acetic acid and subjected to a solvothermal reaction. The mass ratio of 2,2'-dithiodibenzoic acid to urea was 1:0.80-1.35. The products were then separated to obtain carbon quantum dots CD5 with aggregation-induced emission properties.

[0020] Preparation of CD6: 2,2'-dithiodibenzoic acid, ammonium sulfate and copper salt were dissolved in glacial acetic acid and subjected to a solvothermal reaction. The mass ratio of 2,2'-dithiodibenzoic acid, ammonium sulfate and copper salt was 1:0.12-1.50:0.003-0.125. The products were then separated to obtain carbon quantum dots CD6 with aggregation-induced emission properties.

[0021] Preparation of CD7: 2,2'-dithiodibenzoic acid and melamine were dissolved in glacial acetic acid and subjected to a solvothermal reaction. The mass ratio of 2,2'-dithiodibenzoic acid to melamine was 1:0.20-0.60. The products were then separated to obtain carbon quantum dots CD7 with aggregation-induced emission properties.

[0022] Preparation of CD8: Pyromellitic acid, o-phenylenediamine and copper salt were dissolved in an ethanol aqueous solution and subjected to microwave reaction. The mass ratio of pyromellitic acid, o-phenylenediamine and copper salt was 1:0.30-0.75:0.0003-0.025. The products were then separated to obtain carbon quantum dots CD8 with aggregation-induced emission properties.

[0023] Preparation of CD9: 2,2'-dithiodibenzoic acid and urea were dissolved in glacial acetic acid and subjected to a solvothermal reaction. The mass ratio of 2,2'-dithiodibenzoic acid to urea was 1:0.12-1.25. The products were then separated to obtain carbon quantum dots CD9 with aggregation-induced emission properties.

[0024] Preferably, the ethanol-water solution in the preparation process of CD1 is prepared by mixing ethanol and water in a volume ratio of 18 to 22:1.

[0025] Preferably, the microwave reaction in the preparation process of CD1 is carried out under the condition of microwave power of 500W to 900W and the reaction time is 8min to 12min.

[0026] Preferably, the product separation process in the preparation of CD1 includes the following operations: adding the reaction solution to dimethyl sulfoxide (DMSO), centrifuging, taking the supernatant and adding it to water, and then filtering.

[0027] Preferably, the solvothermal reaction in the preparation process of CD2 is carried out at a temperature of 180℃ to 210℃ for a reaction time of 9h to 12h.

[0028] Preferably, the product separation process in the preparation of CD2 includes the following operations: washing the reaction solution with boiling water and then filtering it.

[0029] Preferably, the copper salt used in the preparation of CD3 is at least one of copper nitrate, copper sulfate, copper chloride, and copper acetate.

[0030] Preferably, the solvothermal reaction in the preparation of CD3 is carried out at a temperature of 160℃ to 180℃ for a reaction time of 8h to 12h.

[0031] Preferably, the product separation process in the preparation of CD3 includes the following operations: washing the reaction solution with boiling water and then filtering it.

[0032] Preferably, the solvothermal reaction in the preparation of CD4 is carried out at a temperature of 170℃ to 200℃ for a reaction time of 8h to 12h.

[0033] Preferably, the product separation process in the preparation of CD4 includes the following operations: washing the reaction solution with boiling water and then filtering it.

[0034] Preferably, the solvothermal reaction in the preparation process of CD5 is carried out at a temperature of 170℃ to 200℃ for a reaction time of 8h to 12h.

[0035] Preferably, the product separation process in the preparation of CD5 includes the following operations: washing the reaction solution in boiling water and then filtering it.

[0036] Preferably, the copper salt used in the preparation of CD6 is at least one of copper nitrate, copper sulfate, copper chloride, and copper acetate.

[0037] Preferably, the solvothermal reaction in the preparation process of CD6 is carried out at a temperature of 180℃ to 210℃ for a reaction time of 9h to 12h.

[0038] Preferably, the product separation process in the preparation of CD6 includes the following operations: washing the reaction solution in boiling water and then filtering it.

[0039] Preferably, the solvothermal reaction in the preparation process of CD7 is carried out at a temperature of 160℃ to 180℃ for a reaction time of 8h to 10h.

[0040] Preferably, the product separation process in the preparation of CD7 includes the following operations: washing the reaction solution in boiling water and then filtering it.

[0041] Preferably, the copper salt used in the preparation of CD8 is at least one of copper nitrate, copper sulfate, copper chloride, and copper acetate.

[0042] Preferably, the ethanol-water solution in the preparation process of CD8 is prepared by mixing ethanol and water in a volume ratio of 18 to 22:1.

[0043] Preferably, the microwave reaction in the preparation process of CD8 is carried out under the condition of microwave power of 500W to 900W and the reaction time is 8min to 12min.

[0044] Preferably, the product separation process in the preparation of CD8 includes the following operations: adding the reaction solution to dimethyl sulfoxide, centrifuging, taking the supernatant and adding it to water, and then filtering.

[0045] Preferably, the solvothermal reaction in the preparation process of CD9 is carried out at a temperature of 160℃ to 180℃ for a reaction time of 8h to 12h.

[0046] Preferably, the product separation process in the preparation of CD9 includes the following operations: washing the reaction solution with boiling water and then filtering it.

[0047] Preferably, the fluorescent sensor array tag in step 1) is prepared by a method including the following steps: dispersing carbon quantum dots with aggregation-induced emission properties in an organic solvent to prepare a carbon quantum dot dispersion, cutting a hydrophobic film into multiple sensor response areas, then sequentially stacking and bonding the uncut hydrophobic film, cotton fiber paper, and the hydrophobic film with the cut sensor response areas, then dropping the carbon quantum dot dispersion onto the surface of the cotton fiber paper exposed in the sensor response areas, and then drying to obtain the fluorescent sensor array tag.

[0048] Preferably, the organic solvent is at least one selected from ethanol, methanol, dimethyl sulfoxide, dimethylformamide, and tetrahydrofuran.

[0049] Preferably, the concentration of the carbon quantum dot dispersion is 2 mg / mL to 5 mg / mL.

[0050] Preferably, the hydrophobic film is one of a hydrophobic organic film or a carbon fiber film.

[0051] Preferably, the amount of carbon quantum dot dispersion added to the surface of the cotton fiber paper exposed in the sensor response area is 100 μL to 130 μL.

[0052] Preferably, the drying is carried out at a temperature of 40℃ to 50℃ for a drying time of 5h to 10h.

[0053] Preferably, the freshness in step 2) is divided into three categories: fresh, slightly fresh, and spoiled.

[0054] Preferably, the deep convolutional neural network model in step 3) is MobileNet v3 large.

[0055] A system for detecting the freshness of aquatic products, used to implement the above-mentioned method for detecting the freshness of aquatic products, comprises:

[0056] The image acquisition and processing module is used to acquire images of the fluorescent sensor array tag at regular intervals and simultaneously acquire the TVB-N value of the aquatic product sample, and determine the freshness of the aquatic product sample based on the TVB-N value.

[0057] The model building and training module is used to build a deep convolutional neural network model for freshness detection, and to train the deep convolutional neural network model based on the collected fluorescent sensor array label images and the corresponding freshness, so as to obtain the trained deep convolutional neural network model.

[0058] The results output module is used to input real-time acquired fluorescence sensor array image data into the trained deep convolutional neural network model and output the freshness classification results of aquatic products.

[0059] The beneficial effects of this invention are: This invention constructs a detection system based on a fluorescent sensor array tag with AIE characteristics and a lightweight deep convolutional neural network, which can realize the real-time visual detection of the freshness of aquatic products. It has the advantages of simple operation, no damage to aquatic products, and rapid detection, and is suitable for large-scale application. Attached Figure Description

[0060] Figure 1 This is a flowchart of the aquatic product freshness detection method of the present invention.

[0061] Figure 2 The images show the fluorescence emission and excitation spectra of CD1 to CD9 in Example 1.

[0062] Figure 3 This is a SEM image of the sensor response region of the fluorescent sensor array tag in Example 1.

[0063] Figure 4 This is a graph showing the response results of the fluorescent sensor array tag in Example 1 to different volatile deterioration marker gases.

[0064] Figure 5 The graph shows the response results of the fluorescent sensor array tag in Example 1 to ammonia gas of different concentrations.

[0065] Figure 6The image shows the Euclidean distance test results of the fluorescent sensor array tag in Example 1 after responding to ammonia gas of different concentrations.

[0066] Figure 7 This is a graph showing the fluorescence color change of the fluorescent sensor array tag in Example 1 during the storage of prawns.

[0067] Figure 8 The diagram shows the ROC curve and prediction set confusion matrix of the MobileNet v3 large model in Example 1. Detailed Implementation

[0068] The present invention will be further explained and described below with reference to specific embodiments.

[0069] Example 1:

[0070] A method for detecting the freshness of aquatic products, the steps of which are as follows (flowchart shown) Figure 1 (as shown)

[0071] 1) Fabrication of fluorescent sensing array tags:

[0072] a) Nine types of carbon quantum dots with aggregation-induced emission properties:

[0073] Preparation of CD1: 1.0492 g of pyromellitic acid (PDMA) and 0.4520 g of o-phenylenediamine (o-PDA) were dissolved in 20 mL of ethanol-water solution (ethanol:water volume ratio of 20:1) and then placed in a microwave reactor and reacted for 10 min under microwave power of 600 W. The reaction solution was then added to 15 mL of dimethyl sulfoxide (DMSO), centrifuged, and the supernatant was added to 100 mL of deionized water. The mixture was then vacuum filtered through a filter with a pore size of 0.22 μm to obtain carbon quantum dots CD1 with aggregation-induced emission properties.

[0074] Preparation of CD2: 0.3 g of 2,2'-dithiodibenzoic acid (DTSA) and 0.15 g of ammonium sulfate were added to 10 mL of glacial acetic acid and sonicated for 5 min. The mixture was then transferred to a 20 mL reaction vessel and placed in an electric heating drying oven at 200 °C for 10 h. After natural cooling to room temperature, the reaction solution was added to 1.5 L of boiling water and then vacuum filtered through a 0.22 μm filter to obtain carbon quantum dots CD2 with aggregation-induced emission properties.

[0075] Preparation of CD3: 0.25 g of 2,2'-dithiodibenzoic acid, 0.1 g of urea and 0.1 mL of 0.1 mol / L copper nitrate aqueous solution were added to 20 mL of glacial acetic acid and sonicated for 8 min. The mixture was then transferred to a 50 mL reaction vessel and placed in an electric heating drying oven at 180 °C for 10 h. After natural cooling to room temperature, the reaction solution was added to 1.5 L of boiling water and then vacuum filtered through a 0.22 μm filter to obtain carbon quantum dots CD3 with aggregation-induced emission properties.

[0076] Preparation of CD4: 0.3 g of 2,2'-dithiodibenzoic acid and 0.1 g of urea were added to 40 mL of glacial acetic acid and sonicated for 8 min. The mixture was then transferred to a 100 mL reaction vessel and placed in an electric heating drying oven at 180 °C for 10 h. After natural cooling to room temperature, the reaction solution was added to 1.5 L of boiling water and then vacuum filtered through a 0.22 μm pore size filter to obtain carbon quantum dots CD4 with aggregation-induced emission properties.

[0077] Preparation of CD5: 0.3 g of 2,2'-dithiodibenzoic acid and 0.24 g of urea were added to 40 mL of glacial acetic acid and sonicated for 8 min. The mixture was then transferred to a 100 mL reaction vessel and placed in an electric heating drying oven at 190 °C for 10 h. After natural cooling to room temperature, the reaction solution was added to 1.5 L of boiling water and then vacuum filtered through a 0.22 μm pore size filter to obtain carbon quantum dots CD5 with aggregation-induced emission properties.

[0078] Preparation of CD6: 0.3 g of 2,2'-dithiodibenzoic acid, 0.15 g of ammonium sulfate and 0.1 mL of 0.2 mol / L copper nitrate aqueous solution were added to 10 mL of glacial acetic acid and sonicated for 8 min. The mixture was then transferred to a 20 mL reaction vessel and placed in an electric heating drying oven at 200 °C for 10 h. After natural cooling to room temperature, the reaction solution was added to 1.5 L of boiling water and then vacuum filtered through a 0.22 μm filter to obtain carbon quantum dots CD6 with aggregation-induced emission properties.

[0079] Preparation of CD7: 0.8 g of 2,2'-dithiodibenzoic acid and 0.3 g of melamine were added to 40 mL of glacial acetic acid and sonicated for 5 min. The mixture was then transferred to a 100 mL reaction vessel and placed in an electric heating drying oven at 180 °C for 10 h. After natural cooling to room temperature, the reaction solution was added to 1.5 L of boiling water and then vacuum filtered through a 0.22 μm pore size filter to obtain carbon quantum dots CD7 with aggregation-induced emission properties.

[0080] Preparation of CD8: 0.9 g of pyromellitic acid, 0.5 g of o-phenylenediamine and 0.1 mL of 0.2 mol / L copper nitrate aqueous solution were stirred and dissolved in 20 mL of ethanol aqueous solution (ethanol:water volume ratio of 20:1). The solution was then placed in a microwave reactor and reacted for 10 min under microwave power of 600 W. The reaction solution was then added to 15 mL of dimethyl sulfoxide (DMSO), centrifuged, and the supernatant was added to 100 mL of deionized water. The solution was then vacuum filtered through a filter with a pore size of 0.22 μm to obtain carbon quantum dots CD8 with aggregation-induced emission properties.

[0081] Preparation of CD9: 0.25 g of 2,2'-dithiodibenzoic acid and 0.1 g of urea were added to 20 mL of glacial acetic acid and sonicated for 8 min. The mixture was then transferred to a 50 mL reaction vessel and placed in an electric heating drying oven at 180 °C for 10 h. After natural cooling to room temperature, the reaction solution was added to 1.5 L of boiling water and then vacuum filtered through a 0.22 μm pore size filter to obtain carbon quantum dots CD9 with aggregation-induced emission properties.

[0082] b) Fixation of carbon quantum dots on the substrate surface: Carbon quantum dots (CD1~CD9) with aggregation-induced emission properties were stirred and dispersed in ethanol to prepare a carbon quantum dot dispersion with a concentration of 3 mg / mL. Nine sensor response regions were cut out of a carbon fiber film in a 3×3 array. The uncut carbon fiber film, circular cotton fiber paper (1 cm in diameter) and the carbon fiber film with the cut sensor response regions were stacked and bonded in sequence. 110 μL of carbon quantum dot dispersion was dropped onto the surface of the cotton fiber paper exposed in the sensor response region. The paper was then transferred to an oven and dried at 40 °C for 6 h to obtain a fluorescent sensor array tag.

[0083] 2) Pat the surface moisture of the prawns (purchased from the supermarket) dry with laboratory paper. Place 65g ± 0.5g of prawns at the bottom of a 130mm diameter cuvette and attach a fluorescent sensor array tag to the top inside the cuvette. Place the cuvette at 4℃ for 5 days. Every day, collect image data of the fluorescent sensor array tag using a smartphone under a 365nm UV lamp. Randomly select three boxes of samples and determine the TVB-N value of the response samples (prawns) using a Kjeldahl nitrogen analyzer according to "GB 5009.228-2016 National Food Safety Standard - Determination of Volatile Basic Nitrogen in Food". Determine the freshness grade of the aquatic product samples according to the standard (fresh: TVB-N value <15mg / 100g; less fresh: 15mg / 100g≤TVB-N value≤25mg / 100g; spoiled: TVB-N value>25mg / 100g) based on the TVB-N value.

[0084] 3) Establish the MobileNet v3 large model and train the MobileNet v3 large model based on historically acquired fluorescent sensor array tag images and their corresponding freshness to obtain the trained MobileNet v3 large model.

[0085] 4) Input the real-time acquired fluorescent sensor array tag images into the trained MobileNet v3 large model to output the aquatic product freshness classification results.

[0086] Performance testing:

[0087] 1) The fluorescence emission spectrum and excitation spectrum of the carbon quantum dots (CD1~CD9) with aggregation-induced emission properties in this embodiment are as follows: Figure 2 (As shown in the figure, a to i correspond to CD1 to CD9 respectively.)

[0088] Depend on Figure 2 It can be seen that: CD1 has an excitation wavelength of 354 nm and an emission wavelength of 480 nm; CD2 has an excitation wavelength of 366 nm and an emission wavelength of 455 nm; CD3 has an excitation wavelength of 374 nm and an emission wavelength of 473 nm; CD4 has an excitation wavelength of 370 nm and an emission wavelength of 465 nm; CD5 has an excitation wavelength of 375 nm and an emission wavelength of 475 nm; CD6 has an excitation wavelength of 369 nm and an emission wavelength of 470 nm; CD7 has an excitation wavelength of 372 nm and an emission wavelength of 445 nm; CD8 has an excitation wavelength of 343 nm and an emission wavelength of 477 nm; and CD9 has an excitation wavelength of 370 nm and an emission wavelength of 475 nm. In summary, CD1 to CD9 all have good aggregation-induced emission characteristics.

[0089] 2) The scanning electron microscope (SEM) image of the sensor response region of the fluorescence sensing array tag in this embodiment is shown below. Figure 3 (SEM images of the sensor response regions where CD1 to CD9 dispersions were dropped, corresponding to A to I respectively; the blank control group is an SEM image of the unmodified cotton fiber paper surface.)

[0090] Depend on Figure 3 It can be seen that the sensor response area (modified with CDs) of the fluorescent sensor array tag can be observed to contain aggregate particles in the form of sheets, needles, and petals, indicating that CDs formed aggregates on the surface of cotton fiber paper, and the fluorescent sensor array tag was successfully prepared.

[0091] 3) Feasibility verification of using fluorescent sensor array tags to detect volatile spoilage marker gases in aquatic products: At room temperature, fluorescent sensor array tags were placed in a 1L cylindrical sealed container and reacted with ammonia (NH3), dimethyl sulfoxide (DMSO), ethyl acetate (EtOAc), ethanol (EtOH), hydrogen sulfide (H2S), acetic acid (HAc), formaldehyde (HCHO), tetrahydrofuran (THF), and water (H2O) at concentrations of 500ppm, respectively. After 1 hour, images of the fluorescent sensor array tags were acquired using a smartphone under ultraviolet light, and the color change of the fluorescent sensor array tags was measured by the Euclidean distance ΔE.

[0092] The formula for calculating gas concentration is as follows:

[0093] C=(V 挥发性腐败标志物 ×D 挥发性腐败标志物 ×W 挥发性腐败标志物 ) / (M 挥发性腐败标志物 ×V 容器 )×22.4×10 -9 In the formula, C is the concentration of volatile spoilage markers (in ppm), and V 挥发性腐败标志物 D represents the volume (in μL) of volatile spoilage markers. 挥发性腐败标志物 W is the density (in mg / L) of volatile spoilage markers. 挥发性腐败标志物 M represents the mass fraction of volatile spoilage markers. 挥发性腐败标志物 V represents the molar mass (in g / mol) of a volatile spoilage marker. 容器 The volume of the container is expressed in liters (L).

[0094] The formula for calculating the Euclidean distance ΔE of color is as follows:

[0095] In the formula, ΔR, ΔG, and ΔB are the differences in the R, G, and B color channels before and after the reaction, respectively.

[0096] Meanwhile, the required volumes of ammonia gas with concentrations of 5 ppm, 50 ppm, 100 ppm, 200 ppm, 300 ppm, 400 ppm, and 500 ppm were calculated using the gas concentration calculation formula. At room temperature, fluorescent sensor array tags were placed in a 1 L cylindrical sealed container and reacted with ammonia gas of different concentrations. After 1 hour, images of the fluorescent sensor array tags were acquired using a smartphone under ultraviolet light, and the change in the Euclidean distance ΔE of the color was calculated. This was done to explore whether the fluorescent sensor array tags could detect ammonia gas of different concentrations and whether they could be used to detect the freshness of aquatic products.

[0097] The characteristic images were reconstructed from the feature values ​​of the fluorescent sensor array tag and the volatile putrefactive gas to be tested. The response results of the fluorescent sensor array tag to different volatile putrefactive gases are as follows: Figure 4 As shown, the response results of the fluorescent sensor array tag to different concentrations of ammonia gas are as follows: Figure 5 As shown, the Euclidean distance test results after the fluorescent sensor array tag responds to different concentrations of ammonia gas are as follows: Figure 6 As shown.

[0098] Depend on Figure 4 It can be seen that the feature image changes with the type of volatile putrefactive gas to be tested. Compared with other gases, the fluorescence intensity of the fluorescent sensor array tag after ammonia treatment changed significantly, indicating that the fluorescent sensor array tag has excellent responsiveness to ammonia.

[0099] Depend on Figure 5 It can be seen that as the concentration of ammonia increases, the difference in the response images of the fluorescent sensor array tags continuously increases.

[0100] Depend on Figure 6 It can be seen that the Euclidean distance ΔE increases continuously with the increase of ammonia concentration, and... Figure 5 The results are consistent.

[0101] In summary, the fluorescent sensor array tag in this embodiment can be used to detect the freshness of aquatic products.

[0102] 4) The fluorescence color change of the fluorescent sensor array tag in this embodiment during the storage of prawns is as follows: Figure 7 As shown.

[0103] Depend on Figure 7 It can be seen that the freshness level of different aquatic product samples can be effectively distinguished according to the standard (fresh: TVB-N value <15mg / 100g; less fresh: 15mg / 100g≤TVB-N value≤25mg / 100g; spoiled: TVB-N value>25mg / 100g).

[0104] 5) Validation of deep convolutional neural network models: Six deep convolutional neural network models were selected, and all were adapted to the three-class classification task through transfer learning:

[0105] The MobileNet v3 large model, as a representative of lightweight networks, maintains a low number of parameters and computational complexity while utilizing optimized depthwise separable convolution and efficient activation function design, thus possessing both speed and accuracy advantages. It is the best performing model among the six deep convolutional neural network models.

[0106] The MobileNet v3 small model is a lightweight version of the series, with a more compact structure, making it suitable for running in resource-constrained environments, but its classification accuracy is slightly lower than that of the MobileNet v3 large model.

[0107] The EfficientNet b0 model employs a composite scaling strategy to achieve a balance between network depth, width, and resolution, enabling it to achieve good performance with a relatively small number of parameters.

[0108] The ResNet 18 model, as a typical residual network, effectively alleviates the gradient vanishing problem by introducing residual connections. It performs stably when training deep networks and is suitable for medium-sized classification tasks.

[0109] The VGG 16 model, as an early classic deep convolutional neural network, has a regular stacked convolutional and fully connected layer structure. It has a large parameter scale and is prone to overfitting, but it has strong feature extraction capabilities.

[0110] The GhostNet 1x model is a new type of lightweight network that generates feature maps efficiently while maintaining low computational overhead through the Ghost module, making it suitable for mobile or embedded devices.

[0111] In the embodiments of this invention, all six deep convolutional neural network models mentioned above adjust the final classification layer to have 3 output nodes to adapt to the specific requirements of three-class classification tasks.

[0112] During model training, the Cross-Entropy Loss function was consistently used as the loss function for the three-class classification task. The Adam optimizer was selected, and the initial learning rate was set to 0.001. The training epochs were 100, the batch size was 32, and an early stopping mechanism (patience=10) was introduced to automatically stop training when the accuracy on the validation set no longer improved after a certain number of consecutive epochs, in order to avoid overfitting.

[0113] The final classification layer was modified accordingly for different deep convolutional neural network models:

[0114] In the VGG 16 model, the original classifier's 6th fully connected layer is replaced with a fully connected layer with 3 output nodes (4096→3).

[0115] In ResNet 18, the fully connected layer (fc) is replaced with a fully connected layer with an output dimension of 3.

[0116] In the MobileNet v3 small model, the third layer of the classifier module is replaced with a fully connected layer with 3 output nodes (1024→3).

[0117] In the MobileNet v3 large model, the third layer of the classifier module is replaced with a fully connected layer with 3 output nodes (1280→3).

[0118] In the EfficientNet b0 model, the first layer of the classifier is replaced with a fully connected layer with 3 output nodes (1280→3).

[0119] In the GhostNet 1x model, the classifier is replaced with a fully connected layer with 3 output nodes (1280→3).

[0120] By setting the hyperparameters and adjusting the structure as described above, we can ensure that each model can adapt to the requirements of the three-class classification task under transfer learning conditions, while taking into account both training efficiency and model performance.

[0121] The classification comparison of different models is shown in the table below:

[0122] Table 1. Comparison of different models

[0123] Model types accuracy Accuracy Recall rate F1 - Score AUC EfficientNet b0 0.9672 0.9682 0.9672 0.9671 0.9922 GhostNet 1x 0.9262 0.9311 0.9262 0.9245 0.9890 ResNet 18 0.9262 0.9360 0.9262 0.9239 0.9844 VGG 16 0.9180 0.9207 0.9180 0.9164 0.9788 MobileNet v3 small 0.9180 0.9222 0.9180 0.9163 0.9839 MobileNet v3 large 0.9754 0.9755 0.9754 0.9753 0.9974

[0124] To further reveal the specific performance of the MobileNet v3 large model across different freshness categories, the classification performance of three sample classes—"fresh," "slightly fresh," and "spoiled"—will be analyzed, as shown by the ROC curves. Figure 8 In example a), the AUC values ​​of all three classes of samples are close to 1.0, indicating that the model has good discriminative ability for each class of samples.

[0125] Confusion matrix ( Figure 8 b) shows the model's predicted distribution for each category. It can be seen that the model is most accurate in identifying "fresh" samples, with all 37 samples correctly classified. However, there is some confusion between the less fresh and spoiled samples (2 "less fresh" samples were misclassified as "spoiled" and 1 "spoiled" sample was misclassified as "less fresh"), suggesting that the model still has some room for optimization when dealing with adjacent categories.

[0126] The classification results of the MobileNet v3 large model in the fresh, slightly fresh, and spoiled categories are shown in the table below:

[0127] Table 2. Classification results of the MobileNet v3 large model in the fresh, semi-fresh, and spoiled categories.

[0128] grade accuracy Accuracy Recall rate F1 - Score Sample size Fresh 1.0 1.0 1.0 1.0 37 Second fresh 0.9473 0.9729 0.9473 0.9599 38 corruption 0.9787 0.9583 0.9872 0.9684 47

[0129] Further validation of the above analysis is achieved by combining the classification results in Table 2. The MobileNet v3 large model achieves a classification accuracy of over 94% across the three classes of samples, with F1-scores all exceeding 95%. The "fresh" category shows the highest scores across all metrics, reaching 1.0, indicating that the model is more accurate and sensitive in identifying samples with this feature. Overall, all model metrics exceed 0.94%, demonstrating excellent classification ability and stability.

[0130] Example 2:

[0131] A freshness detection system for aquatic products, used to implement the freshness detection method for aquatic products in Example 1, comprises the following components:

[0132] The image acquisition and processing module is used to acquire images of the fluorescent sensor array tag at regular intervals and simultaneously acquire the TVB-N value of the aquatic product sample, and determine the freshness of the aquatic product sample based on the TVB-N value.

[0133] The model building and training module is used to build a deep convolutional neural network model for freshness detection, and to train the deep convolutional neural network model based on the collected fluorescent sensor array label images and the corresponding freshness, so as to obtain the trained deep convolutional neural network model.

[0134] The results output module is used to input real-time acquired fluorescence sensor array image data into the trained deep convolutional neural network model and output the freshness classification results of aquatic products.

[0135] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for detecting the freshness of aquatic products, characterized in that, Includes the following steps: 1) A fluorescent sensing array tag is obtained by arranging various carbon quantum dots with aggregation-induced emission properties in an array and fixing them on the substrate surface; 2) Place the aquatic product samples into a preservation box, fix the fluorescent sensor array tag to the inside of the box lid, seal the preservation box and store it at low temperature. Then, collect images of the fluorescent sensor array tag at regular intervals and simultaneously measure the TVB-N value of the aquatic product samples. Determine the freshness of the aquatic product samples based on the TVB-N value. 3) Establish a deep convolutional neural network model, and train the deep convolutional neural network model based on historically acquired fluorescent sensor array tag images and their corresponding freshness to obtain the trained deep convolutional neural network model. 4) Input the real-time acquired fluorescent sensor array tag images into the trained deep convolutional neural network model to output the aquatic product freshness classification results; Step 1) The fluorescent sensing array tag contains 9 types of carbon quantum dots with aggregation-induced emission properties, arranged in a 3×3 array. The nine carbon quantum dots with aggregation-induced emission properties are prepared by a method including the following steps: Preparation of CD1: Pyromellitic acid and o-phenylenediamine were dissolved in an ethanol aqueous solution and subjected to microwave reaction. The mass ratio of pyromellitic acid to o-phenylenediamine was 1:0.30-0.

75. The products were then separated to obtain carbon quantum dots CD1 with aggregation-induced emission properties. Preparation of CD2: 2,2'-dithiodibenzoic acid and ammonium sulfate were dissolved in glacial acetic acid and subjected to a solvothermal reaction. The mass ratio of 2,2'-dithiodibenzoic acid to ammonium sulfate was 1:0.12 to 1.

50. The products were then separated to obtain carbon quantum dots CD2 with aggregation-induced emission properties. Preparation of CD3: 2,2'-dithiodibenzoic acid, urea and copper salt were dissolved in glacial acetic acid and subjected to a solvothermal reaction. The mass ratio of 2,2'-dithiodibenzoic acid, urea and copper salt was 1:0.15~2.50:0.004~0.

125. The products were then separated to obtain carbon quantum dots CD3 with aggregation-induced emission properties. Preparation of CD4: 2,2'-dithiodibenzoic acid and urea were dissolved in glacial acetic acid and subjected to a solvothermal reaction. The mass ratio of 2,2'-dithiodibenzoic acid to urea was 1:0.25-1.

00. The products were then separated to obtain carbon quantum dots CD4 with aggregation-induced emission properties. Preparation of CD5: 2,2'-dithiodibenzoic acid and urea were dissolved in glacial acetic acid and subjected to a solvothermal reaction. The mass ratio of 2,2'-dithiodibenzoic acid to urea was 1:0.80-1.

35. The products were then separated to obtain carbon quantum dots CD5 with aggregation-induced emission properties. Preparation of CD6: 2,2'-dithiodibenzoic acid, ammonium sulfate and copper salt were dissolved in glacial acetic acid and subjected to a solvothermal reaction. The mass ratio of 2,2'-dithiodibenzoic acid, ammonium sulfate and copper salt was 1:0.12-1.50:0.003-0.

125. The products were then separated to obtain carbon quantum dots CD6 with aggregation-induced emission properties. Preparation of CD7: 2,2'-dithiodibenzoic acid and melamine were dissolved in glacial acetic acid and subjected to a solvothermal reaction. The mass ratio of 2,2'-dithiodibenzoic acid to melamine was 1:0.20-0.

60. The products were then separated to obtain carbon quantum dots CD7 with aggregation-induced emission properties. Preparation of CD8: Pyromellitic acid, o-phenylenediamine and copper salt were dissolved in an ethanol aqueous solution and subjected to microwave reaction. The mass ratio of pyromellitic acid, o-phenylenediamine and copper salt was 1:0.30-0.75:0.0003-0.

025. The products were then separated to obtain carbon quantum dots CD8 with aggregation-induced emission properties. Preparation of CD9: 2,2'-dithiodibenzoic acid and urea were dissolved in glacial acetic acid and subjected to a solvothermal reaction. The mass ratio of 2,2'-dithiodibenzoic acid to urea was 1:0.12-1.

25. The products were then separated to obtain carbon quantum dots CD9 with aggregation-induced emission properties.

2. The method for detecting the freshness of aquatic products according to claim 1, characterized in that: The microwave reaction in the preparation process of CD1 is carried out at a microwave power of 500W to 900W for a reaction time of 8 to 12 minutes; the solvothermal reaction in the preparation process of CD2 is carried out at a temperature of 180℃ to 210℃ for a reaction time of 9 to 12 hours; the solvothermal reaction in the preparation process of CD3 is carried out at a temperature of 160℃ to 180℃ for a reaction time of 8 to 12 hours; the solvothermal reaction in the preparation process of CD4 is carried out at a temperature of 170℃ to 200℃ for a reaction time of 8 to 12 hours; and the solvothermal reaction in the preparation process of CD5 is carried out at a temperature of 170℃ to 200℃ for a reaction time of 170℃ to 200℃ for a reaction time of 8 to 12 hours. The preparation of CD6 is carried out at a temperature of 180℃ to 210℃ for a reaction time of 8h to 12h; the preparation of CD7 is carried out at a temperature of 160℃ to 180℃ for a reaction time of 8h to 10h; the preparation of CD8 is carried out at a microwave power of 500W to 900W for a reaction time of 8min to 12min; the preparation of CD9 is carried out at a temperature of 160℃ to 180℃ for a reaction time of 8h to 12h.

3. The method for detecting the freshness of aquatic products according to claim 1, characterized in that: The copper salt used in the preparation of CD3 is at least one of copper nitrate, copper sulfate, copper chloride, and copper acetate; the copper salt used in the preparation of CD6 is at least one of copper nitrate, copper sulfate, copper chloride, and copper acetate; and the copper salt used in the preparation of CD8 is at least one of copper nitrate, copper sulfate, copper chloride, and copper acetate.

4. The method for detecting the freshness of aquatic products according to any one of claims 1 to 3, characterized in that: Step 1) The fluorescent sensor array tag is prepared by a method including the following steps: carbon quantum dots with aggregation-induced emission properties are dispersed in an organic solvent to form a carbon quantum dot dispersion, and multiple sensor response areas are cut out of a hydrophobic film. Then, the uncut hydrophobic film, cotton fiber paper, and the hydrophobic film with the cut sensor response areas are stacked and bonded in sequence. Then, the carbon quantum dot dispersion is dropped onto the surface of the cotton fiber paper exposed in the sensor response area, and then dried to obtain the fluorescent sensor array tag.

5. The method for detecting the freshness of aquatic products according to claim 4, characterized in that: The organic solvent is at least one of ethanol, methanol, dimethyl sulfoxide, dimethylformamide, and tetrahydrofuran.

6. The method for detecting the freshness of aquatic products according to claim 4, characterized in that: The hydrophobic film is one of hydrophobic organic film and carbon fiber film.

7. The method for detecting the freshness of aquatic products according to any one of claims 1 to 3, characterized in that: Step 3) The deep convolutional neural network model is MobileNet v3 large.

8. A system for detecting the freshness of aquatic products for implementing the method for detecting the freshness of aquatic products according to any one of claims 1 to 7, characterized in that, The components include: The image acquisition and processing module is used to acquire images of the fluorescent sensor array tag at regular intervals and simultaneously acquire the TVB-N value of the aquatic product sample, and determine the freshness of the aquatic product sample based on the TVB-N value. The model building and training module is used to build a deep convolutional neural network model for freshness detection, and to train the deep convolutional neural network model based on the collected fluorescent sensor array label images and the corresponding freshness, so as to obtain the trained deep convolutional neural network model. The results output module is used to input real-time acquired fluorescence sensor array image data into the trained deep convolutional neural network model and output the freshness classification results of aquatic products.

Images