Nondestructive testing system and method for quality of fresh mutton

Abstract

The invention provides a nondestructive testing system and method for the quality of fresh mutton. The system comprises a flexible spectral impedance bimodal sensor and an operation upper computer, the method comprises the steps of impedance signal and spectral signal acquisition, working mode selection and mutton quality grading model identification. According to the invention, by adopting the flexible sensor, nondestructive and accurate detection of mutton is realized; data sampling is carried out through the flexible spectral impedance bimodal sensor, and enrichment and accuracy of classified data are achieved; through the mutton quality grading model, comprehensive and accurate perception of the mutton quality under different working environments and scenes is realized; and the Bluetooth wireless transmission module is connected with an operation upper computer, so that real-time monitoring and visualization of the mutton quality are realized.

Description

Technical Field

[0001] The present invention relates to the field of food quality detection, and in particular to a system and method for nondestructive detection of fresh mutton quality. Background Art

[0002] Meat has gained an important position in the human diet due to its important nutritional value, such as protein, fat and trace elements. With the improvement of people's quality of life, fresh meat is very popular among consumers. However, fresh meat is not easy to preserve for a long time. During transportation and storage, it is easily spoiled by the effects of temperature, humidity, oxygen, microorganisms and enzymes, which is a key factor threatening the safety and quality of fresh meat.

[0003] According to investigations and literature reading, the current perception of the quality grade of fresh meat mostly relies on manual sensory evaluation, physical and chemical index testing, and rigid sensor non-destructive testing. Manual sensory evaluation requires a lot of manpower and time, and cannot be comprehensive, with large errors; physical and chemical index testing will damage fresh mutton; rigid sensor non-destructive testing needs to be carried out on a specific production line, and cannot achieve convenient and rapid quality grade testing. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a system and method for nondestructive detection of fresh mutton quality, which adopts flexible sensors to achieve nondestructive and accurate detection of mutton; uses flexible spectral impedance dual-modal sensors to perform data sampling to achieve rich and accurate classification data; uses a mutton quality grading model to achieve comprehensive and accurate perception of mutton quality in different working environments and scenarios; and uses a Bluetooth wireless transmission module to connect to an operating host computer to achieve real-time monitoring and visualization of mutton quality.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A nondestructive detection system for the quality of fresh mutton, comprising: a flexible spectral impedance dual-mode sensor and an operating host computer; the flexible spectral impedance dual-mode sensor comprises: a sensor front-end unit and a sensor back-end unit; the sensor front-end unit comprises: a PMDS flexible substrate and a LIG serpentine electrode; the sensor back-end unit comprises: a PI flexible substrate, a flexible Cu circuit, a PDMS protective layer, a main control module, a voltage conversion module, an impedance conversion module, an alligator clip, a spectrum acquisition module, an I2C multi-channel expansion module and a Bluetooth wireless transmission module; the sensor front-end unit and the sensor back-end unit are connected through the alligator clip;

[0007] The flexible spectral impedance dual-modal sensor is connected to the operating host computer through the Bluetooth wireless transmission module; the LIG serpentine electrode is attached to the surface of the PMDS flexible substrate; the flexible Cu circuit is attached to the surface of the PI flexible substrate; the main control module, the voltage conversion module, the impedance conversion module, the spectrum acquisition module, the I2C multi-channel expansion module, the Bluetooth wireless transmission module and the alligator clip are fixed on the flexible Cu circuit by soldering; the PDMS protective layer is spin-coated on the main control module, the voltage conversion module, the impedance conversion module, the spectrum acquisition module, the I2C multi-channel expansion module, the Bluetooth wireless transmission module and the alligator clip; the alligator clip is connected to the LIG serpentine electrode;

[0008] The LIG serpentine electrode is used to receive the excitation signal generated by the impedance conversion module, and send a transmission pulse to the target fresh mutton surface according to the excitation signal, and receive the reflected pulse from the target fresh mutton surface and send the reflected pulse to the impedance conversion module; the impedance conversion module is used to generate the excitation signal, and perform DFT transformation on the reflected pulse to obtain an impedance signal, and send the impedance electrical signal to the main control module; the spectrum acquisition module is used to irradiate the target fresh mutton surface, and convert the reflected light from the target fresh mutton surface into an electrical signal to obtain a spectrum signal, and send the spectrum signal to the main control module through the I2C multi-channel expansion module; the main control module is used to control the impedance conversion module to generate the excitation signal and use the Bluetooth wireless transmission module to send the impedance signal and the spectrum signal to the operating host computer; the voltage conversion module is used to convert the 5V voltage provided by the external lithium battery into 3.3V.

[0009] Preferably, the preparation process of the flexible impedance sensor includes:

[0010] Importing the pre-designed electrode pattern into the laser direct writing device;

[0011] Using the laser direct writing device to perform laser-induced graphene on the PI film to obtain the LIG serpentine electrode;

[0012] Spin-coat a mixed solution of PDMS solution and curing agent on the PI film, and place the PI film on a heating table at 60° C. for curing for 2 h;

[0013] Transferring the LIG serpentine electrode on the cured PI film to the PMDS flexible substrate to obtain the sensor front-end unit;

[0014] Importing a pre-designed circuit pattern into the laser direct writing device;

[0015] Using the laser direct writing device to perform laser etching on the PI / Cu film to obtain the flexible Cu circuit;

[0016] Fix the main control module, the voltage conversion module, the impedance conversion module, the spectrum acquisition module, the I2C multi-channel expansion module, the Bluetooth wireless transmission module and the alligator clip on the surface of the flexible Cu circuit by soldering;

[0017] Uniformly apply a mixed solution of PDMS solution and curing agent on the surface of the flexible Cu circuit, and place the flexible Cu circuit on a heating table at 60° C. for curing for 2 hours to obtain the sensor back-end unit;

[0018] The sensor back-end unit is connected to the sensor front-end unit by using the alligator clip to obtain the flexible spectral impedance dual-mode sensor.

[0019] Preferably, a method for nondestructive testing of fresh mutton quality comprises:

[0020] Fixing the flexible spectral impedance dual-mode sensor on the surface of the target fresh mutton;

[0021] The impedance signal and the spectrum signal are collected based on fixed time intervals by operating the host computer and the Bluetooth wireless transmission module;

[0022] Selecting a working mode through an adaptive mode mode; the working modes include: impedance mode, spectral mode and fusion mode;

[0023] The impedance signal and the spectral signal are input into a mutton quality grading model for quality perception to obtain a target quality detection result.

[0024] Preferably, a method for constructing a mutton quality grading model comprises:

[0025] The mutton purchased at the same time was cut into the same shape, randomly grouped, and stored separately under the same environmental conditions to obtain several groups of test samples;

[0026] Using the flexible spectral impedance dual-mode sensor to collect data from each group of test samples at fixed time intervals to obtain a test spectral impedance original data set, and saving the test spectral impedance original data set to a preset knowledge base module;

[0027] The TVB-N value, pH value, hardness, water loss rate, color difference and protein content of the collected data of each group in the experimental spectral impedance original data set are measured to obtain a mutton quality characterization data set and a mutton quality classification data set;

[0028] A mutton spectral impedance grading network is constructed according to the mutton quality characterization data set and the mutton quality classification data set; the mutton spectral impedance grading network includes: an impedance information grading network layer, a spectral information grading network layer and a spectral impedance dual-modal information fusion grading network layer; the spectral impedance dual-modal information fusion grading network layer includes: a data preprocessing layer, a spatial feature extraction layer, a temporal feature extraction layer, a feature fusion layer, a quality prediction layer, a regulation mechanism layer and an error fusion regulation layer; the impedance information grading network layer is used to monitor the mutton quality of the impedance signal collected under the impedance mode; the spectral information grading network layer is used to monitor the mutton quality of the spectral signal collected under the impedance mode; the spectral impedance dual-modal information fusion grading network layer is used to collect the mutton quality under the fusion mode. The impedance signal and the spectral signal are used to monitor the quality of mutton; the data preprocessing layer is used to preprocess the impedance signal and the spectral signal; the spatial feature extraction layer is used to extract spatial features from the impedance signal and the spectral signal after preprocessing; the temporal feature extraction layer is used to extract temporal features from the impedance signal and the spectral signal after spatial feature extraction; the feature fusion layer is used to weightedly fuse the spatial feature extraction data and the temporal feature extraction data of the impedance signal and the spectral signal; the quality prediction layer is used to predict the quality of mutton according to the fusion result characteristics of the feature fusion layer; the control mechanism layer is used to feedback and adjust the abnormal results predicted by the quality prediction layer; the error fusion control layer is used to feedback and adjust the fusion errors of the feature fusion layer;

[0029] The mutton spectral impedance grading network is optimized to obtain the mutton quality grading model.

[0030] Preferably, the calculation formulas of the impedance information hierarchical network layer and the spectrum information hierarchical network layer are respectively:

[0031] and

[0032] in, 0≤k λ* ≤1; 0≤g λ* ≤1;L final The first final level of the impedance information grading network layer; G final A second final lamb level for the spectral information classification network layer; L n is the nth item of data in the first final mutton grade; G n is the nth item of data in the second final mutton grade; L kis a set of lamb impedance quality level thresholds; as well as are the impedance thresholds of mutton at the fresh level, mutton at the sub-fresh level, and mutton at the spoiled level; G g is a set of threshold values ​​for mutton spectral quality levels; as well as are the spectral thresholds of the mutton at the fresh level, the mutton at the sub-fresh level, and the mutton at the spoilage level respectively; → represents mapping; k λ* represent as well as Any of the following; g λ* represent as well as Any one of; Z 1 and Z 2 are respectively the upper limit and lower limit of the impedance threshold under the mutton freshness level; Z 3 and Z 4 are respectively the upper limit and the lower limit of the impedance threshold under the sub-fresh grade of the mutton; Z 5 and Z 6 are respectively the upper limit and the lower limit of the impedance threshold under the mutton spoilage level; S 1 and S 2 are the upper and lower limits of the spectral threshold under the mutton freshness level respectively; S 3 and S 4 are the upper and lower limits of the spectrum threshold under the sub-fresh grade of the mutton respectively; S 5 and S 6 are respectively the upper limit and lower limit of the spectral threshold under the mutton spoilage level.

[0033] Preferably, the calculation formula of the spectral impedance bimodal information fusion hierarchical network layer includes:

[0034] F Z =f space (X Z ),

[0035] F S =f spacce (X S ),

[0036] H Z =f time (F Z ),

[0037] H S =f time (F S ),

[0038]

[0039] scorc(H S )=W a H S ,

[0040] F fusion =α Z ·H Z +α S ·H S ,

[0041] Y pred =f pred (F fusion ,W o )·+b o ;

[0042] Among them, F Z and F S are the impedance data features extracted through spatial features and the spectral data features extracted through spatial features; H Z and H S are the impedance data features extracted through time series features and the spectral data features extracted through time series features; α Z and α S are impedance fusion weight and spectrum fusion weight respectively; F fusion is the fusion feature; pred is the prediction result; f space ( ) is the spatial feature extraction function; f time ( ) is the time series feature extraction function; X Z is the preprocessed impedance data; X S To preprocess the spectral data; F Z 、F S are impedance space characteristics and spectral space characteristics respectively; score( ) is the score calculation function; H Z[j] , H S[j] are impedance time series characteristics and spectral time series characteristics respectively; n is the window size; W a is the score weight; W o is the prediction weight; b o is the prediction bias; f pred ( ) is the prediction function.

[0043] Preferably, it also includes: a control mechanism layer; the calculation formula of the control mechanism layer includes:

[0044] F feedback =g(Y true ,Y pred ),

[0045]

[0046] F′ fusion =α′ Z ·H Z +α′ S ·H′ S ;

[0047] Among them, F feedback is the feedback signal; g( ) is the feedback function; α Z , α S are impedance first optimization fusion weight and spectrum first optimization fusion weight respectively; F' fusion Y is the fusion feature after regulation; true , Y pred are the true quality and the predicted quality respectively; Y true,i , Y pred,i are the i-th values ​​in the true quality and the predicted quality respectively; η is the learning rate.

[0048] Preferably, it further comprises: an error fusion control layer; the calculation formula of the error fusion control layer comprises:

[0049] E fusion =||Y true -Y pred || 2 ,

[0050]

[0051] α″ Z =α′ Z ·R.

[0052] α S =α′ S ·R.

[0053] F″ fusion =α″ Z ·H Z +α″ S ·H S ;

[0054] Among them, E fusion is the fusion error value; R is the regulatory factor; α″ Z , α″ S are the second optimized fusion weight of impedance and the second optimized fusion weight of spectrum respectively; F″ fusion is the fusion feature after the error fusion control layer; β is the control sensitivity coefficient.

[0055] Preferably, the calculation formula of the mutton quality grading model is:

[0056]

[0057] in, is the final loss rate; is the loss function; μ is the fusion error weight coefficient; τ is the feedback weight coefficient.

[0058] The present invention discloses the following technical effects:

[0059] The present invention provides a system and method for nondestructive detection of fresh mutton quality. By adopting a flexible sensor, the defect of damage caused by conventional detection technology is solved, and nondestructive and accurate detection of mutton is achieved; data sampling is performed by a flexible spectral impedance dual-modal sensor, which solves the problem of obtaining data sources and the problem of accuracy in collecting different signals under different environments, and achieves rich and accurate classification data; through a mutton quality grading model algorithm, the problem of insufficient expression of mutton quality by a single mode is solved, and comprehensive and accurate perception of mutton quality is achieved; by using a Bluetooth wireless transmission module to connect to an operating host computer, the problem of real-time monitoring of mutton quality is solved, and real-time monitoring and visualization of mutton quality are achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] 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 will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0061] Figure 1 A schematic diagram of the structure of a nondestructive detection system for fresh mutton quality provided by an embodiment of the present invention;

[0062] Figure 2 A schematic diagram of a LIG serpentine electrode provided in an embodiment of the present invention;

[0063] Figure 3 A schematic diagram of a flexible Cu circuit provided by an embodiment of the present invention;

[0064] Figure 4 A schematic diagram of the system module structure provided by an embodiment of the present invention;

[0065] Figure 5 A schematic diagram of a non-destructive testing process for fresh mutton quality provided by an embodiment of the present invention;

[0066] Figure 6 A schematic diagram of the construction process of a mutton quality grading model provided in an embodiment of the present invention;

[0067] Figure 7 A flowchart of mutton quality prediction and grading provided by an embodiment of the present invention;

[0068] Description of reference numerals:

[0069] 1-Flexible spectral impedance dual-mode sensor, 2-Operation host computer. DETAILED DESCRIPTION

[0070] 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.

[0071] The purpose of the present invention is to provide a system and method for nondestructive detection of fresh mutton quality. By adopting a flexible sensor, nondestructive and accurate detection of mutton is achieved; data sampling is performed by a flexible spectral impedance dual-modal sensor to achieve rich and accurate classification data; a mutton quality grading model is used to achieve comprehensive and accurate perception of mutton quality in different working environments and scenarios; and real-time monitoring and visualization of mutton quality are achieved by connecting an operating host computer using a Bluetooth wireless transmission module.

[0072] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0073] Figure 1 A schematic diagram of the structure of a nondestructive detection system for fresh mutton quality provided by an embodiment of the present invention is shown in FIG. Figure 1 As shown, the present invention provides a nondestructive detection system for the quality of fresh mutton, comprising: a flexible spectral impedance dual-mode sensor 1 and an operating host computer 2; the flexible spectral impedance dual-mode sensor 1 comprises: a sensor front-end unit and a sensor back-end unit; the sensor front-end unit comprises: a PMDS flexible substrate and a LIG serpentine electrode; the sensor back-end unit comprises: a PI flexible substrate, a flexible Cu circuit, a PDMS protective layer, a main control module, a voltage conversion module, an impedance conversion module, an alligator clip, a spectrum acquisition module, an I2C multi-channel expansion module and a Bluetooth wireless transmission module; the sensor front-end unit and the sensor back-end unit are connected via the alligator clip;

[0074] The flexible spectral impedance dual-mode sensor 1 is connected to the operating host computer 2 via a Bluetooth wireless transmission module; the LIG snake-shaped electrode is attached to the surface of the PMDS flexible substrate (reference Figure 2 ); the flexible Cu circuit is attached to the surface of the PI flexible substrate (reference Figure 3); the main control module, the voltage conversion module, the impedance conversion module, the spectrum acquisition module, the I2C multi-channel expansion module, the Bluetooth wireless transmission module and the alligator clip are fixed on the flexible Cu circuit by soldering; the PDMS protective layer is spin-coated on the main control module, the voltage conversion module, the impedance conversion module, the spectrum acquisition module, the I2C multi-channel expansion module, the Bluetooth wireless transmission module and the alligator clip; the alligator clip is connected to the LIG serpentine electrode;

[0075] refer to Figure 4 The LIG serpentine electrode is used to receive the excitation signal generated by the impedance conversion module, and send a transmission pulse to the target fresh mutton surface according to the excitation signal, and receive the reflected pulse from the target fresh mutton surface, and send the reflected pulse to the impedance conversion module; the impedance conversion module is used to generate an excitation signal, and perform DFT transformation on the reflected pulse to obtain an impedance signal, and send the impedance electrical signal to the main control module; the spectrum acquisition module is used to irradiate the target fresh mutton surface, and convert the reflected light from the target fresh mutton surface into an electrical signal to obtain a spectrum signal, and send the spectrum signal to the main control module through the I2C multi-channel expansion module; the main control module is used to control the impedance conversion module to generate an excitation signal and use the Bluetooth wireless transmission module to send the impedance signal and the spectrum signal to the operating host computer 2; the voltage conversion module is used to convert the 5V voltage provided by the external lithium battery into 3.3V.

[0076] Preferably, the operating host computer 2 includes a display function, a process control function, an alarm function and a knowledge base.

[0077] Furthermore, the preparation process of the flexible impedance sensor includes:

[0078] Importing the pre-designed electrode pattern into the laser direct writing device;

[0079] Laser-induced graphene was produced on PI film using a laser direct writing device to obtain LIG serpentine electrodes;

[0080] Spin-coat the mixed solution of PDMS solution and curing agent on the PI film, and place the PI film on a heating table at 60°C for 2 h;

[0081] The LIG serpentine electrode on the cured PI film was transferred to the PMDS flexible substrate to obtain the sensor front-end unit;

[0082] Import the pre-designed circuit pattern into the laser direct writing device;

[0083] Laser etching was performed on the PI / Cu film using a laser direct writing device to obtain a flexible Cu circuit;

[0084] The main control module, the voltage conversion module, the impedance conversion module, the spectrum acquisition module, the I2C multi-channel expansion module, the Bluetooth wireless transmission module and the alligator clip are fixed on the surface of the flexible Cu circuit by soldering;

[0085] The mixed solution of PDMS solution and curing agent was evenly applied on the surface of the flexible Cu circuit, and the flexible Cu circuit was placed on a heating table at 60°C for curing for 2 hours to obtain a sensor back-end unit;

[0086] The sensor back-end unit is connected to the sensor front-end unit by using an alligator clip to obtain a flexible spectral impedance dual-modal sensor 1.

[0087] refer to Figure 5 , a non-destructive detection method for fresh mutton quality, comprising:

[0088] The flexible spectral impedance dual-mode sensor 1 is fixed on the surface of the target fresh mutton;

[0089] The impedance signal and the spectrum signal are collected based on a fixed time interval by operating the host computer 2 and the Bluetooth wireless transmission module;

[0090] The working mode is selected through the adaptive mode mode; the working modes include: impedance mode, spectrum mode and fusion mode;

[0091] The impedance signal and spectral signal are input into the mutton quality grading model for quality perception to obtain the target quality detection results.

[0092] Specifically, a method for constructing a mutton quality grading model comprises:

[0093] The mutton purchased at the same time was cut into the same shape, randomly grouped, and stored separately under the same environmental conditions to obtain several groups of test samples;

[0094] Using the flexible spectral impedance dual-mode sensor 1 to collect data for each group of test samples at fixed time intervals, obtaining a test spectral impedance original data set, and saving the test spectral impedance original data set to a preset knowledge base module;

[0095] The TVB-N value, pH value, hardness, water loss rate, color difference and protein content of each group of collected data in the experimental spectral impedance original data set were measured to obtain the mutton quality characterization data set and mutton quality classification data set;

[0096] A mutton spectral impedance grading network is constructed based on the mutton quality characterization dataset and the mutton quality classification dataset; the mutton spectral impedance grading network includes: impedance information grading network layer, spectral information grading network layer and spectral impedance dual-modal information fusion grading network layer; the spectral impedance dual-modal information fusion grading network layer includes: data preprocessing layer, spatial feature extraction layer, temporal feature extraction layer, feature fusion layer, quality prediction layer, regulation mechanism layer and error fusion regulation layer; the impedance information grading network layer is used to monitor the mutton quality of the impedance signal collected under the impedance mode; the spectral information grading network layer is used to monitor the mutton quality of the spectral signal collected under the impedance mode; the spectral impedance dual-modal information fusion grading network layer is used to The impedance signal and spectral signal collected under the combined mode are used to monitor the quality of mutton; the data preprocessing layer is used to preprocess the impedance signal and spectral signal; the spatial feature extraction layer is used to extract spatial features from the preprocessed impedance signal and spectral signal; the temporal feature extraction layer is used to extract temporal features from the impedance signal and spectral signal after the spatial feature extraction; the feature fusion layer is used to perform weighted fusion on the spatial feature extraction data and temporal feature extraction data of the impedance signal and spectral signal; the quality prediction layer is used to predict the quality of mutton based on the fusion result features of the feature fusion layer; the control mechanism layer is used to provide feedback and adjustment to the abnormal results predicted by the quality prediction layer; the error fusion control layer is used to provide feedback and adjustment to the fusion errors of the feature fusion layer;

[0097] The mutton spectral impedance grading network was optimized to obtain the mutton quality grading model.

[0098] Preferably, the calculation formulas of the impedance information hierarchical network layer and the spectrum information hierarchical network layer are respectively:

[0099]

[0100] and

[0101] in, 0≤k λ* ≤1; 0≤g λ* ≤1;L final The first final level of the impedance information grading network layer; G final A second final lamb level for the spectral information classification network layer; L n is the nth item of data in the first final mutton grade; G n is the nth item of data in the second final mutton grade; L k is a set of lamb impedance quality level thresholds; as well as are the impedance thresholds of mutton at the fresh level, mutton at the sub-fresh level, and mutton at the spoiled level; G g is a set of threshold values ​​for mutton spectral quality levels; as well as are the spectral thresholds of the mutton at the fresh level, the mutton at the sub-fresh level, and the mutton at the spoilage level respectively; → represents mapping; k λ* represent as well as Any of the following; g λ* represent as well as Any one of; Z 1 and Z 2 are respectively the upper limit and lower limit of the impedance threshold under the mutton freshness level; Z 3 and Z 4 are respectively the upper limit and the lower limit of the impedance threshold under the sub-fresh grade of the mutton; Z 5 and Z 6 are respectively the upper limit and the lower limit of the impedance threshold under the mutton spoilage level; S 1 and S 2 are the upper and lower limits of the spectral threshold under the mutton freshness level respectively; S 3 and S 4 are the upper and lower limits of the spectrum threshold under the sub-fresh grade of the mutton respectively; S 5 and S 6 are respectively the upper limit and lower limit of the spectral threshold under the mutton spoilage level.

[0102] Furthermore, the calculation formula of the spectral impedance bimodal information fusion hierarchical network layer includes:

[0103] F Z =f space (X Z ),

[0104] F S =f spacce (X S ),

[0105] H Z =f time (F Z ),

[0106] H S =f time (F S ),

[0107] scorc(H S )=W a H S ,

[0108] F fusion =α Z ·H Z +α S ·H S ,

[0109] Y pred =f pred (F fusion ,W o )+b o ;

[0110] Among them, F Z and F S are the impedance data features extracted through spatial features and the spectral data features extracted through spatial features; H Z and H S are the impedance data features extracted through time series features and the spectral data features extracted through time series features; α Z and α S are impedance fusion weight and spectrum fusion weight respectively; F fusion is the fusion feature; pred is the prediction result; f space ( ) is the spatial feature extraction function; f time ( ) is the time series feature extraction function; X Z is the preprocessed impedance data; X S To preprocess the spectral data; F Z 、F S are impedance space characteristics and spectral space characteristics respectively; score( ) is the score calculation function; H Z[j] , H S[j] are impedance time series characteristics and spectral time series characteristics respectively; n is the window size; W a is the score weight; W o is the prediction weight; b o is the prediction bias; f pred ( ) is the prediction function.

[0111] Specifically, it also includes: a control mechanism layer; the calculation formula of the control mechanism layer includes:

[0112] F feedback =g(Y true ,Y pred ),

[0113]

[0114] F′ fusion =α′ Z ·H Z +α′ S ·HS ;

[0115] Among them, F feedback is the feedback signal; g( ) is the feedback function; α′ Z , α′ S are impedance first optimization fusion weight and spectrum first optimization fusion weight respectively; F' fusion Y is the fusion feature after regulation; true , Y pred are the true quality and the predicted quality respectively; Y true,i , Y pred,i are the i-th values ​​in the true quality and the predicted quality respectively; η is the learning rate.

[0116] Preferably, it also includes: an error fusion control layer; the calculation formula of the error fusion control layer includes:

[0117] E fusion =||Y true -Y pred || 2 ,

[0118]

[0119] α″ Z =α′ Z ·R.

[0120] α″ S =α′ S ·R.

[0121] F″ fusion =α″ Z ·H Z +α″ S ·H S ;

[0122] Among them, E fusion is the fusion error value; R is the regulatory factor; α″ Z , α″ S are the second optimized fusion weight of impedance and the second optimized fusion weight of spectrum respectively; F″ fusion is the fusion feature after the error fusion control layer; β is the control sensitivity coefficient.

[0123] Furthermore, the calculation formula of the mutton quality grading model is:

[0124]

[0125] in, is the final loss rate; is the loss function; μ is the fusion error weight coefficient; τ is the feedback weight coefficient.

[0126] Preferably, the calculation formula for the adaptive modal mode adjustment includes:

[0127]

[0128] If Score Z / Score S >T, the impedance or spectral signal is disturbed by the environment; Score Z 、Score S is the environmental interference score, T is the environmental interference factor, Z new , S new is the newly collected impedance and spectral data; when both impedance and spectral signals are not disturbed, select the fusion mode; when the spectral signal is disturbed, select the impedance mode; when the impedance signal is disturbed, select the spectral mode; μ Z is the average value of impedance data; μ S is the mean of the spectral data; σ Z is the standard deviation of the impedance data; σ S is the standard deviation of the spectral data.

[0129] refer to Figure 6 ,The construction process of mutton quality grading model includes:

[0130] Fresh mutton purchased at the same time was cut into the same shape, randomly grouped, and stored separately under the same environmental conditions to obtain several groups of test samples;

[0131] Flexible impedance sensors and flexible spectral sensors are used to collect data for each group of test samples at fixed time intervals to obtain the original data set of test spectral impedance, and the original data set of test spectral impedance is saved in the knowledge base module; the specific data set is Among them, Z is impedance data, S is spectral data; Z(t n )、S(t n ) represent the impedance data and spectral data collected at different times respectively;

[0132] For each group of test samples, the TVB-N value, pH value, hardness, water loss rate, color difference, and protein content are measured at fixed intervals to obtain the mutton quality characterization data set and mutton quality classification data set, and save them to the knowledge base module; the specific data sets are:

[0133]

[0134] Among them, TVB-N n , pH n 、Hardness n 、WTRn 、Color n 、Protcin n 、level n They are the specific values ​​of TVB-N value, pH value, hardness, water loss rate, color difference and protein content of mutton.

[0135] Constructing a mutton spectral impedance grading network; the mutton spectral impedance grading network includes: an impedance information grading network layer, a spectral information grading network layer, and a spectral impedance dual-modal information fusion grading network layer;

[0136] The mutton spectral impedance grading network was optimized to obtain the mutton quality grading model.

[0137] Preferably, the data preprocessing layer removes noise from the spectral data by smoothing, and uniformly processes the spectral data and the impedance data to the same dimension. The specific calculation formula of the data preprocessing layer is:

[0138]

[0139] Where S' is the spectral noise removal data, n is the window size, X S , X Z are the pre-processed data of spectrum and impedance, μ Z is the average value of impedance data; μ S' is the average value of the data after removing the noise from the spectrum; σ Z is the standard deviation of the impedance data; σ S' The standard deviation of the data after removing noise from the spectrum.

[0140] Furthermore, for the impedance mode, the impedance information hierarchical network layer specifically includes:

[0141] According to the formula Z fresh = {Z 1 ,Z 2}, Z sub-fresh = {Z 3 ,Z 4}, Z rot = {Z 5 ,Z 6}The impedance data depends on Y level Divided into three levels, Z fresh , Z sub-fresh , Z rot They represent the impedance signals of fresh mutton, sub-fresh mutton and spoiled mutton respectively.

[0142] According to the formula 0≤k λ ≤1, Take the average value of the impedance signal of the same type of mutton and compare it with the average value of the impedance signal of another type of mutton in the same category. Set the threshold coefficient based on the average values ​​of the impedance signals of the two types of mutton. final is the final lamb grade, L k is the threshold parameter for each level, k λ is the impedance threshold coefficient.

[0143] Specifically, for the spectral modality, the spectral information hierarchical network layer specifically includes:

[0144] According to the formula S fresh ={S 1 ,S 2},S sub-fresh ={S 3 ,S 4},S rot ={S 5 ,S 6}The impedance data depends on Y level Divided into three levels, S fresh , S sub-fresh , S rot They represent the spectral signals of fresh mutton, sub-fresh mutton and spoiled mutton respectively.

[0145] According to the formula 0≤k λ ≤1, The impedance signal of the same type of mutton is averaged and compared with the average impedance signal of another type of mutton in the same category. The threshold coefficient is set according to the average impedance signal of the two types of mutton. final The final grade of lamb, G g is the threshold parameter for each level, g λ is the impedance threshold coefficient.

[0146] Preferably, the hierarchical network layer for fusion modality spectral impedance bimodal information fusion specifically includes:

[0147] According to formula F Z =f space (X Z ), F S =f spacce (X S ) extracts the spatial features of the data after impedance and spectral preprocessing to obtain the spatial features; where F Z and F S are the impedance data features extracted through spatial features and the spectral data features extracted through spatial features, respectively. space () is the spatial feature extraction function.

[0148] According to the formula H Z =f time (F Z ), H S =f time (F S ) extracts the time series features of the impedance and spectral preprocessed data to obtain the time series features; where H Z and H S They are the impedance data features extracted through spatial features and the spectral data features extracted through temporal features, respectively. time () is the time series feature extraction function.

[0149] According to the formula scorc(H Z )=W a H Z ,scorc(H S )=W a H S , F fusion =α Z ·H Z +α S ·H S The data extracted from spatial and temporal features are weighted and fused to obtain fusion features; where α Z , α S is the fusion weight, score(H) is the calculated score of the time series feature, and W a To calculate the score weight, F fusion is the fusion feature, and exp is the exponential function.

[0150] According to the formula Y pred =f pred (F fusion ,W o )+b o Quality prediction of fusion features, where Y pred is the prediction result, f pred ) is the prediction function, W o is the prediction weight, b o is the prediction bias. According to formula F feedback =g(Y true ,Y pred ), Introducing a feedback mechanism into the network structure, F feedback is the feedback signal, g) is the feedback function, η is the learning rate, α′ Z and α′ S Adjust the fusion weights for dynamic features, F' fusion is the updated fusion feature.

[0151] According to formula E fusion =||Y true -Y pred || 2 , α″ Z =α′ Z ·R, α″ S =α′ S ·R,F″ fusion =α″ Z ·H Z +α″ S ·H S Fusion error definition and correction of the network,E fusion is the fusion error, R is the control factor, β is the control sensitivity coefficient, α Z and α S is the fusion weight adjusted after error fusion, F″ fusion Fusion features adjusted for fusion errors.

[0152] According to the lamb optimization algorithm The network structure is optimized, is the final loss rate, ) is the loss function, λ is the feedback weight coefficient, and μ is the fusion error weight coefficient.

[0153] Specifically, the thickness of the PDMS flexible substrate of the sensor front end unit is 1 mm to 3 mm, specifically, 2 mm. The thickness of the LIG serpentine electrode 12 is 0.05 mm. The size of the sensor front end unit is 50 mm*50 mm.

[0154] Preferably, the main control module includes: an STM32 microcontroller, a crystal oscillator, a resistor, and a capacitor to control the operation of the entire sensor back-end unit; the voltage conversion module includes: a voltage conversion chip, a capacitor, and a resistor to provide corresponding working voltages for each module of the sensor back-end unit and to stabilize the voltage; the impedance conversion module includes: an impedance conversion chip, a capacitor, and a resistor to stimulate excitation signals of different frequencies to detect the biological dielectric impedance of fresh mutton; the spectrum acquisition module mainly includes: a photoelectric acquisition chip, a flash memory chip, an LED, a capacitor, and a resistor. The flash memory chip stores the underlying program of the photoelectric acquisition chip to ensure its normal operation. The LED is a light source to provide the photoelectric chip with an optical signal that meets the working conditions; an I2C multi-channel expansion module to ensure that the photoelectric acquisition chips arranged in a diagonal array can work simultaneously; the Bluetooth wireless transmission module includes: a Bluetooth transmission chip, a resistor, and a capacitor, which is connected to a Bluetooth wireless host computer to start, receive, and save the collected data; the crocodile clip is used to connect to the sensor front-end unit to constitute the sensor front-end unit. The detection principle is as follows: the front-end unit of the sensor fits tightly on the surface of fresh mutton, the frequency synthesizer inside the impedance conversion chip generates excitation signals of multiple frequencies, and the excitation signals are emitted through the output electrode of the flexible impedance sensing electrode. After the signal passes through the surface of fresh mutton, it is received by the input electrode of the flexible impedance sensing electrode, and the DFT transformation is performed in the DSP module of the impedance conversion chip, and the signal is transmitted to the STM32 microcontroller, and the USART serial port of the STM32 microcontroller transmits the signal to the Bluetooth wireless transmission module, and the Bluetooth wireless transmission module transmits the collected data to the Bluetooth wireless host computer. The flexible spectral impedance dual-mode sensor 1 can detect biological dielectric impedance in the range of 1Hz to 100000Hz. The back-end unit of the sensor fits tightly on the surface of fresh mutton, and the LED irradiation light source is on the surface of fresh mutton, which will be affected by the internal components of mutton. The LED illuminates the surface of the mutton, and the photoelectric collection chip generates corresponding photoelectric signal changes by collecting the light diffusely reflected from the surface of the mutton. The power signal is transmitted to the main control module through the I2C multi-channel expansion module. The main control module transmits the data to the Bluetooth wireless transmission module through its internal serial port, and the Bluetooth wireless transmission module transmits the collected data to the operating host computer 2.

[0155] Optionally, the thickness of the PI flexible substrate of the flexible spectral impedance dual-mode sensor 1 is 0.05mm to 0.35mm, specifically, 0.20mm; the thickness of the flexible Cu circuit is one of 0.035mm, 0.05mm or 0.07mm, for example, 0.05mm; the thickness of the PDMS protective layer can just cover the height of the electronic components. The size of the sensor back-end unit is 120mm*60mm.

[0156] Furthermore, the number of turns, length, width, and overall size of the LIG serpentine electrode were designed in the circuit design software. The electronic components were designed according to the sensor function requirements to meet the circuit system requirements, and the circuit line width was 0.2mm and 0.5mm.

[0157] Preferably, the first laser etching is performed on a flexible Cu film with a thickness of 0.2 mm and a PI substrate using a laser direct writing device at a power of 5 W and a speed of 100 mm / s, and then a second laser etching is performed at the same power and speed to completely etch the Cu film except for the circuit, so that it loses its conductivity.

[0158] Furthermore, the laser-etched flexible Cu circuit was cleaned with deionized water. After cleaning, the flexible Cu circuit was placed on a workbench and soldered using solder paste and a welding gun. After soldering, the circuit was cooled and solidified at room temperature for 3 minutes.

[0159] refer to Figure 7 This embodiment also provides a more specific method for predicting and grading mutton quality, which specifically includes:

[0160] 1): Number the fresh mutton to be monitored, and start the flexible spectral impedance dual-mode sensor 1 corresponding to the number by operating the host computer 2;

[0161] 2): The flexible spectral impedance dual-mode sensor 1 system is initialized to determine whether the sensor is working properly; if it is not working properly, the sensor is restarted for adjustment; if it still does not work properly after adjustment, an alarm is issued to remind manual inspection;

[0162] 3): The flexible spectral impedance dual-mode sensor 1 collects spectral signals and impedance signals of fresh mutton, and collects multiple times in the same period of time to ensure the stability of the data; compares the spectral signal and impedance signal with the knowledge base to determine whether the collected signal is abnormal data; if it is not available, re-collect; if the signal after re-collection is still abnormal data, an alarm is issued and manual inspection is performed;

[0163] Prioritizing, the simultaneous application of the three modalities can more accurately determine the freshness of the current fresh mutton; in the single modal mode, basic quality judgment can be made on the fresh mutton, which is suitable for working environments that cannot meet the requirements of both modalities and require rapid screening of the quality of fresh mutton; in the fusion mode, fresh mutton can be accurately measured, and the quality indicators and quality status of the current mutton can be predicted, which is suitable for high-end restaurants and research-level mutton preparation.

[0164] The database collectively includes: impedance signals and spectral signals of fresh mutton collected in the experiment

[0165] Z={Z(t 1),Z(t 2 ),…,Z(t n )}

[0166] S={S(t 1 ),S(t 2 ),…,S(t n )}

[0167] Quality indicators and quality grades of fresh mutton collected in the experiment

[0168]

[0169] 4): Determine whether the spectral signal and the impedance signal are interfered by the environment. When both are interfered, turn on the fusion mode; when the impedance signal is interfered, turn on the spectral mode; when the spectral signal is interfered, turn on the impedance mode;

[0170] Environmental interference is specifically: when the sensor is in an environment with high external light intensity, the data collected by the spectrum acquisition module will be affected, causing the sensor to work abnormally; when the sensor is subject to electromagnetic interference and is in an environment with poor temperature and humidity, the serpentine sensor electrode will be affected, causing abnormal data;

[0171] The specific formula for signal interference judgment is: If Score Z / Score S >T, the impedance or spectral signal is disturbed by the environment

[0172] 5): After selecting the determined mode, import the data into the mutton quality prediction and grading model for data preprocessing

[0173]

[0174] 6): If it is an impedance mode, it is input into the impedance information hierarchical network layer. The specific formula includes:

[0175] If it is a spectral modality, it is input into the spectral information classification network layer. The specific formula includes:

[0176] If it is a fusion mode, it is input into the spectral impedance dual-modal information fusion hierarchical network layer. The specific formula includes:

[0177] F Z =f space (X Z ), F S =f spacce (X S )

[0178] H Z =f time (F Z ), H S =f time (F S )

[0179]

[0180] scorc(H)=W a H

[0181] F fusion =α Z ·H Z +α S ·H S

[0182] Y pred =f pred (F fusion ,W o )+b o

[0183] Introduce a feedback mechanism into the network. The specific formula includes:

[0184] F feedback =g(Y true ,Y pred )

[0185]

[0186]

[0187] F′ fusion =α′ Z ·H Z +α′ S ·H S

[0188] An error fusion control layer is introduced into the network. The specific formulas include:

[0189] E fusion =||Y true -Y pred || 2

[0190]

[0191] α″ Z =α′ Z ·R, α″ S =α′ S ·R

[0192] F″ fusion =α″Z ·H Z +α″ S ·H S

[0193] The model is optimized according to the adaptive monitoring algorithm. The specific formula includes:

[0194]

[0195] Preferably, since the chemical substances inside fresh mutton can be measured in the visible light / near infrared spectrum and have a fairly stable spectrum, when the LED light illuminates the surface of the mutton, the molecules inside the mutton vibrate and rotate, causing the energy level transition to change its absorption peak. The visible light / near infrared spectrum region expresses the absorption characteristics of key chemical bonds such as CO, C=O, OH and NH, which can reflect the internal quality of the mutton.

[0196] Optionally, since the cell structure inside fresh mutton can be measured and expressed in impedance, and has a fairly stable impedance amplitude and impedance angle, when the sensor front-end unit is close to the surface of mutton, the water loss and structural changes of the cells inside the mutton cause changes in its impedance amplitude and impedance angle, which can reflect the internal quality of the mutton.

[0197] Specifically, the spectral impedance bimodal information network layer uses a spatial feature extraction layer as a feature extractor to capture complex long-term dependencies and local patterns. The spatial feature extraction layer rolls over the input sequence through a sliding window to capture local patterns from different time steps, which helps the model learn the spatial features of the data. The temporal feature extraction layer combines the forward and backward information flows to better capture the long-term dependencies in the sequence data. In addition, the local features in the mutton spectrum and impedance data are crucial for understanding the freshness of mutton, and the spatial feature extraction layer plays a key role in extracting these features. Combined with the temporal feature extraction layer, the model is able to comprehensively consider historical and predictive information and more comprehensively understand the context and dependencies in the time series data. At the same time, the introduction of the attention mechanism enables the model to intelligently select and weight basic features to adaptively adjust the focus on different parts of the data. This approach enables the model to flexibly learn and reconstruct data, which is particularly suitable for dealing with random noise in mutton spectral impedance data, avoiding the need for complex feature engineering. In terms of adaptive anomaly detection, the dynamic adjustment of network structure anomalies is achieved by adopting a feedback mechanism and an error fusion control layer.

[0198] 7): Perform weighted scoring on the output of the spectral impedance bimodal grading network layer and output the quality grade of fresh mutton. The specific formula is:

[0199] S Total =S TVB-N +S pH +SHardness +S WTR +S Color +S Protein

[0200]

[0201] Among them, S Total is the total quality score, S TVB-N , S pH , S Hardness , S WTR , S Color , S Protein are the quality scores in 6), S TVB-N , S pH , S Hardness , S WTR , S Color , S Protein The value of is one of (0, 0.5, 0.75, 1); all scoring criteria rely on the knowledge base module for scoring.

[0202] The beneficial effects of the present invention are as follows:

[0203] The present invention realizes non-destructive and accurate detection of mutton by adopting flexible sensors; realizes rich and accurate classification data by data sampling through flexible spectral impedance dual-modal sensors; realizes comprehensive and accurate perception of mutton quality under different working environments and scenarios through mutton quality grading models; realizes real-time monitoring and visualization of mutton quality by connecting to an operating host computer through a Bluetooth wireless transmission module.

[0204] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0205] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A nondestructive testing system for fresh mutton quality, characterized in that: include: Flexible spectral impedance dual-mode sensor and operating host computer; The flexible spectral impedance dual-modal sensor comprises: a sensor front-end unit and a sensor back-end unit; the sensor front-end unit comprises: a PMDS flexible substrate and a LIG serpentine electrode; the sensor back-end unit comprises: a PI flexible substrate, a flexible Cu circuit, a PDMS protective layer, a main control module, a voltage conversion module, an impedance conversion module, an alligator clip, a spectrum acquisition module, an I2C multi-channel expansion module and a Bluetooth wireless transmission module; the sensor front-end unit and the sensor back-end unit are connected via the alligator clip; The flexible spectral impedance dual-modal sensor is connected to the operating host computer through the Bluetooth wireless transmission module; the LIG serpentine electrode is attached to the surface of the PMDS flexible substrate; the flexible Cu circuit is attached to the surface of the PI flexible substrate; the main control module, the voltage conversion module, the impedance conversion module, the spectrum acquisition module, the I2C multi-channel expansion module, the Bluetooth wireless transmission module and the alligator clip are fixed on the flexible Cu circuit by soldering; the PDMS protective layer is spin-coated on the main control module, the voltage conversion module, the impedance conversion module, the spectrum acquisition module, the I2C multi-channel expansion module, the Bluetooth wireless transmission module and the alligator clip; the alligator clip is connected to the LIG serpentine electrode; The LIG serpentine electrode is used to receive the excitation signal generated by the impedance conversion module, and send a transmission pulse to the target fresh mutton surface according to the excitation signal, and receive the reflected pulse from the target fresh mutton surface and send the reflected pulse to the impedance conversion module; the impedance conversion module is used to generate the excitation signal, and perform DFT transformation on the reflected pulse to obtain an impedance signal, and send the impedance electrical signal to the main control module; the spectrum acquisition module is used to irradiate the target fresh mutton surface, and convert the reflected light from the target fresh mutton surface into an electrical signal to obtain a spectrum signal, and send the spectrum signal to the main control module through the I2C multi-channel expansion module; the main control module is used to control the impedance conversion module to generate the excitation signal and use the Bluetooth wireless transmission module to send the impedance signal and the spectrum signal to the operating host computer; the voltage conversion module is used to convert the 5V voltage provided by the external lithium battery into 3.3V.

2. A nondestructive testing system for fresh mutton quality according to claim 1, characterized in that: The preparation process of the flexible impedance sensor includes: Importing the pre-designed electrode pattern into the laser direct writing device; Using the laser direct writing device to perform laser-induced graphene on the PI film to obtain the LIG serpentine electrode; Spin-coat a mixed solution of PDMS solution and curing agent on the PI film, and place the PI film on a heating table at 60° C. for curing for 2 h; Transferring the LIG serpentine electrode on the cured PI film to the PMDS flexible substrate to obtain the sensor front-end unit; Importing a pre-designed circuit pattern into the laser direct writing device; Using the laser direct writing device to perform laser etching on the PI / Cu film to obtain the flexible Cu circuit; Fix the main control module, the voltage conversion module, the impedance conversion module, the spectrum acquisition module, the I2C multi-channel expansion module, the Bluetooth wireless transmission module and the alligator clip on the surface of the flexible Cu circuit by soldering; Uniformly apply a mixed solution of PDMS solution and curing agent on the surface of the flexible Cu circuit, and place the flexible Cu circuit on a heating table at 60° C. for curing for 2 hours to obtain the sensor back-end unit; The sensor back-end unit is connected to the sensor front-end unit by using the alligator clip to obtain the flexible spectral impedance dual-mode sensor.

3. A non-destructive testing method for fresh mutton quality, characterized in that: The nondestructive detection system for fresh mutton quality according to claim 1, wherein the detection method comprises: Fixing the flexible spectral impedance dual-mode sensor on the surface of the target fresh mutton; The impedance signal and the spectrum signal are collected based on fixed time intervals by operating the host computer and the Bluetooth wireless transmission module; Selecting a working mode through an adaptive mode mode; the working modes include: impedance mode, spectral mode and fusion mode; The impedance signal and the spectral signal are input into a mutton quality grading model for quality perception to obtain a target quality detection result.

4. A method for constructing a mutton quality grading model, characterized in that: The method for nondestructive detection of fresh mutton quality as applied to claim 3 comprises: The mutton purchased at the same time was cut into the same shape, randomly grouped, and stored separately under the same environmental conditions to obtain several groups of test samples; Using the flexible spectral impedance dual-mode sensor to collect data from each group of test samples at fixed time intervals to obtain a test spectral impedance original data set, and saving the test spectral impedance original data set to a preset knowledge base module; The TVB-N value, pH value, hardness, water loss rate, color difference and protein content of the collected data of each group in the experimental spectral impedance original data set are measured to obtain a mutton quality characterization data set and a mutton quality classification data set; A mutton spectral impedance grading network is constructed according to the mutton quality characterization data set and the mutton quality classification data set; the mutton spectral impedance grading network includes: an impedance information grading network layer, a spectral information grading network layer and a spectral impedance dual-modal information fusion grading network layer; the spectral impedance dual-modal information fusion grading network layer includes: a data preprocessing layer, a spatial feature extraction layer, a temporal feature extraction layer, a feature fusion layer, a quality prediction layer, a regulation mechanism layer and an error fusion regulation layer; the impedance information grading network layer is used to monitor the mutton quality of the impedance signal collected under the impedance mode; the spectral information grading network layer is used to monitor the mutton quality of the spectral signal collected under the impedance mode; the spectral impedance dual-modal information fusion grading network layer is used to collect the mutton quality under the fusion mode. The impedance signal and the spectral signal are used to monitor the quality of mutton; the data preprocessing layer is used to preprocess the impedance signal and the spectral signal; the spatial feature extraction layer is used to extract spatial features from the impedance signal and the spectral signal after preprocessing; the temporal feature extraction layer is used to extract temporal features from the impedance signal and the spectral signal after spatial feature extraction; the feature fusion layer is used to weightedly fuse the spatial feature extraction data and the temporal feature extraction data of the impedance signal and the spectral signal; the quality prediction layer is used to predict the quality of mutton according to the fusion result characteristics of the feature fusion layer; the control mechanism layer is used to feedback and adjust the abnormal results predicted by the quality prediction layer; the error fusion control layer is used to feedback and adjust the fusion errors of the feature fusion layer; The mutton spectral impedance grading network is optimized to obtain the mutton quality grading model.

5. The method for constructing a mutton quality grading model according to claim 4, characterized in that: The calculation formulas of the impedance information hierarchical network layer and the spectrum information hierarchical network layer are respectively: and in, 0≤g λ* ≤1;L final The first final level of the impedance information grading network layer; G final A second final lamb level for the spectral information classification network layer; L n is the nth item of data in the first final mutton grade; G n is the nth item of data in the second final mutton grade; L k is a set of lamb impedance quality level thresholds; as well as are the impedance thresholds of mutton at the fresh level, mutton at the sub-fresh level, and mutton at the spoiled level; G g is a set of threshold values ​​for mutton spectral quality levels; as well as are the spectral thresholds of the mutton at the fresh level, the mutton at the sub-fresh level, and the mutton at the spoilage level respectively; → represents mapping; k λ* represent as well as Any of the following; g λ* represent as well as Any item in; Z1 and Z2 are respectively the upper and lower limits of the impedance threshold under the mutton freshness level; Z3 and Z4 are respectively the upper and lower limits of the impedance threshold under the mutton sub-freshness level; Z5 and Z6 are respectively the upper and lower limits of the impedance threshold under the mutton spoilage level; S1 and S2 are respectively the upper and lower limits of the spectral threshold under the mutton freshness level; S3 and S4 are respectively the upper and lower limits of the spectral threshold under the mutton sub-freshness level; S5 and S6 are respectively the upper and lower limits of the spectral threshold under the mutton spoilage level.

6. The method for constructing a mutton quality grading model according to claim 4, characterized in that: The calculation formula of the spectral impedance dual-modal information fusion hierarchical network layer includes: F Z =f space (X Z )、 F S =f spacce (X S )、 H Z =f time (F Z )、 H S =f time (F S )、 scorc(H S )=W a H S 、 F fusion =a Z ·H Z +a S ·H S 、 Y pred =f pred (F fusion ,W o )+b o ; Among them, F Z and F S are the impedance data features extracted through spatial features and the spectral data features extracted through spatial features; H Z and H S are the impedance data features extracted through time series features and the spectral data features extracted through time series features; α Z and α S are impedance fusion weight and spectrum fusion weight respectively; F fusion is the fusion feature; pred is the prediction result; f space () is the spatial feature extraction function; f time () is the time series feature extraction function; X Z is the preprocessed impedance data; X S To preprocess the spectral data; F Z 、F S are impedance space characteristics and spectral space characteristics respectively; score() is the score calculation function; H Z[j] , H S[j] are impedance time series characteristics and spectral time series characteristics respectively; n is the window size; W a is the score weight; W o is the prediction weight; b o is the prediction bias; f pred () is the prediction function.

7. The method for constructing a mutton quality grading model according to claim 6, characterized in that: Also includes: Control mechanism layer; calculation formula of the control mechanism layer include: F feedback =g(Y true ,Y pred )、 F' fusion =a' Z ·H Z +a' S ·H S ; Among them, F feedback is the feedback signal; g() is the feedback function; α' Z , α' S are impedance first optimization fusion weight and spectrum first optimization fusion weight respectively; F' fusion Y is the fusion feature after regulation; true , Y pred are the true quality and the predicted quality respectively; Y true,i , Y pred,i are the i-th values ​​in the true quality and the predicted quality respectively; η is the learning rate.

8. The method for constructing a mutton quality grading model according to claim 7, characterized in that: Also includes: Error fusion control layer; The calculation formula of the error fusion control layer include: AND fusion =||And true -AND pred || 2 、 α″ Z ＝α′ Z ·R、 α″ S ＝α′ S ·R、 F″ fusion =a″ Z ·H Z +a″ S ·H S ; Among them, E fusion is the fusion error value; R is the regulatory factor; α″ Z , α″ S are the second optimized fusion weight of impedance and the second optimized fusion weight of spectrum respectively; F″ fusion is the fusion feature after the error fusion control layer; β is the control sensitivity coefficient.

9. The method for constructing a mutton quality grading model according to claim 8, characterized in that: The calculation formula of the mutton quality grading model is: in, is the final loss rate; is the loss function; μ is the fusion error weight coefficient; τ is the feedback weight coefficient.

Images