Prediction of the Freshness of Grass Carp during Storage with Electric Nose Based on Signal Sequence Merging and Wavelet Transform

,


Introduction
A number of degradative reactions begin in the fsh fesh after the death of a fsh.Tese reactions are caused by chemical, biochemical, and microbial metabolic activity.Various components decompose and new compounds form during fsh spoilage, resulting in changes to the perceived quality (odor, favor, and texture) of the fsh meat [1].
Freshness is very important when assessing the quality of fsh and fshery products [2].It can be described by various indicators based on biochemical changes after slaughter, including the K value [3], total volatile basic nitrogen (TVB-N) [4,5], and thiobarbituric acid contents [6].Te K value, which measures the extent of the breakdown of adenosine triphosphate (ATP) in fsh fesh [7], can be obtained by measuring inosine (HxR), hypoxanthine (Hx), ATP, adenosine diphosphate (ADP), adenosine monophosphate (AMP), and inosinic acid (IMP) with high-performance liquid chromatography (HPLC).A range of methods have been used to measure TVB-N since the measurement method was frst described in Conway and Byrne [8].Generally, the meat sample or an extract of meat is alkalized; the volatile bases are collected and titrated with acid, and then the total bases are calculated in accordance with the amount of acid.Moreover, chromatographic techniques such as GC-MS, SPME-GC-MS, and HPLC have produced convincing results and proved to be suitable for the determination of TVB-N [9][10][11].However, these conventional methods are cumbersome, expensive, time-consuming, and destructive.Terefore, developing a convenient, costefective, rapid, sensitive, and reliable method for freshness determination is necessary.
In recent years, many studies have focused on developing rapid and nondestructive methods to evaluate the freshness of food.Among them, the electric nose (E-nose) is a system composed of a series of chemical sensors with specifc response and appropriate pattern recognition algorithms, which can rapidly detect and recognize single or mixed complex gases while also providing comprehensive odor information for the tested sample [12].Li et al. [4] developed simple mathematical models relating E-nose signals to variations in total viable count (TVC) and TVB-N for packaged pork during refrigerated storage.An input-modifed convolution neural network combined with E-nose and hyperspectral imaging was utilized to evaluate TVB-N of mutton [5].Furthermore, the prediction of K value with E-nose has also been reported in recent literature.Li et al. [3] predicted the freshness of horse mackerel by utilizing the E-nose, electronic tongue (E-tongue), and colorimeter combined with a data fusion strategy and different pattern recognition algorithms.Te square correlation coefcient of the test set for the prediction model of K value was 0.936.Previous studies have successfully showcased the efcacy of E-nose in predicting the freshness of food.However, a signifcant portion of these studies focuses on selecting appropriate pattern recognition algorithms, especially deep learning algorithms, in recent years [13], to improve the accuracy of E-nose prediction for food freshness.For the E-nose, in addition to the pattern recognition algorithm, data processing plays an important role in improving the prediction accuracy.
Te mean value or maximum value is typically used as a feature signal in the analysis of E-nose data [14,15].However, one sensor can only provide one kind of data, which may not necessarily correspond to characteristic responses.Consequently, acquiring more data necessitates the use of additional sensors, but increasing the number of sensors leads to more complex and expensive operations.Terefore, improving the data processing method is necessary to obtain majority of the response data from the Enose and obtain higher prediction accuracy with fewer sensors.Li et al. [16] used discrete wavelet transform (WT) to extract important features from dynamic sensor responses and then evaluated egg storage time and yolk index with satisfactory predictions.Tis inspired us to think about a new idea for predicting the freshness of grass carp (Ctenopharyngodon idella) using an E-nose.Te WT method can remove high frequency noise and compress data.Hence, the features can be selected for determining multicomponent samples by WT, and diferent wavelets give rise to diferent results.As a minute tool for time-frequency dynamic responses, WT has been applied in various felds [17][18][19].Here, WT is tried as a new data processing method to extract feature signal from E-nose.
Tis paper presents a feature extraction method for the purpose of evaluating freshness of grass carp during 4 °C storage using an E-nose efectively.Specifcally, the measured signal sequences of the E-nose are merged and the features are extracted to obtain more useful information from the merged signal sequence as a feature signal by using the WT method.Ten, combined with chemometric methods, both the TVB-N content and the K value can be predicted quantitatively.

Raw Material.
Twenty grass carps were obtained from a local supermarket near the laboratory in Nanchang, China, and each fsh weighed about 2-2.5 kg.Tese grass carps were killed by a sharp blow to the head, and then fsh muscles were dissected carefully from the dorsal and ventral regions of the lateral line and weighed between 50 g and 55 g.Four fllets can be obtained from each grass carp.Immediately, these fllets were stored inside a plastic container at 4 °C prior to determination.
Eight fllets were randomly taken out from the container on days 0, 2, 4, 6, 8, 10, 12, 14, and 16, respectively.Each fllet was homogenized for 1 min and then divided into three parts.One part of the samples was distributed for E-nose sample preparation, another part was used for K value measurement, and the remaining part was utilized for TVB-N content measurement.

TVB-N Content Measurements.
Te TVB-N content in the fllet was measured according to GB/T 5009.44 (2003).Each fllet sample was ground individually, and then 10 g of grounded sample was taken into a conical fask and impregnated with 100 mL distilled water for 30 min; the conical fask was often shaken before fltration to ensure uniform dispersion of the sample within the solution.Subsequently, 5 mL of fltrate was taken and blended with 5 mL of 10 g/L magnesia (MgO), distilled by a Kjeldahl distillation unit for 5 min.Te distillate was absorbed with 10 mL of 20 g/L boric acid solution and titrated with 0.01 M of hydrochloric acid solution.Te TVB-N content was calculated as shown in equation ( 1) by GB/T 5009.44 [20]: In equation (1), X is the TVB-N content (mg/100 g), V 1 is the titration volume for the tested sample (mL), V 2 is the titration volume for the blank (mL), C is the concentration of HCl (mol/L), and m is the weight of ground fllet (g).

Determination of K Value with HPLC.
Te method for measuring the ATP-related compounds with HPLC was modifed and validated based on the reference described by Barat et al. [21] and published in a Chinese journal [22].
Two grams of minced muscle were placed into a centrifuge tube and oscillated under vortex movement with 10 mL of 10% perchloric acid for 2 min and then centrifuged at 10,000 g under cold conditions (4 °C) for 10 min.Te supernatant was taken out, and the sediment was treated with 5.0 mL of 10% perchloric acid repeatedly.Repeat the operation once and put all supernatants together.Te pH of the supernatant was adjusted to 6.5 with 1 mol/L sodium hydroxide solution.It was made up to/diluted to 50 mL with deionized water and fltered through a 0.45 mm membrane flter for further analysis.

2
Journal of Food Biochemistry HPLC analysis was performed by the Agilent 1260 highperformance liquid chromatography equipped with an UVvis absorbance detector (Agilent Technologies, USA).ATP and its related compounds were separated in a Waters C18 column (4.6 mm i.d.×150 mm, 5 μm).Te mobile phase was triethylamine phosphate solution (3.5 mL of phosphoric acid solution and 7.2 mL of triethylamine solution was made up to 1,000 mL with deionized water, and pH was adjusted to 6.5 with triethylamine): methanol � 95 : 5.A sample (20 μL) was injected with a fow rate at 1 mL/min, and peaks were detected at 254 nm.
Te amounts of ATP and its related compounds were determined and calculated based on the extra-calibration curves of standard ATP, ADP, AMP, IMP, HxR, and Hx.K value was defned by the following equation [7] 2 g minced fllet was placed in a 20 mL vial (Alltech, USA).Te vials were sealed with a 20 mm silicone/PTFE magnetic crimp-top cap which was obtained from CNW Technologies GmbH (Dusseldorf, Germany).Te vials were equilibrated for 30 min at 30 °C.In the previous experiment, the incubation condition has been optimized.Te response of the E-nose increased with the incubation temperature and time.Te incubation temperature was selected to be 30 °C because it is a normal temperature.Te response increased quickly with time at 30 °C and became slow after 30 min.So, this incubation condition was chosen.Ten 1,000 μL of headspace gas was injected into the sensor chamber, and it was measured for 200 s, with a 0.5 s sampling interval.Dried air (purity 99.999%) was employed as the carrier gas, and the fow rate was 600 mL/min.After sample analysis, the system was purged for 300 s.

E-Nose Signal Pretreatment.
Te 101 data points (0-50 s) extracted from each sensor of E-nose are selected, and they are normalized between 0 and 1, respectively, by the following equation [23]: where R: the normalized response; R i : the response value at a sensor of E-nose; R min : the minimum response value at a sensor of E-nose; and R max : the maximum response value at a sensor of E-nose.Te merging signal sequence is compressed by WT before analyzing data with statistical techniques.Te different Daubechies (dbN) wavelets and decomposition levels are used, and the optimum dbN mother wavelet and decomposition level are determined for the wavelet transform of E-nose signal.
2.6.Data Analysis.Te genetic algorithm (GA) was performed to evaluate the infuence of the diverse sensors on prediction of the K value.GA is a custom program written by Leardi, executed on Matlab [24].Te evaluation per run of the GA was set to 500, with 300 runs, while the remaining parameters were kept at their default values.
Compressed data were analyzed by partial least squares (PLS), artifcial neural network (ANN), and principal component analysis (PCA).PLS was performed by means of the software SIMCA 14.1 (Umetrics, Umea, Switzerland); ANN and PCA were performed on Matlab 7.0 (MathWorks, Natick, MA, USA).PCA was applied to identify the diferent storage times of chilled grass carp fllet; ANN and PLS were employed to build predictive models for K value and TVB-N content with full cross-validation (leave-one-out cross-validation) approach.As for the performance of the established models, the evaluation indicator systems were mainly related to the square correlation coefcient (R 2 ) and root mean square error of cross-validation (RMSECV), respectively.

TVB-N Content and K Value Changes of Grass Carp during
Storage. Figure 1 presents the changes of TVB-N content and K value of grass carp fllet during storage at 4 °C.Te TVB-N content of grass carp fllet increases as storage time increases as shown in Figure 1(a).Te initial TVB-N content is 15.4 mg/100 g, and the fnal TVB-N content is 38.2 mg/ 100 g when the fesh decayed in the end.Te TVB-N content on day 6 is 20.9 mg/100 g, which indicates that the fllet is beginning to decay.In addition, the error bars reveal that the relative standard deviation (RSD) of measured TVB-N content of each storage group is less than 5%.
Te concept of the K value as a freshness indicator was introduced by Saito et al. [7].Te larger the K value is, the lower the freshness of the fsh is.Generally, the K value of live fsh is between 0 and 10, while K value of fresh fsh is between 15 and 35.Te fsh fesh is not regarded as fresh when the K value is greater than 50 [6].Figure 1 shows an almost linear increase in the K value of grass carp fesh during storage at 4 °C.Te initial K value is 7.72, and the K value is 91.81 when the fesh completely decayed in the end.However, the increase of the K value is signifcant from the sixth day to the eighth day of storage, and the K value exceeds the fresh threshold (50) after the sixth storage day.Te RSD of measured K value of each storage group is less than 10%.

E-Nose Signal.
Te signals extracted from E-nose are shown in Figure 2(a).Each curve represents a diferent sensor response.Tey are denoted as T70/2, PA/2, P30/1, P40/2, LY2/Gh, and LY2/gCTL, respectively, according to their sensors.Diferent sensors have varying sensitivities to gas, resulting in difering responses from each electronic nose sensor, with some exhibiting higher responses than others.Figure 2 Figure 2 shows that the responses of the six sensors vary a lot with time in the beginning of curve and stabilize gradually until the end of the measurement.Each curve contains much information.In most studies, each E-nose sensor only adopts one data point (mean value or max value of signal), and much useful information is wasted.In this paper, full utilization of information included in the signals of each E-nose sensor is attempted.A larger signal variability is observed at the initial period of the curve, indicating that this range may account for the majority of the original information.Terefore, a selection of 101 data points from the frst 50 seconds signal of each sensor was selected as tentative character responses.However, it is crucial to extract feature signal sequences from these selected data points.

Evaluation of Diverse Sensors with GA.
Te GA was used to evaluate the contribution of diverse sensors to prediction.GA can simulate natural selection and evolution and fnd the optimal solution globally by searching through the entire solution space.
In dataset of GA, there is a one-to-one correlation between one direct signal sequence merging and one measured K value.After 300 runs, GA provided a total of 88 feature data points.Te frequency of selection is shown in Figure 3.It can be observed that data points 304-308 and 345 have a high frequency of selection, indicating that data points acquired by P40/2 are more relevant to predicting the K value.Te remaining feature data points, whose selected frequency is nearly equal, are distributed on diverse sensors.So, the data points from all six sensors must be used.

Merging E-Nose Signal
Sequence.In Li et al. [16], WT was used to extract characteristic information from the sensor response of E-nose for evaluating egg storage time and yolk index.Tey decompose the original signal from each sensor, respectively, and obtain a series of coefcients sets.Ten, new corresponding signal was reconstructed from the coefcient sets for qualitative and quantitative analysis.
Similarly, WT was adopted in this paper.However, Figure 2 shows that signal from diferent sensors difers signifcantly.Tus, each selected data sequence from different sensor has to be normalized according to equation (3) unlike the WT.Tese selected data sequences are denoted as T70/2, PA/2, P30/1, P40/2, LY2/Gh, and LY2/gCTL, respectively, according to their sensors.Two signal sequence merging modes are compared: the frst mode is that normalized data sequences are merged from end to head as T70/ 2-PA/2-P30/1-P40/2-LY2/Gh-LY2/gCTL, shown in the top subgraph of Figure 4. Tis mode is referred to as direct signal sequence merging (DSSM).Te beginning of each data sequence has a larger value compared to the end, so a sharp increase appears in the joining of two data sequences.Te second mode is to reverse the normalized data sequences from the 3 sensors (PA/2, P40/2, LY2/gCTL), followed by combining them with the other three data arrays, with the linking sequence same as the DSSM mode.Tis mode is called the reversed signal sequence merging (RSSM) mode.Te top subgraph of Figure 5 shows that the junctions of two 4 Journal of Food Biochemistry data sequences of RSSM mode are smoother compared with the DSSM mode.Tese two signal sequence merging modes are adopted in the following data analysis in order to evaluate the efect of two merging modes on the prediction of K value and TVB-N content.
3.5.Extracting Feature Signal with WT.WT was developed for the analysis of a merged signal sequence with Matlab 7.0.Te data compression process is used to reduce data size.Apart from data compression, WT is also expected to minimize noise and other unwanted contents present in the signal [25].
Te merged signal sequence could be decomposed into two sets of coefcients with the best mother wavelet and scale.Te coefcients are obtained by convolving the merged signal sequence with the low-pass flter for approximation and with the high-pass flter for detail.Te coefcients contain a series of A j set and a series of D j set where A represents the approximation coefcients, D represents the detail coefcient, and j represents the level of decomposition.Te A j set and D j set retain the low-frequency and highfrequency content of the signal, respectively.Te diferent scales of the approximation coefcients and detail coefcient are shown in Figure 4 (DSSM mode) and Figure 5 (RSSM mode).Tere are 606 data points in A 0 , and the number of data points reduces by half when the decomposition level increases by 1. Finally, the data size is reduced by WT, with only 19 data points in the A 5 and D 5 sequence.
In this study, the optimization of WT parameters is performed using diferent signal sequence merging modes, approximation coefcients, and detail coefcients of different scales.Meanwhile, several mother wavelets are tried and evaluated.Te frst-order wavelet transform of the Daubechies' family (db1) is selected as mother wavelet by comparing and computing diferent wavelet bases.
3.6.PCA.PCA has been widely applied in the felds of pattern recognition and multivariate calibration.Te newly generative variables are utilized to represent the original ones after processing by PCA, which efectively simplifes  However, the D-D 2 and R-D 2 models showed no distinct grouping.Te predictive efect of the K value and TVB-N content of grass carp fesh is compared by using the A j sets and D j sets based on WT with the ANN and PLS.Journal of Food Biochemistry

Prediction of K Value and TVB-N Content.
Te original data and all feature data from WT using approximation coefcients and detail coefcients of diferent scales are compared while doing analysis using PLS and ANN.PLS is positioned as a statistical technique for prediction renowned for their efectiveness in estimating structural equation models with latent variables [26].ANN is a powerful, efcient, and nonlinear method with pattern recognition abilities, which makes it perfectly suited for the extraction of feature information from large data, especially due to complex biological, environmental, and instrumental variations [27].Compared with other techniques, such as covariance and regression, ANN and PLS have better predictive ability and are frequently used in the prediction of texture, freshness, and safety of food [27][28][29].For example, Basile et al. [27] employed nondestructive NIR spectroscopy and machine learning techniques to predict the texture parameters and total soluble solids content of grape.Te results indicated that the multivariate models, created by constructing ANN and applying PLS regressions, displayed improved predictive capabilities following the removal of uninformative spectral ranges.Several architectures of the network were investigated to predict the K value and TVB-N content.ANNs used in this paper are backpropagation (BP) and radial basis (RB) neural networks.Te RB training algorithm is faster than BP for practical problems.Te only optimized parameter is SPREAD.Despite various modifcations to SPREDA, the predictive performance for both K value and TVB-N content remains subpar.It is proven that the RB model is not suitable in this study.
In this paper, the BP-ANN adopts the steepest gradient descent backpropagation training algorithm which updates the network weights and biases along the direction of the negative of the gradient."Squashing" functions such as sigmoid transfer functions, that compress an infnite input range into a fnite output range, are usually used in the hidden layers.If the input values are large, the gradient can have a very small magnitude and therefore cause small changes in the weights and biases, even though the weights and biases are far from their optimal values.
Te PLS uses X (measured K value or TVB-N content) to construct a model of Y (predicted K value or TVB-N content), where the objective is to predict the latter from the former for new samples in the prediction set.
Te two models are further evaluated with full crossvalidation using a "leave-one-out" technique.RMSECV and R 2 were adopted to evaluate the prediction ability of K value and TVB-N content.Te small RMSECV means the prediction model is better.As for R 2 , the closer it is to 1, the better the model is.A model performs well when the value of R 2 is in the range of 0.82-0.90; the model performs inaccurately while the value of R 2 is lower than 0.82; and the value of R 2 higher than 0.90 shows excellent performance of the model.

TVB-N Content.
Te prediction results of PLS and ANN models based on signal sequence merging (DSSM and RSSM modes) and WT for TVB-N content are shown in Table 1.It is interesting that the prediction results with A j sets are better in ANN models, in which the best model is the DSSM-A 3 model (R 2 is 0.9683, RMSECV is 1.6582 mg/ 100 g); on the contrary, the prediction results with D j sets are better in PLS models, in which the best model is the DSSM-D 2 model (R 2 is 0.9685, RMSECV is 1.6521 mg/ 100 g).Te best ANN and PLS models are based on the DSSM mode.Most values of R 2 are higher than 0.90, except for D 2 and D 3 models built with ANN.PLS, in addition to the RSSM-D 5 model, demonstrated strong performance in predicting TVB-N content, exceeding an R 2 of 0.9.Furthermore, no signifcant diferences were observed between DSSM and RSSM models in predicting TVB-N content, except for a few models.
According to Table 1, the ANN model was built using the A j sets, while the PLS model was built using the D j sets.Figures 7(a)-7(d) visualize linear relationships between the predicted and measured TVB-N content with DSSM-ANN-A 3 , DSSM-ANN-A 5 , DSSM-PLS-D 2 , and DSSM-PLS-D 4 models, respectively.Te four models have a good coefcient of determination.Although R2 of the DSSM-ANN-A3 model is the highest in the ANN model, the R2 of DSSM-ANN-A5 model is still higher than 0.90 (R2 is 0.9474, RMSECV is 2.1423 mg/100 g).Tus, it can be inferred that the DSSM-ANN-A5 model retains the original signal features, with a dataset size of 19 data points signifcantly less than the 76 data points in the DSSM-ANN-A3 model.Te gradient descent training algorithm in the ANN model was too slow for practical problems.It cost 8 h to complete a performance when using merged signal sequence with 606 data points, and the run time decreased with fewer data numbers.It only cost 31 and 5 min to complete a performance when using DSSM-ANN-A 3 and DSSM-ANN-A 5 models, respectively.Although the running time of PLS model based on SIMCA is almost not afected by the number of data points, compared with the DSSM-D 2 model, the DSSM-D 4 model can better distinguish the storage time of chilled fsh fllets in PCA plots, and the R 2 of DSSM-PLS-D 4 model still exceeds 0.9 (R 2 is 0.9606, RMSECV is 1.8437).Finally, DSSM-ANN-A 5 and DSSM-PLS-D 4 models are selected as prediction models for measuring TVB-N content.

K Value.
Te prediction results of PLS and ANN models based on signal sequence merging (DSSM and RSSM modes) and WT for K value are shown in Table 2. PLS models consistently demonstrate highly benefcial prediction results, with all R 2 exceeding 0.9 and no infuence observed on the R 2 due to varying coefcient sets or signal sequence merging modes.Furthermore, the best model is the RSSM-D 2 model (R 2 is 0.9898, RMSECV is 4.7357).Te K value prediction models with A j sets, as established by ANN, perform well except the A5 model, which exhibits a low R2 possibly due to the limited number of data points.On the other hand, excellent performance is exhibited by the K value prediction model with D j sets in ANN models, except for D 1 .In this case, the best model is the DSSM-D 4 model (R 2 is 0.9731, RMSECV is 7.6938).Compared with A j set,  Journal of Food Biochemistry the K value prediction result with D j sets in ANN model has better performance.Diferent from the TVB-N content prediction model, the performance of the K value prediction model is not signifcantly diferent between the DSSM and RSSM modes.
Figures 7(e)-7(h) visualize linear relationships between the predicted and measured K value with DSSM-ANN-D 4 , DSSM-ANN-D 5 , RSSM-PLS-D 2 , and RSSM-PLS-A 0 models, respectively.Figures 7(d) and 7(e) indicate that the predicted K value is much greater than the measured K value on day 6, and greater deviations occur between predicted and measured K values on day 8. Te reason is that the K value increases swiftly from day 6 (mean K value: 21.38) to day 8 (mean K value: 69.49) as shown in Figure 1(b).Parts of the samples may not be analyzed with HPLC and E-nose meantime.Measurement errors caused by analyzing time will be great.Te DSSM-ANN-D 5 model, with fewer data points and R 2 great than 0.9 (R 2 is 0.9503, RMSECV is 10.4359), is selected as the K value prediction model.On the other hand, compared to the RSSM-PLS-D 2 model, the RSSM-PLS-A 0 model more efectively diferentiates the storage time of chilled fsh fllet in PCA plots, while still maintaining an R 2 greater than 0.9 (R 2 � 0.9897, RMSECV � 4.7495).Finally, DSSM-ANN-D 5 and RSSM-PLS-A 0 models are selected as prediction models for measuring K value.
In a similar work conducted by Huang et al. [14], the TVB-N content of pork meat was measured by integrating near infrared spectroscopy (NIRS), computer vision (CV), and E-nose techniques.PCA was employed to achieve data fusion, and the prediction model for TVB-N content was built using BP-ANN.Te results revealed the outstanding performance of the data fusion model of NIRS, CV, and E-nose (RMSECV � 2.73 mg/100 g, R 2 � 0.9527).However, this method of data fusion is complex, expensive, time-consuming, and not suitable for rapid nondestructive testing requirements.Moreover, it is worth noting that the maximum response value response of each sensor was extracted as the characteristic variable, resulting in data waste.Consequently, the prediction performance of the E-nose model based on BP-ANN was found to be poor (RMSECV � 5.97 mg/100 g, R 2 � 0.6495).Conversely, in this study, only an E-nose was used to measure chilled fsh fllet, and an innovative electronic nose signal data preprocessing method was employed, which ultimately established excellent prediction models for the freshness of grass carp during storage.

Conclusions
In order to predict the freshness of chilled grass carp fesh, a signal pretreatment method was developed based on two types of signal sequence merging modes, with wavelet transform applied.Te PCA analytical results reveal that the diferent scales and two kinds of signal sequence merging modes of approximation coefcients and detail coefcients can be used to distinguish the grass carp fllets of diferent storage time.In addition, successful utilization of PLS and ANN was achieved to build prediction model for measuring TVB-N content and K value of grass carp fllet.Te DSSM-ANN-A 5 and DSSM-PLS-D 4 models are fnally chosen as the TVB-N content prediction models, while the DSSM-ANN-D 5 and RSSM-PLS-A 0 models are fnally selected as the K value prediction models.Even though the number of data points is reduced to 19 (A 5 and D 5 ) from 606 (A 0 ), they still show excellent performance of the model.Tere was no signifcant difference between the DSSM and RSSM models in the results of the established prediction model.Tis study proves that WT can condense and extract feature E-nose signal efectively.Not only can this E-nose signal preprocessing method be used to predict TVB-N content and K value, but it also has the potential capacity for other rapid determination with E-nose.

Figure 1 :
Figure 1: Te changes of TVB-N content (a) and K value (b) of grass carp fllet during storage at 4 °C.

Figure 6 Figure 4 :Figure 5 :
Figure 4: Direct merging signal sequence and its wavelet transform at all the fve scales using db1.A j represents approximation coefcients, D j represents detail coefcient, and j represents the decomposition level.

Table 1 :
Prediction results of PLS and ANN models based on signal sequence merging and WT for TVB-N content in grass carp fllet.Te four models demonstrate outstanding performance in TVB-N content predicition and were subsequently discussed collectively.R 2 and RMSECV are highlighted in bold to facilitate quick identifcation of the models by other researchers.

Table 2 :
Prediction results of PLS and ANN models based on signal sequence merging and WT for K value in grass carp fllet.Te four models demonstrate outstanding performance in K value predicition and were subsequently discussed collectively.R 2 and RMSECV are highlighted in bold to facilitate quick identifcation of the models by other researchers.