Accurate prediction of salmon freshness under temperature fluctuations using the convolutional neural network long short-term memory model

Highlights

• The salmon freshness under temperature fluctuations could be accurately predicted.

• The CNN_LSTM model was proposed to improve the prediction performance compared to microbial kinetic equations.

• The CNN_LSTM model could mine the inherent relation of variable temperature during storage.

Abstract

Freshness prediction was a research hotspot in the field of food science. The current microbial kinetic equations could predict the freshness under certain fixed temperature conditions, but they were no longer effective when the temperature was fluctuated. To solve this problem, this paper used deep learning techniques to mine the inherent relation of variable temperature during storage and proposed a novel model named CNN_LSTM (convolutional neural network_ long short-term memory). The model didn't need to fit the parameters of a fixed equation, and it had the advantage of predicting freshness within a range of temperature fluctuations. The results showed that CNN_LSTM could get better prediction results than classic microbial kinetics methods such as logistic equation, Gompertz equation and Arhenius equation under fixed temperature conditions. When the temperature fluctuated, the model could still accurately predict total viable counts (TVC) under variable temperature conditions, with the determination coefficient (R²) greater than 0.95 and the root mean square error (RMSE) less than 0.2. In addition, the model had the potential to predict freshness under different change factors besides temperature fluctuations, which provided a new prospect for freshness prediction.

Introduction

Salmon is delicious, nutritious and rich in protein, but it is easy to spoil and deteriorate. According to reports, every year 30% of aquatic products cannot be consumed due to spoilage and deterioration, and this loss accounts for 25% of the total agricultural losses(Ghaly et al., 2010). For these aquatic products, freshness is one of the most important indicators to evaluate their quality. Studying the changing law of freshness during some storage conditions, and predicting the freshness or shelf life of aquatic products are the current research hotspots(Jedermann et al., 2017).

Freshness is generally affected by the activities of microorganisms, chemistry and enzymes(Ocaño-Higuera et al., 2011; Wu et al., 2019b). During the process, temperature is the most important factor affecting the growth of microorganisms and enzymes(Jay et al., 2008). Most freshness prediction models are established based on microbial kinetics under some temperature conditions. In the process of establishing microbial kinetic models, current researchers have used the knowledge of mathematics, statistics, and microbiology to simulate microbial metabolism(Baranyi and Roberts, 1995; Liu et al., 2021; Nielsen and Villadsen, 1992). At present, the microbial kinetic models can be divided into first-level model, second-level model and expert system(Buchanan, 1993; Ren et al., 2022). The first-level model mainly describes the relationship between microbial growth and time, including Gompertz equation(Baranyi and Roberts, 1995; Ren et al., 2022), logistic equation(Dalgaard, 1995), etc. The second-level model describes the impact of environmental factors on the growth of microorganisms, such as the Arrhenius equation(Guo et al., 2018; Peleg et al., 2012).

However, these kinetic models have some disadvantages. Although these models can obtain good prediction results at a fixed temperature, they cannot adapt to the real storage environment of aquatic products, because temperature fluctuations are almost inevitable during the cold chain transportation and storage of aquatic products. When the temperature fluctuates, the kinetic model is difficult to adapt to the complex and changeable environment factors, because the kinetic equation often has limited parameters, fixed expressions and other defects, so it is urgent to use new theory to predict the freshness.

Some researchers have applied new algorithms for risk assessment of food safety(Lin et al., 2021). Some deep learning and neural network models have been proved to be feasible to apply to food safety early warning(Geng et al., 2021a, 2021b). Therefore, In order to solve the above problems, this paper proposed a new idea based on deep learning to predict freshness. Deep learning is a special neural network that contains multiple hidden layers of multi-layer perceptrons(Bengio et al., 2017; LeCun et al., 2015). Deep learning can combine underlying feature functions to form a complex, multi-layer, and nonlinear deep neural network to perform high-level representation of data. The network can solve data feature expressions that cannot be represented by traditional mathematical equations (LeCun et al., 2015). In addition, deep learning has the advantages of automatic extraction of data features and strong anti-interference ability, which can effectively improve the shortcomings of microbial kinetic equations.

Therefore, the purpose of this paper is to establish a prediction model through deep learning to predict the freshness of salmon under temperature fluctuations. Total viable counts (TVC) was measured as an indicator of freshness and a deep learning model called CNN_LSTM was proposed to predict freshness. In order to highlight the advantages of deep learning model, several typical microbial dynamics models have also been built and compared. The result showed that this method could solve the temperature fluctuating fitting that couldn't be done by traditional kinetic models, and realize the prediction of salmon freshness under arbitrary temperature conditions.