On-line ammonia nitrogen measurement using generalized additive model and stochastic configuration networks
Introduction
With the continuous improvement of aquaculture technology, large-scale and intensive aquaculture has flourished. In order to increase the density of breeding and yield per unit area, the amount of artificially fed bait has been increasing, and the ecological balance of water becomes the key to the success of aquaculture [1]. The concentration of ammonia nitrogen in water is one of the important indicators of aquaculture water quality. Its toxic components directly affect the healthy growth of breeding organisms and determine the yield and quality of breeding organisms, so the real-time detection of ammonia nitrogen concentration is an urgent problem to be solved in aquaculture.
At present, the main methods for the detection of ammonia nitrogen concentration include Nessler's reagent, molecular absorption spectroscopy, hypobromite oxidation, electrochemical methods and so on [2]. These methods can effectively detect the ammonia nitrogen concentration, but they are all offline, and can’t be used for real-time monitoring [3]. Although there are some multi-parameter water quality analyzers, such as America’s YSI 6820/6920, German WTW Multi 9630 and so on, they still have problems such as high cost and high maintenance cost, and most instruments are suitable for fresh water instead of seawater [4]. In recent years, soft measurement methods have been used to monitor the ammonia nitrogen concentration in sewage [5], [6] and achieved better results, intelligence methods have been used, such as support vector machine [7] and fuzzy neural networks [8]. The aquaculture water environment is a complex ecosystem. Changes in water quality are often the result of a combination of many factors. While paying attention to the quality and safety of breeding organism products, it is necessary to ensure the safety of breeding organisms and avoid unnecessary losses.
Researches also tried to use soft measurement method in aquaculture. Deng et al [9] used radial basis function (RBF) neural networks to establish the ammonia nitrogen concentration prediction model by using water temperature, pH, dissolved oxygen, nitrate nitrogen content and nitrite nitrogen content as auxiliary variables. As the content of nitrate and nitrite are also hard to be measured online. Therefore, the practicality of this method is not strong. To improve this method, Li et al. and Yu et al. used water temperature, pH, conductivity and so on as the auxiliary variables and artificial intelligence methods, such as support vector machines (SVM) [10], stochastic configuration networks (SCN) [11] and Random vector functional-link (RVFL) [12] to predict ammonia nitrogen. As most of these models were based on data driven methods and due to the “black box” characteristics, the established models do not have the ability to interpret and are also difficult to guarantee accuracy. So far there is no successful on-line sensors for the ammonia nitrogen measurement.
Hybrid modeling with part of the model from first principles and part of the model inferred from data, is a good way to take advantages of making full use of mechanistic knowledge and apply the data-driven method to compensate model error for domain problem solving [13].
The mechanism analysis uses the generalized additive model (GAM). The compensation model uses the random neural network SCN. This method can effectively use the relationship between the auxiliary variables and the dominant variable, and the established model has certain explanatory power. This model can overcome the difficulties of the back propagation (BP) and RBF networks, such as slow learning process, falling to the local minimum, and poor generalization capability.
In order to create an on-line soft measurement method of the ammonia nitrogen, especially for the aquaculture water quality control, in this paper, we make the following contributions:
- (1)
We propose the on-line sensor of the ammonia nitrogen for aquaculture water quality. It uses the mechanism model and data driven compensation.
- (2)
We establish an intensive aquaculture system in our laboratory to evaluate our soft measurement method. Based on this system, aquaculture water parameters can be measured. The proposed ammonia nitrogen model is verified by simulation experiments.
The remainder of this paper is organized as follows: Section 2 describes the intensive aquaculture process and mechanism of ammonia nitrogen reaction in aquaculture water. Section 3 introduces the hybrid modelling strategy of ammonia nitrogen on-line measurement, details the mechanism analysis model and error compensation model. Section 4 reports the intensive aquaculture system in laboratory and our experimental results, and Section 5 concludes this paper.
Section snippets
Intensive aquaculture process
The intensive recirculating aquaculture system uses biological reaction and physical filtration methods to remove residual bait feces, ammonia nitrogen, nitrite and other harmful substances contained in the aquaculture water, and recycle the purified water by disinfection, aeration, temperature adjustment and so on to reuse [14]. It involves intervention in the growing process, such as water aeration and supplemental feeding. The process mainly includes biological filtration, disinfection and
Strategy of on-line measurement
From the above mechanism analysis, we can see that there is a nonlinear relationship between the ammonia nitrogen concentration and each influence factor, and the overall constitutes a logarithmic relationship. GAM is based on nonparametric regression and smoothing techniques. It consists of nonparametric additive terms. Each additive term is estimated using a single smoothing function and can explain how the response variable changes with the explanatory variable. Therefore, based on the
Experiment system
In order to realize the proposed ammonia nitrogen soft measurement model, an intensive aquaculture hardware measuring system is developed as shown in Fig. 4. It is mainly used to control the circulation of water and monitoring the water quality parameters. This system uses liquid level sensor, pH sensor, temperature and conductivity integrated sensor to collect data. The obtained signal is transmitted to the PLC controller, then to the computer which is used to implement the proposed soft
Conclusions
In this paper, we solve the on-line measurement problem of the ammonia nitrogen concentration. We use hybrid modelling strategy, which uses the mechanism model and data driven compensation model. The laboratory intensive circulating aquaculture system is designed and developed. The relevant water quality parameters are obtained by the proposed method. From the experimental results, we can see the mechanism model can get the internal relation and data driven model can improve the performance
CRediT authorship contribution statement
Wei Wang: Conceptualization, Methodology, Writing - original draft. Yao Jia: Methodology. Wen Yu: Supervision, Writing - review & editing. Hongshuai Pang: Software. Kewei Cai: Software, Visualization.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (32)
- et al.
New developments in recirculating aquaculture systems in Europe: A perspective on environmental sustainability
Aquacult. Eng.
(2010) - et al.
Data-driven soft measurements in the process industry
Comput. Chem. Eng.
(2009) - et al.
Process modelling of an intensive aquaculture system
Aquacult. Eng.
(1993) - et al.
Fish bioenergetics and growth in aquaculture ponds: II. Effects of interactions among, size, temperature, dissolved oxygen, unionized ammonia and food on growth of individual fish
Ecol. Modell.
(1985) - et al.
Multivariable dynamic modeling for molten iron quality using online sequential random vector functional-link networks with self-feedback connections
Inform. Sci.
(2015) - et al.
Insights into randomized algorithms for neural networks: Practical issues and common pitfalls
Inform. Sci.
(2017) - et al.
Construction of prediction intervals for carbon residual of crude oil based on deep stochastic configuration networks
Inf. Sci.
(2019) - et al.
Comparison of Three Methods for Determination of Ammonia Nitrogen in Surface Water
Meteorol. Environ. Res.
(2019) - et al.
Detection methods of ammonia nitrogen in water: A review
Trends Anal. Chem.
(2020) - et al.
Determination of ammonia nitrogen in natural waters: Recent advances and applications
Trends Environ. Anal. Chem.
(2019)
Intelligent modeling approach to predict effluent quality of wastewater treatment process
Wastewater Water Qual.
Cost effective smart system for water pollution control with underwater wireless sensor networks: A simulation study
Comput. Syst. Sci. Eng.
Analysis of ammonia nitrogen content in water based on weighted least squares support vector machine algorithm
J. Software Eng. Appl.
Soft sensing of effluent ammonia nitrogen using rule automatic formation-based adaptive fuzzy neural network
Desalin. Water Treat.
Cited by (15)
Prediction of X-ray fluorescence copper grade using regularized stochastic configuration networks
2024, Information SciencesP<sup>2</sup>CA-GAM-ID: Coupling of probabilistic principal components analysis with generalised additive model to predict the k−barriers for intrusion detection
2023, Engineering Applications of Artificial IntelligencePortable device for on-site detection of ammonia nitrogen
2022, Information Processing in AgricultureCitation Excerpt :Hence, in the real environment, it is no need to consider the measurement effect of the device when the pH change. The research has shown that the measurement of the ammonia nitrogen concentration is theoretically influence by temperature [34]. Hence, it is necessary to study the exact relationship between its stability results and ambient temperature.
Broad stochastic configuration network for regression
2022, Knowledge-Based SystemsCitation Excerpt :The input parameters of hidden layer nodes are configured by the supervision mechanism and the output parameters are calculated by Moore–Penrose inverse. SCN, with a series of advantages of lower human intervention, higher learning efficiency and stronger generalization ability, has been applied in the fields of online measurement problems [7], industrial process modeling [8], vibration signal fault diagnosis [9] and power big data processing [10]. Since it was demonstrated in 2017, a lot of improved models and variants have been developed.