| Total nitrogen is one of the most critical water quality parameters in municipal wastewater,which can effectively reflect the pollution status and eutrophication degree of water,and its online detection is highly crucial.However,under the complex composition of municipal wastewater and harsh detection conditions in wastewater treatment plants,the traditional total nitrogen detection method based on the oxygen digestion process cannot guarantee its detection accuracy and has to experience a tedious chemical reaction process,which limits its application.Therefore,based on intelligent detection techniques,such as compressed sensing,weakly supervised learning,and attention mechanisms,this thesis proposes time-varying spectral analysis for the oxidative digestion process and constructs a new detection architecture for rapid online detection of total nitrogen in municipal wastewater treatment plants from sensing principle,hardware implementation,and intelligent algorithms,and develops online total nitrogen detection prototype and the multi-parameter water quality monitoring station.The contributions and innovations of this thesis are listed as follows:(1)To solve the problem that the online detection period of total nitrogen in urban wastewater treatment processes is too long to meet the needs of the optimized operation,this thesis proposes a new analysis method based on the time-varying spectrum during the oxidative digestion process,which breaks the traditional detection framework of total nitrogen.UV-vis spectrum is first utilized for observing spectrum variations in the initial oxidative digestion reaction and accurately capturing the transformation process of nitrogenous substances.Subsequently,to effectively extract crucial information features of the time-varying spectrum and shorten the required oxidative digestion reaction time,the thesis proposes a spatial permutation combination population analysis(SPCPA)method for the spectral variable selection based on model population analysis.SPCPA measures the importance of temporal spectral variables from both global and local perspectives and efficiently explores the variable space to find the best combination.(2)To solve the problem that the detection of the time-varying spectrum requires equipment with complex structure,huge volume,and expensive optical components and therefore cannot be used in practice,this thesis proposes an online reconstruction method for time-varying spectrum based on compressed sensing.First,a photosensitive sensor is used to replace the complex and expensive UV-visible spectrophotometer to obtain the compressed light intensity signal during oxidative digestion processes.Then,a reconstruction method for compressed spectrum signal is proposed.In terms of the hardware,source spectrum encoding is invented to realize intelligent spectrum editing at the laser transmitting side,which solves the difficulties of traditional compressed sensing in hardware implementation.In terms of the software,the incremental compressed spectrum reconstruction method is developed,which efficiently learns the sparse basis of the original time-varying spectrum,and can accurately reconstruct the time-varying spectrum during the oxidative digestion process,thus overcoming the hardware obstacles to online detection of the time-varying spectrum.(3)To solve the problem that sample labels are difficult to obtain in practical municipal wastewater treatment plants and the model accuracy degrades because of the insufficiency of labeled samples,total nitrogen model construction methods based on weakly supervised learning are explored.Double regularized structure graph learning is first developed to precisely measure the similarity of different samples from both temporal and spectral dimensions and establish the accurate detection model with the utilization of unlabeled samples.Furthermore,considering sample labels obtained in practice tend to be contaminated by noises,multiregularized robustness semi-supervised learning is proposed,which can utilize the temporal and spectral similarity of different samples to correct outliers of labeled training samples to realize the purpose of constructing accurate detection model by using a few samples with inaccurate labels.(4)Regarding the problem of insufficient calibration samples for online model calibration,sample label generation methods based on multi-sensor information fusion are proposed to automatically generate the total nitrogen concentration of calibration samples based on easily measurable water quality parameters,and thereby reducing the manual calibration cost of the model and ensuring its long-term stability.Firstly,a recursive neural network label generation algorithm based on multi-phase attention mechanism is proposed,which combines the one-dimensional convolutional neural network with the attention mechanism to dynamically explore the spatial-temporal coupling relationship between water quality parameters,and achieve accurate label generation.Secondly,in order to address the problem of insufficient samples in practical application,a lightweight label generation method is proposed based on the cross-coupling attention mechanism,which can efficiently learn the topological network structure of different water quality parameters,and achieve accurate label generation in small sample application scenarios.Based on the above theoretical research,the prototype of total nitrogen rapid detection based on the time-varying spectrum has been developed,and a multiple water parameter monitoring experimental station was established to realize the functions of multi-parameter simultaneous online monitoring,remote access control,and historical data visualization.The application results show that the method proposed in this thesis can effectively shorten the online detection period of total nitrogen,improve the detection accuracy of total nitrogen under harsh working conditions,and provide a reliable solution for realizing the rapid and accurate online detection of total nitrogen in municipal wastewater treatment plants. |
PDF Full Text Request