Interval Prediction Of Activated Sludge Volume Index Based On Microscopic Image Analysi

Activated sludge is widely used in wastewater treatment,and biodegradation and flocculation are the key steps to ensure good sludge-water separation and efficient and stable operation of wastewater treatment plants.Sludge Volume Index(SVI)is a parameter that characterizes the settling and compression performance of activated sludge.The occurrence of sludge expansion often leads to sludge loss,deterioration of effluent quality,cost increase and loss of activated biomass,which seriously endangers the operation of wastewater treatment plants.Quantitative image based on analysis and processing is an effective means of monitoring activated sludge performance and sludge volume indices.Floc and filamentous bacteria image segmentation and morphological feature extraction are the basis of activated sludge phase contrast microscopic image processing and analysis,which directly affects the performance of sludge expansion prediction models.In this paper,a deep learning model based on chunk segmentation is proposed to address the problem of difficult segmentation of microscopic targets of filamentous bacteria in the complex background of activated sludge phase contrast micrographs.For the problem of weak reliability of SVI point estimation due to uncertainty of multiple microscopic images characterizing one activated sludge sample,the SVI interval prediction method based on a subset of morphological features of flocs and filamentous bacteria is proposed.The main research directions of this paper include the following:1.Activated sludge samples were collected at a city wastewater plant in Shenyang,and36 color activated sludge microscopic images were acquired for each sample collected using an image acquisition device in order to characterize the activated sludge samples.The acquired images need to be labeled by the professional labeling software Label Me to construct the dataset.2.The traditional image segmentation method to deal with activated sludge phase contrast microscopic images often has problems of over-segmentation,under-segmentation,and even segmentation failure.Aiming at the poor segmentation effect of filamentous bacteria microscopic targets in activated sludge microscopic images,this paper adopts the Deep Lab V3+ sludge phase contrast microscopic image segmentation model based on the Mobile Net V2＿CBAM backbone network.This method divides a high-resolution phase contrast microscopic image into multiple small areas with a certain overlap rate.Based on the Deep Lab V3+ network structure,the improved lightweight model Mobile Net V2＿CBAM is used as the backbone network to segment each block separately,and then each target area is divided.The split images are stitched back to their original resolution.The experimental results show that the proposed lightweight deep learning segmentation model with chunking strategy has a certain degree of improvement in segmentation accuracy,recall,pixel accuracy and Io U performance metrics and a significant reduction in model size compared to the network model without chunking.3.For a single microscopic image is difficult to characterize an activated sludge sample,a set of color images are systematically scanned for a given slide area along the xy-axis from the upper left to the lower right,and multiple images are jointly characterized for an activated sludge sample.Since it is difficult to guarantee that multiple images cover the same area during the scanning process of different sample image acquisition,the uniqueness of image acquisition directly affects the accuracy and reliability of SVI point prediction.In this paper,we propose an SVI interval prediction method based on a subset of morphological features of flocs and filamentous bacteria.The method is to first construct a subset of features using a combination of random forest feature importance ranking and Pearson correlation coefficient for feature selection and redundancy removal operations,and then quantify the uncertainty associated with point prediction by using Lower Upper Bound Estimation(LUBE)to use the subset of features as input for interval prediction of SVI.The loss function of conventional LUBE is nonlinear,complex and non-differentiable,and gradient descent-based algorithms cannot be used for its minimization.A loss function is used to make it possible to train LUBE(GD-LUBE)using Gradient Descent(GD).The experimental results show that the GD-LUBE method based on feature subsets can construct PIs with narrower interval widths and appropriate coverage probabilities,and GD-LUBE outperforms other algorithms in terms of prediction interval coverage probabilities and average prediction interval widths when the same feature subsets are used as inputs,thus providing more valuable information to the wastewater treatment plant.