Illumination-invariant Froth Image Color Measuring With Its Applications In Soft Measurement Of Flotation Concentrate Grade

Concentrate grade is the most important quality evaluation indicator of mineral flotation processes.However,so far,it is difficult to detect the concentrate grade in real time.In the actual monitoring of industrial flotation processes,the concentrate grade detection mainly depends on manual sampling and laboratory assays,bringing a several-hour delay of concentrate grade detection.Thus,only a limited number of grade data can be detected in the routine process monitoring.Therefore,flotation productions cannot be adjusted effectively and timely due to the unavailable online detection of concentrate grades.It is difficult to ensure the stability of concentrate index,and it is easy to cause low grade of concentrate or low recovery of mineral resources in the mineral processing.In recent years,machine vision-based flotation process monitoring has been referred to as an essential tool to achieve stable and optimal production of industrial flotation processes,because considerable studies have shown that the visual properties of froth surfaces contain a large amount of information related to the flotation operation conditions,as well as the beneficiation index.For example,the surface color of the flotation froth can be referred to as the most direct and immediate indicator of the concentrate grade.As a result,machine vision-based soft measurement and process monitoring of concentrate grade has attracted broad attention both at home and abroad.It is known to all that the froth surface color feature is a highly-vulnerable image feature due to the light interferences.Influenced by the open environment in froth image acquisition,uncertain light sources,dust and mist in the air,as well as the daily changing of light intensity and incident angles,can cause serious influence to the froth color.Thus,it is difficult to extract color features to truly or effectively reflect the concentrate grade in the froth layer.In addition,the redundancy and noise disturbance will inevitably exist in the froth image characteristics and the parameter variables.The dynamic and non-linear behavior of the process sequence data makes it difficult to directly apply traditional soft sensor models to the soft measuring of concentrate grade.Therefore,focusing on problems in machine vision-based soft measuring of concentrate grade in the flotation process monitoring,the main contributions can be summarized as follows.1.Aimed at solving the problem of inaccurate froth color measurement due to the color deviation in froth image acquisition,an illumination-invariant froth color measuring method is proposed by an introduced Wasserstein distance-based structure-preserving Cycle GAN,termed WDSPCGAN.WDSPCGAN is composed of two GANs.Through game training of the two GANs,WDSPCGAN can map the color of froth images to that of froth images under the reference light condition,while maintaining the invariance of their spatial structures and surface textures.The proposed method has been verified in an actual flotation process.Extensive confirmatory and comparative results show that WDSPCGAN can realize the illumination invariance characteristics of froth image under various unknown light conditions while keeping its structure unchanged,providing effective and objective information for the on-line monitoring of key indicators of mineral flotation process.2.Aimed at dealing with difficulties in soft sensor modeling due to the information redundancy and noise contamination in flotation process monitoring,a STA-APSNFIS-based soft measurement method based on the State Transition Algorithm(STA)and Adaptive Pre Sparse Neural Fuzzy Inference System(APSNFIS)is presented for the online detection of concentrate grade in flotation process monitoring.STA-APSNFIS improves the traditional ANFIS model by using a pre-sparse neural network,reducing the redundant data and a series of noises generated in industrial process monitoring,thereby it reduces the complexity of the fuzzy inference system model and speeds up its convergence and online detection efficiency.To avoid falling into the local optimum during the model training to achieve an optimal model,the STA optimization algorithm is adopted to replace the traditional gradient descent algorithm to learn the APSNFIS model parameters.Extensive comparative results demonstrate that the proposed approach has better stability and effectiveness than traditional ANFIS models,as well its variants,such as PSO-ANFIS,GA-ANFIS,and the existing soft sensor models.3.Taking an actual industrial lead-zinc ore flotation process as a specific research and application object,the proposed computational color constancy and the soft sensor modeling methods were applied to the automatic monitoring of zinc flotation process.A machine vision-based on-line zinc concentrate grade monitoring system was established in the lead-zinc flotation field.The practical application results have shown that the proposed method can correct the color characteristics of froth images to that of reference illumination in real time and effectively monitor the grade of zinc concentrate in froth layer,laying the foundation for the subsequent optimal control of zinc flotation process.