Research On Key Technologies Of Eggshell Crack Detection And On-line Detection Based On Acoustic Method

Egg is the one of the most nutritious food in human’s life and the raw material of other foods. Egg takes a very important position in food industry. However, egg would be infected by bacteria vulnerably, if there are cracks on the eggshell, which would be harmful to human health. Eggshell crack detection has become the most important procedure in egg grading line. Conventional method of eggshell crack detection usually operated by human. Its low detection accuracy and speed hardly meet the increasing capacity of egg grading system. Eggshell crack detection has become the bot-tleneck of egg quality detection line in China, because it is far more difficult for automation than other stages. Two main methods which applied in eggshell crack detection were machine vision and acoustic method. MOBA (Netherlands) and other companies has applied acoustic method in eggshell crack detection because of its fast detection speed and relatively simple algorithm.In our research, a retractable actuator was developed for tapping the egg, and a diver circuit and CompactRIO system were designed for driving the actuator and acquisition of force and acoustic signal. The best impact force for eggs with different sizes was researched and the model between egg size and impact force was established. Time and frequency domain features that were used and customized by previous researchers were extracted. F-ratio was used to evaluate the effect of every feature on the discrimination of sound and cracked eggs, and correlations between features were investigated. Artificial Neural Network was utilized for training and testing the recognition effect of extracted feature set. On-line detection system based on DSP and FPGA was developed for tapping the eggs with different size and speed accurately and acquisition of 24 channels signal simutaniously. Recognition accuracy of on-line detection system was researched.The main results and conclusions were listed as follows:(1) A retractable actuator driven by electromagnet and its driver circuit and signal acquisition system were developed, The best impact force for eggs with different sizes was researched. The development of the actuator and the system include:1) Analysis of the actuator’s structure, operat-ing principle, and influence factors of impact force.2) Designation of the diver circuit for actuator and microphone pre-amplifier circuit.3) Development of signal acquisition system based on Com-pactRIO, which sample and analyze the impact force signal and the acoustic signal. Researchs on the best impact force for eggs with different sizes include:1) Establishment of the model between modified force and parameters.2) Exploration of the upper limit and under limit of impact force.3) Other parameters’ influence on impact force. Upper limit of impact force was set to 30 N guarantee undamage to eggshell, and lower limit was set to 25 N for operating stability of the actuator. In order to control the impact force in suitable range for different eggs, the best pulse widths were calculated. The relationship between pulse width and egg size are linear relationship.(2) Analysis and optimization of features extracted from acoustic signal. Different types of features were extracted from both the time and frequency domains of acoustic signals. F-ratio was used to evaluate the features’ capability to distinguish cracked eggs from sound ones. Furthermore, the correlations between pairs of features were investigated, feature with high F-ratio and low cor-relation with other features were reserved. Neural network was utilized for training all the features and the reserved features. Features from new samples were applied for testing the neural network trained before and the influence of different tapping location and crack location was observed. The result indicated that:1) The total energy E0 has the highest F-ratio 2.058, features with F-ratios lower than 0.1 are disregarded. Pair of features are very strongly or strongly correlated (r≥0.6) in our experiment were observed and the feature with the smaller F-ratio is neglected. Finally, DT、 E0、ADF0、DF1、SR、BE1、BE2、BE4, BE5, BE7 and BER are reserved.2) All features and reserved features acquired from the eggs in training set were used for neural network training sep-arately, the comprehensive accuracy of classification is 99.8%(all features) and 98.8%(reserved features). So, features reduction would not decrease the classification accuracy dramatically.3) The distinguish accuracy of sound eggs were larger than 90%, recognition accuracy of crack egg depends on the crack location and tapping location, accuracy would be much higher (90%) if the crack location is the same as impact location (90%). There is possibility (60%) to distinguish the crack located on blunt end and sharp end if the impact location is on equator and also has possibility (60%) to distinguish the crack located on equator if the impact location is on blunt end and sharp end. It would be difficult (30%) to detect the crack if the crack location and impact location are on the opposite end.(3) On-line eggshell crack detection system was developed. The system includes:general framework, DSP based controlling system, FPGA and DSP integrated signal sampling and analyz-ing system. Achieved tapping the different eggs with the same suitable impact force at different speed,24 channels signal sampling simutaniously, and real time analysis. The running state of on-line detection system was tested, the result indicated that:the system run stably, and satisfied the maximum detection speed (5 eggs/s) demand. The signals from different actuators and from different eggs have good uniformity.(4) Features extracted from acoustic signals when tapping the rotating eggs in on-line detec-tion system were analyzed and compared with the features extracted from acoustic signals in static detection system, recognition accuracy was researched. Recognition effect of ANN model in static system (chapter 2) and on-line system (chapter 4) was measured. Reevaluate and refilter of fea-tures of acoustic signal in on-line detection system were researched and compared with the reserved features in static system, and ANN was applied to measure the recognition accuracy of the reeval-uated and refiltered features. The result indicated that:1) The recognition accuracy of ANN model trained in static system was low (sound egg 53% and cracked egg 66%) when distinguishing the signals from on-line system.2) The recognition accuracy promoted using the ANN model retrained in on-line system (sound egg 87% and cracked egg 82%).3) 9 features were reserved after reevalu-ation and refilter, they are:E0, BE3, BE8, VARt, BER, DT, BE2, SR, BE4.4) Signals from on-line system were tested by the ANN model trained by reevaluated and refiltered features, recognition accuracy was 85.5% for sound eggs and 85.3% for cracked eggs. The recognition accuracy was not obviously promoted, however, much lower than the recognition accuracy in static system.5) It is easier to detect the cracks on equator and blunt end of eggs, but not so easy to detect the cracks on sharp end of eggs.6) Recognition accuracy would not be affected by egg weight and egg shape in on-line system.