Research On Fault Diagnosis Of Marine Equipment Based On Artificial Neural Network

The operating environment of marine equipment is complicated and harsh,which makes it apt to fail.Once a marine equipment breaks down,it can cause irreversible damage,and even cause environmental pollution and ship suspension.Therefore,fault diagnosis technology is particularly important for reducing the fault of marine equipment and improving shipping safety.With the development of sensor and artificial intelligence technologies,a large number of equipment’s operating data are recorded.Data-driven based fault diagnosis technology has been rapidly developed.The artificial neural network based fault diagnosis methods are widely used in the field of fault diagnosis.Most of artificial neural network based fault diagnosis methods assume ideal fault data.However,the collected fault samples in actual operation of marine equipment are defective.Defective fault data can seriously reduce the accuracy of fault diagnosis.This thesis focuses on fault diagnosis of marine equipment with defective fault data.Some fault diagnosis models based on artificial neural network are proposed.The research work is shown as follows:Considering the economic cost and equipment complexity,most manufacturers only install sensors on key parts of the equipment under the premise of ensuring the safety of marine equipment.Besides,some sensors may also break down.Therefore,the collected fault samples contain a small number of features.Fault samples with a small number of features are difficult to fully reflect the fault information,resulting in a decrease in the accuracy of conventional fault diagnosis methods.As the powerful classification and recognition capabilities,multilayer perception is often used for fault recognition with a small number of fault features.The random setting of the initial weights and thresholds of multilayer perceptron and the difficulty of determining the network structure can decrease the accuracy of fault diagnosis.A fault diagnosis model based on multilayer perceptron is proposed.The multilayer perceptron is optimized using genetic algorithm to select optimal weights,thresholds and network structure.However,the standard genetic algorithm is sensitive to initial population,and the exploration ability of the new space generated by the crossover and mutation operators is limited.The basic operations of standard genetic algorithm are improved.Chaotic mappings are introduced into the basic operations of standard genetic algorithm.The initialization of the chaotic population of Chebyshev mapping is given,which reduces the sensitivity of initial population.The Logistic mapping crossover operation and the Chebyshev mapping mutation operation are designed to increase the diversity of the new population.Finally,the fault diagnosis of marine diesel engine combustion chamber and fuel oil supply system is respectively carried out.The results show that the proposed fault diagnosis model effectively improves the accuracy of fault diagnosis.The marine equipment is often in a normal state.The time when the fault occurs and the duration of the fault are uncertain.Therefore,the collected monitoring data are imbalanced.When such imbalanced samples are directly used to train support vector machine,the classification hyperplane of support vector machine will shift and the accuracy of fault diagnosis will be reduced.A two-stage fault diagnosis strategy based on conditional generative adversarial network and support vector machine is proposed.In the first stage,conditional generation adversarial network is used to generate low-proportion fault samples,and the rebalanced samples greatly enrich the diversity of fault samples.In the second stage,support vector machine is trained using re-balanced fault samples to perform fault diagnosis.Tent mapping is introduced to the mutation operation of standard differential evolution algorithm to improve the ability of searching for new spaces.Differential evolution algorithm is used to select the optimal hyperparameters of support vector machine.Finally,the fault diagnosis of marine diesel engine combustion chamber and fuel oil supply system is conducted.The experimental results show that the proposed two-stage fault diagnosis model can effectively identify the faults.When the high-speed rotating marine equipment runs,its monitoring data are huge,and the types of monitoring signals are also diverse,such as vibration signals,current signals,sound,and thermal images.The fault features of such massive and complicated fault signals are difficult to extract,which reduces the algorithm efficiency of the traditional data-driven methods.The commonly used method is deep learning based fault diagnosis method,which converts a certain type of fault signals into gray or RGB images as input samples.However,this method can only reflect the single fault feature information of complicated data,resulting in the decrease in accuracy.A generalized signal processing method is proposed.And on this basis,a fault diagnosis model based on input feature mapping and deep residual network is proposed,which solves the problem that a large number of complicated fault samples are difficult to integrate and fault features are difficult to extract.The proposed signal processing method integrates different types of original signals and transforms them into multi-channel input feature mappings.The multi-channel input feature mappings can fully reflect the feature information of complicated fault samples.The proposed fault diagnosis model can automatically extract and learn the fault feature of the original fault signals.The proposed method is verified through the public rolling element bearing fault datasets.The experimental results show that the proposed method can effectively process massive and complicated fault signals and yield ideal fault diagnosis results.The marine equipment often runs under different operating conditions,and the collected operating data usually have different distributions.The artificial neural network is trained with the source domain samples.When the source and target domain fault samples have the same distributions,the artificial neural network can effectively identify the target domain samples.On the contrary,when they have different distributions,the fault diagnosis accuracy in the target domain will decrease to a large extent.An improved adversarial discriminant domain adaptation fault diagnosis model is proposed.A label classifier is added to the original structure of the standard adversarial discriminant domain adaptation model and the CORAL loss function is introduced to reduce the dataset bias in the label classifier.The improved adversarial discriminant domain adaptation fault diagnosis model reduces the dataset bias in the feature extractor and label classifier.It solves the problem that the standard adversarial discriminant domain adaptation model cannot handle the dataset bias in the label classifier.The proposed algorithm is verified by multiple transfer tasks in two public rolling element bearing datasets.The results show that the proposed method has a good ability to identify faults with dataset bias.The above methods provide new solutions for fault diagnosis of marine equipment with defective fault data.The results show that it has a potential industrial application prospect.