Research On Robust Self-Organizing Neural Network Based On TS-type Fuzzy Logic And Its Application

With the rise of artificial intelligence in recent years,the modeling methods based on artificial intelligence algorithms such as neural networks have been greatly developed in theory and interpretation,and some practical applications have been solved through neural networks,such as: achievements in computer vision,natural language processing,industrial detection,and time series forecasting are constantly emerging.However,an important reason why general-purpose neural network models are difficult to deploy on a large scale in industrial practice is that data uncertainty and partial truth are widespread during the sampling process of industrial data,manifested by the existence of randomness,ambiguity,and the unpredictability leads to the model not being robust enough in a complex industrial environment,especially when there are multiple disturbances and uncertain factors.Therefore,the robust modeling methods based on intelligent algorithms has important theoretical significance and application reference value.Takagi-Sugeno(TS)fuzzy inference system is based on TS fuzzy logic neural network,which is the most widely used fuzzy neural network modeling method in recent years due to its excellent interpretability,local approximation ability and relatively refined network structure.In this paper,a robust self-organizing neural network architecture based on TS fuzzy logic is provided.Through improving the structure and calculation of the model,we provide the robust modeling methods for three hot application problems in industrial classification detection and regression prediction.The algorithms are implemented and the robustness of the algorithms are verified by two industrial applications.The specific research content is as follows:(1)Aiming at the problems of false alarmings due to outlier samples under normal conditions in industrial classification detection and the difficulty of single model to achieve fault detection and classification recognition.Based on supervised identification and classification,unsupervised detection of abnormalities,we proposes a robust self-organizing detection model based on TS-type fuzzy neural network.An offset is added to the calculation method of the firing strength of fuzzy rules in the model antecedent network,so that the model can effectively suppress the output of outliers.Secondly,a set of self-organizing mechanism towards to TS-type fuzzy neuron(fuzzy rules)is established,and the model's structure is improved by adding,merging,and deleting to enhance the robustness and generalization ability of the model.Finally,in order to solve the problem that the stochastic gradient descent algorithm is easy to fall into the local best points,an adaptive learning algorithm is proposed.Experiments under normal operating conditions and experiments with abnormal data verify that the proposed model not only has a significant improvement in accuracy and robustness,but also can detect false alarmings and detect faults in abnormal modes in the cases of hardware failure and strong environmental interference.(2)In the cases that some sampled signals are greatly lost or partial truth in the actual operation of the classification detection model,such as data loss,data distortion,or signal saturation,it is not only easy to make the the output of fuzzy rules in the consequent network has a large deviation,but also easy to cause the disorder of the priorities of the fuzzy rules in the antecedent network,which leads to the failure or even collapse of the fuzzy system.A robust fault-tolerant recognition model based on TS-type fuzzy neural network is established.Firstly,a new type of weighted activation degree is used to calculate the firing strength of the corresponding fuzzy rule.Secondly,a calculation method for the initial value of the feature representation coefficient is proposed,that is,the weighting factor in the weighted activation degree.Finally,a new type of robust fuzzy rule is designed,which effectively suppress the contribution of some abnormal components to the output of fuzzy rules.It is verified in experiments that the proposed model not only has high recognition accuracy and generalization ability under normal working conditions,but also has excellent fault-tolerant recognition performance in three kinds of robust experiments.(3)Aiming at the uncertainty of some features and non-Gaussian distribution in industrial regression forecasting,a robust time series prediction model based on TS-type recurrent(time-recursive)fuzzy neural network is proposed.Firstly,due to the relative correlation of the time series prediction problem,we propose a state recursive weighted activation degree.Secondly,a novel feature selection strategy is proposed,which fully considers the combination mode between excellent features to provide the most relevant related information to the predicted target,and the algorithm retains the non-linear relationship of information between related features.Finally,through the GEFCom2012 data set we validate that the model has excellent robust prediction performance on the long-term load forecasting in the presence of uncertainty in the data of the power system data(calendar feature non-Gaussian distribution,feature uncertainty during the festival,and temperature feature shifts caused by global warming).