Research On Markov Logic Network Structure Learning Algorithm And Its Application In Indoor Activity Recognition

Currently,indoor activity recognition technology based on sensor data has gradually become a research hotspot in the fields of smart medical care,smart home and other fields.Markov Logic Network(MLN),as an optimal or suboptimal firstorder logic rule set with weights,can construct the relationship between user activities and activity rules through structural learning,further completing indoor activity recognition.However,as the types of sensors increase day by day,the relationship between indoor user activities and activity rules become more complex.Therefore,indoor activity recognition based on MLN structure learning faces problems,such as poor continuity,slow real-time update,and difficulty in learning crossover and concurrent rules.Focusing on the above problems,this paper studies continuous learning in MLN structure learning and its application in single-user single-activity,real-time learning and its application in single-user multi-activities,crossover and concurrent rule learning and its application in multi-users and multiactivities,combined with the specific application scenarios in indoor activity recognition,the specific work is as follows:1)Continuous learning of MLN structure and its application in singleuser single-activity:to solve the problem that MLN cannot learn new knowledge based on existing knowledge,and its poor continuity of structural learning leads to low learning efficiency,this paper puts forward a cumulative learning-based MLN structure learning algorithm.The method combines the continuous learning characteristics of cumulative learning,so that the newly emerged knowledge will not overwrite the existing knowledge.Instead,the algorithm can carry out learning based on the existing knowledge,overcoming the drawbacks of learning new knowledge from scratch in MLN structure learning.In order to verify the effect of the proposed algorithm,this paper selects a single-user single-activity recognition scenario for experimental verification.The results show that in a scene with 8 types of activities,without affecting the recognition accuracy,the MLN structure learning algorithm based on cumulative learning can shorten the learning time from the original 103 hours to 1 hour,greatly improving the efficiency of indoor activity recognition.2)Real-time learning of MLN structure and its application in single-user multi-activities:centering on the problem that MLN cannot learn in time with scene changes,and its poor real-time structure learning leads to slow structure update,this paper proposes a MLN structure learning algorithm based on probabilistic latent semantic analysis.The method constructs and optimizes the semantic rules of the current scene by considering the real-time characteristics of indoor activities,cutting the real-time data stream by selecting an appropriate sliding window with the sensor position,and carrying out probabilistic latent semantic analysis,thus realizing MLN structure real-time learning and updating.The method is verified experimentally in a single-user multi-activity recognition scenario.The results show that in the recognition of 10 types of activities,the MLN structure learning algorithm based on probabilistic latent semantic analysis can shorten the structure learning and updating time from 19 days to 1 day,solving the problem of slow real-time updating of the structure.3)Crossover and concurrent rule learning of MLN structure and its application in multi-users multi-activities:to address the problem that MLN is difficult to deal with crossover and concurrent rule learning,this paper proposes a MLN structure learning algorithm based on collaborative algorithm.The method first classifies sensor data by using sensor bags and convolutional neural networks to distinguish multiple independent rules in crossed or concurrent activities,and then defines multi-user semantic rules based on application scenarios,uses collaborative algorithms based on user position,action,and trajectory,and combines research contents(1)and(2)to continuously learn and update each MLN structure in real time,realizing the learning and optimization of crossover or concurrent activity rules,and finally builds a multi-user and multi-activity rule set.The method is verified experimentally in a multi-user and multi-activity recognition scenario,and the average accuracy rate can reach 94.75%in an indoor activity data set containing 4 people.