Fall Recognition Based On Surface EMG And Plantar Pressure Signal Fusion

In recent years, aging phenomenon of society is serious. The health of theelderly is becoming social problems that cannot be ignored. Falls is high incidenceevents among the elderly; it can bring tremendous health hazards. With thedevelopment of signal detection and pattern recognition techniques, more and moreresearch about fall detection is taken in domestic and foreign. When falls happened,if there is a device that can promptly issue warning to family members or health careworkers, fallers can get medical help timely and reduce some body injury. Therefore,the study of the elderly fall detection is a meaningful subject with a high applicationvalue. It also reflects the social care for the elderly.The core issue of fall detection is to distinguish between fall and Activities ofDaily Life (ADL), in order to initiate protective device or an alarm to protect theelderly timely. This paper selects surface electromyography combine with plantarpressure signal as the source signal for fall detection: Limb movement is driven byskeletal muscle. When the body is in motion, motor nerve cells in active muscle willgenerate bio-electrical signals, this signal called Surface electromyography (sEMG).It reflects the state of muscle contraction, and describes changes in body posturefrom a biological perspective. Preliminary study found that plantar pressure signalchanges significantly during the dynamic process of fall. It is a new attempt thatcombines sEMG and plantar pressure signal as the source signal to detect fall.This paper takes study on EMG noise reduction, feature extraction, plantarpressure signal acquisition and modeling, pattern classification based on sourcecombination. The specific work done as follows:(1) Propose a new fall detection method and design signal acquisitionprogram.1) This paper introduces application background and technologies of falldetection at home and abroad. Since there is no fall detection study based onmulti-information fusion, an EMG combine with plantar pressure signal methodproposed in this paper.2) According to EMG signal generation mechanism and the location of thelower limb muscles and their major role in the body’s daily activities, after some experiment, this study designs a scientific EMG acquisition program. The studyfinds the relation between human’s activity and the law of variation of plantarpressure signal, then designs an appropriate pressure signal acquisition system,providing a reliable source signal for the subsequent fall detection.(2) Discuss EMG noise reduction method. Since EMG is non-linear andnon-stationary signal, can be mixed with noise easily. Based on different acquisitionmethods, the paper proposes two noise reduction methods:1) There is an aliasing problem of EMG acquired by multiple channels. Tosolve this problem, the paper proposes a method: take a noise reduction on sEMG bysecond generation wavelets first. The second generation wavelets have an improvedthreshold function. Then take the ICA(Independent Component Analysis) separationamong signals by an improved FastICA algorithm.2) For a single sEMG, the paper uses a noise reduction method based onEEMD analysis: First, the EEMD transformation is taken on sEMG. By calculatingthe frequency validity of each IMF component, to determine which component iseffective. Then reconstructs1~6thcomponents which are effective to get areconstructed signal. In the end, the study takes a separation between thereconstructed signal and frequency noise then gets the final denoised signal.This paper takes some experimental analysis to verify whether these twoapproaches are suitable for EMG.(3) Propose two feature extraction methods for fall detection, they are:1) Extract the fuzzy entropy of gastrocnemius and vastus lateralis muscle.2) Extract the approximate entropy and basic scale entropy of gastrocnemiusmuscle.sEMG is a non-stationary biological signal, it is susceptible to be interferedwhen collected from skin. Fuzzy entropy, approximate entropy and basic scaleentropy describe the degree of signal complexity. All of them have anti-noise andanti-jamming ability, and suitable for sEMG. This paper compares the EMG featuredistribution characteristics and effectiveness of two methods.(4) Proposed a pattern classification method and a multi-source signal fusionmethod that are suitable for fall detection.1) Propose a WKFDA(Weighted Kernel Fisher Linear Discriminant Analysis)classification method for classification of falls and ADL. General classificationmethods often get poorly performance for unbalanced data set. Experimental results show that WKFDA overcomes this defect and get a better pattern recognition resultfor the fall and ADL.2) According to the law of variation of plantar pressure signal, when theaction occurs, the study extracts feature from plantar pressure signal. The SVMdecision based on EMG and plantar pressure signal are fused by D-S evidentialreasoning then can get the final pattern recognition result.In this way, the recognition rate of falls and ADL greatly improved. Itprovides a new idea for the fall detection.