聴覚・触覚情報を用いた 対象物内部状態の認識

A Proposal of an Internal-state
Inference System Based on
Multi-modal Sensory Fusion Method
1K.
Hirota, 1N. Iwamatsu, and 1,2Y. Takama
Tokyo Institute of Technology
PREST, Japan Science and Technology Corporation (JST)
1
Striking-based
Internal-state Diagnosis
• Palpation in hospitals, judgment of fruit’s
matureness, inspection of tunnel walls, etc.
• High dependence on the experts’ skill
• Difficulty in transferring implicit knowledge from
experts to novices
• Proposal: Design of automatic internal-state
inference system with low cost
2
Target Objects
EP: Empty Plastic
HP: Half Plastic
FP: Full Plastic
ES: Empty Steel
HS: Half Steel
FS: Full Steel
Steel can
Plastic can
• 2 types of materials (plastic & steel)
• 3 types of water amount (empty, half, full)
3
Proposed System Configuration
Auditory sensor part
(Microphone with amp)
Inference part
Feature
extraction
Target object
(Unknown
internal state)
Reexamination?
Tactile sensor part
(Hammer attached with
Tactile Sensor)
Inference
based on
multi-modal
fusion
4
Developed System
Hammer part
Target object
5
Used Sensors
• Auditory sensor part
– Condenser microphone
(Max 120dB)
– Amplifier (40dB)
• Tactile sensor part
– Accelerometer: MA304Aa (Microstone Inc.)
– Range: ±4G
– Sensitivity:
50±10%(mV/G)
– Frequency:0.8-1KHz
6
Sound wave
Feature Extraction
• Ampmax: maximum
amplitude
• T: time interval from
1st to 2nd stroke
• T1: time interval from
Ampmax to 1/2Ampmax
(last time within T)
• T2: time interval from
Ampmax to 1/4Ampmax
(last time within T)
Tactile info.
7
Inference of
material & amount of water
• Material identification
– Based on the difference of decay time between
steel and plastic cans
– Target is plastic if the following holds, otherwise
steel
• Discriminating amount of water
–
(A: identified material): system’s output
8
Experimental Results
Target’s
state
EP
EP
17
HP
2
FP
0
ES
0
HS
0
FS
0
HP
2
17
2
0
0
0
System’s output
FP
ES
1
0
1
0
18
0
0
19
0
2
0
0
HS
0
0
0
1
17
0
FS
0
0
0
0
1
20
• Accuracy: 80% (total) , 100% (material only)
9
Inference Failure Cases
• E.g. hammer may slip on the target’s
surface
• One of input variables has apparently
different values from average
• No significant difference are found between
membership values : Error detection
• Reexamination is easy because of low-cost
and non-destructive diagnosis
10