Heat Distribution of Current Output Type Artificial Neural Networks IC for the MEMS Microrobot

Heat distribution of the artificial neural networks IC developed for controlling microrobots is described in this paper. We measured the temperature distribution of the designed IC using thermography. As a result, it is found that the major heat source is the current mirror part on the IC. However, it was observed that the part of the artificial neural networks also generated the heat. The heat contribution of the artificial neural network part was evaluated. As a result, the temperature rise of the artificial neural networks part on the IC was calculated to be 0.0535°C per second.


Introduction
Insects inhabit the earth from about 300 million years ago.During that process, insects have experienced many environmental changes and have developed biological functions 1 .The insect has excellent form and function.Therefore, if it is possible to produce micro robots having functions equivalent to those of the insect, it is considered to be possible that the robot deal with various situations and application in various fields 2 .For example, in the medical field, the microrobot that mimics the peristaltic movement of the insect has been studied 3 .
The Insect has a flexible control system with a compact moving mechanism and neural networks of the brain.In order to imitate this, we are applying micro electro mechanical systems (MEMS) technology to create the movement mechanism of the microrobot.MEMS technology is based on the integrated circuit (IC) production process.Also, the conventional program control used as a robot control makes it difficult to deal with unknown situations.Therefore, the control method using the artificial neural networks has been expected as an alternative method of the program control 4 .Neurons of the brain in the living organisms are connected to each other and form networks.By carrying out information transmission using those, the flexible information processing can be performed.By modeling the neuron of the living organism and constructing the artificial neural networks, the robot can judge and execute the required process by themselves like the living organism.We have integrated simple artificial neural networks into a IC and used it as a controlling circuit of the MEMS microrobot 5 .We designed a current output type artificial neural networks IC to drive the MEMS microrobot using shape memory alloy (SMA) for the actuator.In this paper, focusing on heat generation of the artificial neural networks IC, its heat distribution was measured by thermography.We considered the source of the heat generation from the results.

P -229
Artificial muscle wires based on SMA are used in the actuator of the robot.The wire shrinks by joule heat of current flow and extends by cooling.Fig. 1 shows the walking mechanism of the microrobot.The walking motion of the microrobot is generated by the rotational motion of the rotor attached with 4 helical artificial muscle wires.The rotational action of the actuator is obtained from current flows through the artificial muscle wire in the order of A to D. Therefore, the rotor rotates and transmits to the link mechanism of the leg portion.

Artificial neural networks
An artificial neural circuit mimics the function of a neural circuit of the living organism with analog electronic circuits.The neural circuits of the living organism is mainly composed of cell bodies, dendrites, axons and synapses.We modeled the cell body and the synapse.

Artificial neural networks IC
We designed an IC that constructed pulse-type hardware neural networks.Fig. 6 shows the layout pattern of the designed IC.The design rule of the IC was 4 metal 2 poly CMOS 0.35μm.The sizes of the capacitors CG, CM of the cell body model were too large to pattern on the IC.Therefore, it was added on the peripheral circuit.Fig. 7 shows an IC mounted on the circuit board.Fig. 8 shows an example of output waveform of the pulse-type hardware neural networks IC.

Measurement result
Fig. 9 shows measurement results.Mainly two high temperature areas are confirmed.The two areas are the artificial neural networks and the current mirror.In the IC shown in Fig. 9, it was confirmed that the maximum temperature in the artificial neural network was 100.4°C and the maximum temperature in the current mirror was 100.1°C.

Measurement result with reduced influence of current mirror
We focused on the heat generation in the artificial neural networks.In order to confirm that the artificial neural network was generating heat, the load resistance was increased from 12Ω to 110Ω.Fig. 10 shows the measurement result.Load resistance of 110Ω was the highest resistance the IC output the stable pulse.In the IC shown in Fig. 10, it was confirmed that the maximum temperature in the artificial neural networks was 39.1°C and the maximum temperature in the current mirror was 39.3°C.The measurement was performed without applying the power source of the current mirror.In the IC shown in Fig. 11, it was confirmed that the maximum temperature in the artificial neural networks was 29.2°C.Fig. 12 shows the measurement result of the temperature rise at the representative points of the artificial neural networks.The temperature rise in the artificial neural network was confirmed.

Heat generation in artificial neural networks IC
From these results, it is found that the major heat source is current mirror part because the temperature decreased with increasing of the load resistance.However, the high temperature areas of the IC are corresponding to the not only the current mirror but also the artificial neural networks.The temperature of the artificial neural networks portion was higher than that of the neighbor area.Also, we confirmed the heat generation of the artificial neural network even when the power supply of the current mirror is stopped.P -232 artificial neural networks also generates the heat to some extends.

Evaluation of temperature rise per second of artificial neural networks
In the cell body model, when a voltage of VA is applied, the charge accumulates in CG, and the potential difference between MOSFETs MC1 and MC2 disappears.Then, MC2 conducts, the charges of CG and CM are extracted, applied to the MIS3 of the inhibitory synapse model, and flow to GND.It is thought that the heat is generated by the charges extracted from CG, CM.We evaluated the calculated value of the temperature rise per second from the power consumption of the capacitors CG, CM, the mass of the heat generation part, and the specific heat of silicon.The power consumption due to charges extracted from CG and CM can be obtained by equation CG and CM are 4.7μF and 1.0μF respectively.It is assumed that 0.5 V is applied to CG and 1.5 V is applied to CM.Therefore, power consumption of 0.587μJ and 1.13μJ are obtained on each capacitor.Because the artificial neural networks is composed of four cell body models, then, the one cycle power consumption of the whole artificial neural networks is calculated as 6.87μJ.The duration of one cycle is 1.1 seconds.Therefore, the total power consumption per second is 6.25μW.The temperature rise of the artificial neural networks per second can be obtained by equation (2).

] [ C °V VC P T
(2) CV is the specific heat of silicon, 0.73J/gK.V is the volume of heated area.Those are derived from the temperature distribution of the ICs, 6.88×10-5cm3.is the density of the silicon, 2.33g/cm 3 .From these, it is calculated that an artificial neural networks increases 0.0535°C per second.Compared with the calculation result and the measurement result of Fig. 12, the calculated value follows the tendency of the measured value.

Conclusion
The heat distribution of the artificial neural networks IC developed to controlling the microrobot was described in this paper.As a result, heat generation was observed with the artificial neural networks and the current mirror parts.It was found that major heat source was the current mirror part.Focusing on the heat generation of artificial neural networks, we thought that heat generated by charges extracted from CG and CM of the cell body model.The temperature rise per second was evaluated using the power consumption of CG and CM.As a result, it was calculated that the temperature rises of 0.0535°C per second for the IC.

Fig. 2
Fig. 2 shows the circuit diagram of the cell body model.VC in Fig. 2 is the output voltage.The cell body model consists of a capacitor CG and a membrane potential capacitor CM, MOSFETs MC1, MC2, MC3 and MC4.The cell body changes the membrane potential using external stimulation and fires electrical pulses.The cell body model has a refractory period, an analog characteristic for the output pulse, and time varying negative resistance characteristics.The circuit parameters for the cell body model were VA = 3V, MC1, MC2: W/L = 10, MC3: W/L = 0.1, MC4: W/L = 0.3, CG = 4.7μF, CM = 1.0μF.

Fig. 2 .
Fig. 2. Circuit diagram of the cell body model.

Fig. 3
Fig. 3 shows a circuit diagram of an inhibitory synaptic model.When a multiple of cell body models are connected by the synapse model, a synchronization phenomenon occurs at the oscillation timing of the cell body model.The inhibitory synaptic model, particularly, causes anti-phase synchronization.We used the inhibitory mutual coupling to generate driving pulses of the microrobot.The circuit parameters for the inhibitory synaptic model were VDD = 3V, CIS1 = 1 pF, MIS1, MIS2, MIS3, MIS4: W/L = 1.

Fig. 4
Fig.4shows the pulse-type hardware neural networks mimic a central pattern generator of the living organism.The central pattern generator is known as basic rhythm generator of the living organism.As shown in Fig.4, four cell body models are inhibitory mutually coupled using 12 inhibitory synaptic models.As a result, a four-phase anti-phase synchronous waveform is generated.

Fig. 5
Fig. 5 shows a circuit diagram of the current mirror circuit.The shape memory alloy actuator requires the electrical current to generate movement.Therefore, the current mirror circuit is connected to the pulse type hardware neural networks in order to convert the voltage into the current.The circuit parameters for the current mirror circuit were VDD = 4V, MS1: W/L = 40, MS2: W/L = 1, MON: W/L = 66.The number of stages of the current mirror varied by the designed IC.

Fig. 8 .
Fig. 8. Example of the output waveform of the pulse type hardware neural networks IC 4. Measurement of heat generation 4.1.Measurement condition3V was applied to the artificial neural networks using a power source and 4V was applied to the current mirror.A 12Ω load resistance was connected as output instead of SMA.The measurement was carried out for about 30 seconds after the voltage input.The room temperature was 20°C.

Fig. 11 .
Fig. 11.Measurement of heat distribution of IC without input of current mirror.

Fig. 12 .
Fig. 12. Measurement result of temperature rise of artificial neural networks.
The temperature rise almost corresponded to the pulse timing.So, it is suggested the 0