A MEMS neural network of three neurons is built to perform simultaneous acceleration sensing and classification tasks as shown in Fig.1.
Each MEMS neuron consists of one suspended proof mass supported by meandering tethers. Each mass has arrays of electrodes, extending outwards forming parallel plate electrostatic actuators with their adjacent electrodes. Depending on their purpose, the electrode arrays are referred to as “softening electrodes” or “coupling electrodes”. The force-displacement characteristics of the electrode arrays make the movement of the proof masses highly nonlinear replicating the nonlinear behavior of a neuron in a typical recurrent neural network (RNN). 10 In this analogy, as explained in the supplementary material, softening electrodes, which resemble the bias terms in a typical neural network, induce a negative electrostatic stiffness which softens the structures. Moreover, coupling electrodes, which resemble the neural network weights, electrostatically couple the elements together. The coupling electrodes also provide sophisticated multi-directional interaction mechanisms between the proof masses. For example, as shown in the red arrow in Fig.2a, the upward movement of mass 3 (M3), reduces the gap between the interacting electrodes between M2 and M3. This in turn, with respect to M2, will produce a high downward electrostatic force that pushes mass 2 down.
Activity recognition
Detecting and understanding activity recognition of standing and sitting (ARSS) is important as sitting for long hours is linked to a variety of musculoskeletal health problems. As a result, users' sitting behaviors can be improved by using continuous stance monitoring. Wearable devices that run sophisticated algorithms on inertial measurements have been proposed in the literature to provide such monitoring.11 However, wearable devices in general, suffer from the tight power budget requirements to run such techniques continuously in real-life setting scenarios. Next, we present the use of the MEMS neural network hardware to overcome this challenge.
Observing publicly available data12 reveals that the acceleration profile, along the vertical direction of a sitting activity, resembles a triangular signal with a positive then a negative (PN) profile, whereas the stand activity starts with an opposite pulse (i.e. negative then positive (NP) profile pulse). Thus, as shown in Fig. 2, one way of performing ARSS is by determining the direction of the acceleration change; if the acceleration is positive then negative, then the action is sitting. On the other hand, if the acceleration is negative then positive, then the action is standing.
Fig 2b is an illustration of the operation of the MEMS neural hardware as an acceleration classifier for ideal ARSS. In this implementation, M1 should hit its bottom stopper to indicate an NP signal (Fig.2b→c), or M3 should hit its upper stopper to denote a PN signal (Fig.2d→f). Each action is considered as a switch closing two different simple circuits representing the different types of the applied signal. To operate the device for ARSS, softening electrodes and comb-drive actuators of elements 1 and 3 have the same bias voltage as their respective masses. Therefore, exerted force from these two components in each element is equal to zero. Softening electrodes of M2 are biased, resulting in having the lowest stiffness in the network. Using this configuration, the classification steps for the NP signal are as follows: (1) as the downward (negative) acceleration part is applied to the fixed components, the suspended parts of the structure will move upward due to inertia. M2, with the lowest stiffness, will dominate the movement and pulls-in toward its upper stopper. (2) This sudden movement will also decrease the gap size in interactive electrodes 2-1 and, (3) produce a large electrostatic force pulling M1 down, but not enough to cause it to pull-in (Fig2.b). (4) Once the positive acceleration starts, M1 moves further down leading to its pull-in toward its lower stopper (Fig.2c). Note due to bistability (memory), M2 will not release when the positive acceleration is applied. Moreover, applying only positive acceleration by itself is not enough for M1 to pull in. The sequence of events will be opposite when detecting the PN acceleration signal performed by M2 and M3 as shown in Fig.2d to Fig.2f. Specifically, when a PN signal is applied, M2 experiences a downward pull-in due to positive acceleration, reducing the gap in interacting electrodes 2-3. The following negative acceleration results in an upward pull-in for M3.
The experimental testing of a microfabricated MEMS neural unit is presented in Fig.3. A schematic diagram showing the applied electrical connections is shown in Fig. 3a. Fig. 3b Shows the device at rest under an optical microscope. Fig. 3c and d show the final response of the system when a complete NP and PN acceleration are applied, respectively. As expected, the figure shows that for the NP signal, M2 is pulled in upward and M1 is pulled in downward. In this case, the pull-in of M1 indicates the detection of an NP signal. Whereas when the PN signal is applied, M2 is pulled in downward and M3 is pulled upward. The pull-in of M3 indicates the detection of a PN signal. To evaluate the ability of the neural computing unit to reject false signals, two acceleration signals were applied. A positive acceleration signal with a positive slope then a negative slope and a negative acceleration signal but with a negative slope then a positive slope as shown in Fig. 3e and Fig. 3f. These acceleration signals were generated by rotating the device either 90 or -90 degrees from the horizontal position. These false signals were selected as they have some of the characteristics that may confuse with the signals of interest such as the decrease and then increase or the increase then decrease behavior. But they are different as they don’t have the acceleration sign change from positive to negative or from negative to positive. The right parts of Fig. 3e and Fig. 3f show that under the applications of those false signals, while M2 pulled in upward or downward, none of the other devices pulled in. This indicates that the neural computing unit has correctly rejected those signals.
Signal classification
The same hardware was configured differently by changing the bias voltages to perform a different classification problem. In this new configuration, the goal is to distinguish between gradually ramping (triangle) and abruptly changing (step/square) input signals.13 Fig. 4 illustrates the operation of the described network as an electrical signal classifier. Compared to the previous acceleration classification, the signal to be classified is entered as an input voltage to comb-drive actuators of M1 and M3 and the status of M2 determines the signal class; downward pull-in of M2 signals the detection of an abrupt signal, and upward pull-in of M2 signals the detection of a ramp signal. To achieve such operation, the computing device is configured as follows: (1) Bias voltages applied to all the parallel plate softening electrodes should be set to bring all three elements close to their instability (pull-in) point. (2) M3 is biased in such a way that its potential difference with respect to M2 and consequently their interaction force is stronger compared to that of M1 and M2 (M2 is grounded and Vb3 > Vb1). (3) However, with similar input applied to the two comb-drive actuators, comb-drive 1 exerts more force on M1 compared to comb-drive 3 on M3 because of the lower voltage difference between M3 and comb-drive 3 (Vinput – Vb1 > Vinput – Vb3 ). Under such conditions, if a large enough abrupt voltage is applied to the input comb-drives, both M1 and M3 would pull in almost simultaneously (M1 pulls downward and M3 pulls upward). In this situation, however, the stronger coupling (larger interacting force) between M3 and M2 pulls M2 down towards its lower stopper indicating the detection of an abrupt increase in the input signal (Fig. 4b). On the other hand, when a gradually increasing signal is applied to the two comb-drives, at one-point M1 pulls in first since exerted force from comb-drive 1 is larger and creates greater gradual displacement compared to comb-drive 3 (Fig. 4c). Downward pull-in of M1 decreases the gap size in the interacting electrodes 2 − 1 which results in a larger attraction force that is enough to push M2 out of stability and pull it in the upward direction hitting its upper stopper. M2 (output neuron) pulling upwards indicates the detection of a ramp signal. As the input voltage keeps increasing(Fig. 4d), eventually M3 pulls in upward, however, its exerted force on M2 from their interacting electrodes will not be adequate to pull M2 out of the upward pull-in (system has memory). Figure 5 shows the validation of this operation using the MEMS neural network hardware.