The hot film flow sensor utilizes a thermal metallic thin-film deposited on a flexible substrate that is mounted on the surface of the aircraft wing. The thermal element serves as both Joule heater and temperature sensor. Under a constant bias power and zero flow rates, the thermal element achieves a steady-state temperature, which means the heat transfer system reaches equilibrium. If an external flow passes around the thermal element, the element experiences forced convective cooling. Accordingly, the temperature of the thermal element decreases, then the resistance of the element changes, and thus provides the information of the flow speed that governs the cooling rate.

The micro hot film sensors used in this work were homemade [10]. We used standard flexible PCB, i.e., polyimide (PI) film, as the substrate of the sensor array and body of the electric wiring. PI has advantages of excellent thermal isolation and flexibility. Thermal element films (Cr/Ni/Pt) were sputtered on the substrate, and a parylene film was deposited on the sensor and served as encapsulation. The fabrication process is shown in Figure 2. During the fabrication and before parylene coating, the sensors were processed through an annealing treatment (160 °C for 3 h) and electric treatment (50 mW for 4 h) for aging. The temperature coefficients of resistance (TCR) of the fabricated hot film sensors were tested to be about 2,000 ppm/K, with linear coefficients higher than 0.99, and the resistances of the sensors were about 100 ohm.

Hot film flow sensors can be operated either in constant voltage (CV), constant current (CC) and constant temperature difference (CTD) modes. In this work, the CTD mode, a scheme of which is shown in Figure 3, was used considering the superiorities of its sensitivity and dynamic response [11–13]. In the figure R_h is a thermal flow sensor, R_c is a temperature compensating sensor, R_tb is used to adjust the Joule heating level of R_h [14], R_a and R_b are the rest legs of the bridge.

In CTD mode, a feedback is employed to maintain a constant temperature difference for the thermal flow sensor related to the ambient temperature except for very high frequency fluctuations [12–14]. Feeding U back to the top of the bridge restores the flow sensor's resistance to the original value via adjusting the Joule heating. Under the CTD mode, single sensor has a monotone relationship with the flow speed. The sensors' readings (including hot film sensors 1, 2, 3, 4, hot film sensors I, II, III, IV, and the Pitot tube) versus flow speeds tested in a wind tunnel are shown in Figure 4, where the flow speeds were set from 5 to 18 m/s (about 2 m/s per step). The results of Figure 4 reveal that the Pitot tube connecting with a pressure sensor had lower sensitivity at low airspeed and becomes more sensitive when the flow speed increases, while the hot film flow sensor exhibited the opposite tendency.

Connecting the flow sensors into the CTD circuit, the time constants of the sensor systems could be optimized to be less than 10 ms by setting the ratio R_a/R_b as 10 and R_h/R_b at about 1. In the calibration measurements, the sensitivity of the hot film flow sensors were estimated to be around 4 V²/(m/s)ⁿ at the overheat ratio of 10% based on the sensor model U² = A + B·Vⁿ [10], where U is the reading of the sensor, V is the flow speed, A, B and n are the parameters determined through experimental calibration. In this study, n was tested to be about 0.2. The errors of flow measurements were less than 1 m/s. The resolution of the sensors reached 0.1 m/s, and the power consumption was tested to be about 20 mW at an overheat ratio of 10%.

For a certain aircraft, the flow field around the wing surface depends on the free stream air speed V_∞, the angle of attack α and the angle of sideslip β [8,9,14–17]. Therefore it is reasonable to detect these flight parameters by measuring the flow field around the wing surface. The flow field information around the wing contains two main components: the flow speed and the wall shear stress. Although, the both information is useful for inferring the flight parameters, we paid more attention to the flow speed in this study considering the low Prandtl number of about 0.7∼0.8 for air, which is the ratio of velocity boundary layer to thermal boundary layer. For an aircraft with a certain profile, the flow speed of the surface flow at several specific points on the wing surface can characterize the flow field around the wing, and thus can be used to deduce the aerodynamic parameters of the aircraft. The sheer stresses were detected using hot film flow sensors. Figure 5 depicts the relationship between the sensor readings and the three aerodynamic parameters.

It is generally impossible and unnecessary to measure all points on the surface, so we need to find out an appropriate sensor arrangement using as few sensors as possible to detect the aerodynamic parameters with least power and load. The work by Lian et al. [2] indicated that on the front edge of the wing, flow separation and vortexes seldom occurred, in addition, the flow field there changed violently with the change of air speed and flight aerodynamic angles, all of these implied that it was a good choice to mount the sensors on the front edge of the wing. In our work, two sensor arrangements shown in Figure 1 were employed. In the arrangement (1), hot film sensors 1, 2, 3 and 4 were mounted symmetrically on the top left, top right, bottom left and bottom right of the front edge of the wing, respectively. In the arrangement (2), hot film sensors I, II, III and IV were mounted on the top center, bottom center of the front edge, and the left side, right side of the vertical tail, respectively. Besides the hot film sensors, we also introduced a Pitot tube connecting with a pressure sensor in the system for helping to detect the airspeed.

The readings of five sensors (four hot film sensors and a Pitot tube) denoted as [V₁, V₂, V₃, V₄, V_p] were acquired, which were used for deducing the three flight parameters: air speed V_∞, angle of attack α and angle of sideslip β. Readings of the other sensor arrangement is denoted as [V_I, V_II, V_III, V_IV, V_p]. The relationship between the sensors' readings and flight parameters can be formulated by: (1) [ V 1 , V 2 , V 3 , V 4 , V P ] T = f 1 ( V ∞ , α , β ) (2) [ V I , V II , V III , V IV , V P ] T = f 2 ( V ∞ , α , β )

Intuitively, the relationship between the sensors' readings and the three flight parameters is a multiple input multiple output (MIMO) coupling system. As a matter of fact, f is generally a complex multivariate and nonlinear function, which plays a crucial role in the detection approach. An approximation of f can be obtained by various means, such as an analytical method using intrinsic fluid dynamic models, simulation, experiment-based identification or a mixture of the above. Theoretically, the analytical method is too complicated and inefficient (different airfoil results in different f); simulation is feasible in some sense, but is generally less valid; in contrast, the experimental identification is more reliable and effective. In this paper, we adopted experiment-based model identification method to develop f, by using a 3-layers BP neural network shown in Figure 6 as the model structure considering that it had been theoretically proved three layers of neural network could solve arbitrarily complicated nonlinear mapping problems [18]. For the application on small UAVs and MAVs, which requires both accuracy and simplicity, a 3-layer BP neural network with 9 neurons in the hidden layer was used, where the number of hidden neurons was determined through experimental testing. In Figure 6, normalization function f_in, denormalization function f_out, sigmoid function of the hidden layer f_hid and transfer matrix w^ih, w^ho constitute the model structure of the function f.

The experiments were performed in a wind tunnel, as shown in Figure 7. The definition of the aerodynamic parameters, airspeeds V_∞, angles of attack α and angles of sideslip β, are also shown.

In the experiment, a MAV with two sets of sensor arrangements mounted in the way described above was used, and the outputs of the sensors at different airspeeds V_∞, angles of attack α and angles of sideslip β in the ranges of 0∼28 m/s (4∼5 m/s per step), 0°∼20° (2° per step) and 0°∼20° (2° per step) respectively were measured and recorded using a data acquisition unit.

The parameters of the BP network were determined through training, in which the output readings of the sensors were used as input samples of the network while the corresponding data of the actual flight parameters were used as the output samples of the network. The training rates were set various from 0.01 to 0.001 depending on the training accuracy. The hot film sensors were operated in the CTD modes and a temperature compensation approach [14] was used.

Data samples acquired in the experiments were used to carry out the model training. After 850,000 training epochs, the sum-squared network error reached less than 0.01 for the sensor arrangement 1 and less than 0.05 for the sensor arrangement 2, as shown in Figure 8.

The performances of the trained neural networks were evaluated by comparing the actual flight parameters with the model-based flight parameters calculated from the sensors. The comparative results are shown in Figure 9, where the blue solid lines represent the calculated flight parameters by using the network model with the sensor measurements as its inputs, and the green dashed lines represent the actual flight parameters correspondingly. The detailed test processes shown in Figure 9 are: firstly, the airspeed was set from 0 to 28 m/s (4∼5 m/s per step) while the angle of attack and the angle of sideslip were zero; afterwards, the angle of sideslip was set from 0° to 20° (2° per step) at different flow airspeed varied from 5 to 28 m/s (5 m/s per step) while the angle of attack kept at zero; finally, the angle of attack was set from 0° to 20° (2° per step) at different airspeed varied from 5 to 28 m/s (5 m/s per step) while the angle of sideslip kept at zero.

From the results in Figure 9, it can be seen that the calculated flight parameters by the sensor measurements fit the actual flight parameters very well. The measurement errors of air speed, angle of attack and angle of sideslip were approximately 0.27 m/s, 0.87°, 0.27° for the sensor arrangement 1, and 0.37 m/s, 1.44°, 0.93° for the arrangement 2. The measurement errors of the arrangement 1 were smaller than that of the arrangement 2 because the flow field around the front edge of the wing was much more stable than that around the vertical tail.