Research on Icing Classification Method Based on Wind Tunnel Test with Double Impact Surface Probe

Facing supercooled large water droplet environment, an effective ice detection method is a prerequisite to implement the avoidance strategy and get out of the icing environment of SLD as soon as possible. Fiber-optic icing sensors were arranged on the double impact surface probe. The probe was used for icing wind tunnel test. Different machine learning algorithms were used to establish the classification method of icing conditions based on multi-sensor ice thickness information fusion. An appropriate algorithm was selected for the classification method to detect icing conditions. The icing classification method based on SVM could effectively distinguish the conventional water droplet icing condition from the SLD icing condition, and it has significant potential on aviation industry application.


Introduction
The icing probe is an important component of the aircraft icing protection system [1][2][3].At present, ice detection facing SLD environment is a difficult point in engineering practice in the world, including the capabilities and limitations of detection equipment and engineering tools, the impact of special water droplet breakup characteristics on icing, etc. [4][5][6][7][8][9][10].Existing studies have proposed a variety of identification technologies for SLD environment [11][12][13].According to the different detection methods, SLD icing detection technology can be divided into direct detection and indirect detection [14][15][16].Among them, the probe technology used in indirect detection has significant advantages in flexibility and feasibility.Goodrich, Boeing and COMAC proposed in the patents of 2017-2019 [17][18][19], the effective detection range of the SLD environment can be expanded by continuously arranging the icing probes at multiple points on the airframe, so that the detection of the SLD icing is realized through the information fusion of multiple probes, but it is limited by the location and quantity of the probe.The requirement for installation position is high, and there is a risk of missed inspection.Rosemount Company of the United States has designed a special diversion and column configuration based on the shape of the resonant sensor [20].The conventional water drop and the SLD are respectively frozen on different probes by creating an airflow vortex or flow around the resonant probe, so as to realize SLD icing condition detection, but the special flow channel method has the problems of complex probe structure and easy icing blockage of the flow channel.Xiao et al. [21] proposed a three-cylinder column configuration.The three probes are arranged along the same axis, and the diameters of the three probes are sequentially increased, so that the requirements of water drop icing 2 detection with various diameters are met.Shi et al. [22] proposed an improved method of column configuration, in which columns with different diameters are nested together.A photoelectric sensor is introduced to realize ice detection, but the freezing coefficient of the multi-column probe method is too small.The icing conditions are not easy to distinguish.On the basis of previous scholars' research, Chen et al. [23] proposed a double impact surface probe, which has a simple structure.In recent years, scholars have studied the icing characteristics of this kind of probe.It shows that the icing characteristics under the conditions of conventional water droplets and large water droplets are obviously different.Huang et al. [24] studied the parametric design of the double impact surface probe and clarified the key geometric parameters of the probe to distinguish the SLD environment.
In this paper, based on the double impact surface probe, the optical fiber sensor was arranged to carry out the icing wind tunnel test.At the same time, the icing conditions classification method based on multi-sensor ice thickness information fusion of different machine learning algorithms was studied.And the classification method could effectively distinguish the conventional water droplet icing condition from the SLD icing condition by selecting an appropriate algorithm.It provides a theoretical basis for the detector design based on double impact surface probe.

Double Impact Surface Probe
The profile design of the double impact surface probe is shown in figure 1 [24].The optical fiber ice sensor installed on the double impact surface probe is shown in figure 2. Sensor 1 is located at the leading edge of the probe.Sensors 2 and 3 are on the first and second impact surfaces of the probe, respectively.

Icing Wind Tunnel Test
The icing wind tunnel test was carried out in the 0.3m × 0.2m guided icing wind tunnel of China Aerodynamics Research Center in Mianyang, Sichuan.The wind tunnel was designed with a maximum wind speed of 210 m/s, air flow temperature ranging from room temperature to -40 ℃, liquid water content ranging from 0.2 to 3.0 g/m3, and MVD ranging from 10 to 300 μm.The test conditions in table 1 were selected to carry out icing wind tunnel tests under four different working conditions.Condition 1 and Condition 2 were conventional water droplet icing conditions.Condition 3 and Condition 4 were SLD icing conditions.
The current wind tunnel test conditions cannot really simulate SLD icing condition.The biggest difference between SLD icing condition and conventional water droplet icing condition is that the former has water droplet diameter greater than 100 μm.The maximum MVD for conventional water droplet icing condition is 50μm.From the perspective of principle verification, this icing wind tunnel test for SLD icing condition only considered the MVD larger than 70 μm, the MVD less than 40 μm and LWC of SLD icing condition temporarily was not considered.During the test, the double impact surface probe was fixed on the platform of the test section of the icing wind tunnel through a fixture.A signal acquisition and processing system of the optical fiber icing sensor was connected with an external computer, the upper computer software recorded the output voltage signals of the three fiber ice sensors and converted the voltage signals into ice thickness information through the pre-calibration curve of each sensor.Each curve in figure 3 corresponds to the output voltage or ice thickness of a sensor under a working condition, and a sensor outputs 10 dates per second, so the number of data points of a curve under different working conditions varies with the length of icing time of the working condition.After removing the data that has not been frozen at the beginning of the experiment, the number of data groups under different working conditions is shown in table 2. Each group of data includes the ice thickness and voltage values of the sensors at three positions of the probe at the same time, totaling 6 dates.Recording the actual icing thickness at the positions where the optical fiber icing sensors were arranged on the front edge of the probe, the first impact surface and the second impact surface.The results were shown in table 3.

Research on Classification Method of Icing
From the above wind tunnel test data, it can be seen that the probe has the ability to distinguish between conventional water droplet and SLD icing because of the obvious difference of icing at different positions on the probe surface under conventional water droplet icing and SLD icing condition.However, due to the icing wind tunnel test conditions, the actual icing meteorological conditions are more diverse and complex, the traditional method by setting the threshold of the sensor output is difficult to achieve effective detection of unknown condition.Machine learning algorithms have the advantage of stronger generalization ability in classification and recognition, so this paper established icing conditions classification method based on machine learning algorithms.
Distinguishing between conventional water droplets and SLD depends on the different icing characteristics at different positions of the probe.By fusing the ice thickness information output by the icing sensors at the three positions, the icing classification method can be realized by effectively utilizing the icing characteristics of the difference.Therefore, this paper spliced the ice thickness information of three sensors to form a three-dimensional ice thickness feature vector [T1, T2, T3].
The establishment and test of the classification method for icing conditions firstly need to divide the training data and the test data.It not only need to reflect the distinction of the classification method for different icing conditions, but also need to reflect the credibility of the results.Therefore, the data of one working condition was randomly selected from the data of 1 and 2 working conditions of the conventional water drop and 3 and 4 working conditions of the SLD for two groups.The data of two working conditions were used as the training set, and the data of the other two working conditions were used as the test set.It ensured that the working conditions of the training set and the test set contain the data of conventional water droplets and SLD.
Because the decision tree method has good interpretability as a tree structure method, it is helpful to analyze the distinguishing characteristics of the probe for conventional water droplets and SLD.Therefore, this paper first used the decision tree method based on CART algorithm to establish the classification method of icing conditions.The input feature of the classification method was a three-dimensional ice thickness feature vector, and the output was a corresponding conventional water drop or SLD icing condition label.The test result statistics of the classification method were shown in table 4. It is not difficult to see from table 4 that the classification effect of the decision tree is greatly affected by the way the data is divided.In the case of training sets of conditions 2, 4 and 1, 4, the recall rate of conventional water drops is high and the accuracy rate of SLD is high.This is because the ice thickness at the leading edge (sensor 1) of condition 4 is very small, therefore, the decision tree can distinguish conventional water droplets from SLD only by judging the ice thickness of the leading edge on the training set.However, in the test set, Case 3, which belongs to the same SLD but has smaller droplets, has thicker ice thickness at the leading edge than Case 4 due to less flooding, resulting in most of the data of Case 3 being misidentified as conventional droplets.In the case of training set of condition 1 and condition 3, the precision of conventional water drop is high and the recall of SLD is high.This is because the particle size is very small in working condition 1, and no ice will be formed on the second impact surface (sensor 3).Therefore, the decision tree can distinguish conventional water droplets from SLD only by judging the ice thickness on the second impact surface in the training set.However, in the test set, it belongs to the condition 2 of conventional water droplets with larger particle size.There is a small amount of icing on the second impact surface due to the larger diameter of the droplet, causing most of the data for Condition 2 to be misidentified as SLD.
When the training set is condition 2 and condition 3, the classification result is the best.This is because condition 2 includes a special case where a conventional water droplet has a small amount of ice on the second impact surface, The condition 3 includes the special case that the SLD has thick ice on the leading edge, and it is not difficult to see from figure 4. The data of conditions 2 and 3 almost include the data of conditions 1 and 4 respectively, so they have better classification performance.
Because the decision tree has the best classification performance when the training set is conditions 2 and 3 and the test set is conditions 1 and 4, in order to find the most discriminative criterion between conventional water drop and SLD, it is necessary to analyze its specific structure more deeply.The performance of the decision tree in the face of all data sets is shown in figure 4.
As can be taken from figure 4, the decision tree first distinguishes between most conventional water droplets and SLD conditions in combination with the icing condition of the second impact surface (No.3),the remaining conditions are then separated by the icing on the leading edge (No.1).And this process shows why the decision tree has the best classification performance, because the whole decision tree combines the differences between the impact characteristics of conventional water droplets and SLD and the overflow characteristics in the process of these two distinctions.And these two differences are also the goals of the probe design.In contrast, the decision tree of the other three cases using only the difference in impact characteristics or the difference in flooding characteristics, due to the incomplete consideration in the judgment, the miscarriage of justice inevitably occurs.Although the decision tree analysis is consistent with the wind tunnel test results, that is to say, only by integrating the ice thickness information of multiple sensors can the SLD icing conditions be accurately distinguished.However, this also shows that the generalization ability of the decision tree is limited under the current icing conditions and limited data.A more robust approach to icing condition classification is required.Therefore, the SVM method with linear kernel function and the RF method were tested in the same way.That is, the input feature of the classification method was the 3D ice thickness feature vector, and the output was the corresponding icing condition label.The results were shown in table 5.As can be seen from table 5, the SVM has the best effect, while the classification effect of the RF fluctuates due to the fact that the training samples allocated to the subtree are random each time during training.Among the results of RF, the best and most stable is when the training set is conditions 2, 3 and the test set is conditions 1, 4. At this point, the highest and lowest accuracy is 100%.The reason is that the conditions 2 and 3 are at the critical point of the conventional water droplet and the SLD icing condition, and the classifier trained according to the critical condition will perform well when facing the conventional water droplet with smaller particle size and the SLD with larger particle size.It is not difficult to see from table 5 that SVM has strong generalization ability and performs well in four combinations of training set and test set.In order to analyze the reason why SVM based on linear kernel function has the best performance, select the classification results when the training set is working conditions 1 and 3 and the test set is working conditions 2 and 4 for display, as shown in figure 5.It can be seen from figure 5 that, firstly, conventional water droplets and SLD have a good property of being linearly separable in the 3D ice thickness feature space by themselves, this is due to the well differentiated design of the probe; secondly, the SVM algorithm maximizes the margin between the sample and the decision surface, which endows SVM with good generalization.
Therefore, based on the above analysis, it can be seen that in the case of icing conditions and limited data, by fuse that ice thickness information output by multiple sensors and adopt a classification method based on SVM, the conventional water droplet icing condition and the SLD icing condition can be effectively distinguished.

Conclusion
In this paper, based on the double impact surface probe, the optical fiber sensor was arranged, and the icing wind tunnel test data was carried out.Based on the machine learning algorithm, the icing classification method was studied.The main conclusions are as follows: (1) The icing wind tunnel test results of double impact surface shape probe show that the icing situation of conventional water droplets is obviously different from that of SLD.Thereby having strong classification of icing conditions.
(2) fuse that ice thickness information output by a plurality of sensors, the conventional water droplet icing condition and the SLD icing condition can be effectively distinguished.
At present, the ice classification method mainly explored based on a small amount of icing wind tunnel test data;On this basis, the next step is to combine more abundant and complete wind tunnel test data in the face of engineering problems.In view of different water droplet distribution, different flight speeds and other conditions, in-depth analysis and excavation are carried out.

Figure 1 .
Figure 1. Outline drawing of probe (Dimensions are in mm).

Figure 2 .
Figure 2. Installation position of optical fiber ice sensor on the probe.
Working conditions MVD μm LWC g/m3 Wind speed m/s Temperature ℃

Figure 3
shows the probe icing physical map under four working conditions and the curves of voltage and ice thickness at icing time of 50s.In the figure, No.1 in the figure represents the No.1 sensor installed at the leading edge of the probe, No.2 represents the No.2 sensor installed on the first impact surface, and No.3 represents the No.3 sensor installed on the second impact surfaces.

Figure 3 .
Figure 3. Physical diagram of probe icing under four working conditions and curves of voltage and ice thickness in 50s.

Figure 4 .
Figure 4.The performance of the decision tree on all data sets when the training set is under working conditions 2 and 3.

Figure 5 .
Figure 5. SVM decision surface for distinguishing between conventional water droplets and SLD.

Table 1 .
Icing Wind Tunnel Test Conditions.

Table 2 .
Number of data groups under different working conditions.

Table 3 .
Icing Wind Tunnel Test Results.

Table 4 .
Statistics of decision tree results.

Table 5 .
Statistics of SVM and RF classification results.