- Email: [email protected]
Robustness to Sensor Damage of a Highly Redundant Gas Sensor Array
Available online at www.sciencedirect.com
ScienceDirect Procedia Engineering 87 (2014) 851 – 854
EUROSENSORS 2014, the XXVIII edition of the conference series
Robustness to sensor damage of a highly redundant gas sensor array L. Fernandez a,b, A. Gutierrez-Galvez a,b, *, S. Marco a,b a
Institut de Bioenginyeria de Catalunya, Barcelona, Spain b Universitat de Barcelona, Spain
Abstract In this paper we study the role of redundant sensory information to prevent the performance degradation of a chemical sensor array as the number of faulty sensors increases. The large amount of sensing conditions with two different types of redundancy provided by our sensor array makes possible a comprehensive experimental study. Particularly, our sensor array is composed of 8 different types of commercial MOX sensors modulated in temperature with two redundancy levels: 1) 12 replicates of each sensor type for a total of 96 sensors, and 2) measurements using 16 load resistors per sensors for a total of 1536 redundant measures per second. The system is trained to identify ethanol, acetone and butanone using a PCA-LDA model. Test set samples are corrupted by means of three different simulated types of faults. To evaluate the tolerance of the array against sensor failure, the Fisher Score is used as a figure of merit for the corrupted test set samples projected on the PCA-LDA model. © by Elsevier Ltd. by This is an open © 2014 2014 Published The Authors. Published Elsevier Ltd.access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/). Peer-review under responsibility of the scientific committee of Eurosensors 2014. Peer-review under responsibility of the scientific committee of Eurosensors 2014 Keywords: Gas ensor arrays; sensor redundancy, MOX sensors, large sensor arrays.
1. Introduction Redundant sensory information plays an important role providing robustness against sensor failure in chemical sensor systems both artificial and biological. The olfactory system is endorsed with a large number (around 10000) of olfactory receptor neurons of the same type providing a large amount of sensory redundancy. This allows the olfactory system to overcome the faulty reading of damaged olfactory receptor neurons. We can certainly take lessons from such an efficient system to build artificial chemical sensor systems. The chemical sensor array
* Corresponding author. Tel.: +34 93 4039158; fax: +34 93 4021148. E-mail address: [email protected]
1877-7058 © 2014 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/3.0/). Peer-review under responsibility of the scientific committee of Eurosensors 2014 doi:10.1016/j.proeng.2014.11.287
852
L. Fernandez et al. / Procedia Engineering 87 (2014) 851 – 854
community has studied the important role of redundancy to improve the performance of artificial chemical sensory systems [1]. However, it has not been until recently that large sensor arrays have become technologically available. Therefore, the advantages of sensor redundancy have not been properly tested with a large amount of redundant sensory units to date. In this work, we propose to address this issue with the data obtained with a large sensor array of MOX sensors. 2. Materials and methods With the aim of providing experimental proof of the need for a large number of independent sensory units to overcome the harmful effect of sensor damage, the following comprehensive study with a highly redundant gas sensor array was conducted. 2.1. Highly redundant MOX sensor array
Our sensor array is composed of 8 different types of commercial MOX sensors, modulated in temperature (100 temperature conditions) and with two redundancy levels: 1) 12 replicates of each sensor type for a total of 96 sensors, and 2) measurements using 16 load resistors per sensors for a total of 1536 redundant measures per second. The grand total of measurements per analyte is 153600.
Fig. 1. Total amount of measurements that the sensor array acquires per analyte
2.2. Redundancy testing experiments The robustness of the array is studied simulating three kinds of faults. Fault A consists in emulating a total lack of sensitivity to any odorant and concentration in a sensory element [2]. Fault B mimics the behavior of an array feature that has been poisoned so that their response becomes saturated to a fixed level. Finally, fault C corresponds to a variation in the sensory element sensitivity despite being exposed to the same odorant and concentration conditions [3]. In our experiments, we train the system to identify 3 analytes (ethanol, acetone and butanone) dosed at a 6 different concentrations (20, 40, 60, 80, 100 and 120 ppm). Each of these experiments is repeated 10 times, where the collection of experiments is randomized. After data acquisition, sensor voltage readings are converted to resistance values. Samples corresponding to odorant concentrations of 20, 60 and 100 ppm are used as a training set to create a PCA-LDA model (90 samples), while the experiments with concetrations of 40 and 80 ppm are used for testing (60 samples). Test set samples are corrupted and projected on the PCA-LDA model, and the stability of the
853
L. Fernandez et al. / Procedia Engineering 87 (2014) 851 – 854
predicition is evaluated by means of the Fisher Score. The process is repeated for an increasing number of randomly selected faulty units and the value of the Fisher Score is computed. 3. Results and Conclusions Figure 2a-2c shows the Fisher scores of the PCA-LDA model as the number of faulty sensors increases for each type of sensor damage considered (Section 2.2). Faults A and B provide a similar behavior on the Fisher score evolution starting to decrease after more than a half of the sensors are damaged. The effect of fault C is more evident in the Fisher scores results and the decrease starts before half of the sensors have been affected. To explain these results, we have plotted in Figure 2d the scatter plot of the PCA-LDA results for the worst-case scenario: all sensors fail. The results for fault C are multimodal as opposed to those of fault A and B. This explains the worst behavior of the model with fault C.
(b) 1000
1000
900
900
800 700 600 500
(c)
train test fault A 0
2 4 log10(Number of faulty sensors)
Fisher Score
Fisher Score
(a)
800 700 600 500
(d)
0 2 4 log10(Number of faulty sensors) 5
x 10 1
900
700
train test fault C
600 500
0
2 4 log10(Number of faulty sensors)
LDA 2
Fisher Score
1000
800
train test fault B
acetone
ethanol
0 -1
butanone -2
0 2 LDA 1
4 5
x 10
Fig.2.Fisher Score values for the test set samples along the number of faulty sensors considering: (a) fault type A (magenta triangles), (b) fault type B (blue triangles) and (c) fault type C (green triangles). Note as the Fisher Scores of the training (black squares) and the fault free test set (red samples) are included as reference values. (c) It shows the projections of the test set samples on the PCA-LDA model when the totality of the sensory elements fail (same color and marker notation as in the previous figures)
A further comparison has been carried out between the results of the three fault types to better comprehend differences between them. We have computed the percentage of sensors damaged that are needed to reduce the Fisher score to 90%, 50%, and 10% of the value without faulty sensors. The results are shown in Table 1. This
854
L. Fernandez et al. / Procedia Engineering 87 (2014) 851 – 854
results show that the fault C starts falling before those of faults A and B, however it degrades in a more graceful manner since we need up to 99.8 % of the sensors damaged to reach 50% of the Fisher score.
Table 1: Number of independent faulty array elements needed to achieve a reduction on the Fisher Score corresponding the 90%, 50% and 10% of its corresponding zero fault value, for each fault type.
Fisher Score Percentage
Percentage of Sensory Units Damaged Value
Fault A
Fault B
Fault C
90
794
63.4
62.6
27.2
50
441
84.1
83.8
99.8
10
90
96.6
96.5 Not Achieved
We can conclude that the high redundancy of the array endows the system with a robust behavior against sensor damage for three different types of sensory damages tested.
4. References [1] Di Natalte C, D’Amico A, Davide F, Redundancy in sensor arrays. Sensors & Actuators: A: Physical (1993), vol. 37-38, pp. 612-617. [2] Fonollosa J, Vergara A, Huerta R, Algorithmic mitigation of sensor failure: Is sensor replacement really necesary?, Sensors & Actuators: B.Chemical (2013) vol. 183, pp. 211-221. [3] Padilla M, Perera A, Montoliu I, Chaudry A, Persaud K, Marco S, Fault detection, identification, and reconstruction of faulty chemical sensors under drift conditions, using Principal Component Analysis and Multiscale-PCA, in Proc. IEE Joint Conference of Neural Networks (2010), pp. 1-7.
ScienceDirect Procedia Engineering 87 (2014) 851 – 854
EUROSENSORS 2014, the XXVIII edition of the conference series
Robustness to sensor damage of a highly redundant gas sensor array L. Fernandez a,b, A. Gutierrez-Galvez a,b, *, S. Marco a,b a
Institut de Bioenginyeria de Catalunya, Barcelona, Spain b Universitat de Barcelona, Spain
Abstract In this paper we study the role of redundant sensory information to prevent the performance degradation of a chemical sensor array as the number of faulty sensors increases. The large amount of sensing conditions with two different types of redundancy provided by our sensor array makes possible a comprehensive experimental study. Particularly, our sensor array is composed of 8 different types of commercial MOX sensors modulated in temperature with two redundancy levels: 1) 12 replicates of each sensor type for a total of 96 sensors, and 2) measurements using 16 load resistors per sensors for a total of 1536 redundant measures per second. The system is trained to identify ethanol, acetone and butanone using a PCA-LDA model. Test set samples are corrupted by means of three different simulated types of faults. To evaluate the tolerance of the array against sensor failure, the Fisher Score is used as a figure of merit for the corrupted test set samples projected on the PCA-LDA model. © by Elsevier Ltd. by This is an open © 2014 2014 Published The Authors. Published Elsevier Ltd.access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/3.0/). Peer-review under responsibility of the scientific committee of Eurosensors 2014. Peer-review under responsibility of the scientific committee of Eurosensors 2014 Keywords: Gas ensor arrays; sensor redundancy, MOX sensors, large sensor arrays.
1. Introduction Redundant sensory information plays an important role providing robustness against sensor failure in chemical sensor systems both artificial and biological. The olfactory system is endorsed with a large number (around 10000) of olfactory receptor neurons of the same type providing a large amount of sensory redundancy. This allows the olfactory system to overcome the faulty reading of damaged olfactory receptor neurons. We can certainly take lessons from such an efficient system to build artificial chemical sensor systems. The chemical sensor array
* Corresponding author. Tel.: +34 93 4039158; fax: +34 93 4021148. E-mail address: [email protected]
1877-7058 © 2014 Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/3.0/). Peer-review under responsibility of the scientific committee of Eurosensors 2014 doi:10.1016/j.proeng.2014.11.287
852
L. Fernandez et al. / Procedia Engineering 87 (2014) 851 – 854
community has studied the important role of redundancy to improve the performance of artificial chemical sensory systems [1]. However, it has not been until recently that large sensor arrays have become technologically available. Therefore, the advantages of sensor redundancy have not been properly tested with a large amount of redundant sensory units to date. In this work, we propose to address this issue with the data obtained with a large sensor array of MOX sensors. 2. Materials and methods With the aim of providing experimental proof of the need for a large number of independent sensory units to overcome the harmful effect of sensor damage, the following comprehensive study with a highly redundant gas sensor array was conducted. 2.1. Highly redundant MOX sensor array
Our sensor array is composed of 8 different types of commercial MOX sensors, modulated in temperature (100 temperature conditions) and with two redundancy levels: 1) 12 replicates of each sensor type for a total of 96 sensors, and 2) measurements using 16 load resistors per sensors for a total of 1536 redundant measures per second. The grand total of measurements per analyte is 153600.
Fig. 1. Total amount of measurements that the sensor array acquires per analyte
2.2. Redundancy testing experiments The robustness of the array is studied simulating three kinds of faults. Fault A consists in emulating a total lack of sensitivity to any odorant and concentration in a sensory element [2]. Fault B mimics the behavior of an array feature that has been poisoned so that their response becomes saturated to a fixed level. Finally, fault C corresponds to a variation in the sensory element sensitivity despite being exposed to the same odorant and concentration conditions [3]. In our experiments, we train the system to identify 3 analytes (ethanol, acetone and butanone) dosed at a 6 different concentrations (20, 40, 60, 80, 100 and 120 ppm). Each of these experiments is repeated 10 times, where the collection of experiments is randomized. After data acquisition, sensor voltage readings are converted to resistance values. Samples corresponding to odorant concentrations of 20, 60 and 100 ppm are used as a training set to create a PCA-LDA model (90 samples), while the experiments with concetrations of 40 and 80 ppm are used for testing (60 samples). Test set samples are corrupted and projected on the PCA-LDA model, and the stability of the
853
L. Fernandez et al. / Procedia Engineering 87 (2014) 851 – 854
predicition is evaluated by means of the Fisher Score. The process is repeated for an increasing number of randomly selected faulty units and the value of the Fisher Score is computed. 3. Results and Conclusions Figure 2a-2c shows the Fisher scores of the PCA-LDA model as the number of faulty sensors increases for each type of sensor damage considered (Section 2.2). Faults A and B provide a similar behavior on the Fisher score evolution starting to decrease after more than a half of the sensors are damaged. The effect of fault C is more evident in the Fisher scores results and the decrease starts before half of the sensors have been affected. To explain these results, we have plotted in Figure 2d the scatter plot of the PCA-LDA results for the worst-case scenario: all sensors fail. The results for fault C are multimodal as opposed to those of fault A and B. This explains the worst behavior of the model with fault C.
(b) 1000
1000
900
900
800 700 600 500
(c)
train test fault A 0
2 4 log10(Number of faulty sensors)
Fisher Score
Fisher Score
(a)
800 700 600 500
(d)
0 2 4 log10(Number of faulty sensors) 5
x 10 1
900
700
train test fault C
600 500
0
2 4 log10(Number of faulty sensors)
LDA 2
Fisher Score
1000
800
train test fault B
acetone
ethanol
0 -1
butanone -2
0 2 LDA 1
4 5
x 10
Fig.2.Fisher Score values for the test set samples along the number of faulty sensors considering: (a) fault type A (magenta triangles), (b) fault type B (blue triangles) and (c) fault type C (green triangles). Note as the Fisher Scores of the training (black squares) and the fault free test set (red samples) are included as reference values. (c) It shows the projections of the test set samples on the PCA-LDA model when the totality of the sensory elements fail (same color and marker notation as in the previous figures)
A further comparison has been carried out between the results of the three fault types to better comprehend differences between them. We have computed the percentage of sensors damaged that are needed to reduce the Fisher score to 90%, 50%, and 10% of the value without faulty sensors. The results are shown in Table 1. This
854
L. Fernandez et al. / Procedia Engineering 87 (2014) 851 – 854
results show that the fault C starts falling before those of faults A and B, however it degrades in a more graceful manner since we need up to 99.8 % of the sensors damaged to reach 50% of the Fisher score.
Table 1: Number of independent faulty array elements needed to achieve a reduction on the Fisher Score corresponding the 90%, 50% and 10% of its corresponding zero fault value, for each fault type.
Fisher Score Percentage
Percentage of Sensory Units Damaged Value
Fault A
Fault B
Fault C
90
794
63.4
62.6
27.2
50
441
84.1
83.8
99.8
10
90
96.6
96.5 Not Achieved
We can conclude that the high redundancy of the array endows the system with a robust behavior against sensor damage for three different types of sensory damages tested.
4. References [1] Di Natalte C, D’Amico A, Davide F, Redundancy in sensor arrays. Sensors & Actuators: A: Physical (1993), vol. 37-38, pp. 612-617. [2] Fonollosa J, Vergara A, Huerta R, Algorithmic mitigation of sensor failure: Is sensor replacement really necesary?, Sensors & Actuators: B.Chemical (2013) vol. 183, pp. 211-221. [3] Padilla M, Perera A, Montoliu I, Chaudry A, Persaud K, Marco S, Fault detection, identification, and reconstruction of faulty chemical sensors under drift conditions, using Principal Component Analysis and Multiscale-PCA, in Proc. IEE Joint Conference of Neural Networks (2010), pp. 1-7.