For Q1: The effect of other parameters can also be considered in a similar way as to the temperature. Depending on the availability of the parameter labels, one of the two proposed techniques, or a combination of the two, can be adopted. 

The focus of the paper is on experimental data from a real structure, and the tests are quite demanding. The current tests on a glider wing have only considered the temperature factor, in order to propose a general framework considering SPO robust under environmental variations. The LVV facility would allow tests at different wind speeds, so tests could be carried out, but they would require to be scheduled and resourced, and this is unfortunately not possible in the immediate future.

Some discussion has been added to the paper including in lines 42, 45 -47.

For Q2: In this paper, a support vector machine (SVM) for binary classification is introduced as a simple example to show the optimisation criterion and realise the classification tasks.  If more than two states need to be considered, a multi-class SVM or other multi-classification algorithms can be used; this does not affect the proposed framework. An appropriate criterion can be adopted as the optimisation objective according to project requirements. 

If only data sets without labels for different states are available, an unsupervised algorithm and corresponding criterion can be taken to replace the classification algorithm and the criterion used in the proposed SPO framework.

If only the data for the healthy state are available, the possible solution is to make the data from the selected sensors contain as much useful information as possible. The definition of ‘useful information’ can be determined by application requirements. Another possibility is to model the structure of interest and validate the model on healthy-state data. Based on this, sensors can be optimised to better calibrate the model. These are substantial developments that form part of our planned future work.

In real applications, the SHM process would have an ‘operational evaluation’ stage, in which the likely sites and types of damage would be assessed: e.g., by looking for ‘hot spots’ like stress concentration. Once the potential locations of the damage are identified, one can try and generate data by modelling or by using proxies.

The changes in the paper include those in lines 201-205, 271-272.

For Q3: These typos have been corrected.

Yes, it can. The aim of this study is to establish a general framework to consider a certain environmental effect in the design process of a sensor system. After the sensor system is put into use, the healthy state data should be collected at different temperatures for a period of time firstly. Then, the continuously collected data can be used for robust feature extraction using the methods mentioned in the article. The time interval for extracting features should be set according to the regulations of the project. In addition, the algorithm used for damage identification can be adjusted according to specific requirements.

Yes, it is. A set of healthy state measurements with frequency variations caused by temperatures can be used as a baseline to extract features.  In fact, it is possible to remove temperature (or other benign) effects, even without temperature measurements. There is a very large body of work on this topic – data normalisation. However, this paper looks at a more fundamental problem; if the sensor network delivers robust data, there is much less work upstream in removing temperature effects etc.