Authors: Bo Chen and Chuanzhi Zang.
Source: Expert Systems with Applications 38 (2011) 1650-1658
In this paper, authors present an unsupervised structural damage pattern recognition approach based on the fuzzy clustering and the artificial immune pattern recognition (AIPR). This method has been tested using a benchmark structural proposed by the IASC-ASCE Structural Health Monitoring Task Group. Results show the feasibility of using the hybrid AIPR method for the unsupervised structural damage pattern recognition.
1. For Fig. 3, there are “Meet injection criterion?” and “Meet replacement criterion?” two steps. What is the point of using these two steps?
Answer: The memory cell set needs to have a high diversity. If the candidate memory cell is very different with existing memory cells (measured by rms), then it is “injected” into the memory cell. Otherwise, the candidate cell is compared with existing memory cells. If the candidate cell is “better” (has higher affinity value than the matched memory cell) and it is really close with the matched memory cell, then it replaces the matched memory cell (the total number of memory cells doesn’t change). Or if it is better than the matched memory cell but they are not that close, the candidate memory cell is added to the memory cell set (we have one more memory cell now). To sum up, if the candidate memory cell is quite “different”, or has better “performance”, it will be added into the memory cell set. If there’s an existing memory cell similar with it, this memory cell will get deleted.
2. In Section 2.4.1 “For a training antigen ag, the affinity between the antigen and each antibody ab that is in the same pattern as the antigen is calculated.” How can we know the antigen’s label in advance? Isn’t this an unsupervised learning algorithm?
Answer: This is an unsupervised learning algorithm. A fuzzy clustering algorithm is used to classify all feature vectors into several patterns (using representative point). The antigen’s label is generated by the classification. But we still don’t know the “real label” of this antigen.
3. In Section 3 “In the experimental study, a total of 15 accelerometers were used …, three accelerometers for each level.” This means we have 5 levels for the structural. But the benchmark structural only has 4 levels.
4. In Section 3 “To test the memory cells generated by the immune learning process, the previously created 580 feature vectors were reused with pattern labels. These feature vectors were classified by the memory cells to five clusters.” What is the criterion used for this classification? Is it the nearest neighbor criterion? Is it the Euclidian distance? If yes, is there a threshold? For example, if the unknown antigen’s affinity values with every memory cell are all lower than the threshold, this unknown antigen will not get classified.
5. In Section 4 “The pattern recognition success rate rises rapidly when the MCRT value is greater than 0.85. The reason is that more candidate memory cells are injected into the memory cell set.” A higher MCRT value means less replacement. How can this leads to a higher success rate?
Answer: As explained in Question 1, if we keep other criteria the same, and only change the MCRT to a higher value, less replacement will happen. But the candidate memory cells which do not replace existing memory cells are not abandoned, instead they are injected into the memory cell set. Therefore we actually have more memory cells when the MCRT value is higher.