Artificial-immune-system approach enables monitoring structural health

Wireless sensor networks are emerging as useful tools for structural-health monitoring (SHM). While such networks offer a number of advantages, SHM systems face many challenges. Major issues include provision of sustainable, long-term, and autonomous monitoring, active detection and identification of structural damage, and development of adaptive approaches that can adjust to changing monitoring conditions.

We are employing an artificial-immune-system (AIS) approach to developing sensor networks¹ for self-diagnostic and adaptive SHM settings. Our goal is to introduce desirable attributes into these systems, such as adaptation, learning capability, and memory mechanisms. The natural immune system protects living organisms from invading antigens through collaboration of B and T immune cells (see Figure 1). B and T cells are the major types of lymphocytes that mediate the adaptive immune response. They are responsible for recognizing and eliminating pathogenic agents. The surface receptor (antibody) of an immune cell can recognize and bind antigens. When a B cell's receptor recognizes a nonself antigen and receives a stimulation signal from a helper T cell, the former is activated. It proliferates and differentiates into antibody-secreting (plasma) and memory cells. An antibody binding to an antigen will mark it for destruction. The virally infected cells are eventually eliminated by killer T cells.

Figure 1. Lymphocytes (such as B and T cells) recognize and eliminate antigens in the natural immune system. APC: Antigen-presenting cell. TH, TK: Helper, killer T cell.

The AIS approach is well suited to address SHM problems. First, AIS-based SHM systems can automatically manage structure-monitoring tasks by generating and distributing monitoring agents (mimicking B cells). Second, they actively distribute specialized agents to sites where they are needed. Finally, the natural immune system's ability to adapt to foreign attacks using clonal selection has great value and applicability in SHM sensor networks. Selective generation of mobile monitoring agents is essential for producing enough specialized agents in resource-constrained sensor networks.

We have designed an AIS-based SHM sensor-network framework (see Figure 2). A group of autonomous mobile agents monitors a structure's health by patrolling a sensor network. Each agent's ‘antibody’ is an algorithm that recognizes a certain type of structural-damage pattern. A mobile-agent system, Mobile-C,^2–4 supports the generation, migration, and management of the agents. When an agent's antibody recognizes a damage pattern at a sensor node, the agent contacts the knowledge base. If the damage is confirmed, the agent is activated. Its antibody will be cloned and mutated through a clonal-selection process. An alert agent is subsequently generated to send a warning message to remote (human) operators. At the same time, mobile agents carrying cloned antibodies migrate to locations close to the damage site, where they conduct careful damage diagnosis and generate a report for human assessment.

The system's knowledge base performs the role of helper T cells. Since the antibody of the activated mobile agent does not need to fully match the damage pattern, the mutated antibodies may have higher affinities with the pattern. The antibody with the highest affinity is used to update the memory-cell set corresponding to the damage pattern.

Figure 2. Artificial-immune-system-based SHM sensor-network architecture.

The structural-damage patterns are represented by feature vectors extracted from the structure's dynamic-response time series.⁵ Vectors are obtained on the basis of coefficients associated with an auto-regressive (AR) model of the time series. We have tested the effectiveness of the feature vectors using experimental data from a benchmark structure proposed by the American Society of Civil Engineers. The acceleration signals of the normal pattern, as well as the four benchmark-structure damage patterns, are shown in Figure 3, while the feature vectors of the corresponding data patterns projected onto the first two principal components are included in Figure 4.

Figure 3. Acceleration signals of the American Society of Civil Engineers (ASCE) benchmark structure.

Figure 4. Feature vectors of the experimental data of the ASCE structure representing normal and damage patterns.

We employ an artificial-immune-pattern recognition method for structural-damage classification.^5,6 Damage-pattern recognition is achieved by establishing memory-cell sets, each of which is responsible for recognizing one type of damage pattern. These sets evolve into immune learning processes that are stimulated by the antigens of the damage patterns. Continuous refinement of memory cells allows the sets to adapt to changes in monitoring conditions. The memory cells for the benchmark structure's five data patterns are shown in Figure 5. To classify a time series for an unknown damage pattern, we calculate the affinities between the time series' feature vectors and the memory cells. The unknown pattern is classified as belonging to that associated with the memory cell with which the time series has the highest affinity.

Figure 5. Memory cells for the five data patterns of the ASCE structure.

In summary, adaptive SHM is critical for a quick response to operational and environmental changes to structures. Our AIS-based system provides adaptive damage-pattern recognition through memory-cell evolution. Active SHM is achieved by distributing mobile monitoring agents to sites where they are needed, and monitoring is managed automatically by the mobile-agent system. Our future work will explore self-healing sensor networks based on immune models.