SBIR-STTR Award

An artificial neural network approach is proposed to characterize frequency response functions (FRFs) obtained in structural dynamics testing of large systems including launch vehicles. A modal test of large structures often result in the measurement of thousands of FRFs. Currently, FRF measurement problems caused by accelerometer debonding, signal dropouts, etc. often must be detected by experienced analysts manually reviewing reams of test data. The neural network approach provides an automated mechanism well suited to “screening” FRF measurement data for patterns indicative of problems thus significantly reducing the labor in the review process while increasing the accuracy and reliability of the data. The technical objective of this Phase I activity is to establish the feasibility of this neural network approach and to determine what other FRF characteristics can be identified using this technique. This activity will develop and configure the neural network architecture including the number of inputs required, how many hidden layers of neurons are needed, and number of outputs requested. The FRF training data will be assembled and employed to support the neural network “learning” process. The developed architecture will be applied to several test cases to quantify system accuracy as a measure of innovation feasibility.Commercial Applications:It is currently envisioned that a successful neural network based FRF char- acterization tool can be fashioned (and marketed) in such a way as to complement existing software tools in the experimental modal analysis market. While the current crop of modal analysis tools are quite effective for linear systems, it is the objective of this research to develop a user friendly tool that provides insight into systems that possess certain nonlinear characteristics. Additionally, it is anticipated that the FRF screen for sensor debonds, dropouts, etc. can significantly reduce the amount of effort required in determining if and when a sensor problem does arise. If this innovation proves reliable, the implementation flexibility of a neural network based FRF characterization tool provides a wide range of marketing alternatives. For example, current modal analysis software providers could be offered a software emulation of the network which would be exercised within their systems. Alternatively, by using dedicated neural computing boards, a hardware implementation of the network could be bundled with current signal analyzers for a relatively low price. This freedom of implementation (in software or hardware) will allow bd Systems the capability of pursuing both hardware and software portions of the experimental modal analysis

An artificial neural network approach is proposed to characterize frequency response functions (FRFs) obtained in structural dynamics testing of large systems including launch vehicles. A modal test of large structures often result in the measurement of thousands of FRFs. Currently, FRF measurement problems caused by accelerometer debonding, signal dropouts, etc. often must be detected by experienced analysts manually reviewing reams of test data. The neural network approach provides an automated mechanism well suited to ÒscreeningÓ FRF measurement data for patterns indicative of problems thus significantly reducing the labor in the review process while increasing the accuracy and reliability of the data. The technical objective of this Phase I activity is to establish the feasibility of this neural network approach and to determine what other FRF characteristics can be identified using this technique. This activity will develop and configure the neural network architecture including the number of inputs required, how many hidden layers of neurons are needed, and number of outputs requested. The FRF training data will be assembled and employed to support the neural network ÒlearningÓ process. The developed architecture will be applied to several test cases to quantify system accuracy as a measure of innovation feasibility.Commercial Applications:It is currently envisioned that a successful neural network based FRF char- acterization tool can be fashioned (and marketed) in such a way as to complement existing software tools in the experimental modal analysis market. While the current crop of modal analysis tools are quite effective for linear systems, it is the objective of this research to develop a user friendly tool that provides insight into systems that possess certain nonlinear characteristics.Additionally, it is anticipated that the FRF screen for sensor debonds, dropouts, etc. can significantly reduce the amount of effort required in determining if and when a sensor problem does arise. If this innovation proves reliable, the implementation flexibility of a neural network based FRF characterization tool provides a wide range of marketing alternatives. For example, current modal analysis software providers could be offered a software emulation of the network which would be exercised within their systems. Alternatively, by using dedicated neural computing boards, a hardware implementation of the network could be bundled with current signal analyzers for a relatively low price. This freedom of implementation (in software or hardware) will allow bd Systems the capability of pursuing both hardware and software portions of the experimental modal analysis market.Note: no official Abstract listing exists of selected NASA Phase II SBIR projects for this year. Hence, this abstract is modified by idi from relevant Phase I data. The specific Phase II work statement and objectives may diff