Reliability Analysis of Real-Time Multitasking Software Based on Neural Network - Advanced Research on Computer Education, Simulation and Modeling - page 350

execution time and t i is period of task. The parameters C i and t i are determined by
designers before development, which are known before test.
Neural network model to evaluate reliability of real-time multitasking software can
be established based on its characteristics. Improvements should be conducted on
above mentioned neural network model. Firstly it needs to substitute B s , T s and N s
with code number B i , test time T i and failure number N i of each task module. Sec-
ondly add P i as input of neural network. RBFNN of real-time multitasking software is
shown in Fig. 2, which presents reliability model of a task module. P i , B i , T i , N i and T
are inputs of network and module failure rate

Φ

is the output. As modules of soft-
ware are developed in same environment, we can train neural network based on test
data in software module test report, and then to analyze reliability of other modules
with trained neural network. The failure rate of whole system is the sum of all mod-
ules' failure rate.

i

P i

B i

Output failure rate Φ

i

∑

T i

…

N i

T

Fig. 2. Neural network model of real-time multitasking software reliability

4 Case Study

Test report of some real-time multitasking software is shown in Table 1 [4]. The
software was divided into 14 task modules. DS model method [5] was used to com-
pute failure rate of each module after 10 days operation. DS model algorithm is cur-
rently accepted method to study reliability of real-time multitasking software, the
computation result of which has considerable accuracy. As module number of the
software is relatively little and there is little data for training, so it have relatively poor
network computing accuracy. To address this problem and expand training data, neu-
ral network simulation method was used to produce data with same statistical princi-
ple based on original data to expand sample number [6, 7]. The former 12 group of
reliability data was expanded. Then expanded data was used to train neural network
model of general software and real-time multitasking software. Two trained network
was then sued to perform simulation computation on remaining two group data, the
results of which are shown in Table 2. We can know from simulation results that the
proposed model reduced error greatly and improve accuracy significantly compared
with that of general software neural network model.