CN104111887A - Software fault prediction system and method based on Logistic model - Google Patents

Abstract

The invention discloses a software fault prediction system and method based on a Logistic model. The system comprises a test cycle fault management module, a Logistic model measurement module and a quality analysis and evaluation module. The method is as follows: the test cycle fault management module is adopted to divide a system testing process according to certain cycles, and the Logistic model measurement module is adopted to measure the number of faults found in each cycle; the Logistic model is utilized to work out the sum of faults contained in software and the number of faults found in the next cycle; the quality analysis and evaluation module is adopted and used for judging whether the software meets the issue requirement with the specific value of the number of faults found in the software to the sum of faults contained in the software as credibility of the software; the sum of faults in the software is calculated through the three steps, whether the software needs to continue to be tested is further judged, and the judgment result is used for quality evaluation of the next cycle.

Description

Software fault prediction system and method based on Logistic model

Technical field

The present invention relates to a kind of failure prediction technology of software, say that more specifically the fault of utilizing Logistic model to contain software is total and predict in the number of faults of next test period discovery, and then assessing the technology whether final software product has reached the quality requirements of regulation.

Background technology

Software can constantly be found fault in system testing process, and the solution of fault may be introduced new problem again, if fully just releasing not of a version test may cause the quality of version too low, to user, causes damage; If the time to a version test is too abundant, will cause so-called excessive test, cause the waste of test resource.How is the quality of software in test process evaluated so? when can issue meet the requirement to quality? the solution of this problem just depends on the failure prediction technology of software.

Software fault prediction technology is a kind of technology that still undetected failure is estimated in to software before software issue, can help tester before software issue, to adjust in time Test Strategy, find more multiple faults, and then reduce software maintenance cost, effectively improve the quality of software.At present more influential failure prediction method has failure prediction model based on bayes method and the failure prediction model based on random forests algorithm.

Forecast model based on bayes method comprises Bayesian network and Naive Bayes Classifier, and the basis of Naive Bayes Classifier is assumed to be: between given desired value, attribute, condition of reciprocity independence and each example x are represented by the proper vector of property value.Objective function carries out value according to possible the value in existing set V, according to a series of obtain the training sample set of objective function and newly the proper vector of example predict the desired value of new example.Menzies carries out classification performance contrast according to the forecast model of NB Algorithm and decision Tree algorithms, and the model that the model of taking experimental result after logarithm process to show that NB Algorithm builds to data builds than decision Tree algorithms has better classification performance.The very capable of the uncertain processing problem of Bayesian network can carry out fusion and the expression of multiple information greatly efficiently, has been widely used in software fault prediction.The people such as Fenton propose software fault prediction model suitable in different Life Models and leave over failure prediction model based on Bayesian network, and Bayesian network is analyzed in the effect of life cycle early prediction fault.

Calendar year 2001 LEO Breiman has proposed random forest (Random Forest, RF) algorithm, in forest, the training sample of decision tree carries out stochastic sampling generation to original training sample, in random forest, the segmentation candidates property set of decision tree internal node is a nonvoid proper subset of all properties, this subset is from all characteristic attributes, to choose at random the attribute of some as candidate attribute collection, use attribute division measure function concentrates the classification capacity of attribute to differentiate to candidate attribute, finally chooses optimum Split Attribute.The failure prediction of random forests algorithm is carried out metric software project by the value of a plurality of property sets of software for calculation project, training dataset is carried out to equilibrating processing, during structure Random Forest model, according to the predictablity rate of model and recall ratio, screen the forecast model that meets performance index.

The ultimate principle of above-mentioned software fault prediction is if the current module of developing has similar software quality attribute to certain module of having developed before, illustrates that current software module has similar failure prone to the module of having tested before.Therefore before will using in software fault prediction, the failure logging of the software module of exploitation is predicted current module.This failure prediction method has two problems: one, the similarity of software module is not easy tolerance, if can not find similar module, cannot carry out failure prediction; Two, due to developer's technical progress, similar module may show different fault trends, causes prediction to be lost efficacy.

Summary of the invention

In view of this, the technical matters that the present invention solves is to provide a kind of software fault prediction system and method based on Logistic model, the method does not rely on other software module, system testing process was divided according to certain cycle, measure out the number of faults of finding in each cycle, then as input, utilize the number of faults that fault is total and next test period is found containing in Logistic software, the ratio of the fault sum that the number of faults having been found that with software and software contain is as the confidence level of software, if software does not meet the requirements of confidence level, needing to continue test goes down, if software has reached the confidence level requiring,, without continuing test, can issue.

Software fault prediction system based on Logistic model, comprises three modules: test period fault management module, Logistic model metrics module, quality analysis and evaluation module;

Test period fault management module: divide test period, and add up the number of faults of finding in each test period, this module is responsible for the input of confidence level expected results simultaneously;

Logistic model metrics module: be responsible for the iterative algorithm of Logistic model, calculate the fault sum containing in software and estimate with next test period the number of faults of finding;

Quality analysis and evaluation module: be responsible for the calculating of software reliability, and according to result of calculation, whether software can be issued and be judged; The number of faults of finding in number of faults by next test period of calculating in Logistic model relatively and actual test, evaluates output evaluation result to the quality of a test period.

The Forecasting Methodology of the software fault prediction system based on Logistic model comprises the steps:

A, employing test period fault management module are divided described system testing process according to certain cycle, and tolerance

Go out the number of faults of finding in each cycle;

The number of faults that B, the fault sum that utilizes Logistic model to calculate to contain in software and next test period are found;

The ratio of the fault sum that C, employing quality analysis and evaluation module contain with the number of faults having been found that in software and software is as the confidence level of software, for judging whether software has reached issue requirement;

By above-mentioned three steps, calculate the fault sum of software, and then whether software is needed to proceed test judge, and for the quality assessment in next cycle.

Described step B further comprises:

B1, the number of faults input Logistic model in A step;

B2, utilize iterative algorithm to calculate the maximal value of Logistic model, as the estimation of the fault sum containing in software;

B3, calculate the number of faults that next test period finds.

Described step C further comprises:

The number of faults that C1, use have been found that calculates the confidence level of software divided by fault sum;

C2, according to the situation of reaching of software reliability, judge software whether need to proceed test;

C3, when next test period finishes, by relatively finding predicted value and the actual value of fault, the quality condition of this test period is evaluated.

Beneficial effect

Whether the quality assessment of software test procedure and version meet issue requirement is a difficult problem of industry always, and the present invention is by having shown that to the analysis of software fault discovery procedure Logistic model is in the value aspect failure prediction.Comprise two aspects: one,, by the forecast analysis to test result, after each end cycle, can carry out quality assessment; Two, the confidence level of software is measured, while issuing for version, the quality of whole version issue is examined.The value of these two aspects has completed the evaluation to the supervision of test process and test result.

Accompanying drawing explanation

Fig. 1 is the auxiliary failure prediction system FPBL the general frame of the present invention.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein, only in order to explain the present invention, is not intended to limit the present invention.

Software fault prediction technology based on Logistic model of the present invention comprises three steps: the first step, system testing process was divided according to certain cycle, and measure out the number of faults of finding in the test period of having carried out; Second step, the number of faults that the fault of utilizing Logistic model to calculate to contain in software sum and next test period are found; The 3rd step, the quality analysis of software and evaluation, the ratio of the fault sum containing with the number of faults having been found that in software and software is as the confidence level of software, for judging whether software has reached issue requirement; When next test period finishes, by relatively finding predicted value and the actual value of fault, the quality condition of this test period is evaluated.

With reference to figure 1, be the general frame of the auxiliary failure prediction system FPBL of the present invention, this system comprises three modules: test period fault management module, Logistic model metrics module, quality analysis and evaluation module.

Test period fault management module: be responsible for to divide test period, and add up the number of faults of finding in each test period, this module is responsible for the input of confidence level expected results simultaneously;

Logistic model metrics module: nucleus module, be responsible for the iterative algorithm of Logistic model, calculate fault sum and the next test period in software, contained and estimate the number of faults of finding;

Quality analysis and evaluation module: be responsible for the calculating of software reliability, and according to result of calculation, whether software can be issued and be judged; The number of faults of finding in number of faults by next test period of calculating in Logistic model relatively and actual test, evaluates output evaluation result to the quality of a test period.

Below in conjunction with auxiliary failure prediction system FPBL explanation specific embodiment of the invention process.

For realizing the described first step, namely complete the test period fault tolerance in FPBL, we are test period according to being 5 working days, calculate the number of faults in each cycle after tested, in order to guarantee the accuracy of subsequent estimation, our suggestion has the data in three cycles at least, just can predict.Number of faults in each cycle is measured according to the weighted value of the fault order of severity, fault is divided into according to the order of severity: serious, important, general, slight, its corresponding weighted value is respectively 15,9,3,1, is below the fault metric data of certain large software after having carried out seven period measurings:

Week1	Week2	Week3	Week4	Week5	Week6	Week7
							962	1596	1955	2344	2450	2549	2615

User is that confidence level reaches 95% to the quality requirements of this software, and this value can be inputted by interface, also test period fault metric module manage.

For realizing second step, namely the Logistic model metrics in auxiliary failure prediction system FPBL, need to carry out some conversion to this model, so that carry out iterative computation.

N-S graph, is commonly called as " S curve ", by Verhulst, in 1845, is proposed, and fundamental purpose is at that time the growth of simulation population, and its general form is:

Wherein through simplification progressively, the most frequently used form was at present afterwards:

$y=\frac{1}{a+{\mathrm{be}}^{-x}}$

Logistic Growth Curve Model also referred to as narrow sense.

The precondition that Logistic Growth Curve Model is set up: 1. when population is less, what natural resources was enough enriches, and the growth of population does not suffer restraints, and therefore supposes that the natural growth of population is constant; 2. when population is increased to some, the restriction that the Factors on Human mouths such as natural resources, environmental baseline increase is more and more significant, therefore supposes that the natural growth of population can reduce along with the increase of population.

Software test procedure and Logistic model carry out analogy: (1) by objective factors such as the natural resources in the size of code analogy population increase model changing in software, environmental baselines, the size of code of variation comprises two aspects: (a) the code difference of current version and its baseline version; (b) the code difference that the modification of the Bug finding causes.These codes that changed are objective factors that fault produces, and are equivalent to an outside environment.The starting stage of test, due to the variable quantity of code very large (being equivalent to the state that resource is very abundant), the rate of growth of fault is unfettered, can be assumed to a constant; (2) the current size of population of number of faults analogy of finding, this quantity changes along with the variation of size of code, is subject to again its restriction simultaneously.When test proceeds to certain stage, the variation of version is more and more less, and the variable quantity of code is more and more less, finds that the speed of fault will suffer restraints this time, is downward trend.More than analyze and can find out that the fault that software test procedure is found meets the precondition of Logistic model, so can measure with Logistic model.

The basic thought of algorithm is a default N _m, try to achieve the optimal estimation of r, then, using r as known, obtain N _moptimal estimation, like this alternate cycles iteration until convergence.

Note: so have:

Because there being model error, should replace with the following equation with error

Thereby at N _munder known conditions, can obtain parameter r least-squares estimation:

With replacing above formula by following formula:

$N_t=N_m/(1+(\frac{N_m}{N_0}-2)e^{-\mathrm{rt}})+{\mathrm{\ϵ}}_t$

σ in formula _tfor independence, average is 0, waits the stochastic variable of equation, note

$y_t\≡\frac{(e^{\mathrm{rt}}-1)N_0{N}_t}{N_0{e}^{\mathrm{rt}}-N_t},{\mathrm{\σ}}_t\≡\frac{N_0{e}^{\mathrm{rt}}+N_m-N_0}{N_0{e}^{\mathrm{rt}}-N_t}$

Y _t=N _m+ σ _tε _t, thus can moral limits to growth N under r known conditions _mmaximum likelihood estimate:

$\widehat{N_m}=\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\frac{1}{{{\mathrm{\σ}}_i}^2}{y}_i\right)/\left(\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\frac{1}{{{\mathrm{\σ}}_i}^2}\right)$

Here σ _iin contain and estimated parameter N _m, in iterative solution method, we can be with last iterative value N _mreplace it.

Optimal estimation N _m, the specific algorithm step of r is as follows:

1. get initial value N _m, r precision is ε, N _msubstitution:

try to achieve r ⁽⁰⁾

2. make k:=1, a:=N _m ⁽⁰⁾, b:=r ⁽⁰⁾.B substitution wherein ${\mathrm{\σ}}_t\≡\frac{N_0{e}^{\mathrm{rt}}+N_m-N_0}{N_0{e}^{\mathrm{rt}}-N_t},$ Try to achieve N _m ^(k), N _m ^(k)substitution try to achieve r ^(k).

If | N _m ^(k)-a|+|r ^(k)-b|≤ε, stops, and now has N _m ^*=N _m ^(k), r ^*=r ^(k); Otherwise turn " 3 ".

3.a:=N _m ^(k), b:=r ^(k)k:=k+1, turns " 2 ".

By above algorithm, this step has been exported the number of faults N of software _mnumber of faults N with next test period _t+1.

The 3rd step, according to N _mand N _t+1carry out quality analysis and the evaluation of software.

Suppose the test of having carried out I (I>=3) the individual cycle, will predict and quality analysis I+1 cycle now.The prediction of fault odd number is since the 4th cycle, and we at least will have the data in three cycles, and while being less than three data, the deviation that predicts the outcome is too large, cannot use.By the I a having counted value, carry out matching, according to the method for second step, obtain the parameters value of Logistic computing formula, using the formula calculated value of t=I+1 as the prediction minimum value min in I+1 cycle, using the prediction maximal value max of formula maximal value as I+1 cycle, the fault sum of finding to I+1 cycle so should drop in interval [min, max].The step of quality analysis and evaluation is as follows:

1. the quality analysis of I+1 period measuring after finishing.After I+1 period measuring finishes, suppose that the fault of discovery adds up to Bugs, formulate such rule: (1) if Bugs>=max*0.98, I+1 period measuring quality is excellent; (2) if min<=Bugs<0.98*max, I+1 period measuring quality is good; (3) if Bugs<min, I+1 period measuring quality is defective.

2. the confidence level in I (I>=3) individual cycle is calculated.With the fault sum of I actual discovery of cycle of cut-off divided by the predicted value of fault sum in software the confidence level as the software in I cycle, i.e. Ri=N (i)/N _m.The flow process whether software meets issue requirement is as follows:

A) given confidence level R ', the value of R ' is according to the requirement of software is determined;

B) carry out the test of I wheel, and calculate the confidence level R that this takes turns test software afterwards _i;

C) if R _i>=R ', stops test, and software can be issued; If R _i<R ', I=I+1, turns (b)

By above-mentioned three steps, can complete quality assessment and the confidence evaluation of software test procedure, the present invention is by having shown that to the analysis of software fault discovery procedure Logistic model is in the application aspect failure prediction.Comprise two aspects: one,, by the forecast analysis to test result, after each end cycle, can carry out quality examination; Two, whether confidence level meets the judgement of issue condition for version, has stopped the wasting of resources that excessive test causes simultaneously.These two aspects are actually the evaluation to the supervision of test process and test result.

To sum up, the present invention has calculated maximal value and the rate of change of Logistic model by iterative algorithm, and then the number of faults of the fault that software is contained sum and next test period discovery is predicted, these two predicted values can ensure the test mass of software, and effectively reduce software test cost.

It is above-mentioned that only with preferred embodiment, the present invention will be described, non-so limit to the bright interest field of Ben Fa-7-, utilize the fault sum that Logistic model contains software to predict it is core concept of the present invention with the number of faults of next test period discovery, therefore, in the situation that not departing from inventive concept, the equivalence that all utilizations instructions of the present invention and accompanying drawing content are done changes, and all reason is with being contained within the scope of claim of the present invention.

Claims (4)

1. the software fault prediction system based on Logistic model, is characterized in that described system comprises three modules: test period fault management module, Logistic model metrics module, quality analysis and evaluation module;

Test period fault management module: divide test period, and add up the number of faults of finding in each test period, this module is responsible for the input of confidence level expected results simultaneously;

Logistic model metrics module: be responsible for the iterative algorithm of Logistic model, calculate the fault sum containing in software and estimate with next test period the number of faults of finding;

Quality analysis and evaluation module: be responsible for the calculating of software reliability, and according to result of calculation, whether software can be issued and be judged; The number of faults of finding in number of faults by next test period of calculating in Logistic model relatively and actual test, evaluates output evaluation result to the quality of a test period.

2. a Forecasting Methodology for the software fault prediction system based on Logistic model as claimed in claim 1, is characterized in that described method comprises the steps:

A, employing test period fault management module are divided described system testing process according to certain cycle, and tolerance

Go out the number of faults of finding in each cycle;

The number of faults that B, the fault sum that utilizes Logistic model to calculate to contain in software and next test period are found;

The ratio of the fault sum that C, employing quality analysis and evaluation module contain with the number of faults having been found that in software and software is as the confidence level of software, for judging whether software has reached issue requirement;

By above-mentioned three steps, calculate the fault sum of software, and then whether software is needed to proceed test judge, and for the quality assessment in next cycle.

3. the Forecasting Methodology based on the software fault prediction system based on Logistic model claimed in claim 2, is characterized in that described step B further comprises:

B1, the number of faults input Logistic model in A step;

B2, utilize iterative algorithm to calculate the maximal value of Logistic model, as the estimation of the fault sum containing in software;

B3, calculate the number of faults that next test period finds.

4. the Forecasting Methodology based on the software fault prediction system based on Logistic model claimed in claim 2, is characterized in that described step C further comprises:

The number of faults that C1, use have been found that calculates the confidence level of software divided by fault sum;

C2, according to the situation of reaching of software reliability, judge software whether need to proceed test;

C3, when next test period finishes, by relatively finding predicted value and the actual value of fault, the quality condition of this test period is evaluated.