CN113568831B

CN113568831B - Self-supervision deep learning type defect positioning method based on metamorphic test

Info

Publication number: CN113568831B
Application number: CN202110852911.7A
Authority: CN
Inventors: 胡安林; 付凌峰; 雷晏
Original assignee: Chongqing University
Current assignee: Chongqing University
Priority date: 2021-07-27
Filing date: 2021-07-27
Publication date: 2023-07-04
Anticipated expiration: 2041-07-27
Also published as: CN113568831A

Abstract

The invention discloses a self-supervision deep learning type defect positioning method based on metamorphic test, which comprises the steps of firstly constructing metamorphic relation, generating source test cases and follow-up test cases on the basis of a given test case set without test predictors, then collecting coverage information and test output, carrying out aggregation and merging on the former, carrying out metamorphic relation verification on the latter to obtain a test result of total coverage information in each metamorphic relation group, integrating the test result into a coverage matrix and a violation vector, inputting the coverage matrix and the violation vector into a deep neural network model for training, learning the model, finally finishing prediction of suspicious values of codes by the model, and sequencing codes of a defect program according to the suspicious values to finish defect positioning. The method overcomes the limitation of deep learning type defect positioning, can be applied to the situation of test prediction missing, realizes the defect positioning oriented to the label-free data set in a self-supervision mode, and remarkably expands the application range.

Description

Self-supervision deep learning type defect positioning method based on metamorphic test

Technical Field

The invention relates to the field of defect positioning, in particular to self-supervision deep learning type defect positioning for a label-free data set.

Background

Software, a pure abstract knowledge product with a high degree of inherent complexity, is unavoidably error-free. In spite of the large amount of manpower and material resources put into practice in an effort to reduce the vulnerability of the software, the BUG is inevitably present in all kinds of software having a large scale. Thus, debugging also becomes a critical part throughout the software development and maintenance process, and does require a lot of labor and considerable time. To alleviate this burden, the academia has devised a number of techniques for automatic defect localization to help developers find BUGs more efficiently. In recent researches in this field, deep-Learning-based Fault Localization (DLFL) has gained much attention by virtue of its excellent localization effect. The DLFL takes the popular supervised deep learning model as the kernel, can effectively identify and digest patterns hidden in program execution information, and further utilizes the learned patterns to evaluate suspicious values of each line of sentences, which represents the responsibility of each line of sentences to program failure. Therefore, DLFL not only gives better effect, but also has stronger generalization capability than some other leading edge defect localization methods (e.g., spectrum-based defect localization).

Fig. 1 shows the basic flow of DLFL. Given a target program (i.e., defect-located object, defect-containing program) P containing N lines of code, a program containing M test cases t ₁ ,t ₂ ,…,t _M ]Test case set of (a) and test prophetic [ o ] that must be prepared before DLFL begins ₁ ,o ₂ ,…,o _M ]Wherein o _i Store P at t _i The expected output of the upper run. In the batch execution (batch execution), all test cases are input one by one to the target program for execution. Test case t _i Is taken to and corresponds to the execution output of test prophetic o _i Comparing to obtain the execution result y _i To indicate whether the execution of the program has failed. At the same time, the runtime information concerning which statements (the term "statement" is used to represent a line of code in the present invention) have been executed is collected and summarized by the relevant tool to obtain x _i Leaving for subsequent investigation. Next, DLFL uses matching pairs of execution results (i.e., pass or fail) and runtime information (i.e., statement overlays), (x _i ,y _i ) I e {1,2, …, M }, taking the former as a label and the latter as a sample, internalizing the relationship between the sample and the label by means of a deep learning network, training the model. Still further, DLFL evaluates the suspicious value of each line statement (i.e., the likelihood that the line code contains BUG) using a trained model, calculates [ s ] ₁ ,s ₂ ,…s _N ]Wherein s is _i The likelihood that the ith statement contains a defect is measured. In the last step, DLFL sorts all sentences from big to small according to suspicious values, and outputs a ranking list [ l ] of all sentences ₁ ,l ₂ ,…,l _N ]Wherein l _i Pointing to the code line number where the i-th suspicious statement is located.

However, as noted by the dashed box in FIG. 1, the label is a requirement that the DLFL be operational. In other words, supervised learning neural networks internal to DLFL require large scale tagged data sets. This fact allows DLFL to be applied only to those cases where test predictions are readily available. Without the tag, the deep learning model cannot learn which samples will cause program failure and what samples will not cause program failure, and thus cannot learn the association between runtime information and program behavior. In fact, the problem of test foreword missing is often present in practical work, such as numerical calculation type software and machine learning algorithms. Often, accurate or precise test languages are not available at all, or are costly, although possible. This pre-requisite for supervised learning then makes such an efficient technique impractical or impractical in many scenarios.

Disclosure of Invention

In view of the above problems in the prior art, an object of the present invention is to: the problem of applicability of deep learning type defect positioning on unlabeled data sets is solved, so that the method can be applied to the situation that the data sets lack labels.

In order to solve the technical problems, the invention adopts the following technical scheme: a self-supervision deep learning type defect positioning method based on metamorphic test comprises the following steps:

s101: constructing an metamorphic relation MR of the target program P according to priori knowledge;

s102: acquiring a test case set t, and picking all common M' group source test cases S= { S from the test case set t according to the MR requirement ₁ ,s ₂ ,…,s _M′ (s is therein _i ＝{s _i1 ,s _i2 … } represents the i-th set of source test cases, s _ij Representation s _i The j-th test case in the list is used for each source test case group s _i Generating a corresponding subsequent test case group f for MR _i ＝{f _i1 ,f _i2 …, where a set of corresponding source test cases and subsequent test cases are considered as the smallest test unit, i.e., an metamorphic test group MTG, the test case contained in the ith MTG is s _i ∪f _i ；

S103: all test cases in each MTG are input into the target program P one by one for execution, and execution output of each source test case and subsequent test cases in each MTG is collected

Sentence coverage matrix during execution>

Indicating the coverage obtained by executing the j-th source test case in the i-th set of source test case sets,/->

Representing a coverage vector obtained by executing a j-th test case in the i-th subsequent test case group; if it is to->

Or->

Unified use->

Representing, then the element therein->

Indicating whether the kth sentence is covered in the execution of test case c, covering +.>

Taking 1, uncovered->

0, where k has a value in the range of [1, N]N is the total number of defective program code statements;

s104: for each MTG, test case set s _i ∪f _i And corresponding execution output o _i Using MR as a parameter to perform inspection to judge whether the MR is violated or not, and obtaining the test result of the MTG

Finally, a violation vector consisting of M' line test results is obtained>

Here m= { m ₁ ,m ₂ ,…,m _M′ And represents all M' MTGs, M _i Represents the ith MTG;

s105: coverage matrix for each MTG

Conversion to obtain and violate vector y ^(m) Aligned coverage matrix X ^(m) ；

S106: coverage matrix X obtained using S105 ^(m) And S104, the violation vector y ^(m) Respectively training a DLFL model as a sample and a label to obtain a trained model F;

s107: for the target program P, the following n×n identity matrix X is constructed as an auxiliary coverage matrix:

each row represents the coverage condition of one virtual test case on N sentences, and i virtual test cases only cover the ith sentence;

s108: inputting the auxiliary coverage matrix X into a model F to obtain suspicious value vectors

The suspicious value of each line of the N lines of code containing the defective program P is +.>

As an improvement, the coverage matrix of each MTG is set in S105

Conversion to obtain and violate vector y ^(m) Aligned coverage matrix X ^(m) The following formula is used:

wherein, the liquid crystal display device comprises a liquid crystal display device,

indicating whether the kth sentence is covered in the execution of test case c; will s _i ∪f _i The coverage condition of each test case in the kth sentence is logically or operated to obtain the total coverage condition of the ith MTG in the sentence +.>

The coverage conditions of all N sentences form coverage vectors under the MTG

Finally, all M 'coverage vectors form a coverage matrix X of M' ×N ^(m) 。

Compared with the prior art, the invention has at least the following advantages:

the method of the invention uses the metamorphic relation to replace the test case to describe the expected behavior of the program, and on the other hand, the method defines the metamorphic label by checking whether the predefined metamorphic relation is violated. By adjusting the two aspects, the deep learning type defect positioning can work normally on the original unlabeled data set.

According to the method, the DLFL method is expanded from supervised learning to self-supervised learning by combining with metamorphic testing, so that a labeled data set is no longer a necessary condition for applying the DLFL.

The method of the invention uses the metamorphic relation to replace the test case to describe the expected behavior of the program, and on the other hand, defines the metamorphic tag by checking whether the predefined metamorphic relation is violated. By adjusting the two aspects, the deep learning type defect positioning can work normally on the original unlabeled data set.

Drawings

Fig. 1 is a workflow of DLFL.

Fig. 2 is a coverage matrix (samples) and failure vectors (labels).

Fig. 3 is a deep learning model of DLFL.

Fig. 4 is a secondary coverage matrix.

Fig. 5 is a basic flow chart of the method of the present invention.

Detailed Description

The present invention will be described in further detail below.

The DLFL is a model for learning a localization model capable of recognizing a statement potentially responsible for a program failure by using an artificial intelligence neural network including a plurality of hidden layers. Specifically, the DLFL takes as input information collected from executing test case sets, then learns complex nonlinear relationships between statement behavior and program failures using a neural network, and after training a model, the trained model is used to evaluate suspicious values for each line of statements containing defects, resulting in a sorted list of all statements in descending order of suspicious values. This section will only show the necessary details in this process.

The information collected from the execution of each test case is divided into two parts, runtime information and execution results, more precisely, overlay information (i.e., whether it is overlaid) and test results (i.e., whether it is dead).

After the program has completed a complete execution on a test case, those statements that have been executed at least once are said to be covered by the test case, and those that have not been executed are said to be uncovered. The covered state of all statements constitutes a binary vector, where a 1 means the covered state and a 0 means covered.

The test results indicate whether the program is running correctly. Typically, the test results are obtained by comparing the program output with test predictions that match the test case. If the two are not identical, the failure is called, and if the two are identical, the passing is considered. For the purpose of establishing a correlation between a defect statement and a failure result, the present invention indicates failure by 1 and passes by 0.

In summary, a target program containing N rows of sentences, and a test case set containing M test cases, can generate coverage vectors and test results as shown in FIG. 2, wherein x _ij ＝[test caseicoversj-th statement]，y _i ＝[test case i results in failure]。[·]Called Ai Fosen brackets, which is essentially a function of the value range {0,1} and takes 1 if the condition in the brackets is satisfied and takes 0 if it is not satisfied). All coverage vectors constitute a coverage matrix (i.e., training samples) and all test results constitute failure vectors (i.e., training labels).

Fig. 3 mainly presents the deep learning process in DLFL. As illustrated in the figures, the present invention will introduce three representative DLFL techniques, which are also used in this experiment. These are CNN-FL using convolutional neural network, RNN-FL using recurrent neural network, and MLP-FL using multi-layer perceptron, respectively. These three neural networks have the same interface, i.e. they all have an input layer with N nodes and an output layer with only 1 node. During the training process, they all read in sample inputs row by row from the coverage matrix, updating the network weights so that the loss of difference between the output of the network and the labels in the failure vector is minimized. Thus, complex nonlinear relationships between coverage information and test results can be learned. The DLFL can then generate a model that is trained, i.e., a probabilistic-based classifier. Given an arbitrary coverage vector, it can predict how likely the coverage situation will cause program failure.

To determine suspicious values using the trained model, the DLFL first needs to construct an auxiliary coverage matrix as a test set. Fig. 4 shows the structure of this auxiliary coverage matrix. The system consists of coverage information of N virtual test cases, wherein each test case covers a unique line of sentences. By sending it to the trained model for prediction, the likelihood of any single line statement causing program failure can be calculated, i.e., how likely each line statement, once covered, will cause program failure. This probability assessment is essentially equivalent to the suspicious value of the statement containing the defect. Therefore, the prediction of the auxiliary coverage matrix by the model is taken as a suspicious value vector, and the defect degree of each row of sentences is reflected. Finally, the DLFL sorts all sentences according to the suspicious values, and a sentence list to be debugged, which is sorted according to the priority, is generated.

The metamorphic test (Metamorphic Testing, MT for short) is a practical solution to the problem of missing predictions for tests in the software testing field. How to apply MT to determine if the function of the program is functioning properly under a specific input in the event of a test prediction loss.

First, a program and a test case are given. The goal is to check the functional implementation of the program to determine if it is at least able to function properly for a given single test case. In the metamorphic test, it is first necessary to identify at least one MR, an attribute of the program, a necessary basis for functional integrity. In the MT process, several test cases associated with each other may be required, each of which plays a different role in the MR. Some of them are source test cases, which are provided by the original test case set; the remainder is the subsequent (follow-up) test cases, which are generated during the MT process. Taking a given single test case as a source test case, a subsequent test case can be generated according to the association between different roles defined in the MR. If the input and output of these test cases reveal a violation of the MR, then a conclusion can be drawn that the program failed, otherwise the program passed. Because the source test case and the subsequent test case always exist and are used in one entity, they are collectively referred to as a metamorphic test group (Metamorphic Testing Group, abbreviated as MTG). Supplementary explanation is a bit: one MTG may have multiple source test cases inside it, as long as MR needs.

A self-supervision deep learning type defect positioning method based on metamorphic test comprises the following steps:

s101: constructing an metamorphic relation MR of the target program P according to priori knowledge; the metamorphic relationship MR described herein is constructed by those skilled in the art from a priori knowledge of the target procedure, and is accomplished without the need for inventive effort, and for ease of understanding, several rules are presented herein that will be followed by those skilled in the art in constructing MR:

the metamorphic relationship MR described herein is constructed by those skilled in the art from a priori knowledge of the target procedure and is often only subject to specific analysis. In general, in order to be able to construct an actually effective MR, the following principles need to be followed:

a) Correctness. The MR must reflect the correct behavior of the program function, in other words, if the program function is implemented correctly, it must exhibit the properties that are prescribed in the MR.

b) And subsequent test cases are easy to generate. The metamorphic test technology is introduced to solve the problem of missing test predictions at low cost, and if the complexity of the process of generating the subsequent test cases based on MR is too high, even if the problem can be solved finally, the goal of low cost cannot be achieved, and even the problem cannot be solved.

c) MR with multiple variable elements or complex expression is preferentially selected. In general, the more complex the MR's manifestation, the more stringent it's specifications for the functional behavior of the program, and thus the higher the error detection capability, the more likely it is to expose the defects of the program.

d) Preference is given to MR which can produce the only subsequent test case. This criterion requires that no random variables exist in the MR so that once the source test case is determined, the subsequent test case can be uniquely determined. The role of this criterion is to simplify the process of subsequent MR verification, which is advantageous for automation.

MR with a wide application range is preferentially selected. Source test cases acceptable for different MRs are different. If a MR can only generate subsequent test cases in a few source test case configurations, then the data that can be generated by it is very small and therefore of low practical value.

S102: firstly, a test case set t is obtained, wherein each test case is an input acceptable by a target program P, a storage form is a text file, and data with generalization can be analyzed from the text file when MR related operation is carried out [ the process belongs to the prior art, and the description is omitted here. Then, according to the MR requirement, all the common M' group source test cases S= { S are selected from the test case set t ₁ ,s ₂ ,…,s _M′ (s is therein _i ＝{s _i1 ,s _i2 … } represents the i-th set of source test cases, s _ij Representation s _i The j-th test case in the list is used for each source test case group s _i Generating a corresponding subsequent test case group f for MR _i ＝{f _i1 ,f _i2 …, where a set of corresponding source test cases and subsequent test cases are considered as the smallest test unit, i.e., an metamorphic test group MTG, the test case contained in the ith MTG is s _i ∪f _i 。

Here, each source test case set S is illustrated in S102 _i Generating a corresponding subsequent test case group f for MR _i ＝{f _i1 ,f _i2 … the procedure is as follows:

taking the max defect program for maximum value as an example, let MR be expressed as max (a, b) +max (c, d)>-max (-b, -a) -max (-d, -c) and where a noteqb, where the terms to the left of the equation constitute the set of source test cases { (a, b), (c, d) }, the terms to the right of the equation constitute { (-b, -a), (-d, -c) } of the set of subsequent test cases, the MR requirement for the source test cases is a noteqb. A source test case group s meeting the requirements is selected _i = { (2, 5), (3, 7) }, to obtain the corresponding subsequent test case group, s needs to be first _i Substituting the MR expression to obtain max (2, 5) +max (3, 7)>-max(-b,-a)-max(-d, -c). Then, a=2, b=5, c=3, d=7 can be obtained according to the mapping relation between each variable and the source test case in the formula. Substituting the values of these variables into the equation, max (2, 5) +max (3, 7)>-max (-5, -2) -max (-7, -3), the subsequent test case group f contained on the right side of the equation can be obtained _i ＝{(-5,-2),(-7,-3)}。

Sentence coverage matrix during execution>

Or->

Unified use->

Representing, then the element therein->

Taking 1, uncovered->

0, where k has a value in the range of [1, N]N is the total number of defective program code statements. []Called Ai Fosen brackets, which is essentially a function of the value range {0,1} and takes 1 if the condition in the brackets is satisfied and takes 0 if it is not satisfied).

Acquisition execution output o in the present invention _i Sum sentence coverage matrix

The method of (1) is as follows: the method for collecting and executing output comprises two steps, wherein the first step is that the program output is redirected, the content of the program output to stdout is redirected to a file to obtain the original output, and the second step is that useful data in the original output is tidied. The method of collecting the coverage matrix of sentences consists of two steps as well, the first step is to collect the original coverage information of the program running by using a coverage tool such as Emma, GCOV, jaCoCo, and the second step is to sort out the useful data in the original coverage information, and the sentences get the coverage matrix.

Here, by way of example, test case set S in S102 _i ∪f _i And corresponding execution output o _i The procedure for examination using MR as a parameter was as follows: continuing the example above based on the defect program max, existing s _i ＝{(2,5),(3,7)}，f _i = { (-5, -2), (-7, -3) }, discussed below in two cases, showing two results possible with MR examination.

a)

O is set to _i Substituting MR expression max (a, b) +max (c, d)>-max (-b, -a) -max (-d, -c), yielding 5+7>- (-2) - (-3), further simplifying the obtainable 12>5, the true proposition, states that the original expression is at o _i Is established, and it is concluded that the MTG of group i does not violate MR.

b)

O is set to _i Substituting MR expression max (a, b) +max (c, d)>-max (-b, -a) -max (-d, -c), yielding 5+3>- (-5) - (-7), further simplifying the availability of 8>12, a false proposition, explaining that the original expression is at o _i Is not true, and it is concluded that the MTG of group i violates MR.

Finally, a violation vector consisting of M' line test results is obtained>

Here m= { m ₁ ,m ₂ ,…,m _M′ And represents all M' MTGs, M _i Represents the ith MTG.

S105: coverage matrix for each MTG

Conversion to obtain and violate vector y ^(m) Aligned coverage matrix X ^(m) ，

The following formula is specifically used:

indicating whether the kth sentence is covered in the execution of the test case c (see definition in S103); will s _i ∪f _i The coverage condition of each test case in the kth sentence is logically or operated to obtain the total coverage condition of the ith MTG in the sentence +.>

The coverage of all N sentences constitutes the coverage vector +.>

Finally, all M 'coverage vectors form a coverage matrix X of M' ×N ^(m) 。

s107: for the target program P, constructing the following N X N identity matrix X as an auxiliary coverage matrix

Wherein each row represents the coverage of one virtual test case on N sentences, and i virtual test cases only cover the ith sentence. So configured, the next step is to independently evaluate the relative magnitude of each row of statements responsible for violating MR by means of the matrix exploiting the non-linear relationship learned by model F.

Example 1: assuming that the program under test is sqrt (x), its function is to output its square root for arbitrary input x, furthermore assume a given measurementThe test case is x=2. In this simple scenario, the test language deficiency problem is apparent, because, on the one hand

The exact value of (a) is difficult to obtain by manual calculation, on the other hand, the existing calculator cannot be relied on, and after all, the exact function of the prescription provided by the calculator cannot be confirmed. Therefore, it is necessary to resort to metamorphic testing. It can be seen that the mathematical properties of many square roots include this term, namely +.>

It is now used as a strip MR. Then, according to the MR requirement, in addition to the source test case of x=2, a random number y needs to be generated as one subsequent test case, say y=7, and then another subsequent test case is defined, where z=x·y=14. It is then now possible to determine whether the implementation of the target program is correct by simply comparing the product of sqrt (2) and sqrt (7) with the sqrt (14). If the two are not equal, the program fails, otherwise, the program passes.

Input: the defective program P consisting of n=12 lines of codes is as follows, and the function of the program is to output the maximum value of two real numbers. Wherein line 5 is the actual location of the defect;

test case set t= { t composed of m=2 test cases ₁ ,t ₂ Use case inputs and expected outputs are shown in the following table.

Numbering device	Input device	Expected output
			t
₁	2 2	2
			t ₂	3 1	3

And (3) outputting: suspicious values for each of N lines of code of defect program P

S101: from the domain knowledge related to the procedure, MR is constructed: the two numbers input are exchanged for the position, the maximum value is unchanged. The functional interface of the original program is denoted as max (x, y) →z, i.e. two real numbers x and y are input, and the calling program max (x, y) gets the maximum value z of the two. Then MR may be expressed as max (x, y) =max (y, x).

S102: according to MR requirement, all M' =2 groups of source test cases S= { S are selected from the original test case set t ₁ ,s ₂ (s is therein ₁ ＝{s ₁₁ }，s ₁₁ ＝(2,2)，s ₂ ＝{s ₂₁ }，s ₂₁ = (3, 1). And use each source test case set s _i Generating corresponding subsequent test case groups for MR to obtain f ₁ ＝{(2,2)}，f ₂ = { (1, 3) }. A set of related source test cases and subsequent test cases are considered as the smallest test unit, namely, an metamorphic test group (MTG for short), the test cases contained in the 1 st MTG are { (2, 2), (2, 2) }, and the test cases contained in the 2 nd MTG are { (3, 1), (1, 3) }.

S103: all test cases are input into P one by one for execution, and the execution output and execution process of each source test case and subsequent test case in each MTG are collectedThe sentence coverage matrix in (a) is respectively obtained

And->

The detailed values are as follows.

o ₁ ＝{2,2}

o ₂ ＝{1,3}

S104: for each MTG, test case set s _i ∪f _i And corresponding test output o _i Using MR as a parameter to perform inspection to judge whether the MR is violated or not, and obtaining the test result of the MTG

Finally get violation vector->

Specific deduction procedures and results are as follows.

S105: for each MTG, according to the existing coverage matrix

The following conversion is performed

The final result is the violation vector y ^(m) Aligned coverage matrix

The detailed values are as follows.

S106: learning the following model using deep neural networks

F:X ^(m) →y ^(m)

And obtaining a trained model F.

In this example, a multiple regression linear model y=f (X) =x·w without bias was used, and model weights were learned as follows.

w＝[0 0 0 0 1 0 0 0 0 0 0 0] ^T

S107: constructing an auxiliary coverage matrix as a test set as follows

S108: the auxiliary coverage matrix X is input into a model F to obtain suspicious value vectors

The suspicious value of each of the N lines of code containing defective program P, i.e +.>

Specific deduction procedures and results are as follows.

It can be seen that the suspicious value corresponding to the line 5 code is 1, the rest codes are 0, and in fact, the line 5 code is just the defect, so the defect positioning is successful.

Finally, it is noted that the above embodiments are only for illustrating the technical solution of the present invention and not for limiting the same, and although the present invention has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that modifications and equivalents may be made thereto without departing from the spirit and scope of the technical solution of the present invention, which is intended to be covered by the scope of the claims of the present invention.

Claims

1. The self-supervision deep learning type defect positioning method based on metamorphic test is characterized by comprising the following steps of: the method comprises the following steps:

s102: acquiring a test case set t, and picking all common M' group source test cases S= { S from the test case set t according to the MR requirement ₁ ，s ₂ ，…，s _M′ (s is therein _i ＝{s _i1 ，s _i2 … } represents the i-th set of source test cases, s _ij Representation s _i The j-th test case in the list is used for each source test case group s _i Generating a corresponding subsequent test case group f for MR _i ＝{f _i1 ，f _i2 …, where a set of corresponding source test cases and subsequent test cases are considered as the smallest test unit, i.e., an metamorphic test group MTG, the test case contained in the ith MTG is s _i ∪f _i ；

Sentence coverage matrix during execution>

Representing an overlay vector obtained by executing a j-th source test case in the i-th set of source test case sets,/->

Or->

Unified use->

Representing, then the element therein->

Taking 1, uncovered->

s104: for each ofMTG, test case group s _i ∪f _i And corresponding execution output o _i Using MR as a parameter to perform inspection to judge whether the MR is violated or not, and obtaining the test result of the MTG

Finally, a violation vector consisting of M' line test results is obtained>

Here m= { m ₁ ，m ₂ ，…，m _M′ And represents all M' MTGs, M _i Represents the ith MTG;

s105: coverage matrix for each MTG

S106: coverage matrix X obtained using S105 ^(m) And S104, the violation vector y ^(m) Respectively training unbiased multiple regression linear models as a sample and a label to obtain a trained model F;

s108: the auxiliary coverage matrix is formedObtaining suspicious value vector by X input model F

2. The self-supervised deep learning defect localization method based on metamorphic testing of claim 1, wherein: the coverage matrix of each MTG is used in the S105

indicating whether the kth sentence is covered in the execution of the test case c; will s _i ∪f _i The coverage condition of each test case in the kth sentence is logically or operated to obtain the total coverage condition of the ith MTG in the sentence +.>

The coverage of all N sentences constitutes the coverage vector +.>

Finally, all M 'coverage vectors form a coverage matrix X of M' ×N ^(m) 。