Data summarization method, system and fault diagnosis method based on deviation degree analysis

Abstract

The application relates to the technical field of data processing, and discloses a data abstract method and system based on deviation degree analysis and a fault diagnosis method, which comprises the following steps: S1. After a target algorithm, to-be-abstracted test data, the minimum register unit stored data bit number and the operation logic involved in the to-be-tested step in the target algorithm are acquired, the allocation weight corresponding to the data stored in each minimum register unit is obtained; S2. The to-be-abstracted test data is split into a plurality of groups in units of the data bit number of the minimum register unit, and each group of data is subjected to class average pooling processing according to a preset precision requirement; after each group of data is subjected to the class average pooling processing, an average pooling data is obtained; S3. The data abstract result corresponding to the to-be-abstracted test data after the to-be-tested step is obtained. The application solves the problem that the prior art cannot be used to represent high-bit-width operation results, and has the characteristics of significantly reducing the bit width of data.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, more particularly, to a data digest method and system based on deviation degree analysis and a fault diagnosis method. BACKGROUND

[0002] With the development of computer technology and mobile internet technology, the types of algorithms are also increasing, which provides convenience for users. In the development process of various types of algorithms, in order to ensure the normal and stable operation of the algorithm, the functions of the algorithm need to be tested, and in this process, the developer needs to use data for various tests, among which the part of data or code error in the operation process is particularly troublesome, and multiple data comparisons are often needed to detect the deflection degree of the initial data or the influence degree of a certain code in the calculation process on the whole process. In order to avoid the influence of long data on the test efficiency, the data can be summarized to reduce the data amount in the test process and improve the test efficiency.

[0003] In the process of summarizing the data, if any content of the input information is changed, even if one bit is changed, the output will change unpredictably, that is, any information of the input will affect the final output, so it can be used to prevent the data from being privately modified. The existing digest function generally mixes the data, creates a hash value, and usually mixes redundant information in the processing process and then performs lossy compression, losing the original representation of the data. It is generally used for digital signature of identity authentication or verification of information integrity.

[0004] In the case of preserving a certain amount of information, there are generally the following methods for processing high-bit-width data into lower-bit-width data, data compression and hash function. The essence of digest processing is also a hash function, and the general hash function has direct addressing method, digital analysis method, square middle method, folding method, random number method, and remainder method. Because the loss of data is too large and there is no certain representation, these methods are only used as a symbol and cannot be used to convert high-bit-width data into lower-bit-width numerical representation. Generally, the data compression algorithm needs to consider both compression and decompression, and the algorithm is relatively complex when compressing, especially the processing time of the algorithm with a compression rate below 10%. Even if the compression algorithm has the advantages of being able to decompress and losing less useful information, it is not suitable for scenes that need to improve the comparison efficiency and detect the deviation degree in the test time. Therefore, a data digest processing method with low power consumption, which can preserve a certain amount of effective information and has a compression rate below 10%, is needed. The compression rate here is only a description of the performance of the present application in the shortening of data bit width, which is similar to lossy compression, but it is essentially a hash function processing idea.

[0005] The present application discloses a summary generation method, comprising: obtaining probabilities of common generation words and domain words based on the content of an input text and business information corresponding to the domain of the input text; and generating summary information corresponding to the input text based on the probabilities of the common generation words and the domain words and corresponding vocabularies. The present application also discloses a text summary generation method, comprising: obtaining probabilities of common generation words, copy words and domain words corresponding to an input text in real time based on the encoded input text; and obtaining summary information based on the probabilities of the three and corresponding vocabularies.

[0006] In the comparison and error analysis of the high-bit-width operation result of the algorithm, the existing summary processing cannot be used to represent the high-bit-width operation result because of the mixed redundant information and the disturbance of the original data position, and if a compression function is used, there is a problem of insufficient compression rate and complex compression process affecting the analysis efficiency, therefore, how to invent a summary processing method and device capable of representing the high-bit-width operation result with low-bit-width data on the basis of preserving the original meaning of the data and analyzing the deviation degree of the initial data or the test algorithm process is a technical problem to be solved in the technical field. SUMMARY

[0007] The present application provides a data summary method, system and fault diagnosis method based on deviation degree analysis, which has the characteristics of significantly reducing the bit width of the data.

[0008] To achieve the above-mentioned purposes of the present application, the technical solutions adopted are as follows:

[0009] A data summary method based on algorithm process deviation degree analysis, comprising the following steps:

[0010] S1. After obtaining the target algorithm, the data to be summarized, the data bit number stored in the minimum register unit and the operation logic involved in the target step of the target algorithm, the allocation weight corresponding to the data stored in each minimum register unit is obtained according to the operation logic involved in the target step of the target algorithm of each minimum register unit;

[0011] S2. The data to be summarized is divided into several groups in units of the data bit number of the minimum register unit, and the class average pooling processing is performed on each group of data according to the preset accuracy requirement to obtain the average pooling data corresponding to each group of data; after the class average pooling processing of each group of data, an average pooling data is obtained;

[0012] S3. The data summary result corresponding to the data to be summarized after the test step is obtained according to the allocation weight and the average pooling data corresponding to each minimum register unit.

[0013] Preferably, the allocation weight corresponding to the data stored in each minimum register unit is determined according to the operation logic involved in the execution of the target step of the target algorithm of each minimum register unit, comprising:

[0014] The following judgment process is executed for each minimum register unit in sequence:

[0015] If the operation logic involved in the execution of the target step of the target algorithm of the current minimum register unit is loop calculation, the weight coefficient corresponding to the data stored in the current minimum register unit is determined according to the number of loops executed in the target step of the target algorithm;

[0016] If the operation logic involved in the execution of the target step of the target algorithm of the current minimum register unit is parallel calculation, the weight coefficient corresponding to the data stored in the current minimum register unit is determined according to the number of operations and the first complexity of the operation logic in the target step of the target algorithm; the first complexity is based on operation time;

[0017] If the operation logic involved in the execution of the target step of the target algorithm of the current minimum register unit is single calculation, the weight coefficient corresponding to the data stored in the current minimum register unit is determined according to the number of operations and the second complexity of the operation logic in the target step of the target algorithm;

[0018] The weight corresponding to the data stored in each minimum register unit is determined according to the weight coefficient corresponding to each minimum register unit.

[0019] Further, the test data to be summarized is split into a plurality of groups in units of the number of data bits stored in the minimum register unit, and the class average pooling data corresponding to each group of data is obtained by performing class average pooling processing on each group of data according to a predetermined precision requirement, comprising:

[0020] The data stored in the minimum register is split into a plurality of groups in units of the number of data bits specified by the predetermined precision requirement;

[0021] The binary data stored in each group is converted into decimal data respectively;

[0022] For each minimum register unit, the average value of the decimal data corresponding to the binary data in all groups split is calculated;

[0023] According to the size relationship between the decimal data corresponding to the binary data in each group and the average value, a value is assigned to each group;

[0024] For each minimum register unit, the values assigned to each group are concatenated to form the class average pooling data corresponding to the current minimum register unit.

[0025] Further, the value of each group is assigned according to the size relationship between the decimal data corresponding to the binary data in each group and the average value, including:

[0026] If the decimal data corresponding to the binary data in the current group is greater than the average value, the current group is assigned "1";

[0027] If the decimal data corresponding to the binary data in the current group is less than the average value, the current group is assigned "0".

[0028] Further, the data summary result is obtained according to the distribution weight and the class average pooling data, including: adopting a target mathematical model, the data summary result is obtained according to the distribution weight and the class average pooling data;

[0029] The target mathematical model is:

[0030] (k1p1+k2p2+……k n p n ) / (k1+k2+……k n )

[0031] Wherein, p1, p2, ……p n is the class average pooling data corresponding to the first, second, ……, n minimum register unit, k1, k2, ……k n is the weight coefficient corresponding to the data stored in the first, second, ……, n minimum register unit.

[0032] Further, according to the data bit number specified by the preset precision requirement, it is judged whether the class average pooling data, the distribution weight and the data summary result meet the requirement of the data bit number,

[0033] If it meets, the subsequent steps are continued to be executed;

[0034] If it does not meet, the obtaining step of the data is re-executed.

[0035] Further, the operation logic involved in the algorithm execution process of the data stored in each minimum register unit is kept consistent.

[0036] Further, the data summary method obtains at least two data summary results, and a standard data summary result is obtained by benchmarking experiment, and the method comprises:

[0037] The deviation between each data summary result and the standard data summary result is judged;

[0038] If the deviation between the i-th data summary result and the standard data summary result is greater than the deviation between the j-th data summary result and the standard data summary result, then in the algorithm logic, the algorithm failure corresponding to the i-th data summary result is earlier than the algorithm failure corresponding to the j-th data summary result.

[0039] A data summary system based on deviation degree analysis comprises a data preprocessing module, an expression analysis module and a data processing module.

[0040] The expression analysis module is configured to obtain an allocation weight corresponding to the data stored in each minimum register unit according to the operation logic involved in the execution of the target step of the target algorithm by each minimum register unit.

[0041] The data preprocessing module is configured to split the test data to be summarized into a plurality of groups in units of the data bit number of the minimum register unit, and perform class average pooling processing on each group of data according to a preset accuracy requirement to obtain average pooling data corresponding to each group of data.

[0042] The data processing module is configured to obtain a data summary result corresponding to the test data to be summarized after the test step according to the allocation weight corresponding to each minimum register unit and the average pooling data.

[0043] Preferably, a fuzzy analysis module is provided, which is configured to perform fuzzy analysis on the deviation degree caused by the deviation of the test data to be summarized and the deviation of the algorithm execution process after obtaining the data summary result.

[0044] The present application has the following advantages:

[0045] The present application discloses a data summary method based on deviation degree analysis, which analyzes the operation logic of each minimum register unit in the algorithm process according to the data stored in the minimum register unit and the data obtained after the expression preprocessing of the test data to be summarized. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a flowchart of a data digest method based on deviation degree analysis according to the present application.

[0047] Figure 2 is a module diagram of a data digest system based on deviation degree analysis according to the present application.

[0048] Figure 3 is a schematic diagram of the interaction process in a module of a data digest system based on deviation degree analysis according to the present application.

[0049] Figure 4 is a schematic diagram of the present application in a data digest method based on deviation degree analysis according to the present application, taking the sha256 hash algorithm as an example in embodiment 2. DETAILED DESCRIPTION

[0050] The present application will be described in detail below in conjunction with the drawings and specific embodiments.

[0051] Embodiment 1

[0052] As shown in the drawings, Figure 1 the present application provides a data digest method based on deviation degree analysis, comprising the following steps:

[0053] S1. After obtaining the target algorithm, the data to be digested, the minimum register unit storage data bit number, and the operation logic involved in the target step of the target algorithm, according to the operation logic involved in the target step of the target algorithm of each minimum register unit, the data stored in each minimum register unit is obtained Corresponding allocation weight;

[0054] S2. The data to be digested is divided into several groups in units of minimum register unit data bits, and each group of data is processed by class average pooling according to the preset precision requirement, to obtain the average pooling data corresponding to each group of data; After each group of data is processed by class average pooling, an average pooling data is obtained;

[0055] S3. According to the allocation weight and the average pooling data corresponding to each minimum register unit, the data digest result corresponding to the data to be digested after the test step is obtained.

[0056] Embodiment 2

[0057] The present application provides another data digest method based on deviation degree analysis, comprising the following steps:

[0058] S1. After obtaining the target algorithm, the test data to be summarized, the minimum register unit storage data bit number and the operation logic involved in the target step of the target algorithm, the allocation weight corresponding to the data stored in each minimum register unit is obtained according to the operation logic involved in the execution of the target step of the target algorithm of each minimum register unit;

[0059] S2. The test data to be summarized is divided into several groups in units of the data bit number of the minimum register unit, and the average pooling data corresponding to each group of data is obtained by performing the average pooling processing on each group of data according to the preset precision requirement; after the average pooling processing, one average pooling data is obtained for each group of data.

[0060] S3. The data summary result corresponding to the test data to be summarized after the target step is obtained according to the allocation weight corresponding to each minimum register unit and the average pooling data.

[0061] In the embodiment, the target algorithm is the sha256 hash algorithm.

[0062] In one specific embodiment, in step S2, the allocation weight corresponding to the data stored in each minimum register unit is obtained according to the operation logic involved in the execution of the target step of the target algorithm of each minimum register unit, including:

[0063] The following judgment process is performed for each register in sequence:

[0064] If the operation logic involved in the execution of the target step of the target algorithm of the current minimum register unit is loop calculation, the weight coefficient corresponding to the data stored in the current minimum register unit is determined according to the number of loops executed in the target step of the target algorithm;

[0065] If the operation logic involved in the execution of the target step of the target algorithm of the current minimum register unit is parallel calculation, the weight coefficient corresponding to the data stored in the current minimum register unit is determined according to the number of operations and the first complexity of the operation logic in the target step of the target algorithm; the first complexity is a standard of operation time;

[0066] If the operation logic involved in the execution of the target step of the target algorithm of the current minimum register unit is single calculation, the weight coefficient corresponding to the data stored in the current minimum register unit is determined according to the number of operations and the second complexity of the operation logic in the target step of the target algorithm;

[0067] The weight corresponding to the data stored in each minimum register unit is determined according to the weight coefficient corresponding to each minimum register unit.

[0068] In one specific embodiment, the to-be-digest test data is split into a plurality of groups in units of the number of data bits stored in the minimum register unit, and each group of data is subjected to class average pooling processing according to a preset precision requirement to obtain class average pooling data corresponding to each group of data, including:

[0069] Splitting the data stored in the minimum register into a plurality of groups in units of the number of data bits specified by the preset precision requirement;

[0070] Converting the binary data stored in each group into decimal data;

[0071] For each minimum register unit, the decimal data corresponding to the binary data in all groups is averaged;

[0072] According to the size relationship between the decimal data corresponding to the binary data in each group and the average value, a value is assigned to each group;

[0073] For each minimum register unit, the values assigned to each group are concatenated to convert into class average pooling data corresponding to the current minimum register unit.

[0074] In one specific embodiment, the value assigned to each group according to the size relationship between the decimal data corresponding to the binary data in each group and the average value includes:

[0075] If the decimal data corresponding to the binary data in the current group is greater than the average value, the current group is assigned "1";

[0076] If the decimal data corresponding to the binary data in the current group is less than the average value, the current group is assigned "0".

[0077] In one specific embodiment, the data digest result is obtained according to the allocation weight and the class average pooling data, including: using a target mathematical model to obtain the data digest result according to the allocation weight and the class average pooling data;

[0078] The target mathematical model is:

[0079] (k1p1+k2p2+…k n p n ) / (k1+k2+…k n )

[0080] Where p1, p2, …, p n is the class average pooling data corresponding to the 1st, 2nd, …, nth minimum register unit, k1, k2, …, k nThe weight coefficients corresponding to the data stored in the first, second, …, and nth minimum register units.

[0081] In one embodiment, the number of data bits specified by the preset precision requirement is used to determine whether the average-pooling data, the assigned weight, and the data summary result meet the data bit requirement.

[0082] If the requirement is met, the subsequent steps are continued.

[0083] If the requirement is not met, the data obtaining step is re-executed.

[0084] In one embodiment, the operation logic involved in the data stored in each minimum register unit during the algorithm execution is consistent.

[0085] In one embodiment, the data summary method obtains at least two data summary results and performs a benchmark experiment to obtain a standard data summary result. The method comprises:

[0086] determining the deviation between each data summary result and the standard data summary result.

[0087] If the deviation between the ith data summary result and the standard data summary result is greater than the deviation between the jth data summary result and the standard data summary result, the algorithm failure corresponding to the ith data summary result is earlier than the algorithm failure corresponding to the jth data summary result in the algorithm logic.

[0088] In this embodiment, the 256-bit data in the sha256 algorithm needs to be shortened to 8 bits. After considering the precision requirement, the bit width needs to be shortened to 8 bits, which meets the minimum register unit standard for each register unit. The 32-bit register unit is subjected to average-pooling, the data in the register unit is divided into several groups of data with each four bits as a unit, and the average is taken. The group data greater than or equal to the average is regarded as 1, and the group data less than the weighted average is regarded as 0. The group data of each register unit is spliced as the preprocessed data p1, p2, …, p n . The basis is that the data in each register is saved in binary form, and 1 and 0 represent the opening and closing of the minimum circuit unit, respectively. Therefore, the average-pooling operation can be regarded as artificially specifying the binary representation of the register power consumption performance with the average number. Although the data in each 32-bit register is output simultaneously, it is subjected to 57, 58, 59, 60, 61, 62, 63, and 64 times of loop calculation, respectively. Therefore, 57, 58, 59, 60, 61, 62, 63, and 64 can be used as the weights k1, k2, …, k n for weighted processing, and the obtained weighted average is only used as a measurement standard.

[0089] In the process of testing the IP of a chisel implementing the sha256 algorithm, the inventor often needs to compare the generated 256-bit hash value with the predicted 256-bit hash value and modify the operation path of some modules. It is found that manual comparison of hash values is very inefficient before the cause of the error in the partial code is resolved. After the error is found, other improvement experiments are carried out, which require frequent transmission and comparison of hash values. Because of the high bit width of the 256-bit hash value, the improvement scheme loses its advantage. In the process of finding various solutions, it is found that there is no suitable method for this scenario. Data compression is difficult to solve the problem of speed and power consumption. Wanting to improve efficiency increases power consumption and reduces efficiency. The general hash function can only get a flag value, and its processing process is completely unrelated to the algorithm. The value obtained loses its flag significance. Therefore, it is urgent to develop a new abstract processing method that can be applied to this scenario. This method must have the following characteristics: low power consumption, representative meaning, significantly reducing the bit width of data, and reflecting the deviation of data in the algorithm running process.

[0090] After a series of investigations, it is found that the pooling concept in image processing can preserve and process more effective information to a certain extent to save smaller images. However, the premise of pooling is that each number must have a certain meaning, and the influence of each part of the number on the whole is similar, and the final result cannot change the specific meaning represented by the number. In the analysis of the operation process of the algorithm and the exploration of the meaning of the numerical value, it is found that no matter what algorithm, the data of a bit will not be processed alone, but will be divided into different registers as a unit, and then participate in various operation formulas and operation logic. The concept of 1 and 0 in digital circuits also has certain similarity compared to the color of each pixel point in the image, and the specific numerical value also reflects the power consumption performance of each switch. Because there is no linear relationship in the value process, only fuzzy analysis can be used in the description of the deviation. However, in terms of representation, the value obtained by this method has the characteristics of a hash function while having certain effective information of the original data, and meets the previous requirement of reducing the data bit width. In the previous sha algorithm, it can be used to abstract the 256-bit hash value into an 8-bit feature value. The basis of each step of processing is given in the specific implementation method. Because this method cannot meet the one-to-one mapping relationship, there will be a certain probability of collision. However, because it refers to the hash function processing method and is related to the operation process, even a small change in a bit will be reflected in the final result according to the actual weight participating in the operation after multiple operation paths. Therefore, the probability of different values colliding to produce the same result objectively exists but is relatively small.

[0091] The method is born from the deviation analysis of sha256 algorithm, but after investigation and analysis of other algorithms, it is found that the general loop compression processing algorithm can be basically applied without too many changes, for example, sha-1 and similar md5, and if it is other algorithm, it needs to be analyzed according to the expression of the specific digest test data, and the general analysis process is given in the specific implementation method. Moreover, the method can be converted into a specific circuit system through a circuit module, and the logic in the process of analyzing data and algorithm is as shown in Figure 3 The detailed implementation process is given in the specific implementation method, the actual power consumption of the circuit is small, the implementation logic of multiple judgments and data processing is relatively simple, the running speed is much higher than that of the compression function under the same compression rate (the compression rate is only compared with the bit width), and compared with the value process of other hash functions at present, the biggest advantage is that it is born from the analysis of the specific algorithm process, and is closely related to the algorithm running logic, so that the method and device can improve the algorithm test efficiency and fuzzy analysis deviation when performing digest processing.

[0092] The present application discloses a data digest method based on algorithm process deviation analysis, which adopts the minimum register grouping of the set data, analyzes the operation logic of each register in the algorithm process by processing the data through average pooling and the expression of the obtained digest test data.

[0093] As Figure 4As shown, to shorten the data bit width, the method adopted in this invention first splits the 256-bit high-width data obtained during the algorithm testing process into data bits stored in the smallest register unit. In this embodiment, the number of bits is 4. The weight coefficient corresponding to each register in the high-width data is obtained; in this embodiment, all registers have the same weight. Then, the test data to be digested is split into several groups based on the data bits stored in the smallest register unit, and each group of data is subjected to average pooling processing according to a preset precision requirement, resulting in average pooling data of 7.25 for each group. Based on the relationship between the decimal data corresponding to the binary data in each group and the average value of 7.25, each group is assigned a value of 0 or 1. Finally, using the target mathematical model, based on the assigned weights and the average pooling data, an 8-bit data digest result is obtained. Thus, this invention achieves the purpose of shortening the data bit width. Since the data digest result processing is closely related to the algorithm process and does not contain redundant information, the probability of encountering the same result is extremely low. Therefore, it can, to a certain extent, characterize the operation result for comparison and analysis during algorithm testing.

[0094] From the perspective of fuzzy analysis deviation, this method and device, under the premise of analyzing the calculation formula in the algorithm process, through a series of processing on the results before and after data deviation or algorithm code improvement, finally obtains a difference that can not only reflect the length of the path taken by the data or code with deviation error in the algorithm operation process, that is, the larger the value, the earlier the position of the data with deviation or the earlier the logical position of the code with error in the operation process. Therefore, this characteristic can be used to analyze the degree of deviation in the algorithm testing process; and under the premise of setting up a control experiment, it can quantify the degree of impact of a certain code modification on the algorithm process, thereby facilitating developers to test and optimize the algorithm.

[0095] Example 3

[0096] like Figure 2 As shown, a data summarization system based on deviation analysis includes a data preprocessing module, an expression parsing module, and a data processing module;

[0097] like Figure 3 As shown, the expression parsing module is used to obtain the allocation weight corresponding to the data stored in each minimum register unit after acquiring the target algorithm, the test data to be digested, the number of data bits stored in the minimum register unit, and the operation logic involved in the test step in the target algorithm.

[0098] In the embodiment, when calculating and analyzing the operation logic of each register in the algorithm process, it is determined whether to perform loop calculation, if the loop calculation is performed, the distribution weight is obtained through the loop calculation, if not, it is determined whether to perform single calculation, and the distribution weight is obtained through weight comparison.

[0099] The data preprocessing module is used for splitting the to-be-abstracted test data into a plurality of groups in units of data bits of the minimum register unit, performing class average pooling processing on each group of data according to a preset precision requirement, and obtaining average pooling data corresponding to each group of data.

[0100] The data processing module is used for obtaining a data abstract result corresponding to the to-be-abstracted test data after the to-be-abstracted test data passes through the to-be-tested step according to the distribution weight and the average pooling data corresponding to each minimum register unit.

[0101] In the embodiment, the fuzzy analysis module is provided; after the data abstract result is obtained, the fuzzy analysis module is used for performing fuzzy analysis on a deviation degree caused by a deviation of the to-be-abstracted test data and a deviation of an algorithm execution process on an operation result, specifically, a reference experiment is performed, a result obtained by executing an algorithm on the to-be-abstracted test data without modification of the to-be-abstracted test data, without modification of the algorithm execution process, and with small error generated by an algorithm execution environment is taken as a reference value. A plurality of control experiments with different strategies are set to perform a plurality of times of fuzzy analysis. The strategies of the control experiments include, data at different positions of the to-be-abstracted test data are respectively modified and the same algorithm process is executed, an experiment is performed in the algorithm execution environment with small error and an experimental value is obtained; codes at different positions of the algorithm execution process are respectively modified, the unmodified to-be-abstracted test data is executed by using the modified algorithm process, an experiment is performed in the algorithm execution environment with small error and an experimental value is obtained; the unmodified to-be-abstracted test data is executed by using the unmodified algorithm process, experiments are respectively performed in the algorithm execution environment with different errors and experimental values are obtained. A difference between the above experimental values and the reference value is taken as a result of the fuzzy analysis, and a deviation degree of the algorithm is detected. It should be noted that the corresponding control experiment strategy can be selected according to the requirement of the fuzzy analysis.

[0102] Obviously, the above embodiments of the present application are only examples for clearly illustrating the present application, and are not intended to limit the implementation modes of the present application. Any modification, substitution and improvement made within the spirit and principle of the present application should be included in the protection scope of the claims of the present application.

Claims

1. A method for data summarization based on skewness analysis, characterized in that: The method comprises the following steps: S1. After obtaining the target algorithm, the test data to be summarized, the minimum register unit storage data bit number and the operation logic involved in the target step of the target algorithm, the allocation weight corresponding to the data stored in each minimum register unit is obtained according to the operation logic involved in the execution process of the target step of the target algorithm of each minimum register unit; The specific steps include: The following judgment process is performed for each minimum register unit: If the operation logic involved in the execution process of the target step of the target algorithm of the current minimum register unit is loop calculation, the weight coefficient corresponding to the data stored in the current minimum register unit is determined according to the number of loops executed in the target step of the target algorithm; If the operation logic involved in the execution process of the target step of the target algorithm of the current minimum register unit is parallel calculation, the weight coefficient corresponding to the data stored in the current minimum register unit is determined according to the number of operations and the first complexity of the operation logic in the target step of the target algorithm; the first complexity is based on operation time; If the operation logic involved in the execution process of the target step of the target algorithm of the current minimum register unit is single calculation, the weight coefficient corresponding to the data stored in the current minimum register unit is determined according to the number of operations and the second complexity of the operation logic in the target step of the target algorithm; The weight corresponding to the data stored in each minimum register unit is determined according to the weight coefficient corresponding to each minimum register unit; S2. The test data to be summarized is divided into a plurality of groups in units of the data bit number of the minimum register unit, and each group of data is subjected to class average pooling processing according to a predetermined precision requirement to obtain class average pooling data corresponding to each group of data; after each group of data is subjected to class average pooling processing, a class average pooling data is obtained; the specific steps include: The data stored in the minimum register is divided into a plurality of groups in units of the data bit number specified by the predetermined precision requirement; The binary data stored in each group is converted into decimal data; For each minimum register unit, the average value of the decimal data corresponding to the binary data in all groups is calculated; Each group is assigned a value according to the size relationship between the decimal data corresponding to the binary data in the group and the average value; For each minimum register unit, the values assigned to each group are concatenated to form the class average pooling data corresponding to the current minimum register unit; S3. The data summary result corresponding to the test data to be summarized after the target step is obtained according to the allocation weight and the class average pooling data corresponding to each minimum register unit; the specific processing is performed by using a target mathematical model, and the target mathematical model is: (k1p1+k2p2+……k n p n ) / (k1+k2+……k n ) wherein p1, p2, … p n is the class average pooling data corresponding to the 1st, 2nd, … nth minimum register unit, k1, k2, … kn n is the weight coefficient corresponding to the data stored in the 1st, 2nd, … nth minimum register unit.

2. The data summarization method based on deviation analysis according to claim 1, wherein: Each group is assigned a value according to the size relationship between the decimal data corresponding to the binary data in the group and the average value, which includes: If the decimal data corresponding to the binary data in the current group is greater than the average value, the current group is assigned "1"; If the decimal data corresponding to the binary data in the current group is less than the average value, the current group is assigned "0".

3. The data summary method based on deviation degree analysis according to claim 2, characterized in that: whether the class average pooling data, the assigned weight and the data summary result meet the data bit requirement is determined based on the data bit number specified according to the preset accuracy requirement, if yes, the subsequent steps are continued to be executed, if no, the obtaining step of the data is re-executed.

4. The data summarization method based on deviation analysis according to claim 1, characterized in that: the operation logic involved in the algorithm execution process of the data stored in each minimum register unit is kept consistent.

5. A fault diagnosis method based on deviation degree analysis, comprising the following steps: obtaining at least two data summary results based on the data summary method of any one of claims 1-4; and obtaining a standard data summary result based on a benchmark experiment. The method comprises: judging the deviation between each data summary result and the standard data summary result, if the deviation between the i th data summary result and the standard data summary result is greater than the deviation between the j th data summary result and the standard data summary result, then in the algorithm logic, the algorithm failure corresponding to the i th data summary result is earlier than the algorithm failure corresponding to the j th data summary result.

6. A data summarization system based on skewness analysis characterized by: The data summary method according to any one of claims 1-4 is implemented, comprising a data preprocessing module, an expression analysis module and a data processing module. The expression analysis module is used to obtain the assigned weight corresponding to the data stored in each minimum register unit after obtaining the target algorithm, the test data to be summarized, the data bit number stored in the minimum register unit and the operation logic involved in the target step of the target algorithm. The data preprocessing module is used to split the test data to be summarized into several groups in units of the data bit number of the minimum register unit, and perform class average pooling processing on each group of data according to the preset accuracy requirement to obtain the average pooling data corresponding to each group of data. The data processing module is used to obtain the data summary result corresponding to the test data to be summarized after the test step according to the assigned weight and the average pooling data corresponding to each minimum register unit.

7. The data digest system of claim 6, wherein: The fuzzy analysis module is provided, which is used to perform fuzzy analysis on the deviation degree caused by the deviation of the test data to be summarized and the deviation of the algorithm execution process after obtaining the data summary result.

Images