Method and system for constructing anomaly detection model, and analytic hierarchy process and system

Abstract

A method and system for constructing an anomaly detection model, and an analytic hierarchy process and system. The analytic hierarchy process comprises: acquiring a first weight of each of a plurality of factors associated with a target task (S401); at least on the basis of time series data of each factor, using a constructed time series anomaly detection model to obtain a second weight of each of the plurality of factors (S403); and on the basis of the first weight and the second weight, determining a target weight corresponding to each of the plurality of factors (S405). The time series anomalies of factor values can be considered in the process of determining a target weight, avoiding the situation that a large weight is assigned to an anomalous factor in the analytic hierarchy process, significantly improving the effectiveness of analytic hierarchy results.

Description

Method and system for building anomaly detection model and hierarchical analysis method and system

[0001] This application claims priority to Chinese patent application 202311212872.X filed on September 19, 2023 and entitled “Method and system for constructing anomaly detection model and hierarchical analysis method and system”, the entire disclosure of which is incorporated herein by reference. Technical Field

[0002] The present application relates to a method and system for constructing a time series anomaly detection model, a hierarchical analysis method and system, and a computer-readable storage medium for implementing the above method. Background Art

[0003] The analytic hierarchy process is a comprehensive evaluation method for system analysis and decision-making. It is a multi-factor decision-making method proposed on the basis of a full study of the human thinking process, which can reasonably realize the quantitative processing process of qualitative problems.

[0004] The Analytic Hierarchy Process (AHP) constructs a hierarchical structure to transform complex decision-making processes into a pairwise comparison of the importance of several factors, thereby transforming difficult-to-quantify qualitative judgments into actionable comparisons of importance. Specifically, the AHP first calculates the weights of multiple factors through matrix calculations. These weights are then multiplied by their corresponding factor values, and the resulting product of these weights and factor values ​​serves as the basis for decision-making. For example, the AHP can be applied to financial sectors such as enterprise capability assessment, merchant value evaluation, customer demand rating, supplier selection, manufacturing systems, and strategic decision-making. However, in current AHP methods, the calculation of factor weights relies solely on the comparison of the importance of several factors, ignoring the fact that factor values ​​themselves may vary over time. This results in assigning larger weights to anomalous factors during the AHP analysis process, leading to significant errors in the AHP analysis results. For example, when using the hierarchical analysis method to evaluate the value of a merchant, turnover, as one of several factors, may have time-series outliers over time. At this time, if we only rely on the comparison of the importance of several factors and assign a larger weight to the turnover of one of the factors, it will lead to large errors in the analysis results of the hierarchical analysis method, and the output hierarchical analysis results, namely the merchant value, will lose their reference value.

[0005] Summary of the Invention

[0006] In order to solve or at least alleviate one or more of the above problems, the following technical solutions are provided.

[0007] According to a first aspect of the present application, a method for constructing a time series anomaly detection model is provided, the method comprising the following steps: obtaining a time series dataset, the time series dataset comprising time series data of one or more factors; generating a training dataset based on the time series dataset; and training a neural network model using at least the training dataset to construct the time series anomaly detection model, wherein layer bias parameters in the neural network model are harmonized based on entropy information determined by the training dataset.

[0008] According to the method for constructing a time series anomaly detection model described in one embodiment of the present application, obtaining a time series dataset includes: taking the data generated by each factor in each time period of multiple consecutive time periods as the time series data of the factor; and obtaining the time series dataset based on the time series data of each factor of the one or more factors.

[0009] According to the method for constructing a time series anomaly detection model in one embodiment or any of the above embodiments of the present application, generating a training dataset based on the time series dataset includes: converting the time series dataset into a matrix form to generate the training dataset.

[0010] According to the method for constructing a time series anomaly detection model described in one embodiment of the present application or any of the above embodiments, converting the time series dataset into a matrix form to generate the training dataset includes: using the time series data of each factor in the time series dataset as a row of elements or a column of elements in a matrix to generate the training dataset, each element in the row of elements or the column of elements corresponds to the data generated by the factor in each time period.

[0011] According to the method for constructing a time series anomaly detection model described in one embodiment of the present application or any of the above embodiments, the method further includes: judging whether the time series data has an anomaly based on the continuous change trend of the time series data of each factor; determining the time series anomaly label corresponding to the time series data of each factor according to the judgment result to obtain a time series anomaly label set corresponding to the time series data set; and using a loss function to update the trained neural network model based on the time series anomaly label set and the output of the trained neural network model to construct the time series anomaly detection model.

[0012] According to the method for constructing a time series anomaly detection model described in one embodiment of the present application or any one of the above embodiments, the neural network model includes: a first layer, the first layer is constructed using a first activation function and a first layer bias parameter reconciled based on the entropy information determined by the training data set; a second layer, the second layer is constructed using a second activation function and a second layer bias parameter reconciled based on the entropy information determined by the training data set; and an output layer, the output layer is constructed using a softmax function.

[0013] According to the method for constructing a time series anomaly detection model according to one embodiment of the present application or any of the above embodiments, the neural network model is further constructed in the following manner: determining the input of the first activation function based on the first layer weights and the training data set; determining the input of the second activation function based on the second layer weights and the output of the first layer; and determining the input of the softmax function based on the output of the second layer.

[0014] According to the method for constructing a time series anomaly detection model according to one embodiment of the present application or any of the above embodiments, the layer bias parameters in the neural network model are harmonized in the following manner: the layer bias parameters are harmonized by multiplying the entropy information determined by the training data set by the layer bias parameters in the neural network model.

[0015] According to the method for constructing a time series anomaly detection model described in one embodiment or any one of the above embodiments of the present application, the entropy information is determined by the following steps: for each row or column element in the training data set, determining the probability factor of each element in the row or column relative to the elements in the row or column; using the probability factor to determine the information amount of the elements in the row or column; and using the information amount of the elements in the row or column as the entropy information of the time series data corresponding to the factors of the elements in the row or column.

[0016] According to the method for constructing a time series anomaly detection model described in one embodiment of the present application or any of the above embodiments, the probability factor is determined by the following steps: taking the ratio of each element in the row or the column to the sum of the elements in the row or the column as the probability factor.

[0017] According to the method for constructing a time series anomaly detection model according to one embodiment of the present application or any of the above embodiments, the time series anomaly detection model is at least used to correct the weights of one or more factors according to the time series data of one or more factors in the hierarchical analysis method.

[0018] According to a second aspect of the present application, a system for constructing a time series anomaly detection model is provided, the system comprising: a memory; a processor coupled to the memory; and a computer program stored on the memory and running on the processor, wherein the execution of the computer program results in the following operations: obtaining a time series dataset, the time series dataset comprising time series data of one or more factors; generating a training dataset based on the time series dataset; and training a neural network model using at least the training dataset to construct the time series anomaly detection model, wherein the layer bias parameters in the neural network model are harmonized based on entropy information determined by the training dataset.

[0019] According to a third aspect of the present application, a computer-readable storage medium is provided, which includes instructions, and the instructions, when run, execute the steps of the method for constructing a time series anomaly detection model according to the first aspect of the present application.

[0020] According to a fourth aspect of the present application, a hierarchical analysis method is provided, which includes the following steps: obtaining a first weight of each factor in a plurality of factors associated with a target task; obtaining time series data of each factor, and obtaining a second weight of each factor in the plurality of factors based on at least the time series data of each factor using a time series anomaly detection model constructed according to the method for constructing a time series anomaly detection model described in the first aspect of the present application; and determining a target weight corresponding to each factor in the plurality of factors based on the first weight and the second weight.

[0021] According to a fifth aspect of the present application, a hierarchical analysis system is provided, comprising: a memory; a processor coupled to the memory; and a computer program stored on the memory and running on the processor, wherein the running of the computer program results in the following operations: obtaining a first weight of each of a plurality of factors associated with a target task; obtaining time series data of each factor, and obtaining a second weight of each of the plurality of factors based at least on the time series data of each factor using a time series anomaly detection model constructed according to the method for constructing a time series anomaly detection model described in the first aspect of the present application; and determining a target weight corresponding to each of the plurality of factors based on the first weight and the second weight.

[0022] According to a sixth aspect of the present application, a computer-readable storage medium is provided, which includes instructions, and the instructions, when run, execute the steps of the hierarchical analysis method according to the fourth aspect of the present application.

[0023] According to one or more embodiments of the present application, in the process of determining the target weight of each factor among multiple factors associated with the target task, the hierarchical analysis scheme proposed, in addition to determining the first weight of each factor, also determines the second weight of each factor based on the time series data of each factor through the constructed time series anomaly detection model. In this way, in the process of determining the target weight of the factor, the anomalies that may occur in the factor value itself as time changes are fully considered, avoiding the allocation of larger weights to abnormal factors in the hierarchical analysis process, and significantly improving the hierarchical analysis results. According to one or more embodiments of the present application, the scheme for constructing a time series anomaly detection model innovatively proposes an entropy coefficient-harmonized neural network model, which can reconcile the layer bias parameters in the neural network model based on the entropy information determined by the training data set, thereby improving the training efficiency of the time series anomaly detection model and the accuracy of identifying data time series anomalies using the time series anomaly detection model. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The above and / or other aspects and advantages of the present application will become clearer and easier to understand through the following description of various aspects in conjunction with the accompanying drawings, in which the same or similar elements are represented by the same reference numerals. The accompanying drawings include:

[0025] FIG1 shows a schematic diagram of a typical AHP model.

[0026] FIG2 shows a flowchart of a method for building a time series anomaly detection model according to one or more embodiments of the present application.

[0027] FIG3 shows a block diagram of a system for building a time series anomaly detection model according to one or more embodiments of the present application.

[0028] FIG4 shows a flowchart of a hierarchical analysis method according to one or more embodiments of the present application.

[0029] FIG5 shows a block diagram of a hierarchical analysis system according to one or more embodiments of the present application. DETAILED DESCRIPTION

[0030] The present application is described more fully below with reference to the accompanying drawings, which illustrate exemplary embodiments of the present application. However, the present application may be implemented in various forms and should not be construed as being limited to the embodiments set forth herein. The foregoing embodiments are provided to make the disclosure herein comprehensive and complete, so as to more fully convey the scope of protection of the present application to those skilled in the art.

[0031] In this specification, terms such as "comprise" and "include" indicate that in addition to the units and steps directly and clearly stated in the specification and claims, the technical solution of this application does not exclude the situation where it has other units and steps that are not directly or clearly stated.

[0032] Unless otherwise specified, terms such as "first" and "second" do not indicate the order of units in terms of time, space, size, etc., but are only used to distinguish between units.

[0033] Hereinafter, various exemplary embodiments according to the present application will be described in detail with reference to the accompanying drawings.

[0034] FIG1 shows a schematic diagram of a typical AHP model.

[0035] As shown in FIG1 , the AHP model 100 hierarchically processes the multi-factor decision-making problem into three layers, wherein the first layer is the problem overview, the second layer is the factor system, and the third layer is the candidate decision.

[0036] In Figure 1, the first-level problem summary PSD (Problem Simple Description) is used to describe the decision problem to be solved. The second-level factor system includes multiple factors involved in processing the decision problem to be solved. The multiple factors can be represented as a set M = {m1,m2,m3,…,m k ,…,m n}, where m1, m2, m3, …, m k ,…,m n Represents multiple factors, n represents the size of the set M, that is, the number of factors. The candidate decisions of the third layer include multiple candidate decisions involved in processing the decision problem to be solved, and the multiple candidate decisions can be represented as a set TC = {tc1, tc2, tc3, ..., tc k ,…,tc o}, where tc1, tc2, tc3, …, tc k ,…,tc o Represents multiple candidate decisions, and o represents the size of the set TC, i.e., the number of candidate decisions. It should be noted that the processing of the candidate decision layer varies depending on the decision problem. For example, in a single-choice decision problem, a single candidate decision from the candidate decision set TC is output; in a multi-objective evaluation context, each candidate decision in the candidate decision set TC and its corresponding score can be output.

[0037] As an example, referring to Figure 1, the decision problems to be solved described in the first-level problem overview can include enterprise capability assessment, merchant value assessment, customer demand rating, supplier selection, manufacturing system, and strategic decision-making. For example, when the decision problem to be solved described in the first-level problem overview is merchant value assessment, the second-level factor system can be implemented as M = {m1, m2, m3, m4, m5} = {turnover, profit, market share, number of customers, number of transactions}, and the candidate decisions at the third level can be implemented as merchant value scoring.

[0038] In traditional hierarchical analysis methods (AHPs), the process of assigning weights to multiple factors relies solely on comparing the importance of these factors, ignoring the fact that factor values ​​themselves may exhibit anomalies over time. This results in assigning larger weights to anomalous factors during the AHP process, leading to significant errors in the AHP analysis results. For example, when using the AHP method to assess merchant value, turnover, as one of several factors, may exhibit temporal anomalies over time. In this case, if a larger weight is assigned to a statistically inaccurate turnover based solely on comparing the importance of these factors, the AHP analysis results will be significantly inaccurate, making the merchant value assessment unreliable.

[0039] Therefore, it is expected that the importance of the factors themselves can be considered while considering the temporal changes of the factor values ​​during the execution of the hierarchical analysis method. That is, it is expected to propose a hierarchical analysis method that can integrate the temporal change information of the factors to avoid assigning larger weights to abnormal factors during the hierarchical analysis process.

[0040] FIG2 shows a flowchart of a method for building a time series anomaly detection model according to one or more embodiments of the present application.

[0041] As shown in FIG2 , in step S201 , a time series data set is acquired, where the time series data set includes time series data of one or more factors.

[0042] Optionally, in step S201 , data generated by each factor in each of a plurality of consecutive time periods may be used as time series data of the factor, and a time series data set may be obtained based on the time series data of each of the one or more factors.

[0043] In one embodiment, when using the analytic hierarchy process to evaluate merchant value, factors associated with evaluating merchant value may include, for example, turnover, profit, market share, customer base, and number of transactions. For example, for one of the factors, turnover, the accumulated turnover for each of 36 consecutive months may be obtained as the time series data for that factor; for one of the factors, profit, the accumulated profit for each of 36 consecutive months may be obtained as the time series data for that factor. Similarly, for the factors, turnover, profit, market share, customer base, and number of transactions, the data generated for each of 36 consecutive months may be obtained as the time series data for the corresponding factors, thereby obtaining a time series dataset including time series data for multiple factors, namely turnover, profit, market share, customer base, and number of transactions.

[0044] In step S203 , a training dataset is generated based on the time series dataset.

[0045] Optionally, in step S203 , the time series dataset may be converted into a matrix form to generate a training dataset.

[0046] In one embodiment, the time series data of each factor in the time series data set can be used as a row of elements in a matrix to generate a training data set, and each element in the row of elements corresponds to the data generated by the factor in each time period. Exemplarily, for the factors turnover, profit, market share, customer volume, and number of transactions, after obtaining a time series data set including the time series data of multiple factors turnover, profit, market share, customer volume, and number of transactions, the time series data set can be converted into a training data set in the form of a matrix with 5 rows (i.e., the number of factors) and 36 columns (i.e., the number of time periods), wherein a row of elements in the matrix corresponds to the time series data of one factor, and each element in the row of elements corresponds to the data generated by the factor in each time period. In one embodiment, the time series data of each factor in the time series data set can be used as a column of elements in a matrix to generate a training data set, and each element in the column of elements corresponds to the data generated by the factor in each time period.

[0047] As an example, for N factors, the acquired time series data set C can be expressed as C = {C1, C2, ..., C N}, where C1, C2, ..., C N Represent the time series data of each factor respectively. As an example, the time series dataset C can be converted into a matrix form to generate a training dataset by the following formula (1):

[0048] Where t represents the number of consecutive time periods, N represents the number of factors, and each element in the matrix corresponds to the data generated by the factor in each time period. For example, element c 11 Represents the data generated by the first factor in the first time period of t consecutive time periods, element c ij Represents the data generated by the i-th factor in the j-th time period of t consecutive time periods. For example, the time series data C1 of the 1st factor can be converted into the 1st row element in the matrix, the time series data C2 of the 2nd factor can be converted into the 2nd row element in the matrix, and the time series data C N can be converted to the Nth row element in the matrix.

[0049] In step S205, a neural network model is trained using at least the training data set to construct a time series anomaly detection model, and layer bias parameters in the neural network model are adjusted based on entropy information determined by the training data set.

[0050] Alternatively, the neural network model may be implemented as an entropy regression model, which may include a first layer, a second layer, and an output layer, wherein the first layer is constructed using a first activation function and a first layer bias parameter reconciled based on entropy information determined by a training data set, the second layer is constructed using a second activation function and a second layer bias parameter reconciled based on entropy information determined by the training data set, and the output layer is constructed using a softmax function. Alternatively, the input of the first activation function may be determined based on the first layer weights and the training data set, the input of the second activation function may be determined based on the second layer weights and the output of the first layer, and the input of the softmax function may be determined based on the output of the second layer. Alternatively, the layer bias parameters may be reconciled by multiplying the entropy information determined by the training data set by the layer bias parameters in the neural network model.

[0051] In one embodiment, the neural network model can be implemented by the following formula (2):

[0052] Among them, the above three equations represent the first layer, second layer and output layer of the neural network model respectively, Z1 represents the output of the first layer, and Z2 represents the output of the second layer. Represents the output of the neural network model, σ represents the activation function e represents the natural logarithm, W1 represents the weight of the first layer, C i represents the training data set (for example, it can be obtained by referring to the above formula (1), i represents the number of factors), H i Represented by the training data set C i Determine the entropy information, b1 represents the bias of the first layer, W2 represents the weight of the second layer, b2 represents the bias of the second layer, It can be applied to every element of Z2, x∈Z2.

[0053] Optionally, the entropy information Hi can be determined by the following steps: for the training data set C i For each row or column element in the data, determine a probability factor for each element in the row or column relative to the elements in the row or column; determine the information content of the elements in the row or column using the probability factor; and use the information content of the elements in the row or column as the entropy information of the time series data corresponding to the factors of the elements in the row or column. Alternatively, the ratio of each element in the row or column to the total sum of the elements in the row or column can be used as the probability factor.

[0054] In one embodiment, the probability factor of each element relative to the elements in its row can be calculated using the following formula (3):

[0055] in, Represents element c ij The probability factor relative to the element in its row, Represents element c ij The sum of the elements in the row.

[0056] In one embodiment, the entropy information of the time series data of the factor can be calculated by the following formula (4):

[0057] Among them, H i Time series data C representing factors i Entropy information, Represents element c ij The probability factor of the element in its row can be calculated by referring to the above formula (3), Indicates base 2 It is understood that the entropy information of each row element in the matrix form of the training data set as shown in the above formula (1) can be determined using formula (4). For example, for the first row element c of the time series data corresponding to the first factor in the matrix form of the training data set as shown in the above formula (1), 11 ,c 12 ,...,c 1t , the above formula (3) can be used to calculate the first row element c 11 ,c 12 ,...,c 1t Each element in is relative to the element c in its row 11 ,c 12 ,...,c 1tThe probability factor, for example, the element c 11 With the first row element c 11 ,c 12 ,...,c 1t The ratio of the sum of 11 The probability factor p 11 , the element c 12 With the first row element c 11 ,c 12 ,...,c 1t The ratio of the sum of 12 The probability factor p 12 , and so on. After calculating the first row element c 11 ,c 12 ,...,c 1t Each element in is relative to the element c in its row 11 ,c 12 ,...,c 1t The probability factor p 11 ,p 12 ,...,p 1t Afterwards, the first row element c can be determined using formula (4) 11 ,c 12 ,...,c 1y Entropy information, that is, for each probability factor calculate And calculate each Sum to get the first row element c 11 ,c 12 ,...,c 1t Similarly, the entropy information H1 of the i-th row element in the training data set in matrix form as shown in the above formula (1) can be calculated through the above process. i .

[0058] By adjusting the layer bias parameters in the neural network model (for example, the layer biases b1 and b2 in the above formula (2)) based on the entropy information determined by the training data set, there is no need to collect a large number of time series anomaly samples, thereby improving the training efficiency of the time series anomaly detection model and the accuracy of identifying data time series anomalies using the time series anomaly detection model.

[0059] Optionally, the method for constructing a time series anomaly detection model according to one or more embodiments of the present application may further include determining whether the time series data is abnormal based on the continuous change trend of the time series data of each factor, determining the time series anomaly label corresponding to the time series data of each factor based on the judgment result to obtain a time series anomaly label set corresponding to the time series data set, and using a loss function based on the time series anomaly label set and the output of the trained neural network model to update the trained neural network model to construct a time series anomaly detection model. Optionally, the loss function can be used to adjust the parameters of the trained neural network model until the loss function is in a convergence state to obtain a time series anomaly detection model.

[0060] In one embodiment, for N factors, the acquired time series data set C can be expressed as C={C1, C2, ..., C N}, where C1, C2, ..., C N Represents the time series data of each factor respectively. As an example, the time series anomaly label of the time series data of the factor without abnormality can be set to 0 and the time series anomaly label of the time series data of the factor with abnormality can be set to 1, thereby obtaining the time series anomaly label set Y = {y1, y2, ..., y N}, where y1,y2,...,y N Represent C1, C2, ..., C N The timing anomaly label of .

[0061] In one embodiment, the loss function can be constructed by the following formula (5):

[0062] Where N represents the number of factors, y (i) represents the time series anomaly label of the time series data of the i-th factor, represents the output of the trained neural network model of the i-th factor, and ln represents the logarithm with base e.

[0063] The time series anomaly detection model constructed by the method for constructing a time series anomaly detection model according to one or more embodiments of the present application can be used to detect anomalies in time series data. Optionally, the time series anomaly detection model constructed by the method for constructing a time series anomaly detection model according to one or more embodiments of the present application can be used to modify the weights of one or more factors in the hierarchical analysis method based on the time series data of one or more factors. This modification process will be described in detail below.

[0064] According to one or more embodiments of the present application, the method for constructing a time series anomaly detection model innovatively proposes an entropy coefficient-harmonized neural network model, which can harmonize the layer bias parameters in the neural network model based on the entropy information determined by the training data set, thereby improving the training efficiency of the time series anomaly detection model and the accuracy of identifying data time series anomalies using the time series anomaly detection model.

[0065] FIG3 shows a block diagram of a system for building a time series anomaly detection model according to one or more embodiments of the present application.

[0066] As shown in FIG3 , the system 30 for building a time series anomaly detection model includes a communication unit 310 , a memory 320 (e.g., a non-volatile memory such as a flash memory, a ROM, a hard drive, a magnetic disk, an optical disk, etc.), a processor 330 , and a computer program 340 stored on the memory 320 and executable on the processor 330 .

[0067] The communication unit 310 serves as a communication interface and is configured to establish a communication connection between the system 30 for constructing a time series anomaly detection model and an external device or network (eg, a cloud server, a mobile terminal, etc.).

[0068] The memory 320 stores a computer program 340 executable by the processor 330. In addition, the memory 320 may also store data (eg, a time series data set, a training data set, etc.) generated when the processor 330 executes the computer program and data or commands received from the outside via the communication unit 310.

[0069] The processor 330 is configured to execute the computer program 340 to implement the method for building a time series anomaly detection model according to one or more embodiments of the present application.

[0070] FIG4 shows a flowchart of a hierarchical analysis method according to one or more embodiments of the present application.

[0071] As shown in FIG. 4 , in step S401 , a first weight of each factor among a plurality of factors associated with a target task is obtained.

[0072] Optionally, in step S401, a comparison matrix can be constructed based on a pairwise comparison of the importance of multiple factors, the comparison matrix can be processed to determine the priorities of the multiple factors, and the comparison matrix can be used to perform a consistency check on the priorities of the multiple factors, and the first weight of each factor in the multiple factors can be determined using the comparison matrix that passes the consistency check.

[0073] In one embodiment, the elements of each column in the comparison matrix may be summed to obtain the sum of the elements of each column, each element in the comparison matrix may be divided by the sum of the elements of the corresponding column to generate a column-averaged comparison matrix, and the average value of the elements in each row of the column-averaged comparison matrix may be determined as the priority of the multiple factors.

[0074] As an example, multiple factors m1, m2, m3, ..., m associated with the target task can be constructed in an automated manner. k ,…,m n The comparison matrix A is obtained as shown in the following formula (6):

[0075] Among them, i represents the row number of the element in the matrix, j represents the column number of the element in the matrix, and a ii =1,a ji =1 / a ij ,n represents the number of factors (also the dimension of the matrix), a ij Can be based on factor m i Compared to factor m j to determine the degree of importance.

[0076] As an example, after obtaining the contrast matrix A shown in the above formula (6), the elements of each column j in the contrast matrix A can be summed using the following formula (7) to obtain the sum of the elements of each column j:

[0077] Among them, i represents the row number of the element in the matrix, j represents the column number of the element in the matrix, and a ij represents the elements in the contrast matrix A, and n represents the number of factors (also the dimension of the matrix).

[0078] As an example, after obtaining the sum of the elements in each column, each element in the contrast matrix A can be divided by the sum of the elements in the corresponding column to generate a column-averaged contrast matrix A' using the following formula (8):

[0079] Among them, i represents the row number of the element in the matrix, j represents the column number of the element in the matrix, and a ij ' represents the elements in the contrast matrix A' after column averaging, and n represents the number of factors (also the dimension of the matrix).

[0080] As an example, after obtaining the column-averaged contrast matrix A', the average value p of each row element in the column-averaged contrast matrix A' can be determined by the following formula (9): i Taking priority of multiple factors:

[0081] Among them, i represents the row number of the element in the matrix, j represents the column number of the element in the matrix, and a ij ' represents the elements in the contrast matrix A' after column averaging, and n represents the number of factors (also the dimension of the matrix). i After that, the priority of n factors can be expressed as P = [p1, p2, ..., p n ].

[0082] In one embodiment, the maximum eigenvalue of the comparison matrix can be determined based on the priorities of multiple factors, a consistency index of the comparison matrix can be determined based on the maximum eigenvalue of the comparison matrix and the number of factors, a random index of the comparison matrix can be determined based on the dimension of the comparison matrix, and a consistency check can be performed on the comparison matrix based on the consistency index and the random index of the comparison matrix. In one embodiment, when the ratio of the consistency index of the comparison matrix to the random index is less than a predetermined threshold, it can be determined that the comparison matrix passes the consistency check, and then the comparison matrix that passes the consistency check can be processed to determine the priorities of the multiple factors and the priorities of the multiple factors are used as the first weights of the corresponding elements.

[0083] As an example, after obtaining the priority P representing multiple factors, the maximum eigenvalue λ of the comparison matrix A can be determined by the following formula (10): max :

[0084] Among them, i represents the number of rows in the matrix where the element is located, n represents the number of factors (also the dimension of the matrix), and p i represents the i-th element in vector P, [AP] i Represents the i-th element in [AP].

[0085] As an example, in determining the maximum eigenvalue λ of the contrast matrix A max Afterwards, the consistency index CI of the comparison matrix A can be determined by the following formula (11):

[0086] Where n represents the number of factors (also the dimension of the matrix), λ max Denotes the maximum eigenvalue of the contrast matrix A.

[0087] As an example, the random index RI of the comparison matrix A can also be determined according to the dimension n of the comparison matrix A, for example, by using a mathematical table lookup method to determine the random index RI of the comparison matrix according to the dimension n of the comparison matrix A. As an example, after determining the consistency index CI and the random index RI of the comparison matrix A, the consistency ratio CR of the comparison matrix A can be determined by the following formula (12): CR = CI / RI Formula (12)

[0088] For example, when the consistency ratio CR is less than 0.1, it can be determined that the comparison matrix passes the consistency test. Then, the comparison matrix that passes the consistency test can be processed to determine the priorities of the multiple factors (for example, by means of formulas (7) to (9) described above) and the priorities of the multiple factors are used as the first weights of the corresponding elements. For example, the priorities of the determined n factors can be expressed as a priority weight matrix They can be used as the first weights of the n factors respectively.

[0089] In step S403, the time series data of each factor is obtained, and the time series anomaly detection model constructed by the method for constructing a time series anomaly detection model according to one aspect of the present application is used to obtain the second weight of each factor in the multiple factors based at least on the time series data of each factor.

[0090] Optionally, in step S403, the data generated by each factor in each of a plurality of consecutive time periods can be used as the time series data of the factor, and the time series data of each factor can be used as a row of elements or a column of elements in a time series matrix to generate a time series matrix in matrix form, each element in the row of elements or the column of elements corresponds to the data generated by the factor in each time period, and the time series matrix is ​​input into the time series anomaly detection model to obtain a second weight of each of the multiple factors.

[0091] As an example, for the i-th factor, the acquired time series data B i It can be expressed as B i ={b 11 ,b 12 ,b 13 ,…,b 1k ,…,b 1t}, where t represents the number of consecutive time periods, b 11 ,b 12 ,b 13 ,…,b 1k ,…,b 1t Represents the data generated in each time period. As an example, for n factors, the obtained time series matrix B can be expressed as:

[0092] Among them, t represents the number of consecutive time periods, n represents the number of factors, and each element in the matrix corresponds to the data generated by the factor in each time period, such as element b 11 Represents the data generated by the first factor in the first time period of t consecutive time periods, element b ij represents the data generated by the i-th factor in the j-th time period of t consecutive time periods. For example, the first row of elements in the time series matrix B represents the time series data B1 of the first factor, the second row of elements in the time series matrix B represents the time series data B2 of the second factor, and so on.

[0093] As an example, the timing matrix B shown in formula (13) above can be input into the timing anomaly detection model constructed according to the method for constructing a timing anomaly detection model described in one aspect of the present application to obtain the second weight of each factor in the multiple factors, which can be expressed as a timing anomaly weight matrix

[0094] In step S405 , a target weight corresponding to each of the multiple factors is determined based on the first weight and the second weight.

[0095] Optionally, in step S405, the product of the first weight and the second weight may be determined as the target weight corresponding to each of the multiple factors. As an example, the target weight It can be determined by the following formula (14) and formula (15):

[0096] in, represents the first weight of the factor, represents the second weight of the factor, n represents the number of factors, represents the first weight of the i-th factor, represents the second weight of the i-th factor, and represents the target weight of the i-th factor.

[0097] Optionally, the hierarchical analysis method according to one or more embodiments of the present application may further include obtaining a score corresponding to each of the multiple factors, and determining a hierarchical analysis result based on the score of each factor and the target weight corresponding to the factor. Optionally, the hierarchical analysis result may be determined based on the sum of the product of the score of each factor and the target weight corresponding to the factor. As an example, the decision value S obtained as the hierarchical analysis result using the hierarchical analysis method may be determined by the following formula (16):

[0098] in, represents the target weight of the factor, represents the normalized factor score, n represents the number of factors, represents the target weight of the ith factor, and represents the score of the i-th factor.

[0099] The following will take supplier value evaluation as an example to further describe the specific implementation of the hierarchical analysis method according to one or more embodiments of the present application.

[0100] As described above in conjunction with Figure 1, when the decision problem to be solved described in the first-level problem overview is merchant value assessment, the second-level factor system can include 5 factors, namely M = {m1, m2, m3, m4, m5} = {turnover, profit, market share, number of customers, number of transactions}, and the candidate decisions at the third level can be implemented as a value score for the merchant.

[0101] As an example, the comparison matrix A shown in the following formula (17) can be obtained based on m1, m2, m3, m4, and m5 in an automated manner:

[0102] As an example, the comparison matrix A can be processed with reference to the above formulas (7) to (12) and their related descriptions to determine the priorities of the five factors, and the comparison matrix A can be used to perform a consistency check using the priorities of the five factors, and the first weight of each of the five factors can be determined using the comparison matrix that passes the consistency check. For example, the priorities of the five factors can be expressed as a priority weight matrix It can be used as the first weight of the five factors respectively.

[0103] Next, the time series data of the five factors can be obtained, and the time series anomaly detection model constructed by the method for constructing a time series anomaly detection model according to one aspect of the present application can be used to obtain the second weight of each of the five factors based on at least the time series data of each factor. For example, for one of the factors, turnover, the accumulated turnover in each month of 36 consecutive months can be obtained as the time series data of the factor; for one of the factors, profit, the accumulated profit in each month of 36 consecutive months can be obtained as the time series data of the factor. Similarly, for the five factors of turnover, profit, market share, customer volume, and number of transactions, the data generated in each month of 36 consecutive months can be obtained as the time series data of the corresponding factor, thereby obtaining the time series data of the five factors of turnover, profit, market share, customer volume, and number of transactions.

[0104] As an example, referring to the description of step S403 above, the time series data of the above five factors within 36 months can be converted into a matrix form to construct a time series matrix B with 5 rows and 36 columns as shown in the following formula (18):

[0105] As an example, the timing matrix B shown in formula (18) above can be input into the timing anomaly detection model constructed according to the method for constructing a timing anomaly detection model described in one aspect of the present application to obtain the second weight of each of the five factors, which can be expressed as the timing anomaly weight matrix

[0106] Then, the target weight corresponding to each of the five factors can be determined based on the first weight and the second weight of the five factors. Alternatively, the product of the first weight and the second weight of the five factors can be determined as the target weight of each of the five factors. As an example, the target weight It can be determined by referring to the above formula (14) and formula (15)

[0107] After determining the target weight of each of the five factors, the score corresponding to each of the five factors can be obtained, and the hierarchical analysis result can be determined based on the score of the five factors and the target weight corresponding to the factor. Alternatively, the hierarchical analysis result can be determined based on the sum of the product of the score of each factor and the target weight corresponding to the factor. As an example, the decision value S obtained as the hierarchical analysis result using the hierarchical analysis method can be determined using the following formula (19):

[0108] in, represents the normalized scores of the five factors. It represents the target weights of the five factors, and the decision value of 0.184 can be used as the valuation result for evaluating the supplier's value.

[0109] FIG5 shows a block diagram of a hierarchical analysis system according to one or more embodiments of the present application.

[0110] As shown in FIG5 , the hierarchical analysis system 50 includes a communication unit 510 , a memory 520 (e.g., a non-volatile memory such as a flash memory, a ROM, a hard disk drive, a magnetic disk, an optical disk), a processor 530 , and a computer program 540 stored on the memory 520 and executable on the processor 530 .

[0111] The communication unit 510 serves as a communication interface and is configured to establish a communication connection between the hierarchical analysis system 50 and an external device or network (eg, a cloud server, a mobile terminal, etc.).

[0112] The memory 520 stores a computer program 540 that can be executed by the processor 530. In addition, the memory 520 can also store data generated when the processor 530 executes the computer program (e.g., the first weight of the factor, the second weight of the factor, the target weight of the factor, etc.) and data or commands received from the outside via the communication unit 510.

[0113] The processor 530 is configured to execute the computer program 540 to implement the hierarchical analysis method according to one or more embodiments of the present application.

[0114] The present application may also be implemented as a computer-readable storage medium, the computer storage medium including instructions that, when run, execute a method for building a time series anomaly detection model according to one aspect of the present application. Furthermore, the present application may also be implemented as a computer-readable storage medium, the computer storage medium including instructions that, when run, execute a hierarchical analysis method according to one aspect of the present application.

[0115] In the applicable situation, the combination of hardware, software or hardware and software can be used to realize the various embodiments provided by the application. Moreover, in the applicable situation, without departing from the scope of the application, the various hardware components and / or software components set forth herein can be combined into a composite component comprising software, hardware and / or both. In the applicable situation, without departing from the scope of the application, the various hardware components and / or software components set forth herein can be divided into a subcomponent comprising software, hardware or both. In addition, in the applicable situation, it is contemplated that the software component can be implemented as a hardware component, and vice versa.

[0116] Software according to the present application (such as program code and / or data) can be stored on one or more computer storage media. It is also contemplated that the software identified herein can be implemented using one or more general or special computers and / or systems that are networked and / or otherwise. Where applicable, the order of the various steps described herein can be changed, combined into composite steps and / or divided into sub-steps to provide the features described herein.

[0117] The embodiments and examples set forth herein are provided to best illustrate embodiments according to the present application and its specific applications, and thereby enable those skilled in the art to make and use the present application. However, those skilled in the art will appreciate that the above description and examples are provided for ease of illustration and example only. The descriptions set forth are not intended to be exhaustive of all aspects of the present application or to limit the present application to the precise forms disclosed.

Claims

1. A method for constructing a time series anomaly detection model, characterized in that: The method comprises the following steps: Acquire a time series data set, wherein the time series data set includes time series data of one or more factors; Generate a training data set based on the time series data set; as well as A neural network model is trained using at least the training data set to construct the time series anomaly detection model, and layer bias parameters in the neural network model are adjusted based on entropy information determined by the training data set.

2. The method according to claim 1, wherein obtaining a time series data set comprises: taking the data generated by each factor in each of a plurality of consecutive time periods as the time series data of the factor; as well as The time series data set is obtained based on the time series data of each factor of the one or more factors.

3. The method according to claim 1, wherein generating a training data set based on the time series data set comprises: The time series data set is converted into a matrix form to generate the training data set.

4. The method according to claim 2 or 3, wherein converting the time series data set into a matrix form to generate the training data set comprises: The training data set is generated by taking the time series data of each factor in the time series data set as a row of elements or a column of elements in a matrix, wherein each element in the row of elements or the column of elements corresponds to the data generated by the factor in each time period.

5. The method according to claim 1, wherein the method further comprises: Determine whether the time series data is abnormal based on the continuous change trend of the time series data of each factor; Determine the time series anomaly label corresponding to the time series data of each factor according to the judgment result to obtain a time series anomaly label set corresponding to the time series data set; as well as The trained neural network model is updated based on the time series anomaly label set and the output of the trained neural network model using a loss function to construct the time series anomaly detection model.

6. The method according to claim 1, wherein the neural network model comprises: A first layer constructed using a first activation function and a first layer bias parameter reconciled based on entropy information determined from the training data set; a second layer constructed using a second activation function and a second layer bias parameter reconciled based on entropy information determined from the training data set; as well as The output layer is constructed using the softmax function.

7. The method according to claim 6, wherein the neural network model is further constructed by: Determining an input of the first activation function based on the first layer weights and the training data set; determining an input of the second activation function based on the second layer weights and the output of the first layer; and An input of the softmax function is determined based on the output of the second layer.

8. The method of claim 1, wherein the layer bias parameters in the neural network model are adjusted by: The layer bias parameters in the neural network model are adjusted by multiplying the entropy information determined by the training data set by the layer bias parameters.

9. The method of claim 4, wherein the entropy information is determined by: For each row or column of elements in the training data set, determine a probability factor of each element in the row or column relative to the elements in the row or column; Determining the information content of the elements in the row or the column using the probability factor; as well as The amount of information of the elements in the row or the column is taken as the entropy information of the time series data of the factors corresponding to the elements in the row or the column.

10. The method according to claim 9, wherein the probability factor is determined by the following steps: The ratio of each element in the row or the column to the total of the elements in the row or the column is used as the probability factor.

11. The method according to claim 1, wherein the time series anomaly detection model is used at least to modify the weights of one or more factors according to the time series data of the one or more factors in the hierarchical analysis method.

12. A system for building a time series anomaly detection model, characterized in that: The system comprises: Memory; a processor coupled to the memory; and A computer program stored on the memory and running on the processor, the execution of the computer program causing the following operations: Acquire a time series data set, wherein the time series data set includes time series data of one or more factors; Generate a training data set based on the time series data set; as well as A neural network model is trained using at least the training data set to construct the time series anomaly detection model, and layer bias parameters in the neural network model are adjusted based on entropy information determined by the training data set.

13. The system of claim 12, wherein the execution of the computer program causes obtaining a time series data set to include: taking the data generated by each factor in each of a plurality of consecutive time periods as the time series data of the factor; as well as The time series data set is obtained based on the time series data of each factor of the one or more factors.

14. The system of claim 12, wherein the execution of the computer program results in generating a training data set based on the time series data set comprising: The time series data set is converted into a matrix form to generate the training data set.

15. The system according to claim 13 or 14, wherein the execution of the computer program causes the time series data set to be converted into a matrix form to generate the training data set to include: The training data set is generated by taking the time series data of each factor in the time series data set as a row of elements or a column of elements in a matrix, wherein each element in the row of elements or the column of elements corresponds to the data generated by the factor in each time period.

16. The system of claim 12, wherein execution of the computer program further results in the following operations: Determine whether the time series data is abnormal based on the continuous change trend of the time series data of each factor; Determine the time series anomaly mark corresponding to the time series data of each factor according to the judgment result Tag to obtain a time series anomaly tag set corresponding to the time series data set; as well as The trained neural network model is updated based on the time series anomaly label set and the output of the trained neural network model using a loss function to construct the time series anomaly detection model.

17. The system of claim 12, wherein the neural network model comprises: A first layer constructed using a first activation function and a first layer bias parameter reconciled based on entropy information determined from the training data set; a second layer constructed using a second activation function and a second layer bias parameter reconciled based on entropy information determined from the training data set; as well as The output layer is constructed using the softmax function.

18. The system of claim 17, wherein the execution of the computer program further causes the neural network model to be further constructed by: Determining an input of the first activation function based on the first layer weights and the training data set; determining an input of the second activation function based on the second layer weights and the output of the first layer; and An input of the softmax function is determined based on the output of the second layer.

19. The system of claim 12, wherein execution of the computer program further causes layer bias parameters in the neural network model to be tuned by: The layer bias parameters in the neural network model are adjusted by multiplying the entropy information determined by the training data set by the layer bias parameters.

20. The system of claim 15, wherein execution of the computer program further causes the entropy information to be determined by: For each row or column of elements in the training data set, determine a probability factor of each element in the row or column relative to the elements in the row or column; Determining the information content of the elements in the row or the column using the probability factor; as well as The amount of information of the elements in the row or the column is taken as the entropy information of the time series data of the factors corresponding to the elements in the row or the column.

21. The system of claim 20, wherein execution of the computer program further causes the probability factor to be determined by: The ratio of each element in the row or the column to the total of the elements in the row or the column is used as the probability factor.

22. The system according to claim 12, wherein the time series anomaly detection model is used at least to modify the weights of one or more factors according to the time series data of the one or more factors in the hierarchical analysis method.

23. A computer-readable storage medium, characterized in that: The computer storage medium includes instructions, which, when executed, execute the method for building a time series anomaly detection model according to any one of claims 1 to 11.

24. A hierarchical analysis method, characterized in that: The analytic hierarchy process comprises the following steps: Obtaining a first weight of each factor of a plurality of factors associated with the target task; Obtaining time series data of each factor, and obtaining a second weight of each of the multiple factors based at least on the time series data of each factor by using a time series anomaly detection model constructed according to the method for constructing a time series anomaly detection model according to any one of claims 1 to 11; as well as A target weight corresponding to each of the plurality of factors is determined based on the first weight and the second weight.

25. The method according to claim 24, wherein obtaining a first weight of each factor of a plurality of factors associated with the target task comprises: constructing a comparison matrix based on a pairwise comparison of the importance of the plurality of factors; processing the comparison matrix to prioritize the plurality of factors; as well as A consistency check is performed on the comparison matrix using the priorities of the multiple factors, and a first weight of each of the multiple factors is determined using the comparison matrix that passes the consistency check.

26. The method of claim 25, wherein processing the comparison matrix to determine the priority of the plurality of factors comprises: Summing the elements of each column in the comparison matrix to obtain the sum of the elements of each column; Dividing each element in the contrast matrix by the sum of the elements in the corresponding column to generate a column-averaged contrast matrix; as well as An average value of each row element in the column average comparison matrix is ​​determined as the priority of the multiple factors.

27. The method according to claim 25, wherein performing consistency check on the comparison matrix using the priorities of the multiple factors comprises: Determining the maximum eigenvalue of the comparison matrix according to the priorities of the multiple factors; Determining a consistency index of the comparison matrix according to the maximum eigenvalue of the comparison matrix and the number of factors; Determining a random index of the comparison matrix according to the dimension of the comparison matrix; and A consistency check is performed on the comparison matrix based on the consistency index and the random index of the comparison matrix.

28. The method according to claim 27, wherein the method further comprises: In response to a ratio of a consistency index to a random index of the comparison matrix being less than a predetermined threshold, determining that the comparison matrix passes a consistency check; processing the comparison matrix that passed the consistency check to determine the priority of the plurality of factors; as well as The priorities of the multiple factors are used as the first weights of the corresponding elements.

29. The method according to claim 24, wherein the second weight of each factor in the plurality of factors is further obtained by: taking the data generated by each factor in each of a plurality of consecutive time periods as the time series data of the factor; Generate a time series matrix in a matrix form by using the time series data of each factor as a row of elements or a column of elements in the time series matrix, wherein each element in the row of elements or the column of elements corresponds to the data generated by the factor in each time period; and The timing matrix is ​​input into the timing anomaly detection model to obtain a second weight of each of the multiple factors.

30. The method of claim 24, wherein determining a target weight corresponding to each of the plurality of factors based on the first weight and the second weight comprises: A product of the first weight and the second weight is determined as a target weight corresponding to each of the plurality of factors.

31. The method of claim 24, wherein the method further comprises: Obtaining a score corresponding to each of the multiple factors; as well as A hierarchical analysis result is determined based on the score of each factor and the target weight corresponding to the factor.

32. The method according to claim 31, wherein determining the hierarchical analysis result based on the score of each factor and the target weight corresponding to the factor comprises: Based on the sum of the product of the score of each factor and the target weight corresponding to the factor To determine the hierarchical analysis results.

33. A hierarchical analysis system, characterized in that: The hierarchical analysis system comprises: Memory; a processor coupled to the memory; and A computer program stored on the memory and running on the processor, the execution of the computer program causing the following operations: Obtaining a first weight of each factor of a plurality of factors associated with the target task; Obtaining time series data of each factor, and obtaining a second weight of each of the multiple factors based at least on the time series data of each factor using a time series anomaly detection model constructed according to the method for constructing a time series anomaly detection model according to any one of claims 1 to 11; and A target weight corresponding to each of the plurality of factors is determined based on the first weight and the second weight.

34. The system of claim 33, wherein the execution of the computer program causes obtaining a first weight of each of a plurality of factors associated with the target task to include: constructing a comparison matrix based on a pairwise comparison of the importance of the plurality of factors; processing the comparison matrix to prioritize the plurality of factors; as well as A consistency check is performed on the comparison matrix using the priorities of the multiple factors, and a first weight of each of the multiple factors is determined using the comparison matrix that passes the consistency check.

35. The system of claim 34, wherein execution of the computer program causes processing the comparison matrix to determine the priority of the plurality of factors to include: Summing the elements of each column in the comparison matrix to obtain the sum of the elements of each column; Dividing each element in the contrast matrix by the sum of the elements in the corresponding column to generate a column-averaged contrast matrix; as well as An average value of each row element in the column average comparison matrix is ​​determined as the priority of the multiple factors.

36. The system of claim 34, wherein the execution of the computer program causes the consistency check of the comparison matrix using the priorities of the plurality of factors to include: Determining the maximum eigenvalue of the comparison matrix according to the priorities of the multiple factors; Determine a comparison matrix based on the maximum eigenvalue of the comparison matrix and the number of factors Consistency indicators; Determining a random index of the comparison matrix according to the dimension of the comparison matrix; and A consistency check is performed on the comparison matrix based on the consistency index and the random index of the comparison matrix.

37. The system of claim 36, wherein execution of the computer program further results in: In response to a ratio of a consistency index to a random index of the comparison matrix being less than a predetermined threshold, determining that the comparison matrix passes a consistency check; processing the comparison matrix that passed the consistency check to determine the priority of the plurality of factors; and The priorities of the multiple factors are used as the first weights of the corresponding elements.

38. The system of claim 33, wherein the execution of the computer program causes the second weight of each factor in the plurality of factors to be further derived by: taking the data generated by each factor in each of a plurality of consecutive time periods as the time series data of the factor; Generate a time series matrix in a matrix form by using the time series data of each factor as a row of elements or a column of elements in the time series matrix, wherein each element in the row of elements or the column of elements corresponds to the data generated by the factor in each time period; and The timing matrix is ​​input into the timing anomaly detection model to obtain a second weight of each of the multiple factors.

39. The system of claim 33, wherein execution of the computer program causes determining a target weight corresponding to each of the plurality of factors based on the first weight and the second weight to include: A product of the first weight and the second weight is determined as a target weight corresponding to each of the plurality of factors.

40. The system of claim 33, wherein execution of the computer program further results in: obtaining a score corresponding to each of the plurality of factors; and A hierarchical analysis result is determined based on the score of each factor and the target weight corresponding to the factor.

41. The system of claim 40, wherein the execution of the computer program causes the determination of a hierarchical analysis result based on the score of each factor and the target weight corresponding to the factor to include: The hierarchical analysis result is determined based on the sum of the product of the score of each factor and the target weight corresponding to the factor.

42. A computer-readable storage medium, characterized in that: The computer storage medium comprises instructions, which, when executed, execute the hierarchical analysis method according to any one of claims 24-32.

Images