TWI747334B - Fraud measurement detection device, method, program product and computer readable medium - Google Patents

Abstract

A fraud measurement detection device, method, program product and computer readable medium. The device includes a data collection module, a storage module, a data pre-process module, a calibration module and a fraud judgment module. When a measurement interval is judged as suspicious one and in the meantime the distribution is identified as being suspicious also then the judgment module would deduct the fraud data based on distribution. With the invention, real time analysis could be available remotely for the validity of measurement data to avoid fraud data by manual manipulation as well as to reflect truly quality level and detect possible quality problem in advance and reduce redundant inspection cost to save resource.

Description

Device, method, program product and computer readable medium for detecting data fraud

The present invention relates to a device, method, program product, and computer readable medium for machine learning and fraud identification, in particular to a device, method, program product, and computer readable medium for detecting data fraud.

In the past, fraud identification is mostly used in the banking, securities, and insurance industries. For example, US 8,862,526 B2, US 10,115,111 B2, and US 10,325,271 B2 are mainly used to generate a causal model corresponding to the user through the system, and use the causal model to predict the user’s next move. Expected behavior during the event. The Hidden Markov Model (HMM), which is commonly used for speech recognition or joint probability prediction, is used in CN 105608536 A for risk and misprediction in other industries, and it is also used for fraud identification in WO 2019037205 A1. Aspect.

For fraudulent behaviors or intentions of measurement equipment, most of them are detected by hardware means, for example, EP 2205947 A1 and EP 1564533 B1, etc., but the aforementioned patents fail to consider the versatility of the equipment.

Therefore, in order to solve the problem of repeated inspections caused by mistrust in the supply chain, as well as the cost of quality inspection resulting from repeated inspections and shorten the delivery time, the inventor proposes a data fraud detection device for detecting a piece of a measuring device. Measurement behavior, the device for detecting fraudulent data includes: a measurement data capture Take the module, read a measurement data and a measurement interval measured by the measurement equipment; a storage module, the signal is connected to the measurement data acquisition module, the storage module stores the measurement data and The measurement interval; a measurement data processing module, the signal is connected to the storage module, the measurement data processing module calculates its z-score based on a measurement data distribution of the measurement data and a corresponding specification value Distribute and obtain a set of characteristic values; an analog data generation module, a signal connected to the measurement data processing module, the analog data generates a simulated normal data and a simulated fraud data distribution, and obtain the simulated normal data and the simulation A set of simulated characteristic values of fraud data; a measurement data classification module, a signal is connected to the measurement data processing module and the analog data generation module, the measurement data classification module is based on the characteristic value set or the simulated characteristic The value set performs training and feature selection operations to establish a classification model to classify whether the measurement data distribution is a suspicious data distribution; a measurement interval abnormality detection module, the signal is connected to the storage module, the measurement interval is abnormal The detection module establishes an unsupervised learning model based on the measurement interval to determine whether the measurement interval is a suspicious interval; The measurement data classification module and the storage module, when the unsupervised learning model determines that the measurement interval is the suspicious interval, and the classification model classifies the measurement data distribution into the suspicious data distribution, the fraud data presumption model The group uses a hidden Markov model to estimate the corresponding fraud data.

Furthermore, a calibration module signal is connected to the measurement data acquisition module and the fraud data estimation module, the calibration module includes a simple object, and the calibration module obtains a calibration parameter by cooperating with the simple object, and the hidden The Markov model is established by the calibration parameters of the calibration module and the Baum-Welch algorithm and compared with the yield of an initial condition, and the presumption of fraud data is obtained by the hidden Markov model using the Viterbi algorithm. .

Further, the calibration module gives a calibration command, and the calibration module determines the calibration parameter according to the result of executing the calibration command on the simple object.

Wherein, the simulated data generating module uses random numbers to generate the simulated fraud data.

Among them, the measurement data acquisition module uses OCR (Optical Character Recognition) technology to read the measurement data.

Among them, the measurement data classification module uses Support Vector Machine (SVM) and/or Random Forests (Random Forests, RF) algorithms to establish the classification model.

Wherein, the measurement interval anomaly detection module uses Isolation Forests (IF) and/or z-score algorithm to establish the unsupervised learning model.

Wherein, the measurement data classification module uses self-service sampling to obtain the feature value set and the simulated feature value set.

Wherein, the measurement data processing module performs z-score distribution conversion on the measurement data distribution and the specification value.

The inventor further proposes a method for detecting data fraud, including: using a measurement data acquisition module to read a measurement data and a measurement interval measured by a measurement device; using a storage module to store the amount For the measurement data and the measurement interval, a measurement data processing module is used to calculate a feature value set according to a measurement data distribution of the measurement data and a corresponding specification value; a measurement data classification module is used according to The feature value set, or a simulated feature value set that simulates normal data and a simulated fraud data, is trained and feature selection operations are performed to establish a classification model to classify whether the measurement data distribution is a suspicious data distribution; use a quantity The measurement interval anomaly detection module establishes an unsupervised learning model based on the measurement interval to determine whether the measurement interval is a suspicious interval; and when the unsupervised learning model determines that the measurement interval is the suspicious interval, and the When the classification model classifies the measurement data distribution into the suspicious data distribution, a fraud data estimation module uses a hidden Markov model to infer the corresponding fraud data.

Further, a calibration module is used in conjunction with a simple object to obtain a calibration parameter; the hidden Markov model is established by the calibration parameter of the calibration module and the Baum-Welch algorithm and is obtained by comparing the yield rate with an initial condition, And the presumption of fraud data is obtained by the hidden Markov model using the Viterbi algorithm.

Further, the calibration module gives a calibration command, and the calibration module determines the calibration parameter according to the result of executing the calibration command on the simple object.

Wherein, the simulated normal data and the simulated fraud data are a historical measurement data and a historical measurement interval stored in the storage module, or the simulated normal data and the simulated fraud data are generated by an analog data generation module Distribution.

Wherein, the measurement data acquisition module uses OCR technology to read the measurement data.

Wherein, the measurement data classification module uses a support vector machine and/or random forest algorithm to establish the classification model.

Wherein, the measurement interval anomaly detection module uses an isolation forest and/or z-score algorithm to establish the unsupervised learning model.

Wherein, the measurement data classification module uses self-service sampling to obtain the feature value set and the simulated feature value set.

Wherein, the measurement data processing module performs z-score distribution conversion on the measurement data distribution and the specification value.

The inventor further provides a program product for executing the aforementioned data fraud detection method after being loaded into a computer device.

The present inventor further provides a computer-readable medium for performing the aforementioned data fraud detection method after being loaded into a computer device.

According to the above technical features, the following effects can be achieved:

1. The effectiveness of quality inspection measurement data can be analyzed remotely and instantaneously, avoiding fraud by artificially falsifying data, while reflecting the true quality level in real time, detecting possible quality problems early and reducing the resources and costs required for repeated quality inspections.

2. The simulated data generation module generates simulated normal data and simulated fraud data, and the measurement data classification module performs machine learning based on this to establish a classification model, which can shorten the time required to establish a classification model and improve the utility of the data fraud detection device sex.

3. Using an unsupervised learning model to identify suspicious intervals can objectively determine whether the measurement interval is a suspicious interval, and at the same time include the motive of fraudulent behavior and the actual observed suspicious phenomenon as a reference for judging whether it is fraudulent data.

4. The calibration module obtains the calibration parameters of the quality inspector for simple objects, as a parameter reference for establishing the hidden Markov model, and adds options other than the Baum-Welch algorithm to improve the stability of the data fraud detection device.

100: Detect data fraud device

1: Measurement data acquisition module

11: Measurement data

12: Measurement interval

2: Storage module

3: Measurement data processing module

31: Historical measurement data distribution

32: Historical measurement data z-score distribution

33: Nominal value

34: Lower limit

35: upper limit

36: z-score nominal value

37: z-score lower limit

38: z-score upper limit

39: z-score feature value collection

4: Analog data generation module

41: Simulation data z-score distribution

42: Simulated eigenvalue set

5: Measurement data classification module

S501: Perform training

S502: Feature selection operation

51: Classification model

52: Optimal eigenvalue set

6: Measurement interval anomaly detection module

61: Unsupervised learning model

S601: Isolated forest algorithm

S602: z-score algorithm

62: pollution parameters

63: Anomaly score

64: Threshold parameter

7: Calibration module

71: Simple Objects

72: blind test box

73: Correction instruction analysis generator

74: inner diameter

75: outer diameter

76: Column Height

77: Calibration parameters

78: Calibration parameter table

79: Incentive Index

8: Fraud data presumption module

81: Hidden Markov Model

811: State transition matrix

812: Observation Probability Matrix

813: initial state

82: Observation sequence

821: Possible to sample OK/NG collection

[The first figure] is a system block diagram of an embodiment of the present invention.

[The second figure] is the first functional block diagram of the embodiment of the present invention.

[Third Figure] is the second functional block diagram of the embodiment of the present invention.

[Fourth Figure] is the third functional block diagram of the embodiment of the present invention.

[Fifth Figure] is the fourth functional block diagram of the embodiment of the present invention.

[Figure 6] is the fifth functional block diagram of the embodiment of the present invention.

[Figure 7] is a functional block diagram 6 of the embodiment of the present invention.

[Figure 8] is a schematic diagram of the implementation of the calibration module of the embodiment of the present invention.

[Figure 9] is a schematic diagram of the calibration process of the embodiment of the present invention.

[Figure 10] is a functional block diagram of the fraud data estimation module of the embodiment of the present invention.

[Figure 11] is a schematic diagram of the process of inferring fraud data in an embodiment of the present invention.

Based on the above technical features, the main effects of the data fraud detection device, method, program product, and computer readable media of the present invention will be clearly presented in the following embodiments.

Please refer to the first figure, which discloses a data fraud detection device 100 according to an embodiment of the present invention, which is used to implement a data fraud detection method to detect a measurement behavior. It can also be used as a program product or a computer readable medium, which After entering a computer device, the method for detecting data fraud is executed. The device 100 for detecting data fraud includes: a measurement data acquisition module 1, a storage module 2, a measurement data processing module 3, and an analog data generation Module 4, a measurement data classification module 5, a measurement interval anomaly detection module 6, a correction module 7 and a fraud data estimation module 8.

Please refer to the first and second figures. The measurement data acquisition module 1 reads a measurement data 11 and a measurement interval 12 displayed by a measurement device in the quality inspection process. The measurement data capture The acquisition module 1 can be various electronic devices with camera functions, including but not limited to smart phones, tablet computers, etc., and application software can be installed on the measurement data acquisition module 1 to read images and use OCR technology Read the measurement data 11. The measurement interval 12 is the time interval from the last time the measurement data 11 is confirmed to this time the measurement data 11 is confirmed. After the measurement data capture module 1 completes each image capture, it can prompt a quality inspector operating the measurement equipment to confirm whether to continue image capture through voice, image, light, etc. The quality inspector can first preset the number of image captures to be captured, and the measurement data acquisition module 1 automatically proceeds to the completion of image capture after the quality inspector confirms the number of image captures.

The storage module 2 is signally connected to the measurement data acquisition module 1. The measurement data acquisition module 1 obtains the measurement data 11 and the measurement interval 12 and sends them to the storage module 2 as a The historical measurement data is stored, and the historical measurement data stored in the storage module 2 can be classified in subsequent steps.

Please refer to the first and third figures. The measurement data processing module 3 is signaled to the storage module 2. The measurement data processing module 3 can firstly store the historical measurement in the storage module 2 The data generates a historical measurement data distribution 31, calculates the average value and standard deviation, and performs z-score conversion, and obtains a historical measurement data z-score distribution 32. The measurement data processing module 3 can also perform z-score conversion on a preset specification value, such as a nominal value 33, a lower limit value 34, and an upper limit value 35, and obtain a z-score standard value respectively. Weighing 36, a z-score lower limit 37, and a z-score upper limit 38. After the measurement data processing module 3 obtains the historical measurement data z-score distribution 32 formed according to a set measurement quantity (for example, 30 pieces), the historical measurement data distribution 31 can also calculate the historical measurement data. A z-score feature value set 39 of the data z-score distribution 32, the z-score feature value set 39 includes but is not limited to descriptive statistics such as skewness and kurtosis. If the measurement quantity is not reached (for example, less than 30 pieces but greater than 15 pieces), the measurement data processing module 3 can use Bootstrapping to repeatedly sample to the measurement quantity and calculate the characteristic value, and then recurse For example, the central limit theorem is used 10,000 times to take the average value to generate the z-score feature value set 39.

Please refer to the first and fourth figures, and please match the third figure. The analog data generating module 4 is connected to the measurement data processing module 3, and a random number simulation can be used to generate an analog data z-score distribution 41. If the data in the simulation data z-score distribution 41 is lower than the z-score lower limit value 37 or higher than the z-score upper limit value 38, then the method includes but not limited to the following processing: Random sampling, replace with the z-score nominal value of 36, or replace with z=0 to generate data. The analog data generation module 4 counts A simulated feature value set 42 of the simulated data z-score distribution 41 is calculated. The simulated feature value set 42 includes but is not limited to descriptive statistics such as skewness and kurtosis.

Please refer to the first and fifth figures. The measurement data classification module 5 is signaled to connect the measurement data processing module 3 and the analog data generation module 4, and the measurement data classification module 5 is based on the z-score. The feature value set 39 or the simulated feature value set 42 is trained by supervised learning S501 and feature selection operation S502, and then a classification model 51 is established. The classification model 51 can be established using at least a support vector machine or a random forest algorithm. When using the support vector machine algorithm to establish the classification model 51, iterative feature elimination (Recursive Feature Elimination, RFE) and cross validation (Cross Validation, CV) can be used to find an optimal feature value set 52 to establish the classification model 51; When the random forest algorithm is used to establish the classification model 51, the feature importance (Feature Importance) can be calculated, and the feature whose feature importance is greater than the set mean is obtained as the optimal feature value set 52 for training to establish the classification model 51. The classification model 51 can be used to determine whether the optimal feature value set 52 that is newly input for classification is a suspicious data distribution. The simulated data generation module 4 first generates the simulated normal data and the simulated fraud data, and the measurement data classification module 5 then performs machine learning based on this to establish the classification model 51, which can shorten the need to establish the classification model 51 Time improves the practicability of the device 100 for detecting data fraud.

Please refer to the first and sixth figures. The measurement interval anomaly detection module 6 is signaled to the storage module 2. The measurement interval anomaly detection module 6 can use the isolated forest algorithm S601 or z-score calculation Method S602 establishes an unsupervised learning model 61. When the isolated forest algorithm S601 is used to establish the unsupervised learning model 61, the measurement interval anomaly detection module 6 first compares the measurement interval 12 stored in the storage module 2 and a preset pollution parameter 62 ( Contamination Parameter), the contamination parameter 62 is 3%, for example, an abnormality score 63 is calculated, and the measurement interval 12 where the abnormality score 63 is a negative value is set as an abnormal value to establish the unsupervised learning model 61. At the same time, the measurement interval anomaly detection module 6 can use the anomaly score The distribution of 63 is calculated by the quartile, and the third quartile and the first quartile are added or subtracted, such as twice the difference between the third quartile and the first quartile, as The new upper and lower limits of outliers can better define the outliers with relative scores. When using the z-score algorithm S602 to establish the unsupervised learning model 61, first take a threshold parameter (Threshold value) 64, the threshold parameter 64, such as 3, after z-score conversion, the measurement interval 12 exceeds ±3 It is set as an outlier.

Please refer to the fifth and sixth figures, and please match the first to third figures. When the measurement interval 12 is found to be an abnormal value, that is, a suspicious interval, the measurement interval anomaly detection module 6 will Communicate with the fraud data estimation module 8 to enable the measurement data classification module 5 to activate the classification model 51 to obtain the optimum value for the historical measurement data distribution 31 (for example, the past 30) stored in the storage module 2 The feature value set 52 performs judgment and classification. Using the unsupervised learning model 61 to identify the suspicious interval, it is possible to objectively determine whether the measurement interval 12 is the suspicious interval, and at the same time, the motive of the fraudulent behavior and the actually classified distribution of the suspicious data are included in whether it is a fraud data. Judgment reference, for example, if the measurement data 11 exceeds the specification value, the quality inspector is motivated to commit fraud, such as using the specification value to replace the measurement data 11, or repeated use meets the specification The value of the measurement data 11 will lengthen the measurement interval 12 and the historical measurement data distribution 31 will be distorted.

Please refer to the seventh to ninth figures, and please match the first and second figures. The calibration module 7 is connected to the measurement data acquisition module 1, and the calibration module 7 contains a complex number of known geometric dimensions. Simple object 71, a blind test box 72, and a calibration instruction analysis generator 73. The simple object 71 can be, for example, five hollow cylinders with the same appearance and very small differences in size, and have at least three sizes: an inner diameter 74, an outer diameter 75 and a column height 76, and each size is independent of each other. . Before the quality inspector uses the data fraud detection device 100 for the first time, the calibration command analysis generator 73 uses the measurement data acquisition module 1 to give a calibration command, such as sequentially measuring one of the simple objects 71 Size (e.g. the inner diameter 74) several times to search Collect the initial distribution of the measurement interval 12 and send it to the calibration command analysis generator 73, and generate a humanized behavior test item according to the distribution of the measurement interval 12: the calibration module 7 can prompt and perform different levels of An incentive index 79 encourages the quality inspector to randomly select the simple object 71 within a given time (such as m seconds) and measure another size (such as the outer diameter 75) until the measurement result meets the preset designation Condition (for example, there are n measurement results greater than or less than d), the excitation index 79 such as i%, the collected measurement results include the number of times, the measurement data 11 and the measurement interval 12 are determined to be stored in a calibration parameter table 78 A calibration parameter 77 in the QC, the calibration parameter 77 responds to the behavior of the quality inspector to reflect the calibration command, and the calibration parameter 77 can be used as an input to an observation probability matrix 812 of a hidden Markov model 81 [the hidden Markov Please match the model 81 and the observation probability matrix 812 with the tenth figure]. For example, the calibration parameter table 78 can be filled with A, B, C, D... in the column headings, and 6, 7... in the column headings. Each of the correction parameters 77 in the correction parameter table 78 corresponds to A single column heading and a single column heading. The A6 column corresponds to the column heading A and the column heading 6. The value of the A6 column of the calibration parameter table 78 is 0.996, the value of the B6 column is 0.985, and the value of the C6 column is The value is 0.95, the value of the D6 column is 0.1, and these values are used as the calibration parameter 77. In actual implementation, the calibration parameter 77 and the calibration parameter table 78 are not limited to this.

The calculation and judgment process of the correction instruction analysis generator 73 is further illustrated as follows:

1. After receiving the initial measurement interval 12, the overall probability density function p(t) is formed, assuming that p is a normal distribution (μ, σ).

，其中，f為頻率機率密度函數，b為t的倒數。 2. Inverse distribution

, Where f is the probability density function of frequency, and b is the reciprocal of t.

3. Given the measurement interval of 12, such as m seconds, the new distribution will become fm (m μ, m σ).

4. The test with extremely low design probability is: within a given m seconds, complete several tests to obtain at least n results less than d. Check whether the result of n+q is lower than the incentive index 79 when q=3, such as i%. Then calculate the cumulative distribution function P and min m=P(0.95|(n+q)μ,(n+q)σ).

5. For example, if the five dimensions are v, w, x, y, and z in sequence, and w<d<x, assuming that n>5 is the number of inspections in the usual quality inspection batch, so n=6, then complete The probability of is as follows:

。這是指連續6次取得尺寸小於d的物件。 (1)n=6,q=0:

. This refers to obtaining an object with a size smaller than d for 6 consecutive times.

。 (2)n=6,q=1:

.

3.10%。 (3) n=6, q=2:

3.10%.

9.91%。 (4) n=6, q=3:

9.91%.

(5) According to the above calculation results, in this case n=6, q=0, and q=1, 2, or when the test ends more than 9 times in the time, the result will be judged as suspicious. At this time, the calibration parameter 77 respectively refer to the A6 column, B6 column and C6 column of the calibration parameter table 78, and the calibration parameter 77 of the remaining results refer to the D6 column of the calibration parameter table 78.

6. Let's take another case where n=7 to illustrate, the same w<d<x, then the probability of completion is as follows:

。這是指連續7次取得尺寸小於d的物件。 (1)n=7,q=0:

. This refers to obtaining an object with a size smaller than d 7 times in a row.

。 (2) n=7, q=1:

.

。 (3) n=7, q=2:

.

。 (4) n=7, q=3:

.

(5) According to the above calculation results, in this case, n=7, q=0, and q=1, 2, 3, or when the test ends more than 10 times in a time, the result will be judged as suspicious. However, since in this case, the result of checking q=3 is not greater than the incentive index 79 (in this case, it is set to 5%), so the correction instruction is not generated under the incentive index 79.

7. Let's explain another case with n=7. At this time, x<d<y, the probability of completion is as follows:

。這是指連續7次取得尺寸小於d的物件。 (1)n=7,q=0:

. This refers to obtaining an object with a size smaller than d 7 times in a row.

。 (2) n=7, q=1:

.

12.54%。 (3) n=7, q=2:

12.54%.

15.05%。 (4) n=7, q=3:

15.05%.

(5) According to the above calculation results, in this case, the possibility of completing the test is higher, and will not be judged as suspicious. Except that the calibration parameter 77 is set to C7=0.95 when the test is over 10 times in the time, the remaining calibration parameters 77 are A7=B7=D7=0.1.

In some situations, if it is deemed necessary to perform calibration again to confirm the behavior of the quality inspector, the incentive index 79 may be changed to generate a new calibration instruction to perform the calibration again.

Please refer to the seventh, tenth and eleventh figures, and please match the first, fifth and sixth figures. The fraud data estimation module 8 is connected to the correction module 7. The measurement interval is abnormal The detection module 6, the measurement data classification module 5, and the storage module 2. When the measurement interval 12 is found to be the suspicious interval, the measurement interval anomaly detection module 6 communicates with the fraud data estimation module 8 so that the measurement data is divided The class module 5 activates the classification model 51. If the optimal feature value set 52 is also classified in the suspicious data distribution, the fraud data inference module 8 determines the stored measurement interval corresponding to the optimal feature value set 52 12 Convert and output an observation sequence 82 with the unsupervised learning model 61, and then establish the hidden Markov model 81.

而A[1,2]=0.7代表本次抽樣樣本為OK良品而下次為NG不良品的概率為0.7，由於抽樣結果為獨立事件而且兩次抽樣結果互斥，因此每一列的總和為1。該觀測概率矩陣812也可以為2D矩陣，該觀測概率矩陣812的列為狀態，行則為觀測，例如

而B[2,2]=0.9代表抽樣樣本為NG不良品而觀測到該可疑間隔的概率為0.9。 The hidden Markov model 81 (λ) includes a state transition matrix 811 (A), the observation probability matrix 812 (B), and an initial state 813 (Π), usually expressed as λ=(A, B, Π). The state transition matrix 811 may be a 2D matrix. The column of the state transition matrix 811 is the current state, and the row is the next state, which is used to represent the measurement probability result of each and next sampling, for example,

And A[1,2]=0.7 means that the probability that this sample is OK good and NG next time is 0.7. Since the sampling result is an independent event and the two sampling results are mutually exclusive, the sum of each column is 1 . The observation probability matrix 812 may also be a 2D matrix. The columns of the observation probability matrix 812 are states, and the rows are observations, for example

And B[2,2]=0.9 represents that the sampling sample is NG defective and the probability of observing the suspicious interval is 0.9.

做為該初始狀態813，並複 製擴展於各列做為該狀態轉移矩陣811，例如

，以及利用該校正模組7所得的該校正參數77建立對應的該觀測概率矩陣812，以該校正參數表78的B6欄位舉例，則

，其中OK列在本例中可預設為

以建立該隱藏式馬可夫模型81(λ)後再以維特比(Viterbi)演算法 運算該觀測序列82的一可能抽樣OK/NG集合821計算良率。再以相同之該歷史量測數據z-score分布32的良率做為該初始狀態813(Π)而利用Baum-Welch演算法所得到的該隱藏式馬可夫模型81對應的良率與該初始狀態813做比較，決定所使用的最可能之該隱藏式馬可夫模型81及所產生的詐欺數據。藉由該校正參數77，增加Baum-Welch演算法以外建立該隱藏式馬可夫模型81的參考選項，提高該檢測數據詐欺裝置100[該檢測數據詐欺裝置100請搭配第一圖]的穩定性。 Please refer to the seventh, tenth and eleventh figure, and please match the third figure, you can find the z-score distribution of the historical measurement data corresponding to the observation sequence 82 and the lower limit of the z-score 37 And the z-score upper limit 38 to calculate the yield (OK, NG). For example

As the initial state 813, and replicate and expand in each column as the state transition matrix 811, for example

, And use the correction parameter 77 obtained by the correction module 7 to establish the corresponding observation probability matrix 812. Taking the B6 column of the correction parameter table 78 as an example, then

, Where the OK column can be preset as

After establishing the hidden Markov model 81 (λ), the Viterbi algorithm is used to calculate a possible sample OK/NG set 821 of the observation sequence 82 to calculate the yield. Then use the same yield rate of the historical measurement data z-score distribution 32 as the initial state 813(Π), and use the Baum-Welch algorithm to obtain the yield rate corresponding to the hidden Markov model 81 and the initial state 813 makes a comparison to determine the most likely hidden Markov model 81 used and the fraud data generated. With the calibration parameter 77, a reference option for establishing the hidden Markov model 81 other than the Baum-Welch algorithm is added to improve the stability of the data fraud detection device 100 [the detection data fraud device 100 please match the first figure].

Please refer to the first and second figures again. With the detection data fraud device 100, the validity of the measurement data 11 of the quality inspection can be analyzed in real time from a distance, avoiding fraud by artificially falsifying data, and reflecting the real quality in real time. Level, early detection of possible quality problems and reduction of resources and costs for repeated quality inspections.

Based on the description of the above-mentioned embodiments, when one can fully understand the operation and use of the present invention and the effects of the present invention, the above-mentioned embodiments are only the preferred embodiments of the present invention, and the implementation of the present invention cannot be limited by this. The scope, that is, simple equivalent changes and modifications made according to the scope of the patent application of the present invention and the content of the description of the invention, are all within the scope of the present invention.

100: Detect data fraud device

1: Measurement data acquisition module

2: Storage module

3: Measurement data processing module

4: Analog data generation module

5: Measurement data classification module

6: Measurement interval anomaly detection module

7: Calibration module

8: Fraud data presumption module

Claims (18)

A device for detecting data fraud is used to detect a measurement behavior of a measurement device. The device for detecting data fraud includes: a measurement data acquisition module that reads a measurement data measured by the measurement device; and A measurement interval; a storage module, the signal is connected to the measurement data acquisition module, the storage module stores the measurement data and the measurement interval; a measurement data processing module, the signal is connected to the storage module , The measurement data processing module calculates the z-score distribution and obtains a set of characteristic values according to a measurement data distribution of the measurement data and a corresponding specification value; an analog data generation module is connected to the signal A measurement data processing module that generates a simulated normal data and a simulated fraud data distribution, and obtains the simulated normal data and a simulated feature value set of the simulated fraud data; a measurement data classification module , The signal is connected to the measurement data processing module and the simulation data generation module, and the measurement data classification module performs training and feature selection operations according to the feature value set or the simulation feature value set to establish a classification model to classify Whether the measurement data distribution is a suspicious data distribution; a measurement interval anomaly detection module, the signal is connected to the storage module, the measurement interval anomaly detection module establishes an unsupervised learning model based on the measurement interval, To determine whether the measurement interval is a suspicious interval; a fraud data inference module, the signal is connected to the measurement interval anomaly detection module, the measurement data classification module and the storage module, when the unsupervised learning model When determining that the measurement interval is the suspicious interval, and the classification model classifies the measurement data distribution into the suspicious data distribution, the fraud data estimation module uses a hidden Markov model to infer the corresponding fraud data; and a correction model Group, the signal is connected to the measurement data acquisition module and the fraud data estimation module, the correction module The group includes a simple object, the calibration module cooperates with the simple object to obtain a calibration parameter, the hidden Markov model is established by the calibration parameter of the calibration module and the Baum-Welch algorithm, and is compared with the yield of an initial condition It is obtained by comparison, and the presumption of fraud data is obtained by the hidden Markov model using the Viterbi algorithm. For example, the data fraud detection device of claim 1, further, the calibration module gives a calibration command, and the calibration module determines the calibration parameter according to the result of executing the calibration command on the simple object. For example, the device for detecting data fraud in claim 1, wherein the simulated data generating module generates the simulated fraud data by using random numbers. For example, the device for detecting data fraud in claim 1, wherein the measurement data acquisition module uses OCR technology to read the measurement data. For example, the device for detecting data fraud in claim 1, wherein the measurement data classification module uses a support vector machine and/or random forest algorithm to establish the classification model. For example, the data fraud detection device of claim 1, wherein the measurement interval anomaly detection module uses an isolation forest and/or z-score algorithm to establish the unsupervised learning model. For example, the detection data fraud device of claim 1, wherein the measurement data classification module uses a self-service sampling method to obtain the feature value set and the simulated feature value set. For example, the data fraud detection device of claim 1, wherein the measurement data processing module performs z-score distribution conversion on the measurement data distribution and the specification value. A method for detecting data fraud, including: using a measurement data acquisition module to read a measurement data and a measurement interval measured by a measurement device; using a storage module to store the measurement data and the amount For the measurement interval, a measurement data processing module is used to calculate a feature value set based on a measurement data distribution of the measurement data and a corresponding specification value; Use a measurement data classification module to perform training and feature selection operations based on the feature value set, or a simulated feature value set that simulates normal data and a simulated fraud data, and establishes a classification model to classify whether the measurement data is distributed Is a suspicious data distribution; a measurement interval anomaly detection module is used to establish an unsupervised learning model based on the measurement interval to determine whether the measurement interval is a suspicious interval; and a calibration module is used in conjunction with a simple The object obtains a calibration parameter, which is established by the calibration parameter of the calibration module and the Baum-Welch algorithm and compared with the yield of an initial condition to obtain a hidden Markov model; when the unsupervised learning model determines the measurement interval When it is the suspicious interval and the classification model classifies the measurement data distribution into the suspicious data distribution, a fraud data estimation module uses the hidden Markov model to infer the corresponding fraud data using the Viterbi algorithm. For example, the method for detecting data fraud in claim 9, further, the calibration module gives a calibration command, and the calibration module determines the calibration parameter according to the result of executing the calibration command on the simple object. For example, the method for detecting data fraud in claim 9, wherein the simulated normal data and the simulated fraud data are historical measurement data and a historical measurement interval stored in the storage module, or are generated by an analog data generation module Generate the simulated normal data and the distribution of the simulated fraud data. For example, the method for detecting data fraud in claim 9, wherein the measurement data acquisition module uses OCR technology to read the measurement data. For example, the method for detecting data fraud in claim 9, wherein the measurement data classification module uses a support vector machine and/or random forest algorithm to establish the classification model. For example, the method for detecting data fraud in claim 9, wherein the measurement interval anomaly detection module uses an isolation forest and/or z-score algorithm to establish the unsupervised learning model. For example, the method for detecting data fraud in claim 9, wherein the measurement data classification module uses a self-service sampling method to obtain the feature value set and the simulated feature value set. For example, the method for detecting data fraud in claim 9, wherein the measurement data processing module performs z-score distribution conversion on the measurement data distribution and the specification value. A program product used to execute a data fraud detection method such as any one of claim 9 to claim 16 after being loaded into a computer device. A computer-readable medium is used to execute the method for detecting data fraud such as any one of claim 9 to 16 after loading a computer device.

Images