Method and system for checking commodity master data anomalies
By combining unsupervised learning algorithms with manual review, the anomaly detection boundary of product master data is dynamically adjusted, solving the problem of low efficiency in product master data verification in the logistics industry. This enables efficient and accurate anomaly data identification and early warning, adapting to the complexity of the e-commerce environment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SUNING COM CO LTD
- Filing Date
- 2026-02-28
- Publication Date
- 2026-06-09
Smart Images
Figure CN122174096A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data anomaly verification, and more particularly to a method and system for verifying anomalies in commodity master data. Background Technology
[0002] Due to cost constraints, the logistics industry typically collects product master data manually before transferring it into the system. Even with strict process control and manual verification, errors can still occur due to human factors. Product master data has a wide range of applications and high reusability, such as intelligent packaging recommendations, intelligent vehicle recommendations, route planning, pricing strategy determination, and logistics-related cost calculation. Failure to detect errors promptly can trigger a series of problems.
[0003] Chinese patent CN111221810A discloses a method, system, computer equipment, and storage medium for identifying anomalies in product master data. It proposes a method to perform a first anomaly detection on the product master data to remove abnormal product data; then, by setting upper and lower limits for each attribute feature of the product master data, a second anomaly detection is performed on the normal product master data after the first anomaly detection. This patent was applied for by the applicant and has the following shortcomings in practical application:
[0004] First, adopting a single-step progressive anomaly detection logic carries the risk of over-cleaning.
[0005] Second, it failed to consider the huge differences in sample size under different fourth-level categories in the e-commerce environment, ignored the impact of sample size on the accuracy of statistical results, and used a uniform fixed process, which is prone to generating false anomalies. Summary of the Invention
[0006] The technical problem to be solved by the present invention is to address the inefficiency of manual verification of product master data, in light of the aforementioned deficiencies in the prior art.
[0007] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows:
[0008] Firstly, a method for verifying anomalies in product master data includes the following steps:
[0009] Collect the master data of the goods in the inventory at regular intervals every day, and classify the master data of the goods according to the product category;
[0010] An anomaly detection algorithm based on unsupervised learning is used to distinguish the master data of each product category into normal data and undetermined anomaly data.
[0011] Based on the statistical distribution characteristics of normal data, obtain the normal value range of the product master data for the corresponding product category;
[0012] Uncertain abnormal data that falls outside the normal value range are marked as suspected abnormal data;
[0013] Send the normal value ranges and suspected abnormal data of the master data of all product categories to manual review.
[0014] In one implementation, after removing extreme anomalous data from the product master data, unsupervised learning anomaly detection is performed; the extreme anomalous data includes items with empty size attributes; products with no specific category; and virtual products with a weight greater than 1.
[0015] In one implementation, the unsupervised learning anomaly detection algorithm includes the Isolation Forest algorithm and the Greedy Takeaway algorithm; when the sample size of the same product category is greater than or equal to the detection threshold, anomaly detection is performed based on the Isolation Forest algorithm, and when the sample size of the same product category is less than the detection threshold, anomaly detection is performed based on the Greedy Takeaway algorithm.
[0016] In one implementation, when the sample size of the same product category is less than the first boundary threshold, unsupervised learning anomaly detection is not performed on the product master data of that product category. Instead, the maximum and minimum values of the product master data of that product category are directly output as the normal value range, and no suspected anomaly data is output. When the sample size of the same product category is greater than the second boundary threshold, the judgment boundary of the normal value range is widened according to a preset correction coefficient, and the pending anomaly data falling outside the widened normal value range is marked as suspected anomaly data.
[0017] In one implementation, the step of obtaining the normal value range of the product master data for a corresponding product category based on the statistical distribution characteristics of normal data includes:
[0018] Let the interval of normal data be [m, n], and let the samples of the same product category approximately follow a mean of μ and a variance of σ. 2 The normal distribution N(μ,σ) 2 );
[0019] When the sample size of the same product category is greater than or equal to the first statistical threshold, the confidence interval is set to [μ-3σ, μ+3σ] according to the Gaussian distribution. If m < μ-3σ or n > μ+3σ, then the normal value interval is taken as [k1*Q1, k2*Q3] according to the quartile method. If m ≥ μ-3σ and n ≤ μ+3σ, then the normal value interval is taken as [(μ-3σ)*k3, (μ+3σ)*k4]. In the formula, Q1 is the first quartile, Q3 is the third quartile, and k1, k2, k3, and k4 are preset adjustment coefficients.
[0020] When the second statistical threshold is less than the sample size of the same product category and less than the first statistical threshold, the normal value range is taken as [k5*Q1, n], where Q1 is the first quartile, k5 is the preset adjustment coefficient, and n is the maximum value of the normal data.
[0021] When the sample size of the same product category is less than or equal to the second statistical threshold, the normal value range is taken as [k6*Q10, n] according to the decimal method, where Q10 is the first decimal place, k6 is the preset adjustment coefficient, and n is the maximum value of the normal data.
[0022] In one implementation, multiple sets of normal value intervals generated for the same product category within a preset period are extracted; by traversing the normal value intervals of each dimension attribute of the master data in the set, the upper limit of each dimension is updated to the maximum value in the set, and the lower limit of each dimension is updated to the minimum value in the set.
[0023] In one implementation, suspected abnormal data that has not been manually modified is included in normal data after manual review, and the normal value range of the corresponding product master data is readjusted.
[0024] Secondly, a system for checking anomalies in product master data is proposed, including:
[0025] The data acquisition module is used to collect the master data of the goods in the inventory on a daily schedule and to classify the master data of the goods according to the product category;
[0026] The data detection module is used for anomaly detection algorithms based on unsupervised learning to distinguish the master data of each product category into normal data and pending anomaly data.
[0027] The data classification module is used to obtain the normal value range of the master data of the corresponding product category based on the statistical distribution characteristics of normal data; and to mark the pending abnormal data that falls outside the normal value range as suspected abnormal data.
[0028] The user interaction module is used to send the normal value range and suspected abnormal data of the master data of all product categories to manual review.
[0029] In one implementation, a review and feedback module is also included, which is used to review the suspected abnormal data that has not been manually modified and include it in the normal data, and readjust the normal value range of the product master data for the corresponding product category.
[0030] The beneficial effects of this invention are:
[0031] 1. Change the previous passive model of data repair being forced by backend application failures. Through a proactive verification mechanism, provide early warnings before risks are transmitted to the financial end, eliminate the lag and randomness of anomaly detection, effectively curb potential asset losses, and realize the transformation from "post-event remediation" to "pre-event prediction and in-event control".
[0032] 2. For massive amounts of data, by combining rules and algorithms, the scope of verification is narrowed from all data to abnormal data, making manual review more targeted and improving both efficiency and accuracy.
[0033] 3. This invention utilizes an unsupervised anomaly detection algorithm to mine clustering patterns in a multidimensional feature space. By identifying the convergence of feature distributions, it filters out high-confidence normal data that is physically and logically consistent. The statistical boundary established using this purified sample set as a benchmark possesses stronger objectivity and discrimination accuracy. Compared to the method in patent CN111221810A, which establishes statistical boundaries based on the full dataset after initial screening, this solution fundamentally suppresses benchmark shifts caused by background noise by pre-removing potential noise points. This ensures that the statistical threshold accurately locks onto the true anomaly interval, effectively solving the problems of over-cleaning and false positives of outliers caused by overly narrow boundaries in existing technologies.
[0034] 4. This invention combines the "logic recognition" of unsupervised learning anomaly detection algorithms with the "quantitative recognition" of statistics, and improves the ability to capture complex anomaly scenarios through cross-validation.
[0035] 5. This invention introduces an adaptive statistical boundary determination mechanism based on sample size triggering. This hierarchical triggering strategy significantly improves the engineering applicability and cross-category generalization ability of this invention in complex e-commerce logistics environments. Attached Figure Description
[0036] The invention will now be further described with reference to the accompanying drawings.
[0037] Figure 1 This is a data processing flowchart of a method for checking anomalies in commodity master data according to an embodiment of the present invention. Detailed Implementation
[0038] like Figure 1 As shown in the figure, this embodiment proposes a method for checking anomalies in product master data, including the following steps:
[0039] S1: Collect the master data of goods in stock at regular intervals every day, and classify the master data of goods according to the product category;
[0040] Specifically, inventory is sampled daily at midnight; daily updated product master data is collected, including product codes, categories, size identifiers, length, width, height, volume, and gross weight. The latest daily billing data, including billing headers, billing details, and billing operation data, can also be extracted simultaneously. If any anomalies are detected in the product master data, backend data, such as billing data, must also be corrected accordingly. Data is transmitted daily to a HIVE database for storage, and then cleaned and structured.
[0041] Regarding product categories, taking Suning's product management system catalog as an example, the catalog is divided into four levels. Each level contains multiple sub-categories, and these categories are independent and do not affect each other. For example, the first to fourth level categories could be: Self-operated Products - Home Appliances - White Goods - Four-door Refrigerators. Within a fourth-level catalog, the weight and volume of all products will not differ significantly. Each product belongs to its own product catalog.
[0042] S2, an anomaly detection algorithm based on unsupervised learning, distinguishes the master data of each product category into normal data and undetermined anomaly data;
[0043] In this embodiment, 1% of undetermined abnormal data is typically output;
[0044] In one implementation, after removing extreme anomalies from the product master data, unsupervised learning is used for anomaly detection. The extreme anomalies include items with empty size attributes; products with no specific category; and virtual products with a weight greater than 1. These extreme anomalies are placed in an anomaly pool. The anomaly pool is checked and modified daily by warehouse staff.
[0045] Specifically, "unspecified product category" includes situations where the major product category is not specified, or the lowest-level product subcategory is not specified, such as when it is "refrigerator / television / air conditioner / washing machine". It can also include situations where one of the dimensions (length, width, height, volume, gross weight) is empty, 0, a character, or a symbol.
[0046] In one implementation, the unsupervised learning anomaly detection algorithm includes an isolated forest algorithm and a greedy take-away algorithm. When the sample size of the same product category is greater than or equal to the detection threshold, anomaly detection is performed based on the isolated forest algorithm; when the sample size of the same product category is less than the detection threshold, anomaly detection is performed based on the greedy take-away algorithm. In this embodiment, the detection threshold is 256.
[0047] The core idea of the Isolation Forest algorithm is that "outliers are isolated points that are easily isolated." Specifically,
[0048] 1) Randomly select no fewer than Ψ=256 points from the product master data as subsamples and place them into the root node of an isolated tree;
[0049] 2) Randomly specify a dimension, and within the current node's data range, randomly generate a cut point p (the cut point is generated between the maximum and minimum values of the specified dimension in the current node's data) to perform a binary split;
[0050] 3) The selection of this cutting point generates a hyperplane that divides the current node's data space into two subspaces: points less than p in the currently selected dimension are placed in the left branch of the current node, and points greater than or equal to p are placed in the right branch of the current node.
[0051] 4) Recursively perform steps 2 and 3 on the left and right branch nodes of the node, continuously constructing new leaf nodes until there is only one data on the leaf node (no further splitting is possible) or the tree has grown to the set height. (Since the splitting process is completely random, an ensemble method is needed to make the results converge, that is, repeatedly split from the beginning and then calculate the average of the splitting results each time, number of times = 100)
[0052] 5) Calculate the anomaly score s:
[0053] ,
[0054] where h(x) is the path length of x in each tree, that is, the number of splits; c(Ψ) is the average of the path lengths when the given number of samples = Ψ, used to standardize the path length h(x) of the sample x; c(Ψ)=2H_(Ψ - 1)-2(Ψ - 1) / Ψ, where H_Ψ is the harmonic function, equal to 1 + 1 / 2 + … + 1 / Ψ;
[0055] is defined as the Euler - Mascheroni constant, with its value being Ω = 0.5772156649
[0056] If the anomaly score s is close to 1, then it must be an anomaly point;
[0057] If the anomaly score s is much less than Ω, then it must not be an anomaly point;
[0058] If the scores s of all points are around Ω, then there are likely no anomaly points in the sample.
[0059] Preset a threshold y (Ω < y < 1). Finally, define that if the anomaly score is less than the set threshold y, it is not considered an anomaly, otherwise it is considered an anomaly. Among them, the set threshold y is closer to 1 and can be configured and adjusted.
[0060] For the greedy removal algorithm, the basic idea is that "after removing a point, the data fluctuation of the remaining samples becomes significantly more stable, indicating that the removed point is an anomaly". Specifically:
[0061] 1) For all samples with a sample size = n (30 < n < 256), sort all samples in ascending order according to a certain feature, obtaining the sequence { , , ,…, , }
[0062] 2) Calculate the overall standard deviation of the samples = . Where M represents the average value
[0063] 3) Starting from , take samples without replacement. Calculate the standard deviation of the sequence formed by the remaining samples excluding the samples from 1 to the i-th samples { , , …, }, where the standard deviation = , and represents the average value of the remaining samples excluding { , , …, }.
[0064] 4) After each sample is taken, calculate the variance after the sample is taken, and the mutation ratio of the variance compared with the previous variance = ( - ) / . Here, i ∈ [1, n]. Preset the parameter p. When > p or = 0, the operation stops. Otherwise, the operation continues.
[0065] 5) When > p or = 0, it means that the i-th and previous samples may all be abnormally small. If all < p, it means that there are no outliers in the overall samples.
[0066] 6) Similarly, sort the samples in descending order and repeat the calculation starting from step 2. It can be judged whether there are abnormally large samples.
[0067] S3. According to the statistical distribution characteristics of normal data, obtain the normal value range of the product master data for the corresponding product category; mark the pending abnormal data falling outside the normal value range as suspected abnormal data;
[0068] In one implementation, when the sample size of the same product category < the first boundary threshold, unsupervised learning anomaly detection is not performed on the product master data of this product category, but directly output the maximum and minimum values of the product master data of this product category as the normal value range, and no suspected abnormal data is output; when the sample size of the same product category > the second boundary threshold, widen the determination boundary of the normal value range according to the preset correction coefficient, and mark the pending abnormal data falling outside the widened normal value range as suspected abnormal data.
[0069] In this embodiment, the first boundary threshold is set to 30 because, according to empirical statistical standards, a sample size of less than 30 is not statistically significant. The second boundary threshold is set to 100, with a preset correction coefficient of 0.5%. This means that, based on a relatively conservative principle, the neighborhood of the upper and lower limits of the category is expanded outward by 0.5%. In other words, only when the feature value of a sample exceeds the upper limit by (1+0.5%) or is less than the lower limit by (1+0.5%) is it considered to be suspected abnormal data.
[0070] The accuracy of the algorithm's output normal values depends on the size of the effective sample size; the larger the sample, the more accurate the algorithm's results. Therefore, the algorithm's results cannot be directly used to construct upper and lower limits; the existing thresholds must be adjusted using statistical methods such as Gaussian distribution and quartiles.
[0071] The Gaussian distribution (normal distribution) specifically refers to:
[0072] If a random variable X follows a mathematical expectation of μ and a variance of σ... 2 The normal distribution is denoted as N(μ,σ). 2 The 3σ principle of normal distribution refers to the following in the normal distribution curve:
[0073] 1. Approximately 68.3% of the data lies within ±1σ of the mean (μ±1σ).
[0074] 2. Approximately 95.4% of the data lies within ±2σ of the mean (μ±2σ).
[0075] 3. Almost all the data (99.7%) were within ±3σ of the mean (μ±3σ).
[0076] This means that only a very small number of data points will fall outside the range of the mean ± 3σ, and these data points are usually considered outliers.
[0077] The quartile method specifically refers to dividing a dataset into four quartiles, typically used to describe the distribution of data. The first quartile represents 25% of the data points less than that value, the second quartile represents 50% of the data points less than that value, and the third quartile represents 75% of the data points less than that value. For a normal distribution, quartiles are usually represented by Q1, Q2, and Q3.
[0078] In a normal distribution, quartiles can be calculated. Since the normal distribution is symmetric, the median is the second quartile, Q2. The first quartile, Q1, can be calculated using the following formula: Q1 = μ - 0.6745σ, where μ is the mean and σ is the standard deviation. Similarly, the third quartile, Q3, can be calculated using the following formula: Q3 = μ + 0.6745σ.
[0079] In one implementation, the step of obtaining the normal value range of the product master data for a corresponding product category based on the statistical distribution characteristics of normal data includes:
[0080] Let the interval of normal data be [m, n], and let the samples of the same product category approximately follow a mean of μ and a variance of σ. 2 The normal distribution N(μ,σ) 2 );
[0081] When the sample size of the same product category is greater than or equal to the first statistical threshold, the confidence interval is set to [μ-3σ, μ+3σ] according to the Gaussian distribution. If m < μ-3σ or n > μ+3σ, then the normal value interval is taken as [k1*Q1, k2*Q3] according to the quartile method. If m ≥ μ-3σ and n ≤ μ+3σ, then the normal value interval is taken as [(μ-3σ)*k3, (μ+3σ)*k4]. In the formula, Q1 is the first quartile, Q3 is the third quartile, and k1, k2, k3, and k4 are preset adjustment coefficients.
[0082] In this embodiment, the first statistical threshold is 200, k1=0.5, k2=2, k3=0.5, and k4=1.
[0083] When the second statistical threshold is less than the sample size of the same product category and less than the first statistical threshold, the normal value range is taken as [k5*Q1, n], where Q1 is the first quartile, k5 is the preset adjustment coefficient, and n is the maximum value of the normal data.
[0084] In this embodiment, the first statistical threshold is 200, the second statistical threshold is 50, and k5=0.3.
[0085] When the sample size of the same product category is less than or equal to the second statistical threshold, the normal value range is taken as [k6*Q10, n] according to the decimal method, where Q10 is the first decimal place, k6 is the preset adjustment coefficient, and n is the maximum value of the normal data.
[0086] In this embodiment, the second statistical threshold is 50, and k6=0.3.
[0087] The adjustment coefficients involved in the above process are continuously adjusted based on actual operations. In practical applications, fine-tuning can be made based on the theoretical results, taking into account factors such as your own product structure.
[0088] After the above normal value ranges (upper and lower limit table) are generated, they are synchronized to the Product Master Data System (RCS), and then distributed by the latter to the Warehouse Management System (WMS). When warehouse staff collect or modify weights or volumes, if the values exceed the upper or lower limits, the WMS system will remind the staff to verify.
[0089] In one implementation, multiple sets of normal value intervals generated for the same product category within a preset period are extracted; by traversing the normal value intervals of each dimension attribute of the master data in the set, the upper limit of each dimension is updated to the maximum value in the set, and the lower limit of each dimension is updated to the minimum value in the set.
[0090] S4: Send the normal value range and suspected abnormal data of the master data of all product categories to manual review.
[0091] In one implementation, suspected abnormal data that has not been manually modified is included in normal data after manual review, and the normal value range of the corresponding product master data is readjusted.
[0092] The method for detecting the accuracy of an algorithm is to check whether the results calculated by the algorithm have been manually modified. If data such as the weight and volume of a product (excluding other information, such as product descriptions) has been manually modified after the algorithm's recommendation, then the algorithm's recommendation is considered accurate. Furthermore, the modified product data is considered correct and will not be included in the outlier data again. An average accuracy rate is also calculated (average accuracy rate = number of outliers corrected by warehouse staff / total number of outliers extracted by the algorithm). The accuracy rate reflects the model's rationality and serves as a reminder for algorithm adjustments and corrections. For product categories with low accuracy rates, the deviations of so-called outlier samples are examined across all samples to first pinpoint whether the problem lies with experience or the algorithm module, and then determine whether algorithm adjustments are necessary.
[0093] This embodiment also proposes a system for checking anomalies in product master data, including:
[0094] The data acquisition module is used to collect the master data of the goods in the inventory on a daily schedule and to classify the master data of the goods according to the product category;
[0095] The data detection module is used for anomaly detection algorithms based on unsupervised learning to distinguish the master data of each product category into normal data and pending anomaly data.
[0096] The data classification module is used to obtain the normal value range of the master data of the corresponding product category based on the statistical distribution characteristics of normal data; and to mark the pending abnormal data that falls outside the normal value range as suspected abnormal data.
[0097] The user interaction module is used to send the normal value range and suspected abnormal data of the master data of all product categories to manual review.
[0098] In one implementation, a review and feedback module is also included, which is used to review the suspected abnormal data that has not been manually modified and include it in the normal data, and readjust the normal value range of the product master data for the corresponding product category.
Claims
1. A method for verifying anomalies in commodity master data, characterized in that, Includes the following steps: Collect the master data of the goods in the inventory at regular intervals every day, and classify the master data of the goods according to the product category; An anomaly detection algorithm based on unsupervised learning is used to distinguish the master data of each product category into normal data and undetermined anomaly data. Based on the statistical distribution characteristics of normal data, obtain the normal value range of the product master data for the corresponding product category; Uncertain abnormal data that falls outside the normal value range are marked as suspected abnormal data; Send the normal value ranges and suspected abnormal data of the master data of all product categories to manual review.
2. The method for verifying anomalies in commodity master data according to claim 1, characterized in that: After removing extreme anomalies from the product master data, unsupervised learning is then used for anomaly detection. The extreme anomalies include items with empty size attributes, products with no specific category, and virtual products with a weight greater than 1.
3. The method for verifying anomalies in commodity master data according to claim 1, characterized in that: The unsupervised learning anomaly detection algorithm includes the Isolation Forest algorithm and the Greedy Takeaway algorithm. When the sample size of the same product category is greater than or equal to the detection threshold, anomaly detection is performed based on the Isolation Forest algorithm; when the sample size of the same product category is less than the detection threshold, anomaly detection is performed based on the Greedy Takeaway algorithm.
4. The method for verifying anomalies in commodity master data according to claim 1, characterized in that: When the sample size of the same product category is less than the first boundary threshold, unsupervised learning anomaly detection is not performed on the product master data of that product category. Instead, the maximum and minimum values of the product master data of that product category are directly output as the normal value range, and no suspected anomaly data is output. When the sample size of the same product category is greater than the second boundary threshold, the judgment boundary of the normal value range is widened according to the preset correction coefficient, and the pending anomaly data falling outside the widened normal value range is marked as suspected anomaly data.
5. The method for verifying anomalies in commodity master data according to claim 1, characterized in that: Based on the statistical distribution characteristics of normal data, the steps to obtain the normal value range of the product master data for the corresponding product category include: Let the interval of normal data be [m, n], and let the samples of the same product category approximately follow a mean of μ and a variance of σ. 2 The normal distribution N(μ,σ) 2 ); When the sample size of the same product category is greater than or equal to the first statistical threshold, the confidence interval is set to [μ-3σ, μ+3σ] according to the Gaussian distribution. If m < μ-3σ or n > μ+3σ, then the normal value interval is taken as [k1*Q1, k2*Q3] according to the quartile method. If m ≥ μ-3σ and n ≤ μ+3σ, then the normal value interval is taken as [(μ-3σ)*k3, (μ+3σ)*k4]. In the formula, Q1 is the first quartile, Q3 is the third quartile, and k1, k2, k3, and k4 are preset adjustment coefficients. When the second statistical threshold is less than the sample size of the same product category and less than the first statistical threshold, the normal value range is taken as [k5*Q1, n], where Q1 is the first quartile, k5 is the preset adjustment coefficient, and n is the maximum value of the normal data. When the sample size of the same product category is less than or equal to the second statistical threshold, the normal value range is taken as [k6*Q10, n] according to the decimal method, where Q10 is the first decimal place, k6 is the preset adjustment coefficient, and n is the maximum value of the normal data.
6. The method for verifying anomalies in commodity master data according to claim 5, characterized in that: Extract multiple normal value range sets generated for the same product category within a preset period; by traversing the normal value ranges of each dimension attribute of the master data in the set, update the upper limit of each dimension to the maximum value in the set, and update the lower limit of each dimension to the minimum value in the set.
7. The method for verifying anomalies in commodity master data according to claim 1, characterized in that: After manual review, suspected abnormal data that has not been manually modified is included in the normal data, and the normal value range of the corresponding product master data is readjusted.
8. A system for verifying anomalies in commodity master data, characterized in that, include: The data acquisition module is used to collect the master data of the goods in the inventory on a daily schedule and to classify the master data of the goods according to the product category; The data detection module is used for anomaly detection algorithms based on unsupervised learning to distinguish the master data of each product category into normal data and pending anomaly data. The data classification module is used to obtain the normal value range of the master data of the corresponding product category based on the statistical distribution characteristics of normal data; and to mark the pending abnormal data that falls outside the normal value range as suspected abnormal data. The user interaction module is used to send the normal value range and suspected abnormal data of the master data of all product categories to manual review.
9. The method for verifying anomalies in commodity master data according to claim 8, characterized in that: It also includes a review and feedback module, which is used to manually review and include suspected abnormal data that has not been manually modified into normal data, and readjust the normal value range of the corresponding product master data.