Digital integrated quality management system based on multi-source data fusion

Abstract

The invention relates to a digital integrated quality management system based on multi-source data fusion, and belongs to the technical field of industrial internet and quality management. A data acquisition layer of the system obtains real-time and static multi-source heterogeneous data through a multi-source adapter; the data processing layer is used for cleaning, converting and standardizing the acquired data; the intelligent analysis layer performs deep analysis and prediction on the data by using an adaptive quality prediction model, an anomaly detection module and a root cause analysis engine; the application service layer displays a quality trend and an anomaly detection result through a visual billboard, and provides credible tracing and collaborative decision-making functions; and the feedback closed layer adjusts system processing logic according to the decision support data to form closed-loop quality control. According to the method, real-time fusion and efficient utilization of multi-source data are realized through a dynamic routing technology, an adaptive quality prediction model and a block chain evidence storage mechanism, and the intelligent level and decision-making efficiency of quality management are remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of industrial Internet and quality management technology, and relates to a digital integrated quality management system based on multi-source data fusion. Background Art

[0002] As a core link in the digital transformation of the manufacturing industry, industrial quality management is facing the dual challenges of global industrial chain reconstruction and upgrading of consumer demand. With the rise of trends such as "nearshore manufacturing" and "flexible production", the traditional competition model dominated by scale expansion has gradually given way to quality differentiation competition. Enterprises urgently need to build market barriers through quality control capabilities. According to statistics from the China Quality Association, industrial product quality losses account for about 5%-15% of corporate revenue, and the consumer side has stringent requirements for product consistency (such as chip yield) and traceability (such as food supply chain transparency), making it difficult for traditional management models that rely on manual sampling and post-analysis to meet the goal of "zero-defect production."

[0003] Existing industrial quality management generally adopts a composite approach of "sampling inspection + process control + post-event traceability," which relies on a dual-track system of standardized processes and manual experience. In manufacturing, companies often implement batch sampling inspections based on international sampling standards (such as GB / T2828.1AQL). For example, the automotive parts industry typically conducts destructive inspections of 1%-5%, determining the batch pass rate through key dimensional measurements and material performance testing. However, this model carries the risk of "false positives" (one automaker once triggered a large-scale recall due to a hidden defect with a probability of one in a thousand). To improve process stability, some companies have introduced statistical process control (SPC), deploying sensors at key workstations to collect parameters such as temperature and pressure, and relying on XR control charts to provide early warning of abnormal fluctuations. However, in actual applications, rigid rules (such as monitoring only ±3σ thresholds) often lead to false alarm rates exceeding 30% (data from one semiconductor company showed that 68% of alarms were invalid). At the traceability level, most factories use a batch traceability system based on paper work orders and barcode scanning, recording process flow information through the MES (Manufacturing Execution System). However, data granularity typically remains at the batch level, and cross-system data synchronization delays can reach hours (for example, one medical device company experienced a 48-hour delay in batch release due to untimely sterilization data upload). Furthermore, quality improvement relies heavily on expert experience. For example, process parameter adjustments require engineers to iterate through trial and error based on a historical case library. One steel company's rolling process optimization required trial production of 200 coils before determining the optimal parameter combination. Although some companies have implemented QMS (Quality Management Systems) to digitize processes, data silos between systems remain prominent (with an average of seven independent software programs per factory). Quality analysis still relies primarily on offline reporting, with real-time decision-making coverage of less than 15%. This "partial digitization, global fragmentation" situation has led to high quality control costs (defect handling costs account for over 60% of total quality costs), making it difficult to meet the new demands of small-batch customized production.

[0004] In summary, existing industrial quality management solutions still have the following problems: data silos: multi-source heterogeneous data (such as sensor data, ERP records, and manual quality inspection reports) are difficult to integrate; prediction lag: quality anomalies cannot be warned in real time and rely on post-analysis; traceability difficulties: the root cause identification of quality problems is inefficient and there is a lack of reliable records; insufficient human-machine collaboration: the decision-making process relies on experience and lacks data-driven support.

[0005] In the existing technology, a single database or general BI tool is difficult to solve the above problems. There is an urgent need for an integrated system that supports dynamic fusion and closed-loop feedback. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a digital integrated quality management system based on multi-source data fusion, combining dynamic routing technology, adaptive quality prediction model, and blockchain evidence storage mechanism to solve the problems of data silos, prediction lag, and traceability difficulties in traditional quality management.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A digital integrated quality management system based on multi-source data fusion, including data acquisition layer, data processing layer, intelligent analysis layer, application service layer and feedback closure layer, among which,

[0009] The data collection layer includes a multi-source adapter, a real-time stream collection module, and a static data import module. It uses these modules to acquire multi-source heterogeneous data, which includes both real-time and static data. Real-time data is added with a timestamp and data source tag and then pushed to the message queue, while static data is temporarily stored in a buffer table.

[0010] The data processing layer includes a dynamic cleaning module, a dynamic routing module, and a heterogeneous data converter. It removes outliers, fills in missing values, dynamically allocates routes, and converts unstructured data from multi-source heterogeneous data, generating standardized intermediate tables and storing them in the intermediate database.

[0011] The intelligent analysis layer includes an adaptive quality prediction model, an anomaly detection module, and a root cause analysis engine. The intelligent analysis layer predicts and analyzes standardized data. It uses the adaptive quality prediction model to perform time series analysis on standardized data to generate quality prediction results. The anomaly detection module performs pattern recognition on standardized data to generate anomaly detection results. The root cause analysis engine calculates the probability of associated factors based on the anomaly detection results to generate root cause analysis data. The analysis parameters are adjusted based on the quality prediction results and root cause analysis data.

[0012] The application service layer includes a visual dashboard, a trusted traceability module, and a collaborative decision-making terminal; the feedback closure layer adjusts the processing logic of the data processing layer and the intelligent analysis layer based on the decision support data.

[0013] Furthermore, in the data collection layer, the multi-source adapter supports API, MQTT, and JDBC protocols for connecting to sensors, ERP, MES, and manual entry terminal data sources; the real-time stream collection module processes high-speed sensor data streams through Kafka and generates real-time data streams with metadata descriptions; the static data import module batch imports Excel, CSV structured or semi-structured data into SQL Server buffer tables.

[0014] Furthermore, in the data processing layer, the dynamic cleaning module performs the following operations based on the rule engine: for structured data, it uses field-level rules to fill in null values, standardize date formats, and verify uniqueness; for semi-structured data, it uses dynamic Schema mapping rules to perform JSON / XML path parsing and tag completion; for unstructured data, it uses contextual association rules to verify image resolution, extract text keywords, and generate file hash values; the rule engine also sets general rules, including timestamp standardization processing rules, data source credibility weighting rules, and basic missing value processing rules;

[0015] The dynamic routing module assigns processing paths based on data type labels. Structured data directly triggers field-level cleaning rules and is written to the intermediate database. Semi-structured data is parsed into temporary tables and then triggers dynamic schema cleaning rules. Unstructured data uses external tools to extract metadata and is then stored in association with structured data.

[0016] The heterogeneous data converter converts JSON or XML nested data into a relational table through the following steps: parsing the JSON / XML main table fields and inserting them into the main table; expanding the nested array or hierarchical structure to generate a detailed table and normalizing the data through primary and foreign key associations; performing strong type verification during the parsing process and defining field types.

[0017] Furthermore, in the data processing layer, the operations performed include:

[0018] Dynamic Schema Mapping and Lightweight Parsing: This system automatically parses semi-structured data using SQL Server's OpenJSON functions and XPath parsing rules. It generates a field list by extracting metadata header information, infers field types based on sample data, and constructs a virtual intermediate table in memory. For nested structures, it uses a lazy loading mechanism to parse specific content only when a query is accessed. It also generates hash values for image files and establishes a joint index with the primary key of the structured data.

[0019] In-memory computing and column storage optimization: Data is horizontally sharded by time range, loaded into the in-memory column store using SQL Server's In-Memory OLTP technology, and a clustered column store index is created. Numeric columns are pre-allocated using contiguous memory blocks, vectorized aggregate calculations are performed using the SIMD instruction set, and the bit-packing algorithm is applied to compress storage space.

[0020] Dynamic routing and parallelized pipeline: Source type and priority tags are injected during the data collection phase, and data is distributed to real-time or batch processing queues based on a rules engine. Data is sharded based on the last two digits of the device ID hash value, with corresponding CPU threads and memory quotas allocated to the real-time queue and threads and quotas to the batch queue. Multi-channel task synchronization is achieved through a lock-free ring buffer, and atomic compare-and-swap operations are used to update read and write pointers.

[0021] Incremental fusion and partial updates: SQL Server CDC is used to capture incremental data changes, and temporal tables are enabled to store historical version snapshots. Lightweight cleansing is performed on incremental data, and the incremental results are merged with the snapshot table by device ID and batch number using the MERGE statement. During cold data migration, consistent bitmap indexes from row numbers to disk page numbers are maintained, supporting cross-storage tier federated queries.

[0022] Hybrid storage and hot-cold tiering: Column-level access frequency is monitored, and frequently accessed data of the day is retained in the in-memory column storage area. Cold data that has not been accessed for a preset time is page-compressed and then migrated to the disk row storage area. Hot data uses column storage indexes to accelerate aggregate analysis, while cold data is quickly located through bitmap indexes. When it is frequently accessed again, a preload thread is triggered to restore the in-memory copy.

[0023] Furthermore, at the intelligent analysis layer, the adaptive quality prediction model is based on LSTM neural network training historical data. Based on the Keras framework, it has a stacked LSTM structure that gradually extracts time series features through two layers of LSTM. The first layer of LSTM retains the time dimension for processing by the second layer, which aggregates the time series information to generate the final representation. The output layer implements binary classification in a probabilistic form.

[0024] The anomaly detection module uses the isolation forest algorithm to identify data that deviates from the normal pattern in real time. The process is expressed as follows:

[0025] Construct an isolated tree: randomly select subsample data and recursively split it into leaf nodes; the split threshold randomly selects features and split points;

[0026] Calculate the anomaly score: The path length is the number of edges from the root node to the leaf node of the data point; the anomaly score formula is:

[0027] score=2 (-E(h(x)) / c(n))

[0028] Where h(x) is the path length and c(n) is the average path length of the tree;

[0029] The root cause analysis engine locates the factors associated with quality issues based on the Bayesian network. The nodes of the Bayesian network include: root nodes, intermediate nodes, and leaf nodes. The root nodes are equipment fault codes and raw material batches; the intermediate nodes are process parameter deviations; and the leaf nodes are quality defect types. Specific dependencies exist between the nodes.

[0030] Calculate the conditional probability table using historical data:

[0031] P(Surface_Crack=Yes|Temperature_Deviation=High)

[0032] Evidence propagation: When Surface_Crack=Yes is detected, the posterior probability of each root node is calculated in reverse:

[0033] P(Motor_Fault=Yes|Surface_Crack=Yes)

[0034] P(Material_Batch=B123|Surface_Crack=Yes)

[0035] Output: A ranked list of root cause probabilities.

[0036] Furthermore, at the application service layer, the application service layer displays quality prediction results and anomaly detection results through a visual dashboard to generate quality trend data; generates operation logs for quality prediction results through the trusted traceability module and stores them in the blockchain to generate traceability records; obtains manual feedback data based on anomaly detection results through the collaborative decision-making terminal to generate decision recommendations; and generates decision support data based on quality trend data and decision recommendations.

[0037] Furthermore, in the feedback closed layer, the processing logic of the data processing layer and the intelligent analysis layer is adjusted according to the decision support data, including: updating the cleaning rules according to the decision support data through the rule optimization module to generate an optimization rule set; generating improvement tasks according to the decision support data and tracking the execution status through the closed-loop work order system to generate work order results; updating the quality knowledge base according to the work order results to generate feedback data; and adjusting the cleaning logic of the data processing layer and the prediction parameters of the intelligent analysis layer according to the feedback data.

[0038] The beneficial effects of the present invention are:

[0039] The system of the present invention effectively integrates multi-source heterogeneous data from sensors, ERP, MES and manual entry terminals through multi-source adapters, real-time stream acquisition modules and static data import modules, breaking down data silos, achieving comprehensive fusion and integration of data, and providing a rich data foundation for quality management.

[0040] The dynamic cleaning module, dynamic routing module and heterogeneous data converter of the present invention together constitute a powerful data processing capability, which can efficiently remove outliers, fill missing values, dynamically allocate routes and convert unstructured data from multi-source heterogeneous data, generate standardized intermediate tables, and greatly improve data processing efficiency and accuracy.

[0041] The intelligent analysis layer of the present invention adopts an adaptive quality prediction model, anomaly detection module and root cause analysis engine to conduct in-depth analysis and prediction of standardized data. It can detect quality anomalies in advance, accurately locate the root cause of quality problems, and provide strong support for decision-making.

[0042] The application service layer of the present invention displays quality prediction results and anomaly detection results through a visual dashboard, making quality trends clear at a glance; the trusted traceability module uses blockchain technology to record key operation logs, ensuring the immutability and traceability of data, and enhancing the transparency and credibility of quality management.

[0043] The feedback closed layer of the present invention adjusts the processing logic of the data processing layer and the intelligent analysis layer according to the decision support data, forming a closed-loop quality control system. By continuously optimizing the data processing rules and analysis models, the performance and effect of the quality management system are continuously improved.

[0044] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0046] Figure 1 Schematic diagram of the structure of a digital integrated quality management system based on multi-source data fusion according to an embodiment of the present invention. DETAILED DESCRIPTION

[0047] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0048] Among them, the accompanying drawings are only for illustrative purposes and represent only schematic diagrams rather than actual pictures, and should not be understood as limiting the present invention. In order to better illustrate the embodiments of the present invention, some parts of the accompanying drawings may be omitted, enlarged or reduced, and do not represent the dimensions of actual products. For those skilled in the art, it is understandable that some well-known structures and their descriptions may be omitted in the accompanying drawings.

[0049] The same or similar numbers in the drawings of the embodiments of the present invention correspond to the same or similar parts; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "back", etc. indicating directions or positional relationships, they are based on the directions or positional relationships shown in the drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operate in a specific direction. Therefore, the terms describing the positional relationship in the drawings are only used for illustrative purposes and cannot be understood as limiting the present invention. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to specific circumstances.

[0050] See also Figure 1 , which is a digital integrated quality management system based on multi-source data fusion.

[0051] Example

[0052] This embodiment introduces a specific structure and solution of a digital integrated quality management system based on multi-source data fusion. Figure 1 As shown, it at least includes: a data acquisition layer, a data processing layer, an intelligent analysis layer, a reference service layer and a feedback closure layer, wherein the data acquisition layer obtains multi-source heterogeneous data, and the multi-source heterogeneous data includes at least real-time data and static data; the data processing layer cleans and converts the multi-source heterogeneous data to obtain standardized data; the intelligent analysis layer predicts and analyzes the standardized data to obtain quality prediction results and anomaly detection results; the application service layer generates decision support data based on the quality prediction results and anomaly detection results; the feedback closure layer adjusts the processing logic of the data processing layer and the intelligent analysis layer according to the decision support data to form a closed-loop quality control.

[0053] In this embodiment, the data acquisition layer includes a multi-source adapter, a real-time stream acquisition module, and a static data import module. The multi-source adapter is used to connect to different data sources, including sensor data and structured data; the real-time stream acquisition module is used to process high-speed sensor data streams, add timestamps and data source tags, and generate real-time data streams with metadata descriptions; the static data import module is used to batch import structured or semi-structured data; among them, the real-time data stream is pushed to the Kafka queue, and the batch imported structured and semi-structured data is temporarily stored in the buffer table.

[0054] Specifically, the multi-source adapter connects to data sources such as sensors, ERP, MES, and manual input terminals, and supports protocols such as API, MQTT, and JDBC. The real-time stream acquisition module processes high-speed sensor data streams through Kafka. The static data import module batch imports structured and semi-structured data in Excel, CSV, and other formats.

[0055] The data collection layer adds timestamps and data source tags to the original data to form a unified metadata description, and pushes real-time data to the Kafka queue, while static data is temporarily stored in the SQL Server buffer table.

[0056] In this embodiment, the data processing layer includes a dynamic cleaning module, a dynamic routing module and a heterogeneous data converter. The dynamic cleaning module removes outliers and fills missing values in the multi-source heterogeneous data according to preset rules to generate preliminary cleaned data; the dynamic routing module assigns processing paths according to the type labels of the preliminary cleaned data to generate routing allocation data; the heterogeneous data converter converts the unstructured data in the routing allocation data into relational data to generate a standardized intermediate table; the standardized data is generated according to the standardized intermediate table and stored in the intermediate library.

[0057] Specifically, the dynamic cleaning module uses a rules engine (such as SQL Server's SSIS) to remove outliers and complete missing data. First, different rule types are designed to address typical issues with specific data types. Data types are categorized by structure, including structured data, semi-structured data, and unstructured data. Typical issues with structured data include missing values, incorrect field formats, and duplicate primary keys. Therefore, a combination of general rules and field-level rules is designed, such as null value filling (mean / most recent value), date format standardization, and uniqueness verification.

[0058] For semi-structured data, typical problems include redundant nesting levels, missing tags, and confusing formats. Therefore, structural parsing rules and dynamic rules are designed, such as JSON / XML path parsing, dynamic Schema mapping, and tag completion (such as completing the default namespace).

[0059] For unstructured data, typical problems are difficulty in content extraction and missing metadata. Therefore, contextual association rules are designed, such as image resolution verification, text keyword extraction, and file hash value generation.

[0060] The specific rule classification and design principles are shown in Table 1:

[0061] Table 1

[0062]

[0063] For structured data, such as data in SQL tables, a dedicated field-level function is set for processing, which can automatically complete missing fields. For example, it can automatically complete missing batch numbers. The pseudo code is as follows:

[0064] UPDATE ProductionData

[0065] SET BatchID=(SELECT TOP 1BatchID FROM WorkOrder WHERE MachineID=ProductionData.MachineID)

[0066] WHERE BatchID IS NULL;

[0067] The processing logic is to define integrity constraints and cross-table association completion by field through SQL Server stored procedures or SSIS data flow tasks.

[0068] For semi-structured data, such as data in JSON or XML tables, dedicated dynamic parsing rules are set for processing, which can also automatically complete missing fields. The pseudo code is as follows:

[0069] INSERT INTO CleanData

[0070] SELECT

[0071] ISNULL(JSON_VALUE(raw_data,'$.sensor_id'),'DefaultSensor')ASSensorID,

[0072] COALESCE(JSON_VALUE(raw_data,'$.temperature'),'25')AS Temperature

[0073] FROM RawJsonData;

[0074] Its execution logic is to use dynamic Schema mapping technology to parse nested fields by path and complete missing labels with preset default values or associated tables.

[0075] In addition, for the three data types, general rules are designed for general processing across data types, including timestamp standardization processing rules, data source credibility weighting rules, and basic missing value processing rules. Among them, the timestamp standardization processing rule refers to unifying the time fields in all data into the ISO 8601 format (such as YYYY-MM-DD HH:MM:SS), the data source credibility weighting rule refers to dynamically allocating weights according to the data source (such as sensor accuracy level, manual entry frequency), which affects the subsequent analysis results, and the basic missing value processing rule refers to filling numerical null values according to the global mean, and marking categorical null values as "Unknown".

[0076] The dynamic routing module assigns processing flows based on data tags (e.g., data_type=json, priority=high). Specifically, structured data directly triggers field-level cleaning rules and is then written to the intermediate database; semi-structured data is first parsed into a temporary table, and then dynamic Schema rules are applied to trigger general cleaning; unstructured data is processed by external tools, and metadata is extracted and stored in association with structured data.

[0077] The engine configuration of the above rules is specifically set as follows: the rules are stored in the RuleConfig table of SQL Server, and the fields include rule ID, applicable data type, priority, and execution script; through the platform configuration visual rule management interface, it supports binding rule sets by data topics (such as "production quality inspection" and "supply chain").

[0078] Additionally, established rules require optimization and feedback loop processing. Optimization involves adaptive rule adjustment. The system records rule execution results (e.g., "a sensor's data loss rate is >30%") and automatically suggests adjustments to the completion strategy (e.g., switching to the device's historical average). Abnormal data is manually marked on the dashboard, triggering rule library updates (e.g., adding a "supplier batch number verification rule"). The feedback loop involves validation of results. After cleaning, data quality indicators (e.g., completeness and consistency) are written into the quality knowledge base to drive dynamic prioritization of rules.

[0079] Heterogeneous data converters convert unstructured data such as XML and JSON into relational data tables. Specifically, taking the process of converting nested JSON data into a relational table (flattening) as an example, the process is as follows:

[0080] Determine the data structure before conversion (original JSON):

[0081]

[0082] Determine the target structure for conversion (relational table). Split it into two related tables to achieve data normalization, as shown in Table 2 and Table 3. Table 2 is the order master table (OrderMaster) and Table 3 is the order detail table (OrderDetail):

[0083] Table 2

[0084] Order number Customer Name Order Date ORD20231001 Company A 2023 / 10 / 1

[0085] Table 3

[0086] Order number Production order number Order quantity unit price ORD20231001 001 5 29.99 ORD20231001 002 2 89.5

[0087] Determine the specific conversion process, which uses the SQL Server OpenJSON function for processing, mainly including the following steps:

[0088] Step 1: Parse the JSON main table fields; the pseudo code for this step is as follows:

[0089]

[0090] Step 2: Expand the nested items array; the pseudo code for this step is:

[0091]

[0092] The key steps in the above process include nested expansion, primary and foreign key associations, and strong type validation. Specifically, nested expansion uses the OPENJSON function to extract array objects and generate multiple rows of detailed records. Primary and foreign key associations extract the order_id from the root JSON node to ensure the link between the primary and detail tables. Strong type validation defines the field type (such as DECIMAL(10,2)) in the WITH clause to prevent format errors.

[0093] Additionally, this embodiment also provides an example of a process for converting XML hierarchical data into a relational table (flattening), and the process is also as follows:

[0094] Determine the data structure before conversion (original XML), which is represented as:

[0095]

[0096]

[0097] To determine the target structure for conversion (relational table), the hierarchical data needs to be split into three tables to eliminate multi-value dependencies, as shown in Tables 4, 5, and 6. Table 4 is the department table, Table 5 is the employee table, and Table 6 is the skill table.

[0098] Table 4

[0099] Department Name Production

[0100] Table 5

[0101] Employee ID Department Name Name Position E001 Production Zhangsan Yiban E002 Production Lisi Manager

[0102] Table 6

[0103] Employee ID Skill Name E001 CAD E001 PLC E002 Project Management

[0104] To perform the conversion, SQL Server's XML parsing method is used. The process is as follows:

[0105] Step 1: Extract department information:

[0106] INSERT INTO Department(dept_name)

[0107] SELECT

[0108] x.dept.value('@name','VARCHAR(50)')

[0109] FROM@xml_data.nodes(' / Department')AS x(dept);

[0110] Step 2: Parse employee master data:

[0111] INSERT INTO Employee(emp_id,dept_name,name,role)

[0112] SELECT

[0113] emp.value('@id','VARCHAR(10)'),

[0114] Production', -- department name is hard-coded (actually should be obtained dynamically from the XML parent node)

[0115] emp.value('(Name)[1]','VARCHAR(50)'),

[0116] emp.value('(Role)[1]','VARCHAR(50)')

[0117] FROM@xml_data.nodes(' / Department / Employee')AS x(emp);

[0118] Step 3: Expand the multi-valued skill field:

[0119] INSERT INTO Skill(emp_id,skill_name)

[0120] SELECT

[0121] emp.value('@id','VARCHAR(10)'),

[0122] skill.value('.','VARCHAR(50)')

[0123] FROM@xml_data.nodes(' / Department / Employee')AS x(emp)

[0124] CROSS APPLY x.emp.nodes('Skills / Skill')AS y(skill);

[0125] The key technologies in the above process include XPath navigation, multi-value splitting, and dynamic association. XPath navigation locates the node hierarchy (such as / Department / Employee) through the nodes() function. Multi-value splitting uses CROSS APPLY to <skill>The node expands to multiple rows. Dynamic associations are used to associate the employee and skill tables through the @id attribute (in actual scenarios, the department name must be dynamically obtained from the parent node).

[0126] In general, the data processing layer uses dynamic schema mapping technology to automatically match fields and generate standardized intermediate tables. The cleaned data is stored in the SQL Server intermediate database, triggering tasks in the intelligent analysis layer.

[0127] In this embodiment, the intelligent analysis layer includes an adaptive quality prediction model, an anomaly detection module, and a root cause analysis engine. The intelligent analysis layer predicts and analyzes standardized data. The adaptive quality prediction model performs time series analysis on the standardized data to generate quality prediction results. The anomaly detection module performs pattern recognition on the standardized data to generate anomaly detection results. The root cause analysis engine calculates the probabilities of associated factors based on the anomaly detection results to generate root cause analysis data. The analysis parameters are adjusted based on the quality prediction results and the root cause analysis data.

[0128] The adaptive quality prediction model uses historical data trained on an LSTM neural network and dynamically adjusts parameters. This neural network is a sequential model built on the Keras framework and employs a stacked LSTM architecture. It progressively extracts time series features through two LSTM layers. The first LSTM layer retains the time dimension for processing by the second layer, which aggregates the time series information to generate the final representation. The output layer implements binary classification using a probabilistic sigmoid model, making it suitable for supervised learning scenarios where quality is predicted. It consists of the following layers stacked in sequence:

[0129] The first LSTM layer (long short-term memory network layer) contains 64 neuron units (units = 64). Set return_sequences = True to output the complete sequence of each time step (rather than just the result of the last time step) so that the subsequent recurrent layer can be connected.

[0130] The input shape is (timesteps, features), where timesteps represents the time step (sequence length) and features represents the feature dimension of each time step.

[0131] The dropout layer (regularization layer) has a dropout rate of 0.2, which means that 20% of the neuron outputs are randomly blocked during training, and overfitting is suppressed by introducing randomness.

[0132] The second layer, LSTM (nested long short-term memory network layer), contains 32 neuron units (units=32). By default, return_sequences=False, which only outputs the hidden state of the last time step to summarize the timing information.

[0133] The fully connected output layer (Dense layer) contains one neuron unit (units = 1) and uses the sigmoid activation function to compress the output to the interval [0, 1] to represent the probability value of qualified quality (binary classification task).

[0134] The pseudo code of the above network structure is expressed as:

[0135] Python

[0136] model=Sequential()

[0137] model.add(LSTM(units=64,return_sequences=True,input_shape=(timesteps,features)))#First layer LSTM (containing 64 neurons)

[0138] model.add(Dropout(0.2)) #prevent overfitting

[0139] model.add(LSTM(units=32)) #Second layer LSTM (32 neurons)

[0140] model.add(Dense(units=1,activation='sigmoid'))#output layer (quality pass probability)

[0141] The input data for this neural network model must undergo a systematic preprocessing process. For time series data collected by industrial sensors, characteristic dimensions such as temperature and pressure are first Z-score standardized by production batch to eliminate interference from different dimensions on model training. The continuous data is then cut into 60-minute time series segments (timesteps = 60) using a sliding time window to form the sample structure required for supervised learning. The quality inspection result corresponding to the end of each time window (0 for unqualified, 1 for qualified) serves as the training label. The final input data dimension is a three-dimensional tensor, containing the number of samples, 60 time steps, and five sensor monitoring features (number of samples, 60, 5). The output is the product quality acceptance probability value corresponding to the end of each time window (e.g., 0.93), which is used to predict the yield rate of the next batch of products.

[0142] The Adam optimizer is used for model training, with an initial learning rate of 0.001, a batch size of 32, and a basic iteration cycle of 100 times. The tuning strategy sets a dual trigger mechanism: when the AUC indicator of the validation set is lower than 0.85, the time series feature extraction capability is enhanced by stacking a third layer of 16-neuron LSTM network; if the training loss curve fluctuates violently, the learning rate is decayed (such as reducing it to 0.0001) or the Dropout ratio is increased to 0.3 to stabilize the training process. To adapt to the dynamic changes in production line data, the system starts the online learning mechanism every 24 hours. By freezing the weights of the first two layers of LSTM networks and only fine-tuning the terminal fully connected layer, the model parameters are gradually updated using the newly added data. This not only maintains the stability of the core feature extractor, but also enables the prediction layer to track subtle changes in the production process.

[0143] The anomaly detection module uses the isolation forest algorithm to identify data that deviates from the normal pattern in real time. The process is expressed as follows:

[0144] Input Data: Feature selection: Key quality indicators (e.g., dimensional deviation, surface roughness) + environmental parameters (temperature and humidity). Preprocessing: Numerical data is normalized to the range [0, 1]. Categorical features (e.g., device model) are converted to one-hot encoding. Dimensionality: (number of samples, features = 8).

[0145] Processing steps:

[0146] Construct an isolation tree: randomly select subsample = 256 data points and recursively split them into leaf nodes; the split threshold randomly selects features and split points.

[0147] Calculate the anomaly score: Path length: the number of edges from the root node to the leaf of the data point; the anomaly score formula is:

[0148] score=2 (-E(h(x)) / c(n))

[0149] Where h(x) is the path length and c(n) is the average path length of the tree.

[0150] Output result: anomaly score range [0,1], score > 0.65 is marked as anomaly (the threshold can be adjusted dynamically).

[0151] Parameter tuning: Number of trees: default n_estimators = 100, increase to 200 when the data volume is > 100,000; sample size: max_samples = 256, reduce to 128 for high-dimensional data (> 20 dimensions);

[0152] Tuning basis: Maximize F1-score (harmonic mean of precision and recall) through confusion matrix.

[0153] The root cause analysis engine locates the factors associated with quality issues based on the Bayesian network. The nodes of the Bayesian network include:

[0154] Root node: equipment fault code (such as Motor_Fault), raw material batch (Material_Batch);

[0155] Intermediate node: process parameter deviation (such as Temperature_Deviation);

[0156] Leaf node: quality defect type (such as Surface_Crack).

[0157] The dependency between nodes is expressed as: Material_Batch→Temperature_Deviation; Motor_Fault→Temperature_Deviation; Temperature_Deviation→Surface_Crack

[0158] Input Data: Preprocessing: Discretization: Continuous variables (such as temperature deviation) are segmented into Low / Medium / High; Missing Value Handling: Fill in missing values using maximum likelihood estimation (for example, if Material_Batch is missing, use the batch with the highest historical frequency). Format: Each data entry is a combination of discrete states (such as [Motor_Fault = Yes, Material_Batch = B123, ...]).

[0159] Specific process:

[0160] Parameter learning: Calculate the conditional probability table (CPT) from historical data, for example:

[0161] P(Surface_Crack=Yes|Temperature_Deviation=High)=0.78

[0162] Evidence propagation: When Surface_Crack=Yes is detected, the posterior probability of each root node is calculated in reverse:

[0163] P(Motor_Fault=Yes|Surface_Crack=Yes)=0.62

[0164] P(Material_Batch=B123|Surface_Crack=Yes)=0.41

[0165] Output: A ranked list of root cause probabilities (e.g., Motor_Fault: 62%, Material_Batch: 41%).

[0166] Parameter optimization: Structure learning: Optimize network topology using the K2 algorithm (based on score search); CPT adjustment: Recalculate CPT when new data causes a priori probability deviation >10%.

[0167] Overall, at the intelligent analysis layer, when the LSTM prediction of anomaly prediction defective rate suddenly increased to 15% (threshold = 10%), anomaly detection was triggered; then, the isolation forest method was used to locate the abnormal batch, identifying batch BATCH23001 with an anomaly score of 0.72;

[0168] Then, using Bayesian network root cause analysis, it was inferred that the probability of Motor_Fault reached 68%, suggesting that the motor should be repaired first;

[0169] Feedback closure: Maintenance results are updated to the knowledge base, optimizing the LSTM training data distribution. Analysis results are pushed to the application service layer and fed back to the model training library.

[0170] The application service layer of this embodiment includes a visual dashboard, a trusted traceability module, and a collaborative decision-making terminal. The visual dashboard displays quality trends, forecast results, and a KPI dashboard. The trusted traceability module uses blockchain (Hyperledger Fabric) to log key operations and ensure they cannot be tampered with. The collaborative decision-making terminal supports manual annotation of abnormal data, triggering process approval or work order dispatch.

[0171] The application service layer interacts strongly with users, and user operation records are sent back to the feedback closed layer to optimize the model and rule base.

[0172] The application service layer displays quality prediction results and anomaly detection results through a visual dashboard to generate quality trend data; generates operation logs for quality prediction results through a trusted traceability module and stores them in the blockchain to generate traceability records; obtains manual feedback data based on anomaly detection results through a collaborative decision-making terminal to generate decision recommendations; and generates decision support data based on quality trend data and decision recommendations.

[0173] The feedback closed layer in this embodiment includes a rule optimization module and a closed-loop work order module, which adjusts data cleaning rules and prediction model weights according to manual feedback, and adjusts the processing logic of the data processing layer and the intelligent analysis layer according to the decision support data, including: updating the cleaning rules according to the decision support data through the rule optimization module to generate an optimized rule set; generating improvement tasks according to the decision support data and tracking the execution status through the closed-loop work order system to generate work order results; updating the quality knowledge base according to the work order results to generate feedback data; and adjusting the cleaning logic of the data processing layer and the prediction parameters of the intelligent analysis layer according to the feedback data.

[0174] In terms of overall operation, at the data processing layer, dynamic routing and parallelized pipeline technologies significantly improve the efficiency of heterogeneous data processing through intelligent task scheduling. In traditional serial processing, data from different sources and types must be processed sequentially according to a fixed process (e.g., cleansing followed by conversion). This often results in high-priority tasks (such as real-time sensor streams) being blocked by lower-priority tasks (such as historical ERP data). To address this issue, the system employs a label-driven dynamic routing strategy. During the data ingestion phase, multi-source adapters inject metadata tags (e.g., type=sensor for sensor data, priority=high for urgent tasks) into each data entry, forming a comprehensive profile encompassing data source, type, and priority. Subsequently, a rules engine (e.g., SQL Server PolyBase) intelligently distributes data based on a pre-set routing rule table. Real-time sensor data tagged with priority=high is assigned to a real-time processing queue with eight threads and 30% of CPU resources, while low-priority historical data with type=ERP is assigned to a batch processing queue with only four threads and 10% of CPU resources. This differentiated resource allocation mechanism, combined with a parallel pipeline architecture, enables real-time data streams to be processed in milliseconds through independent thread fast channels (delay reduced from 10 seconds to 200 milliseconds), while batch data is executed asynchronously in the background, ultimately achieving a 3-fold increase in data throughput per unit time while ensuring the timeliness of processing critical business data.

[0175] At the data processing layer, dynamic fusion of heterogeneous multi-source data enables efficient cross-format and cross-source data integration through adaptive data structure parsing and resource optimization mechanisms. To address the inefficiencies inherent in traditional solutions, which require predefined fixed table structures and manual parsing code, the system employs dynamic schema mapping technology. First, metadata is extracted from the raw data (e.g., JSON key names and XML tags) to generate a field list. Field types are then automatically inferred based on the first 100 rows of sample data (e.g., continuous numeric values as floating-point types and text as strings). Finally, a virtual intermediate table is constructed in memory (e.g., dynamically parsing nested JSON to generate a temporary table using SQL Server's OPENJSON function). This eliminates the need to develop dedicated adapters for new data formats and reduces parsing time by 70%. Furthermore, the system makes intelligent routing decisions based on data tags (e.g., type=sensor marks sensor streams and priority=high identifies real-time tasks), assigning high-priority data to fast ETL pipelines for cleansing and transformation, while lower-priority tasks are processed asynchronously in batch queues. To accelerate the analysis of massive amounts of data, the in-memory computing engine uses column-based indexing technology to compress and store time series data (such as temperature and pressure sensor streams) in memory, increasing aggregate query speed by five times and reducing memory usage by 40%. This allows complex analysis to be completed without full data persistence. This technology system achieves real-time fusion and efficient utilization of multimodal data through the synergy of dynamic structure adaptation, priority-aware scheduling, and memory optimization.

[0176] A pseudo code example of dynamic fusion mapping is shown as follows:

[0177]

[0178] At the data processing layer, processing efficiency is improved through in-memory computing and columnar storage optimization. This optimization significantly improves the efficiency of massive data analysis by restructuring data storage and computing models. Traditional solutions require data to be persisted to disk before processing. Frequent I / O reads and writes result in latency of up to minutes for operations like aggregation analysis. This performance bottleneck is particularly pronounced when processing massive, correlated queries such as sensor time series data. To address this, the system utilizes in-memory computing technology. First, raw data is sharded by time windows (e.g., "recent 1-hour production data") or business topics. Then, using SQL Server's In-Memory OLTP technology, hot data blocks are loaded directly into in-memory tables, avoiding the performance penalties associated with disk access. Furthermore, columnar indexes are created for in-memory tables to target analytical queries (e.g., statistically average temperature and defect rate distribution). This allows for efficient compression of consecutive values within a column (e.g., pressure and speed), reducing storage space usage by 60%. Aggregation calculations only read relevant column data, rather than entire rows. For example, when executing a query like "Calculate the average temperature per batch," the column storage engine directly reads the compressed temperature column for batch calculations. Combined with the high-speed read and write capabilities of memory, this reduces the response time from minutes to seconds. This synergy between memory residency and columnar processing reduces storage costs while achieving orders of magnitude improvements in analytical performance.

[0179] Specific examples are:

[0180] Creating columnstore indexes to speed up quality analysis

[0181] CREATE COLUMNSTORE INDEX idx_quality ON QualityData(

[0182] batch_id,temperature,pressure,defect_rate );

[0184] Incremental fusion and local update technologies are also employed at the data processing layer. These mechanisms effectively address the resource waste associated with traditional full-data re-fusion by accurately identifying data changes and minimizing the processing scope. To address the duplicate computations (e.g., full data cleansing takes two hours per hour) and storage redundancy caused by periodic full-data processing, the system introduces Change Data Capture (CDC) technology. This technology uses SQL Server's CDC module to monitor data source changes in real time, extracting new records (e.g., real-time sensor readings) or field-level changes (e.g., order status updates), and generating log files containing only incremental data. The system then performs lightweight processing on the incremental datasets—automatically adapting to structural changes based on dynamic schema mapping (e.g., adding a new JSON field without modifying the parsing logic), performing cleansing and transformation operations in parallel, and finally merging the processed results with a snapshot of historical data versions (e.g., yesterday's production data baseline) to form the latest fused view. For example, when a new supplier field is added to the ERP system, the incremental mechanism only needs to perform metadata inference and type matching on the change record, rather than re-parsing all historical data. This on-demand processing mode reduces the data fusion time from 2 hours in full mode to 15 minutes. At the same time, by retaining only incremental logs instead of full copies, storage space requirements are reduced by 75%, achieving significant optimization of resource utilization while ensuring data consistency. Example: Enable CDC to capture changes to the order table

[0185] EXEC sys.sp_cdc_enable_table

[0186] @source_schema=N′dbo′,

[0187] @source_name=N′OrderMaster′,

[0188] @role_name = NULL;

[0189] Additionally, the system in this embodiment adopts hybrid storage and hot and cold tiered storage in terms of storage. In the traditional method, all data is stored in the same storage medium (such as HDD), and high-frequency access data and archived data are mixed and stored, resulting in low I / O efficiency. This embodiment adopts hot and cold data tiered storage to distribute data to different storage levels according to access frequency. The process is as follows: first, a heat analysis is performed to count the frequency of statistical access (such as the number of queries in the past 7 days); then, based on the heat analysis results, hot data caching and cold data archiving are performed. For hot data caching, high-frequency data is stored in SSD or memory (such as quality inspection results in the past 3 days); for cold data archiving, low-frequency data is migrated to object storage (such as obsolete production logs). As a result, in terms of I / O performance, the query speed of hot data is increased by 5 times; in terms of cost optimization, the cost of cold data storage is reduced by 80%.

[0190] Standardized data is stored in the SQL Server master database and divided into tables by subject (such as "production quality table" and "supply chain table"); the original data is backed up to the distributed file system (HDFS) for subsequent traceability.

[0191] In terms of specific implementation, the dynamic data fusion process can be summarized as follows: sensor data (JSON format) is parsed by the multi-source adapter and stored in the SQL Server buffer table; the dynamic cleaning module calls the stored procedure usp_CleanSensorData to remove records that exceed the threshold; the converter links the data to the production work order table based on "device ID-timestamp" to generate the intermediate table Tmp_QualityData.

[0192] The quality prediction and closed-loop feedback process can be summarized as follows: the intelligent analysis layer performs a prediction once an hour and generates a defective rate curve; the workshop director discovers anomalies through the information dashboard and manually marks incorrect prediction points; the feedback closed-loop layer updates the LSTM model weights and generates equipment maintenance work orders.

[0193] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.< / skill>

Claims

1. A digital integrated quality management system based on multi-source data fusion, characterized by: It includes data collection layer, data processing layer, intelligent analysis layer, application service layer and feedback closure layer, among which, The data collection layer includes a multi-source adapter, a real-time stream collection module, and a static data import module. It uses these modules to acquire multi-source heterogeneous data, which includes both real-time and static data. Real-time data is added with a timestamp and data source tag and then pushed to the message queue, while static data is temporarily stored in a buffer table. The data processing layer includes a dynamic cleaning module, a dynamic routing module, and a heterogeneous data converter. It removes outliers, fills in missing values, dynamically allocates routes, and converts unstructured data from multi-source heterogeneous data, generating standardized intermediate tables and storing them in the intermediate database. The intelligent analysis layer includes an adaptive quality prediction model, an anomaly detection module, and a root cause analysis engine. The intelligent analysis layer predicts and analyzes standardized data. It uses the adaptive quality prediction model to perform time series analysis on standardized data to generate quality prediction results. The anomaly detection module performs pattern recognition on standardized data to generate anomaly detection results. The root cause analysis engine calculates the probability of associated factors based on the anomaly detection results to generate root cause analysis data. The analysis parameters are adjusted based on the quality prediction results and root cause analysis data. The application service layer includes a visual dashboard, a trusted traceability module, and a collaborative decision-making terminal; the feedback closure layer adjusts the processing logic of the data processing layer and the intelligent analysis layer based on the decision support data.

2. A digital integrated quality management system based on multi-source data fusion according to claim 1, characterized in that: In the data collection layer, multi-source adapters support API, MQTT, and JDBC protocols for connecting to sensors, ERP, MES, and manual input terminal data sources; The real-time stream acquisition module processes high-speed sensor data streams through Kafka and generates real-time data streams with metadata descriptions; the static data import module batch imports Excel, CSV structured or semi-structured data into the SQL Server buffer table.

3. The digital integrated quality management system based on multi-source data fusion according to claim 1, characterized in that: In the data processing layer, the dynamic cleaning module uses the rules engine to perform the following operations: For structured data, field-level rules are used to fill in null values, standardize date formats, and verify uniqueness; for semi-structured data, dynamic schema mapping rules are used to perform JSON / XML path parsing and tag completion; for unstructured data, context association rules are used to verify image resolution, extract text keywords, and generate file hash values. The rules engine also sets general rules, including timestamp standardization rules, data source credibility weighting rules, and basic missing value processing rules. The dynamic routing module assigns processing paths based on data type labels. Structured data directly triggers field-level cleaning rules and is written to the intermediate database. Semi-structured data is parsed into temporary tables and then triggers dynamic schema cleaning rules. Unstructured data uses external tools to extract metadata and is then stored in association with structured data. The heterogeneous data converter converts JSON or XML nested data into a relational table through the following steps: parsing the JSON / XML main table fields and inserting them into the main table; expanding the nested array or hierarchical structure to generate a detailed table and normalizing the data through primary and foreign key associations; performing strong type verification during the parsing process and defining field types.

4. The digital integrated quality management system based on multi-source data fusion according to claim 3 is characterized by: In the data processing layer, the operations performed include: Dynamic Schema Mapping and Lightweight Parsing: This system automatically parses semi-structured data using SQL Server's OpenJSON functions and XPath parsing rules. It generates a field list by extracting metadata header information, infers field types based on sample data, and constructs a virtual intermediate table in memory. For nested structures, it uses a lazy loading mechanism to parse specific content only when a query is accessed. It also generates hash values for image files and establishes a joint index with the primary key of the structured data. In-memory computing and column storage optimization: Data is horizontally sharded by time range, loaded into the in-memory column store using SQL Server's In-Memory OLTP technology, and a clustered column store index is created. Numeric columns are pre-allocated using contiguous memory blocks, vectorized aggregate calculations are performed using the SIMD instruction set, and the bit-packing algorithm is applied to compress storage space. Dynamic routing and parallelized pipeline: Source type and priority tags are injected during the data collection phase, and data is distributed to real-time or batch processing queues based on a rules engine. Data is sharded based on the last two digits of the device ID hash value, with corresponding CPU threads and memory quotas allocated to the real-time queue and threads and quotas to the batch queue. Multi-channel task synchronization is achieved through a lock-free ring buffer, and atomic compare-and-swap operations are used to update read and write pointers. Incremental fusion and partial updates: SQL Server CDC is used to capture incremental data changes, and temporal tables are enabled to store historical version snapshots. Lightweight cleansing is performed on incremental data, and the incremental results are merged with the snapshot table by device ID and batch number using the MERGE statement. During cold data migration, consistent bitmap indexes from row numbers to disk page numbers are maintained, supporting cross-storage tier federated queries. Hybrid storage and hot-cold tiering: Column-level access frequency is monitored, and frequently accessed data of the day is retained in the in-memory column storage area. Cold data that has not been accessed for a preset time is page-compressed and then migrated to the disk row storage area. Hot data uses column storage indexes to accelerate aggregate analysis, while cold data is quickly located through bitmap indexes. When it is frequently accessed again, a preload thread is triggered to restore the in-memory copy.

5. The digital integrated quality management system based on multi-source data fusion according to claim 1, characterized in that: At the intelligent analysis layer, the adaptive quality prediction model is based on historical data trained on an LSTM neural network. Built on the Keras framework, it uses a stacked LSTM structure, gradually extracting time series features through two LSTM layers. The first LSTM layer retains the time dimension for processing by the second layer, which aggregates the time series information to generate the final representation. The output layer implements binary classification using probabilistic methods. The anomaly detection module uses the isolation forest algorithm to identify data that deviates from the normal pattern in real time. The process is expressed as follows: Construct an isolated tree: randomly select subsample data and recursively split it into leaf nodes; the split threshold randomly selects features and split points; Calculate the anomaly score: The path length is the number of edges from the root node to the leaf node of the data point; the anomaly score formula is: score=2 (-E(h(x)) / c(n)) Where h(x) is the path length and c(n) is the average path length of the tree; The root cause analysis engine locates the factors associated with quality issues based on the Bayesian network. The nodes of the Bayesian network include: root nodes, intermediate nodes, and leaf nodes. The root nodes are equipment fault codes and raw material batches; the intermediate nodes are process parameter deviations; and the leaf nodes are quality defect types. Specific dependencies exist between the nodes. Calculate the conditional probability table using historical data: P(Surface_Crack=Yes|Temperature_Deviation=High) Evidence propagation: When Surface_Crack=Yes is detected, the posterior probability of each root node is calculated in reverse: P(Motor_Fault=Yes|Surface_Crack=Yes) P(Material_Batch=B123|Surface_Crack=Yes) Output: A ranked list of root cause probabilities.

6. The digital integrated quality management system based on multi-source data fusion according to claim 1, characterized in that: At the application service layer, the application service layer displays quality prediction results and anomaly detection results through a visual dashboard, generating quality trend data; The trusted traceability module generates an operation log for the quality prediction results and stores it in the blockchain to generate a traceability record; Obtain manual feedback data based on anomaly detection results through collaborative decision-making terminals to generate decision recommendations; Generate decision support data based on quality trend data and decision recommendations.

7. The digital integrated quality management system based on multi-source data fusion according to claim 1, characterized in that: In the feedback closed layer, the processing logic of the data processing layer and the intelligent analysis layer is adjusted according to the decision support data, including: updating the cleaning rules according to the decision support data through the rule optimization module to generate an optimization rule set; generating improvement tasks according to the decision support data and tracking the execution status through the closed-loop work order system to generate work order results; updating the quality knowledge base according to the work order results to generate feedback data; and adjusting the cleaning logic of the data processing layer and the prediction parameters of the intelligent analysis layer according to the feedback data.

Images