Order auditing method based on large model of manufacturing industry

Abstract

The invention provides an order auditing method based on a large model in the manufacturing industry, which belongs to the technical field of large models and comprises the following steps: step 1, collecting order information and confirmation order information; step 2, preprocessing order data to form a data set required by fine tuning of the model; 3, confirming single data preprocessing, and forming a data set required by model fine tuning; 4, dividing a test set and a training set, and generating a training set in a json format; and step 5, performing SFT fine tuning training. Step 6, scoring standard assignment and model evaluation; and step 7, realizing integration of a front-end system and a rear-end system, and performing real-time notification and early warning. The accuracy, efficiency and intelligent level of order auditing are improved, the manual auditing cost is reduced, and smooth proceeding of the production process is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of large models, and in particular to an order review method based on a large model of the manufacturing industry. Background Art

[0002] With the rapid development of the global manufacturing industry, companies are facing increasingly fierce market competition. To stand out from the competition, companies are seeking breakthroughs in improving production efficiency, reducing costs, and optimizing customer service. In this process, order review, as a critical link in a company's production operations, has a direct impact on production planning, inventory management, customer satisfaction, and ultimately, economic benefits.

[0003] Traditional order review methods rely primarily on manual labor, verifying and analyzing each order to ensure its legality, rationality, and feasibility. However, this approach presents numerous problems when processing large-scale, high-frequency orders: It is inefficient, requiring significant time and human resources for manual review. Especially during peak order periods, reviewers are often overwhelmed, resulting in slow order processing. It is also inaccurate, as manual review is susceptible to subjective factors such as fatigue and lack of experience, leading to errors in review results and, in turn, impacting the accuracy of production plans. It is also costly, as companies expand and order volumes increase, requiring more manpower for order review, leading to rising labor costs. It is also difficult to adapt to complex scenarios: In the manufacturing industry, order types are diverse, involving multiple factors such as product specifications, production processes, and delivery deadlines. Manual review cannot quickly adapt to complex and changing order scenarios, making it prone to errors. It is also low in intelligence: In today's information-based and intelligent era, traditional manual review methods are no longer able to meet the needs of companies for efficient and intelligent production management. Summary of the Invention

[0004] In order to solve the above technical problems, the present invention provides an order review method based on a large model of the manufacturing industry, which realizes rapid and accurate review of orders, thereby improving the production and operation efficiency of enterprises, reducing costs, and improving customer satisfaction. With the continuous maturity of big data technology, enterprises can collect, store and analyze large amounts of order data to provide data support for building a large model for order review. Machine learning algorithms have made significant progress, enabling computers to automatically learn and extract useful information from data, providing intelligent means for order review. In the process of transformation and upgrading of the manufacturing industry, enterprises have an increasingly urgent demand for informatization and intelligence. Order review, as a key link, urgently needs to be transformed into an intelligent one. Faced with market competition, enterprises need to continuously improve production efficiency and reduce costs to maintain their competitive advantage.

[0005] The technical solution of the present invention is:

[0006] An order review method based on a large model of the manufacturing industry includes the following steps:

[0007] Step 1: Collect order information and confirmation information;

[0008] Step 2: pre-process the order data to form the data set required for model fine-tuning;

[0009] Step 3: Confirm single data preprocessing to form the data set required for model fine-tuning;

[0010] Step 4: Divide the test set and training set, and generate the training set in JSON format;

[0011] Step 5, SFT fine-tuning training;

[0012] Step 6: Specify scoring criteria and evaluate the model;

[0013] Step 7: Implement front-end and back-end system integration, real-time notification and warning.

[0014] Further,

[0015] The data sources include manual input and system interface docking.

[0016] Connect with the information systems of various departments and obtain order data in real time through API interface

[0017] Further,

[0018] Step 2 is to convert the raw order data collected in step 1 into a dictionary or object format that is easy to process.

[0019] The order data is preprocessed and cleaned; the text format is unified using string processing functions, and redundant spaces in the text are removed through regular expression replacement; the data is converted into a unified format; for numerical data, it is ensured that they are compared at the same order of magnitude; duplicate records are identified based on unique identifiers, duplicate records are deleted, and the latest or most complete records are retained.

[0020] Further,

[0021] Step 3 specifically includes

[0022] Image denoising: Use OpenCV image processing technology to denoise the image;

[0023] Image enhancement: adjust the brightness, contrast, and clarity of the image to improve the accuracy of subsequent text recognition;

[0024] Image cropping: Based on the structure of the confirmation form, crop the part containing key information;

[0025] Use OCR technology to perform text recognition on the preprocessed image, extract the text information in the image into text format; verify and correct the OCR recognition results; store the extracted text information by line or by cell;

[0026] Text post-processing includes: First, use the language model Bert to correct the recognition results; Second, format restoration, restore the format of the text to ensure that the recognition results are consistent with the original picture; Third, semantic verification, verify the recognition results according to the context semantics;

[0027] Perform standardization processing on the extracted text information to handle missing values or recognition errors.

[0028] Furthermore,

[0029] Use 80% of the samples in the collected dataset as the training set, and the remaining 20% as the test set;

[0030] Generate a training set in json format

[0031] Organize the order and confirmation form information produced in steps 1, 2, 3, and 4 into the following json format: instruction, order information, confirmation form information, output format, where instruction is "<Instruction>Your task is to compare the differences between the order and the confirmation form information one by one, and then summarize them into a list of strings and return;< / Instruction>";

[0032] The output result in the output format is order_id.

[0033] Furthermore,

[0034] For the SFT fine-tuning training, use the AdamW optimizer to start the model and perform tuning.

[0035] Start the model training script to perform supervised fine-tuning training; during the training process, monitor and tune the model performance in the following ways: Logging: Use a logging tool to record the losses and evaluation metrics during the training process; Visualization tool: Use TensorBoard to visualize the training curve and monitor the model performance in real time; Early stopping mechanism: If the validation set loss does not decrease for several consecutive rounds, stop the training early to prevent overfitting.

[0036] Furthermore,

[0037] Integrate the front end with the back-end service. Through REST API or GraphQL, the front end can send requests to the back end and obtain the required data;

[0038] Integrate the real-time notification function, which can automatically send warnings for notification;

[0039] Integrated report generation function supports the generation of detailed audit reports; the report generation tool can be integrated with the system and supports the export of reports to PDF and Excel formats.

[0040] The beneficial effects of the present invention are

[0041] (1) Improve audit efficiency and shorten processing time. By adopting the SFT fine-tuning model as the core technology, the method greatly improves the efficiency of order audit through automation. In the traditional order audit process, manual audit requires a lot of time to compare order information and confirmation sheets. Especially when the order volume is large and the information is complex, the audit process often takes a long time. The present invention uses advanced big data and artificial intelligence technology to quickly complete the extraction of order information and the identification of confirmation sheets, realize the automation of the audit process, and greatly shorten the processing time. This not only improves the work efficiency of the enterprise, but also reduces production delays caused by audit delays, thereby improving customer satisfaction.

[0042] (2) Improve audit accuracy and reduce error rates. Since manual audits are easily affected by factors such as fatigue and distraction, the error rate is relatively high. However, the present invention uses a large model to intelligently compare order information and confirmation sheets, which can effectively identify subtle differences and reduce human errors. In addition, the system can also self-learn and optimize based on historical data and audit rules to continuously improve the accuracy of audits. This highly automated audit method not only reduces problems such as rework and returns caused by audit errors, but also reduces the company's potential economic losses and enhances the company's market competitiveness.

[0043] (3) Optimize resource allocation and reduce operating costs. Traditional order review requires a lot of human resources, but the present invention reduces the necessity of manual participation through automated review, thereby reducing labor costs. At the same time, the efficient operation of the system enables enterprises to arrange production and logistics plans more reasonably, optimize inventory management, and reduce resource waste. In addition, since the present invention adopts big data and artificial intelligence technology, with the continuous advancement of technology and large-scale application, long-term operating costs will be further reduced. In summary, the order review method and system of the present invention bring efficient, accurate, and low-cost review experience to enterprises, which helps enterprises maintain their advantages in the fierce market competition and achieve sustainable development. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a schematic diagram of the workflow of the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0046] This paper proposes an order review method based on a large model of the manufacturing industry, aiming to bring efficient and accurate order review experience to enterprises and help them achieve sustainable development.

[0047] The specific steps include:

[0048] Step 1: Collect order information and confirmation information.

[0049] Step 2: Preprocess the order data to form the data set required for model fine-tuning.

[0050] Step 3: Confirm the single data preprocessing to form the data set required for model fine-tuning.

[0051] Step 4: Divide the test set and training set, and generate the training set in JSON format.

[0052] Step 5: SFT fine-tuning training.

[0053] Step 6: Specify scoring criteria and evaluate the model.

[0054] Step 7: Implement front-end and back-end system integration, real-time notification and warning.

[0055] Described step 1 specifically comprises the following:

[0056] The first step in data collection is to identify the data source. In the manufacturing industry, order data primarily comes from multiple departments, including sales, production, procurement, and logistics. Therefore, it's important to ensure that you obtain complete and accurate order information from these departments, including but not limited to order number, customer number, product number, product name, product specifications, quantity, unit price, total price, delivery date, and order status.

[0057] To ensure comprehensiveness and accuracy of data, we employ a combination of data collection methods. These include: manual entry of key information, such as customer information and order numbers, to ensure accuracy and traceability. System interface integration connects to various departments' information systems and obtains order data in real time through APIs. This approach ensures real-time data accuracy while reducing the error rate associated with manual data entry.

[0058] Described step 2 specifically comprises the following:

[0059] Collect the raw order data from step 1 and convert it into a dictionary or object format that is easy to process.

[0060] 2.1 Data Cleaning

[0061] Use string processing functions to unify text formats. Use regular expression replacement to remove extra spaces in the text, including spaces at the beginning and end of lines, multiple consecutive spaces, etc., to make the text more regular. For missing or abnormal data (such as null values, negative numbers, non-numeric characters, etc.), fill or correct them according to business rules. Convert data to a unified format, such as unifying the date format to YYYY-MM-DD, and the price format to two decimal places. For numerical data, ensure that they are compared at the same order of magnitude to avoid misjudgment due to different units. Identify duplicate records based on unique identifiers such as order numbers and product numbers. Delete duplicate records and retain the latest or most complete records.

[0062] Described step 3 specifically comprises the following:

[0063] Confirmation forms are typically stored in image formats (e.g., PNG, JPEG) or PDF, and contain information such as tables, text, and barcodes. The goal of preprocessing is to extract structured text from the image and compare it with the order information.

[0064] 3.1 Image Preprocessing

[0065] Image preprocessing uses the following methods: Image denoising: Using OpenCV image processing technology to denoise images and remove background noise, watermarks, and other interfering information. Image enhancement: Adjusting the image's brightness, contrast, and clarity improves the accuracy of subsequent text recognition. Image cropping: Based on the structure of the confirmation form, cropping out sections containing key information (such as table areas) to reduce unnecessary computation.

[0066] 3.2 Text Recognition

[0067] Use OCR technology to perform text recognition on the pre-processed image and extract the text information in the image into text format. Verify and correct the OCR recognition results to ensure the accuracy of recognition. For example, for characters that are recognized incorrectly, they can be inferred and replaced based on the context. The extracted text information is stored by row or cell for subsequent parsing. Text post-processing is the last step of OCR technology, and its purpose is to improve the accuracy and readability of the recognition results. The text post-processing steps include: first, using the language model Bert to correct the recognition results. For example, correct the recognized "O" to "0". Second, format restoration, restore the text format (such as paragraphs, tables), and ensure that the recognition results are consistent with the original image. Third, semantic verification, verify the recognition results based on the context semantics. For example, check whether the recognized date and amount are logical.

[0068] 3.3 Data Standardization

[0069] Standardize the extracted text information, such as unifying the date format and unit to ensure consistency with the order information format. Handle missing values or identification errors, such as inferring missing fields through contextual information.

[0070] Described step 4 specifically comprises the following:

[0071] 4.1 Use 80% of the samples in the above collected data set as the training set and the remaining 20% as the test set.

[0072] At the same time, consider the temporal distribution of samples. Generally speaking, use older patent samples as the training set and newer patent samples as the test set. This not only better evaluates the model's generalization performance on new patents but also reflects the dynamic changes in technological development.

[0073] 4.2. Generate training set in json format

[0074] The order and confirmation form information produced through steps 1, 2, 3, and 4 shall be organized into the following JSON format: instruction, order information, confirmation form information, output format, where instruction is "<Instruction>Your task is to compare the differences between the order and confirmation form information item by item, and then summarize them into a list of strings and return.< / Instruction>". The content in the order information includes: applying unit, cloud migration project, original product configuration, current product configuration, and billing start time. The original product configuration and current configuration include the quantity and configuration of cloud hard disks, the quantity and configuration of cloud servers, etc. The output format of the output result is order_id. For example, "The applying unit in the order is 'Rizhao Arbitration Commission', and the applying unit in the confirmation form is 'Office of Rizhao Arbitration Commission', there is a difference. The billing start time in the order is '2022-06-29 00:00:00', and the billing date in the confirmation form is '2022 / 6 / 30', there is a difference."

[0075] Step 5 specifically includes the following:

[0076] 5.1 Place the JSON format training set file generated in step 4 under the / data / NL2SQL directory

[0077] 5.2 Modify the fine-tuning startup parameters. Set pretrained_model to the location of the base model. In this patent, the base model uses qwen_72B. dataset_path is the location of the training set JSON, and output_dir is the location of the output model file. nproc_per_node_num is the number of cards used for this fine-tuning. In this patent, it is set to 3. CUDA_VISIBLE_DEVICES is the cards used for this fine-tuning. In this patent, it is set to '1,2,3,4'. Set the initial learning rate to 3e-5. Gradually increase the learning rate in the first 10% of the training steps. The batch size affects the training speed and stability of the model. In this patent, the batch size is set to 16. The number of training rounds is set to 5 rounds.

[0078] 5.3 Optimizer selection

[0079] Use the AdamW optimizer, and set the weight decay to 0.01 to prevent overfitting. The momentum parameters are β1 = 0.9 and β2 = 0.999, using the default settings of AdamW, which is suitable for most tasks.

[0080] 5.4 Start the model and tuning

[0081] Start the model training script and perform supervised fine-tuning training. During the training process, you can monitor and tune the model performance in the following ways: Logging: Use logging tools (such as logging) to record the loss and evaluation indicators during the training process. Visualization tools: Use TensorBoard to visualize the training curve and monitor the model performance in real time. Early stopping mechanism: If the validation set loss does not decrease within several consecutive rounds, stop training early to prevent overfitting.

[0082] Described step 6 specifically comprises the following:

[0083] The construction of a scoring mechanism is a key step in ensuring that the differences between orders and confirmations can be accurately identified and evaluated. The design of the scoring mechanism requires comprehensive consideration of multiple factors, including the type of difference, severity, impact on the production process, etc. The following is a detailed description of the construction of the scoring mechanism. The full score is 100 points. The scoring index setting consists of the following two parts: Difference point identification accuracy: This indicator measures the system's ability to identify differences between orders and confirmations, including the difference point detection rate and false alarm rate. Difference point severity score: This indicator scores the difference point based on the impact of the difference on order execution, which is divided into three levels: mild, moderate, and severe.

[0084] 6.1 Difference Point Identification Accuracy Scoring

[0085] Discovery rate: Full marks will be awarded if the discovery rate is above 90%, and 1 point will be deducted for every 1% decrease.

[0086] False alarm rate: If the false alarm rate is below 5%, full marks will be awarded, and 2 points will be deducted for every additional 1%.

[0087] 6.2 Difference Severity Scoring

[0088] Minor differences: 1 point for each one found.

[0089] Moderate differences: 3 points for each one found.

[0090] Major differences: 5 points for each one found.

[0091] Described step 7 specifically comprises the following:

[0092] 7.1 Front-end and back-end integration

[0093] The front-end of the user interaction module (such as a web page or mobile app) needs to be integrated with the back-end services (such as order entry service and audit result query service). Through REST API or GraphQL, the front-end can send requests to the back-end and obtain the required data.

[0094] 7.2 Real-time notification and warning

[0095] The user interaction module needs to integrate real-time notification capabilities (such as WebSocket, email notifications, and SMS notifications) to promptly notify relevant personnel when high-risk discrepancies are discovered. For example, if the system identifies a delivery time discrepancy that may lead to production delays, it can automatically send an early warning notification to the production manager.

[0096] 7.3 Report Generation and Export

[0097] The user interaction module needs to integrate report generation capabilities to support the generation of detailed audit reports. Reports can include a list of differences, scoring results, and action suggestions. Report generation tools (such as JasperReports and Apache POI) can be integrated with the system to support exporting reports to formats such as PDF and Excel.

[0098] This invention intelligently compares differences between order systems and confirmations, improving the accuracy, efficiency, and reliability of order review, thereby reducing enterprise operating costs and improving customer satisfaction. Improving review accuracy: In the manufacturing industry, order review is a critical step in ensuring order accuracy. By building a large model, this invention conducts in-depth analysis of the order system and confirmations, automatically identifying differences, effectively avoiding omissions and errors that may occur during manual review, and ensuring the accuracy of order information. Improving review efficiency: Traditional order review relies on manual operations and is time-consuming. This invention leverages big data and artificial intelligence technologies to automate order review, significantly shortening review time and improving review efficiency, helping enterprises quickly respond to market demand. Reducing operating costs: Through intelligent review methods, this invention reduces reliance on manual review and reduces labor costs. Furthermore, it avoids rework and returns caused by order errors, reducing enterprise operating costs. Improving customer satisfaction: Accurate and efficient order review helps ensure product quality and delivery time, thereby improving customer satisfaction. This invention provides enterprises with an order review method and system based on a large model for the manufacturing industry, helping to establish a positive corporate image and enhance customer trust. Optimize internal enterprise management. By optimizing the order review process, this invention facilitates collaboration between departments within an enterprise and improves overall enterprise management. It is easy to implement and scale. Its modular design allows for easy integration with existing enterprise systems, making it simple to implement. Furthermore, its efficient and accurate review results are easily recognized by enterprises.

[0099] The above description is only a preferred embodiment of the present invention and is only used to illustrate the technical solution of the present invention, and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. An order review method based on a large model of the manufacturing industry, It is characterized in that it includes the following steps: Step 1, collect order information and confirmation form information; Step 2, preprocess the order data to form a data set required for model fine-tuning; Step 3, preprocess the confirmation form data to form a data set required for model fine-tuning; Step 4, divide the test set and the training set, and generate a training set in json format; Step 5, perform SFT fine-tuning training; Step 6, specify the scoring criteria and evaluate the model; Step 7, implement the integration of the front-end and back-end systems, and provide real-time notifications and warnings.

2. The method according to claim 1, wherein the data sources include manual input and system interface docking.

3. The method according to claim 2, wherein docking with the information systems of each department, and obtaining order data in real time through the API interface.

4. The method according to claim 1, wherein Step 2 is to convert the original order data collected in Step 1 into a dictionary or object format that is easy to process.

5. The method according to claim 4, wherein in the preprocessing of the order data, data cleaning is performed; the text format is unified by using string processing functions, and redundant spaces in the text are removed through regular expression replacement; the data is converted into a unified format; for numerical data, ensure that they are compared on the same order of magnitude; duplicate records are identified according to the unique identifier, and the duplicate records are deleted, and the latest or most complete records are retained.

6. The method according to claim 1, wherein Step 3 specifically includes Image denoising: Use the image processing technology OpenCV to perform denoising processing on the image; Image enhancement: Adjust the brightness, contrast, and clarity of the image to improve the accuracy of subsequent text recognition; Image cropping: Crop out the part containing key information according to the structure of the confirmation form; Use OCR technology to perform text recognition on the preprocessed image, and extract the text information in the image into text format; Verify and correct the OCR recognition results; Store the extracted text information by row or by cell; The text post-processing includes: First, use the language model Bert to correct the recognition results; Second, format restoration, restore the format of the text to ensure that the recognition results are consistent with the original image; Third, semantic verification, verify the recognition results according to the context semantics; Perform standardization processing on the extracted text information to handle missing values or recognition errors.

7. The method according to claim 1, wherein 80% of the collected data set samples are used as the training set, and the remaining 20% are used as the test set; Generate a training set in json format Sort out the order and confirmation form information produced through Steps 1, 2, 3, and 4 into the following json format: instruction, order information, confirmation form information, output format, where instruction is "<Instruction> Your task is to compare the differences between the order and the confirmation form information one by one, and then summarize them into a list of strings and return; < / Instruction>"; The output result in the output format is order_id.

8. The method according to claim 1, wherein The SFT fine-tuning training uses the AdamW optimizer to start the model and tune it.

9. The method according to claim 8, characterized in that Start the model training script and perform supervised fine-tuning training. During the training process, monitor and tune the model performance through the following methods: Logging: Use logging tools to record losses and evaluation metrics during training. Visualization tools: Use TensorBoard to visualize training curves and monitor model performance in real time. Early stopping mechanism: If the validation set loss does not decrease within several consecutive rounds, stop training early to prevent overfitting.

10. The method according to claim 1, characterized in that The front-end is integrated with the back-end services. Through REST API or GraphQL, the front-end can send requests to the back-end and obtain the required data; Integrated real-time notification function, which can automatically send early warning notifications; Integrated report generation function supports the generation of detailed audit reports; the report generation tool can be integrated with the system and supports the export of reports to PDF and Excel formats.

Images