Method and device for automatically detecting consistency of automobile drawing annotation by using computing power of intelligent computing cloud platform

Abstract

The application provides a kind of intelligent computing cloud platform through computing power and carries out automobile drawing annotation consistency automatic detection method and device, it is related to the technical field of wisdom calculation cloud platform and computing power infrastructure, this method includes: based on the consistency semantic tree of pre-set sample data set is constructed, and the sample data set is divided into training set and verification set;Intelligent computing cloud platform utilizes computing power, based on the training set, the initial model is iteratively trained N times, and the initial model after the Mth iteration training is determined as the final model;The final model is deployed on the intelligent computing cloud platform to perform annotation consistency detection on the received batch of automobile drawings.This way, the model based on the characteristics of the early training can accurately locate the annotation position and identify abnormal annotation in the automobile drawing, reducing the understanding differences in the process of manual annotation consistency review of automobile drawings, improving the detection efficiency and accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent computing cloud platform and computing power infrastructure technology, specifically to an automatic detection method and device for the consistency of automotive drawing annotations using computing power on an intelligent computing cloud platform. Background Technology

[0002] With the rapid development of artificial intelligence technology, the "intelligent computing cloud platform" has emerged.

[0003] "Intelligent computing cloud platform" refers to a cloud computing platform that provides intelligent computing services based on the hardware and software resources of an intelligent computing center.

[0004] "Intelligent computing" refers to computing driven by intelligent computing power (GPU, FPGA, ASIC, etc.), and is a general term for computing theories, methods, architectures, and technical capabilities characterized by data intensity and model-based nature. Its goal is to obtain better results by utilizing specific computing resources and appropriate computing methods according to the needs of the task.

[0005] A "Smart Computing Center" refers to a facility that utilizes large-scale heterogeneous computing resources, including general-purpose computing power (CPU) and intelligent computing power (GPU, FPGA, ASIC, etc.), to provide the necessary computing power, data, and algorithms for artificial intelligence applications (such as the development, training, and inference of deep learning models). A Smart Computing Center encompasses facilities, hardware, and software, and can provide full-stack capabilities from underlying computing power to top-level application enablement.

[0006] "Computing power" is the core of "intelligent computing cloud platform". It is the ability of computer equipment or computing / data center to process information. It is the ability of computer hardware and software to work together to perform a certain computing requirement. It is the computing power to achieve the target output by processing information data. It is a new type of productivity that integrates information computing power, network carrying capacity and data storage capacity. It mainly provides services to society through computing power infrastructure.

[0007] Throughout the entire process of automotive R&D, parts manufacturing, final assembly, inspection, and final acceptance, the consistency of automotive drawing annotations is a core prerequisite for accurately conveying design intent. The combination of "dimensional value + unit + error range" in drawings, such as "50mm±2mm" or "25℃±1℃," not only provides a quantitative basis for processing precision but is also a key principle for avoiding manufacturing deviations. Inconsistencies in annotation such as "50mm±2" or "15℃±2°" (indicating unit mismatch or missing units) can easily lead to inaccurate processing. Therefore, the consistency of automotive drawing annotations directly determines the accuracy of the production end's reproduction of the design scheme. With the deepening of digital transformation in the automotive industry, companies face the dual demands of archiving and reviewing tens of thousands of historical drawings and verifying massive amounts of new drawings. However, the consistency check of automotive drawing annotations, as a core verification step, is deeply mired in a dilemma between efficiency and accuracy. Currently, a large number of automotive technical drawings are still archived in bitmap formats such as scanned copies and photocopies, which commonly suffer from font distortion, image tilting, watermark overlay, and noise interference caused by paper aging. Traditional manual review requires checking unit symbols, error symbols, and numerical relationships in the annotations one by one, which is time-consuming and labor-intensive. Furthermore, due to differences in the understanding of the dimensions of automotive drawings among reviewers, it is impossible to form standardized inspection conclusions, which seriously slows down the progress of digital archiving and production preparation.

[0008] While some automated detection methods exist in existing technologies, none are designed to address the consistency requirements of automotive drawing annotations, and they differ fundamentally from the application scenarios supported by intelligent computing centers. Current automated detection methods are typically based on a simple hardware architecture of a single workstation (CPU + entry-level GPU), generally using traditional image processing libraries such as OpenCV (e.g., Canny / Hough operators). This is vastly different from the application scenarios of automated detection methods in intelligent computing centers. Intelligent computing centers require the integration of kilo-level GPU clusters (e.g., NVIDIA A100 / H100) and high-speed RDMA networks to achieve EFLOPS-level computing power. This architecture requires the computing network to achieve 400Gbps inter-node communication, the storage network to reach 200Gbps bandwidth, and overall latency to be controlled at the sub-microsecond level.

[0009] It is evident that since the emergence of intelligent computing centers, how to use computing power to automatically detect the consistency of automotive drawing annotations in order to improve detection efficiency and accuracy has become an urgent problem to be solved. Summary of the Invention

[0010] This invention provides a method and apparatus for automatically detecting the consistency of automotive drawing annotations using computing power on an intelligent computing cloud platform. This addresses the problem of how to automatically detect the consistency of automotive drawing annotations using computing power to improve detection efficiency and accuracy since the emergence of intelligent computing centers.

[0011] To solve the above problems, the present invention is implemented as follows:

[0012] In a first aspect, embodiments of the present invention provide a method for automatically detecting the consistency of automotive drawing annotations using a smart computing cloud platform based on computing power, the method comprising:

[0013] Step S1: Construct a sample dataset based on a preset consistency semantic tree, and divide the sample dataset into a training set and a validation set. The consistency semantic tree is used to represent the standard style of annotations in automotive drawings. The annotations include a first value, a first unit corresponding to the first value, a plus or minus sign, a second value, and a second unit corresponding to the second value. The consistency semantic tree includes a semantic layer, a formal layer, and a validation layer. The semantic layer is used to define that the categories of the first unit and the second unit are consistent. The formal layer is used to define that the symbols of the first unit and the second unit are consistent. The validation layer is used to define the logical relationship between the first value, the first unit, the second value, and the second unit.

[0014] Step S2: The intelligent computing cloud platform uses computing power to perform N iterations of training on the initial model based on the training set. The initial model after the Mth iteration is determined as the final model. The loss value of the initial model after the Mth iteration in the validation set is less than the loss value of the initial model after the Lth iteration in the validation set. N is a positive integer, M is a positive integer less than or equal to N, L is a positive integer less than or equal to N, and L is not equal to M. The Lth iteration is any one of the N iterations except the Mth iteration.

[0015] Step S3: Deploy the final model on the intelligent computing cloud platform to perform batch annotation consistency checks on the received automotive drawings.

[0016] In one embodiment, step S1 includes:

[0017] Step S11: Based on the consistent semantic tree, construct a standard style set for annotation writing;

[0018] Step S12: Perform random unit replacement processing on the standard style of the annotation to obtain a first sample dataset. The first sample dataset includes abnormal unit category styles and abnormal unit symbol styles. The abnormal category styles indicate that the categories of the first unit and the second unit are inconsistent, and the abnormal unit symbol styles indicate that the symbols of the first unit and the second unit are inconsistent.

[0019] Step S13: Randomly delete characters from the standard style of the annotation to obtain a second sample dataset. The second sample dataset includes a unit missing style, a value missing style, and a plus or minus sign missing style. The unit missing style indicates that the first unit and / or the second unit is missing. The value missing style indicates that the first value and / or the second value is missing. The plus or minus sign missing style indicates that the plus or minus sign is missing.

[0020] Step S14: Perform random character mixing on the standard style of the annotation to obtain a third sample dataset. The third sample dataset includes a first abnormal arrangement style, a second abnormal arrangement style and a third abnormal arrangement style. The first abnormal arrangement style indicates that the first value and the second value are misaligned. The second abnormal arrangement style indicates that the first value and the first unit are misaligned. The third abnormal arrangement style indicates that the second value and the second unit are misaligned.

[0021] Step S15: Construct a sample dataset based on the positive sample set and the negative sample set. The negative sample set includes the first sample dataset, the second sample dataset, and the third sample dataset. The positive sample set is the standard style set of the annotation.

[0022] In one embodiment, step S15 includes:

[0023] Step 151: Perform preset processing on each standard style in the positive sample set and each abnormal style in the negative sample set to obtain the sample data after preset processing. The preset processing includes at least one of adding scanning noise, blurring, grayscale deviation processing, tilting processing, and image overlay background elements. Adding scanning noise is to randomly apply optical character recognition (OCR) noise to the standard style or abnormal style.

[0024] Step 152: Set labels for each of the pre-processed sample data to obtain the sample dataset.

[0025] In one embodiment, step S2 includes:

[0026] Step S21: Construct an initial model with a dual-branch structure. The initial model includes a detection branch and an inference branch. The detection branch is used to output the location information written in the car drawing, the detection results of each anomaly pattern, and the confidence level of the detection results. The inference branch is used to output the deviation degree of each anomaly pattern.

[0027] Step S22: The intelligent computing cloud platform utilizes computing power and adopts a distributed data parallel framework to input a preset number of sample data randomly selected from the training set into the initial model for feature extraction. After that, the detection branch outputs the predicted location information, the detection results of each anomaly pattern, and the confidence level of the detection results. The inference branch outputs the deviation degree of the predicted anomaly pattern. The detection results represent the qualitative detection results of each anomaly pattern, and the deviation degree represents the quantitative detection results of each anomaly pattern.

[0028] Step S23: If the quantitative detection result matches the qualitative detection result, calculate the total loss value of the initial model based on the first loss value, the second loss value, and the third loss value. The first loss value represents the error between the location information of the sample data output by the detection branch and the true location information of the sample data. The second loss value represents the error between the detection result of the sample data output by the detection branch and the true result corresponding to the label of the sample data. The third loss value represents the confidence error of the detection result of the sample data output by the detection branch.

[0029] Step S24: Based on the total loss value, backpropagation is performed using the chain rule to calculate the gradient of the model parameters corresponding to the initial model, and the model parameters are iterated N times according to the gradient and the learning rate to obtain the initial model after N iterations of training, where N is a preset value, the model parameters include convolution kernel weights, bias terms, and regression parameters, and the learning rate is reduced according to the cosine annealing algorithm in each iteration.

[0030] Step S25: If the total loss value of the initial model after the Mth iteration of training is less than the total loss value of the initial model after the Lth iteration of training in the validation set, the initial model after the Mth iteration of training is determined as the final model.

[0031] In one embodiment, step S3 includes:

[0032] Step S31: After converting the final model into an inference format adapted to the intelligent computing cloud platform, deploy it on the GPU inference node of the intelligent computing cloud platform, and encapsulate the final model into a callable API interface. The API interface supports receiving batches of automotive drawings and outputting the detection results of the consistency of the annotations of the automotive drawings.

[0033] In one embodiment, after step S3, the method further includes:

[0034] Step S4: Generate a review report based on the detection results. The review report includes the abnormal style type, the location of the abnormal style in the car drawing, and a screenshot of the abnormal style.

[0035] Step S5: Based on the review report, the test results are reviewed. If it is determined that there are abnormal results in the test results, the abnormal patterns in the car drawings corresponding to the abnormal results are used as new sample data.

[0036] Secondly, embodiments of the present invention provide an automatic consistency detection device for automotive drawing annotations using computing power on an intelligent computing cloud platform, the device comprising:

[0037] The construction module is used to construct a sample dataset based on a preset consistency semantic tree, and divide the sample dataset into a training set and a validation set. The consistency semantic tree is used to represent the standard style of annotations in automotive drawings. The annotations include a first value, a first unit corresponding to the first value, a plus or minus sign, a second value, and a second unit corresponding to the second value. The consistency semantic tree includes a semantic layer, a formal layer, and a validation layer. The semantic layer is used to define that the categories of the first unit and the second unit are consistent. The formal layer is used to define that the signs of the first unit and the second unit are consistent. The validation layer is used to define the logical relationship between the first value, the first unit, the second value, and the second unit.

[0038] The training module is used by the intelligent computing cloud platform to perform N iterations of training on the initial model based on the training set using computing power. The initial model after the Mth iteration of training is determined as the final model. The loss value of the initial model after the Mth iteration of training in the validation set is less than the loss value of the initial model after the Lth iteration of training in the validation set. N is a positive integer, M is a positive integer less than or equal to N, L is a positive integer less than or equal to N, and L is not equal to M. The Lth iteration of training is any one of the N iterations of training except for the Mth iteration of training.

[0039] The deployment module is used to deploy the final model on the intelligent computing cloud platform to perform batch annotation consistency checks on the received automotive drawings.

[0040] Thirdly, embodiments of the present invention provide an electronic device, including: a processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the method described in the first aspect.

[0041] Fourthly, embodiments of the present invention provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the method described in the first aspect.

[0042] Fifthly, embodiments of the present invention provide a computer program product including computer instructions that, when executed by a processor, implement the steps of the method described in the first aspect.

[0043] In this embodiment of the invention, a sample dataset is constructed based on a pre-defined consistent semantic tree, and the sample dataset is divided into a training set and a validation set. Data is synthesized through the consistent semantic tree, solving the problem of sample scarcity and laying the foundation for the model to learn complete features. The intelligent computing cloud platform utilizes computing power to accelerate the iterative process of the initial model through data parallelism, improving training efficiency. The final model is deployed on the intelligent computing cloud platform, and the platform dynamically schedules computing power to achieve parallel detection of batch drawings. Thus, based on the features trained in the early stages, the model can accurately locate annotation positions and identify abnormal annotations in automotive drawings, reducing misunderstandings during the manual review of automotive drawing annotation consistency and improving detection efficiency and accuracy. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of an automatic detection method for consistency of automotive drawing annotations using computing power provided by an embodiment of the present invention;

[0046] Figure 2 This is a structural diagram of an automatic detection device for consistency of automotive drawing annotations using computing power provided by an embodiment of the present invention;

[0047] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] The technical terms involved in this invention will be briefly explained below.

[0050] The "computing power" described in this invention is a new type of productivity that integrates information computing power, network carrying capacity, and data storage capacity, and mainly provides services to society through computing power infrastructure.

[0051] The "computational power" (CP) described in this invention refers to the ability of a data center server to process data and output results. It is a comprehensive indicator of a data center's computing power, encompassing general computing power, supercomputing power, and intelligent computing power. The commonly used unit of measurement is floating-point operations per second (FLOPS, 1 EFLOPS = 10^18 FLOPS), with higher values ​​indicating stronger overall computing power. It is estimated that 1 EFLOPS is approximately the computing power output of 5 Tianhe-2A supercomputers, 500,000 mainstream server CPUs, or 2 million mainstream laptops. The calculation formula is: CP = CP 通用 +CP 智能 +CP 超级 .

[0052] The "Network Power" (NP) mentioned in this invention refers to the performance of data transmission capability of computing facilities, which includes comprehensive capabilities such as network architecture, network bandwidth, transmission latency, intelligent management and scheduling, and involves network transmission within and between data centers. It is a comprehensive indicator for measuring network transmission scheduling capability.

[0053] The "Storage Power" (SP) described in this invention refers to the comprehensive capabilities of a data center in four aspects: data storage capacity, performance, security and reliability, and green and low-carbon operation. It is a comprehensive indicator for measuring the data storage capacity of a data center, including external storage devices such as storage arrays and internal storage devices within servers. The commonly used unit of measurement for storage capacity is exabytes (EB, 1EB = 2^60 bytes), the commonly used unit of measurement for performance is the number of read / write operations per second (IOPS / TB), and the disaster recovery ratio is an important indicator of security and reliability.

[0054] The "computing infrastructure" mentioned in this invention refers to a new type of information infrastructure that integrates information computing power, network carrying capacity, and data storage capacity, enabling centralized computing, storage, transmission, and application of information.

[0055] The "new information infrastructure" mentioned in this invention refers to network infrastructure such as 5G networks, fiber optic broadband networks, backbone networks, international communication networks, and satellite internet; computing infrastructure such as data centers, general computing centers, intelligent computing centers, and supercomputing centers; and new technology facilities such as artificial intelligence.

[0056] The “computing power” mentioned in this invention includes: general computing power, intelligent computing power, and supercomputing power.

[0057] The "general computing power" mentioned in this invention refers to the computing power provided by servers based on CPU (Central Processing Unit) chips, which is used to support basic general computing such as cloud computing and edge computing.

[0058] The "intelligent computing power" mentioned in this invention refers to: a computing platform deployed on a large scale based on dedicated chips such as GPU (Graphics Processing Unit), FPGA (Field Programmable Gate Array), and ASIC (Application Specific Integrated Circuit) for various artificial intelligence innovative applications, such as natural language processing and machine vision.

[0059] The “supercomputing power” mentioned in this invention refers to the computing power provided by high-performance computing clusters such as supercomputers. It utilizes the centralized computing resources of multiple computer systems working in parallel and uses a dedicated operating system to handle extremely complex or data-intensive problems. It is mainly used for computing in cutting-edge scientific fields, such as planetary simulation, drug molecule design, and gene analysis.

[0060] The "intelligent computing center" described in this invention includes, but is not limited to, a "smart computing center." An "intelligent computing center" is a facility that provides the necessary computing power, data, and algorithms for artificial intelligence applications (such as the development, training, and inference of deep learning models) by utilizing large-scale heterogeneous computing resources.

[0061] The "intelligent computing center" mentioned in this invention, also known as an artificial intelligence computing center, is a type of computing infrastructure that provides computing power services, data services, and algorithm services required for artificial intelligence applications, based on artificial intelligence theory and adopting an artificial intelligence computing architecture.

[0062] The “computing resources” mentioned in this invention refer to the technologies and facilities required for the development of the digital society that have the ability to compute, transmit, store and apply information, including but not limited to computing resources such as CPUs and GPUs, network resources such as switches and routers, storage resources such as storage arrays and distributed storage, security resources such as firewalls and intrusion detection systems, and supporting and guaranteeing resources such as wind, fire, water and electricity.

[0063] The "computing power service" mentioned in this invention refers to a service that provides users with on-demand computing power through computing resources, such as computing tasks, model inference, and data processing.

[0064] The "model" mentioned in this invention includes, but is not limited to, "large language model" and "multimodal large model".

[0065] The "large language model" (LLM) described in this invention refers to a language model with a large parameter scale, designed to understand and generate human language. It is trained on a large amount of text data and can perform a wide range of tasks, including text summarization, translation, and sentiment analysis.

[0066] The “Multimodal Large Models” mentioned in this invention refer to models that combine multimodal information such as text, images, videos, and audio for training, including but not limited to multimodal large language models.

[0067] The "automotive drawings" mentioned in this invention refer to engineering technical drawings used throughout the entire process of automotive R&D, production, and testing to convey the design intent, processing requirements, and assembly standards of automobiles and their components. Automotive drawings can include: assembly drawings of the complete vehicle and chassis, assembly drawings, and component drawings. Assembly drawings of the complete vehicle and chassis show the assembly relationships of components in the complete vehicle or chassis, and indicate the relative positions and connection methods of key components. Assembly drawings (such as engine assembly, transmission assembly, and body frame assembly drawings) indicate the assembly dimensions of each sub-component within the assembly. Component drawings (such as drawings of individual parts like bolts, bearings, gears, and body sheet metal parts) indicate the detailed dimensions, tolerances, and processing requirements of the parts. The consistency of annotations in automotive drawings is dense and directly affects processing accuracy; therefore, the automatic detection in this invention mainly targets the consistency detection of annotation units and error units in automotive drawings.

[0068] The "annotation" described in this invention is the core object for automatic detection of automotive drawings. Annotation includes a first value, the first unit corresponding to the first value, a plus or minus sign, a second value, and the second unit corresponding to the second value. The standards for automatic detection of the consistency of annotations in automotive drawings are: national standards GB / T14689 and GB / T4457.4, as well as industry specifications, etc.

[0069] With the deepening of digital transformation in the automotive industry, enterprises face the dual need to archive and review tens of thousands of historical drawings and verify massive amounts of new drawings. Typically, drawings are scanned to obtain image files (formats such as JPG, PNG, TIFF, etc.) for review and archiving. However, since historical drawings are not CAD source files, annotation information cannot be extracted from the geometric layer, rendering traditional CAD review tools unusable. Based on this, this invention provides a method and apparatus for automatically detecting the consistency of automotive drawing annotations using a smart computing cloud platform. This addresses the problem of how to automatically detect the consistency of automotive drawing annotations using computing power to improve detection efficiency and accuracy since the emergence of intelligent computing centers.

[0070] It is important to emphasize that this application is based on a distributed architecture built on an intelligent computing cloud platform that integrates computing power (e.g., a kilocalorie-level GPU cluster and a high-speed network) in an intelligent computing center scenario. Through the deep integration of computing power, algorithms, and data by the intelligent computing cloud platform, a "super intelligent detection system" specifically tailored for consistency detection of automotive drawing annotations has been constructed. Its core advantages are reflected in the breakthroughs in all dimensions of technical architecture, processing power, detection accuracy, and scenario adaptability.

[0071] The intelligent computing cloud platform, through the collaboration of a cluster of 1000-level NVIDIA A100 / H100 GPUs and a high-speed RDMA network, creates an EFLOPS-level (100 quadrillion calculations per second) computing power pool. The communication rate between computing nodes exceeds 400Gbps, the storage network bandwidth reaches 200Gbps, and the overall latency is compressed to sub-microsecond levels. This distributed architecture completely breaks the hardware constraints of traditional single-machine (CPU + entry-level GPU): traditional methods are limited by the memory and computing power of a single GPU, and can only process up to 4K resolution drawings, with a single drawing processing time exceeding 1 second (parallel processing capability of only 1-2 drawings per second); while the intelligent computing center can easily parse 16K ultra-high-definition drawings (accuracy down to 0.01px) and achieve batch processing of 200+ drawings per second through the parallel computing capabilities of GPU tensor cores, adapting to the mass production and batch inspection needs of the automotive industry.

[0072] The intelligent computing cloud platform utilizes a multimodal fusion model combining computer vision (CV), geometric topology analysis, and a physical rule engine, along with the feature association capabilities of graph neural networks (GNNs), to achieve holographic detection of complex assembly drawings.

[0073] For example, the computer vision module is responsible for accurately extracting pixel-level features for consistent annotation from images; the geometric topology analysis module constructs spatial relationship maps between parts (such as the topological relationships of assemblies with 2000+ feature points), avoiding related dimension errors caused by isolated detection; and the physical rule engine verifies whether the annotations meet manufacturing feasibility based on geometric tolerance standards (such as GB / T 1182) (such as whether the cumulative tolerance exceeds the assembly threshold). This collaborative mechanism improves detection accuracy, which is completely unattainable by single-machine methods relying solely on traditional OpenCV operators (such as Canny edge detection and Hough line detection).

[0074] Furthermore, the intelligent computing cloud platform possesses dynamic evolution capabilities, completely overcoming the static limitations of traditional methods: through an online learning system, it can automatically identify newly emerging annotation styles (such as designer handwriting or enterprise-specific simplified annotation rules) without requiring manual model retraining; by leveraging generative adversarial networks to generate adversarial examples for extreme scenarios such as uneven lighting and scanning noise, the model's robustness to complex environments is improved by more than 30%; and the real-time feedback mechanism directly connects the detection results with product lifecycle management and manufacturing execution systems. For example, when the cumulative tolerance of 872 positioning holes in a new energy vehicle battery pack exceeds 0.2mm, the system automatically generates a heatmap to mark the risk area and triggers a design change process, achieving a paradigm shift from passive verification after drawing completion to proactive verification during the design process.

[0075] Please see Figure 1 , Figure 1 This is a flowchart of an automatic consistency detection method for automotive drawing annotations using computing power provided by an embodiment of the present invention, such as... Figure 1 As shown, it includes the following steps:

[0076] Step S1: Construct a sample dataset based on a preset consistency semantic tree, and divide the sample dataset into a training set and a validation set. The consistency semantic tree is used to represent the standard style of annotations in automotive drawings. The annotations include a first value, a first unit corresponding to the first value, a plus or minus sign, a second value, and a second unit corresponding to the second value. The consistency semantic tree includes a semantic layer, a formal layer, and a validation layer. The semantic layer is used to define that the categories of the first unit and the second unit are consistent. The formal layer is used to define that the symbols of the first unit and the second unit are consistent. The validation layer is used to define the logical relationship between the first value, the first unit, the second value, and the second unit.

[0077] In this step, a consistent semantic tree can be established based on GB / T14689, GB / T4457.4, and industry standards to clarify the standard styles corresponding to annotations including the first numerical value, the first unit, the plus or minus sign, the second numerical value, and the combination of the second unit. For example, the standard styles corresponding to annotations can be represented as: 15℃±2℃, 50mm±2mm, etc.

[0078] The consistent semantic tree includes a semantic layer, a formal layer, and a validation layer.

[0079] The semantic layer is used to define that the first and second units are of the same category. For example, in the error annotation of the temperature category, the unit is always the temperature unit, and non-temperature units are mixed with other units (such as angle units, length units, pressure units, etc.).

[0080] The formal layer is used to define that the symbols for the first and second units are consistent. For example, in error annotations for temperature, the unit is always Celsius, rather than mixing Celsius and Fahrenheit.

[0081] The verification layer is used to define the logical relationship between the first value, the first unit, the second value, and the second unit. For example, 100kPa±1MPa, although both are pressure units, have different order of magnitude and the pressure error range (1MPa) is greater than the standard pressure (100kPa), which does not conform to the error logic of the annotation. Therefore, the annotation of 100kPa±1MPa does not conform to the logical relationship required by the verification layer and is a negative example. Similarly, the annotation of 100kPa±1000kPa also does not conform to the logical relationship required by the verification layer and is a negative example.

[0082] In this way, based on the definitions of the semantic layer, formal layer, and verification layer in the pre-defined consistent semantic tree, a quantitative foundation is provided for constructing the sample data in the sample dataset. The compliant annotations (i.e., positive examples, such as 25℃±1℃) and non-compliant annotations (i.e., negative examples, such as 15℃±2°, 50mm±2, etc.) generated based on the consistent semantic tree are integrated into a sample dataset. This dataset is then proportionally divided into a training set for initial model learning and a validation set for evaluating the model's generalization ability during iterations. Stratified sampling ensures that the proportions of positive and negative examples and the distribution of error types are consistent between the two sets of samples.

[0083] Furthermore, the intelligent computing cloud platform can call the graphics engine in parallel across multiple nodes to synchronously draw sample data that closely resembles the style of real drawings. This systematically generates 100,000 samples covering all types of errors, avoiding the scarcity of negative examples and improving the diversity of samples during subsequent model training. This allows the model to learn the characteristics of compliant and non-compliant annotations defined by the consistent semantic tree, thereby improving the comprehensiveness and accuracy of model detection.

[0084] Step S2: The intelligent computing cloud platform uses computing power to perform N iterations of training on the initial model based on the training set. The initial model after the Mth iteration is determined as the final model. The loss value of the initial model after the Mth iteration in the validation set is less than the loss value of the initial model after the Lth iteration in the validation set. N is a positive integer, M is a positive integer less than or equal to N, L is a positive integer less than or equal to N, and L is not equal to M. The Lth iteration is any one of the N iterations except the Mth iteration.

[0085] In this step, a dual-branch initial model can be selected, comprising a detection branch and an inference branch. The detection branch can be based on a YOLOv8n or PP-YOLOE-tiny architecture to detect text regions and error types. The inference branch can be based on a lightweight Transformer (TinyViT+TextFusion) architecture to fuse visual and text features to evaluate the reliability of the detection results. Furthermore, data parallelism is achieved through a smart computing cloud platform using computing power (e.g., multi-GPU clusters, using a distributed training framework such as PyTorch Distributed) to avoid insufficient computing power on a single node. After every K iterations (e.g., K = 5), the performance of the initial model after the Kth iteration is evaluated using a validation set, until a preset N iterations (e.g., N = 100) are completed, at which point iterations cease. Then, the loss values ​​of the N initial models trained after N iterations are compared on the validation set. The initial model trained after the Mth iteration is determined as the final model. The loss value of the initial model trained after the Mth iteration on the validation set is less than the loss value of the initial model trained after the Lth iteration on the validation set. In other words, the initial model trained after the Mth iteration has the smallest loss value, and its training effect is optimal. In this way, the initial model trained after the Mth iteration fully learns all types of error features (such as inconsistencies in the categories of the first and second units defined in the semantic layer, inconsistencies in the signs of the first and second units defined in the formal layer, and errors in the logical relationships between the first numerical value, the first unit, the second numerical value, and the second unit defined in the validation layer). Ultimately, it has the minimum loss value on the validation set, improving the accuracy of detecting consistency in automotive drawing annotations and laying an efficient and accurate model foundation for subsequent batch detection. See the following steps for details.

[0086] In one embodiment, step S2 includes:

[0087] Step S21: Construct an initial model with a dual-branch structure. The initial model includes a detection branch and an inference branch. The detection branch is used to output the location information written in the car drawing, the detection results of each anomaly pattern, and the confidence level of the detection results. The inference branch is used to output the deviation degree of each anomaly pattern.

[0088] Step S22: The intelligent computing cloud platform utilizes computing power and adopts a distributed data parallel framework to input a preset number of sample data randomly selected from the training set into the initial model for feature extraction. After that, the detection branch outputs the predicted location information, the detection results of each anomaly pattern, and the confidence level of the detection results. The inference branch outputs the deviation degree of the predicted anomaly pattern. The detection results represent the qualitative detection results of each anomaly pattern, and the deviation degree represents the quantitative detection results of each anomaly pattern.

[0089] Step S23: If the quantitative detection result matches the qualitative detection result, calculate the total loss value of the initial model based on the first loss value, the second loss value, and the third loss value. The first loss value represents the error between the location information of the sample data output by the detection branch and the true location information of the sample data. The second loss value represents the error between the detection result of the sample data output by the detection branch and the true result corresponding to the label of the sample data. The third loss value represents the confidence error of the detection result of the sample data output by the detection branch.

[0090] Step S24: Based on the total loss value, backpropagation is performed using the chain rule to calculate the gradient of the model parameters corresponding to the initial model, and the model parameters are iterated N times according to the gradient and the learning rate to obtain the initial model after N iterations of training, where N is a preset value, the model parameters include convolution kernel weights, bias terms, and regression parameters, and the learning rate is reduced according to the cosine annealing algorithm in each iteration.

[0091] Step S25: If the total loss value of the initial model after the Mth iteration of training is less than the total loss value of the initial model after the Lth iteration of training in the validation set, the initial model after the Mth iteration of training is determined as the final model.

[0092] In this embodiment, the detection branch can be based on a YOLOv8n or PP-YOLOE-tiny architecture. On one hand, the detection branch locates the specific position of the annotation on the automotive drawing, outputting the annotation's coordinate information, such as the pixel coordinates of 15℃±2℃ in the drawing. It also identifies abnormal annotation styles, such as inconsistent units or missing symbols, directly determining whether the annotation conforms to the rules of the consistency semantic tree. Furthermore, it calculates the confidence level of the output of the detection branch; for example, the probability of inconsistent units is 0.98, used to evaluate the reliability of the detection results. A higher confidence level indicates greater confidence in the model's judgment, providing a quantitative basis for subsequent loss calculation and result selection. On the other hand, the inference branch outputs the deviation degree of each abnormal style to further verify the detection results. Constructing a dual-branch structure for the initial model allows for adaptation to cloud platform training and subsequent deployment requirements, and simultaneously achieves annotation location positioning and anomaly identification while evaluating the reliability of the judgment, improving the accuracy and reliability of the model's detection results.

[0093] After constructing the initial model with a dual-branch structure, the intelligent computing cloud platform utilizes a kilo-calorie GPU cluster and high-speed network resources to batch process training set samples through a distributed data parallel framework. For example, a predetermined number of samples can be randomly selected from the training set (e.g., 128 drawings with annotations each time). After inputting these samples into the initial model, the backbone network (e.g., ResNet-50) extracts key features of the annotations. These key features include, for example, the outlines of the symbols “℃” and “±”, and the font differences between “mm” and “m”. After processing by the detection branch, the predicted location information of the annotations and the detection results of abnormal styles are output. Based on the extracted features, the prediction confidence of the above detection results is output. Simultaneously, the deviation degree of each predicted abnormal style is output through the inference branch to quantitatively analyze key features such as the outlines of the symbols “℃” and “±”, and the font differences between “mm” and “m”, thereby verifying the reliability of the detection results. With the support of distributed computing power, the efficiency of feature extraction and branch inference for single batches of samples is improved, solving the computing power bottleneck of traditional single-machine processing of large-scale samples.

[0094] To comprehensively measure the difference between model predictions and true labels, and to ensure that optimization covers the three dimensions of localization, recognition, and confidence, the total loss of the initial model can be calculated based on the first, second, and third loss values, provided that the quantitative and qualitative detection results match. The first loss value can be calculated using CIoU Loss to measure the deviation between the predicted location of the annotation output by the detection branch and the actual location of the annotation in the sample data, such as coordinate offset and scale difference, ensuring that the model can accurately locate the annotation content in the drawing. The formula for calculating the first loss value is as follows:

[0095]

[0096] L box The first loss value, i The sample data number, b i For the first i The actual location annotated in each sample data point To detect the first branch output i The predicted locations are annotated in the sample data.

[0097] The second loss can be calculated using cross-entropy loss to measure the difference between the anomaly detection results output by the detection branch and the true label of the sample. Examples of anomalies include: inconsistencies in the categories of the first and second units; inconsistencies in the signs of the first and second units; and errors in the logical relationships between the first value, the first unit, the second value, and the second unit. The second loss can be expressed as... L cls .

[0098] The third loss value can be calculated based on the semantic matching term of cosine similarity to measure the deviation between the confidence of the detection branch output and the true label. For example, the model should output high confidence for correctly identified anomaly annotations and low confidence for incorrectly identified ones, avoiding situations where there is a false positive but high confidence. The formula for calculating the third loss value can be found below:

[0099]

[0100] L sem This is the third loss value. f u To test the confidence level of the branch output, f v This represents the confidence level corresponding to the true label.

[0101] The total loss value of the initial model is calculated based on the first loss value, the second loss value, and the third loss value, as shown in the following formula:

[0102]

[0103] L is the total loss value, λ1 is the weight of the first loss value, λ2 is the weight of the second loss value, and λ3 is the weight of the third loss value.

[0104] In this way, the total loss is obtained by weighted summation of the three factors, quantifying the model's prediction accuracy. Then, parameter optimization is initiated based on the total loss: the gradients of all model parameters, such as convolutional kernel weights, bias terms, and regression parameters of the detection branch, are calculated through backpropagation using the chain rule. This is combined with a dynamically adjusted learning rate using cosine annealing, which decays with each iteration to avoid parameter oscillations. This iterative update of the model parameters gradually brings the model's predictions closer to the true labels. The model updated in each iteration can be applied to the validation set for inference, and the validation set loss value is calculated. If the loss value of the initial model after the Mth iteration is less than that of the initial model after the Lth iteration, the initial model after the Mth iteration is the optimal model among the N iterations. This initial model can then be selected as the final model, completing the training and obtaining a model usable for actual detection. Ultimately, this achieves iterative evolution from the initial model to a high-precision final model.

[0105] Step S3: Deploy the final model on the intelligent computing cloud platform to perform batch annotation consistency checks on the received automotive drawings.

[0106] In this step, the final model is deployed on an intelligent computing cloud platform, enabling image-level compliance inspection without CAD dependency. This allows for batch annotation consistency inspection of received automotive drawings (e.g., scanned images or PDF files), determining compliance with drafting standards, and automatically identifying annotation irregularities. It boasts high accuracy, high versatility, and high robustness, showing broad application prospects in industrial design review, engineering archiving verification, and automated drafting quality control, completely eliminating reliance on vector layer data. Simultaneously, it avoids discrepancies in understanding dimensional annotations on automotive drawings among different reviewers, improving inspection efficiency and accuracy.

[0107] Specifically, in one embodiment, step S3 includes:

[0108] Step S31: After converting the final model into an inference format adapted to the intelligent computing cloud platform, deploy it on the GPU inference node of the intelligent computing cloud platform, and encapsulate the final model into a callable API interface. The API interface supports receiving batches of automotive drawings and outputting the detection results of the consistency of the annotations of the automotive drawings.

[0109] In this embodiment, to achieve efficient inference environment adaptation of the final model to the cloud platform, model format conversion and service encapsulation can be performed. Model format conversion converts the model parameters saved during the training phase into a cloud platform-friendly inference format (such as ONNX or TensorRT format), reduces computational load through INT8 quantization compression, and ensures that accuracy loss is controlled within 1% to improve inference speed. Service encapsulation integrates the converted model into the cloud platform's inference framework, encapsulating it as an API interface that can be called via HTTP or HTTPS, and explicitly defining the input of the interface as batch automotive drawing files and the output as structured detection results.

[0110] Furthermore, the intelligent computing cloud platform can dynamically allocate GPU inference nodes based on the batch size of received automotive drawings (e.g., 10,000 drawings) using cluster management tools (such as Kubernetes). For example, one NVIDIA T4 / V100 node can be allocated for every 500 drawings, and each node can deploy multiple model instances. These model instances can be parallel replicas based on the same final model, isolated by Docker containers to achieve high-concurrency processing. The API interface can receive batches of automotive drawings uploaded by users, supporting formats including scanned images (JPG / PNG / TIFF, etc.) and PDF image pages. It automatically extracts image layers from the automotive drawings and then uniformly converts them to grayscale images to reduce the number of channels, lower computational load, and standardize the resolution (e.g., 300 dpi to ensure clear annotation details).

[0111] For example, given the large size of automotive drawings (e.g., A3 drawings with 4961×7016 pixels) and dense annotations, the drawings can be cropped into 640×640 pixel blocks. Simultaneously, the coordinate offset of each cropped image within the original image is obtained, and minor enhancements (such as contrast adjustment and noise reduction filtering) are applied to eliminate the effects of scanning noise or uneven lighting. Then, the images are normalized using the mean and variance from model training. Finally, the processed cropped images are batch-input into model instances on each GPU node to perform parallel annotation consistency detection of the automotive drawings, improving detection efficiency.

[0112] In one embodiment, after step S3, the method further includes:

[0113] Step S4: Generate a review report based on the detection results. The review report includes the abnormal style type, the location of the abnormal style in the car drawing, and a screenshot of the abnormal style.

[0114] Step S5: Based on the review report, the test results are reviewed. If it is determined that there are abnormal results in the test results, the abnormal patterns in the car drawings corresponding to the abnormal results are used as new sample data.

[0115] In this embodiment, a review report containing complete information is generated based on the batch detection results, transforming the raw detection data output by the model into a decision-making basis that reviewers can directly use. The review report clearly specifies the coordinate positions of the anomalies in the automotive drawings (e.g., coordinate positions (300, 500)(400, 800)) and the corresponding screenshots (e.g., img / 12.jpg) to visually mark the problem areas, facilitating quick location by reviewers. Furthermore, based on the verification results of the semantic consistency tree, the error type is labeled (e.g., "mismatch," indicating inconsistent unit symbols or categories), and the representation in the original drawings is attached (e.g., 15mm ± 2°). Additionally, the review report may include the confidence level of the detection results (e.g., 0.97) to characterize the reliability of the detection results with the model. The review report can be shown in Table 1 below:

[0116] Table 1

[0117] Serial Number Coordinates Screenshot location Error Type Original text Confidence 12 (300,500)(400,800) img / 12.jpg mismatch 15mm±2° 0.97

[0118] Then, the test results are reviewed based on the review report. If abnormal results are found in the test results, the abnormal patterns in the car drawings corresponding to the abnormal results are used as new sample data to continuously improve the model performance.

[0119] The specific process is as follows:

[0120] First, professional reviewers verify each annotation anomaly result based on the generated review report, focusing on confirming whether the model has misjudged, such as misclassifying 25℃±2℃ as a symbol inconsistency, and whether there are missed detections, such as 100kPa±1MPa not being identified as a semantic contradiction. If the review finds model detection errors or omissions in annotations, the drawing fragment containing the anomaly annotation is cropped as a new sample, its true error type is labeled, and it is added to the original sample set.

[0121] After cleaning, the newly added samples can be used for subsequent model retraining to specifically optimize the model's ability to identify edge cases. This method of supplementing samples through human feedback compensates for the model's limitations in complex scenarios, enabling the cloud platform to continuously evolve with real-world applications and ultimately achieve long-term improvements in detection accuracy.

[0122] In this embodiment of the invention, a sample dataset is constructed based on a pre-defined consistent semantic tree, and the sample dataset is divided into a training set and a validation set. Data is synthesized through the consistent semantic tree, solving the problem of sample scarcity and laying the foundation for the model to learn complete features. The intelligent computing cloud platform utilizes computing power to accelerate the iterative process of the initial model through data parallelism, improving training efficiency. The final model is deployed on the intelligent computing cloud platform, and the platform dynamically schedules computing power to achieve parallel detection of batch drawings. Thus, based on the features trained in the early stages, the model can accurately locate annotation positions and identify abnormal annotations in automotive drawings, reducing misunderstandings during the manual review of automotive drawing annotation consistency and improving detection efficiency and accuracy.

[0123] In one embodiment, step S1 includes:

[0124] Step S11: Based on the consistent semantic tree, construct a standard style set for annotation writing;

[0125] Step S12: Perform random unit replacement processing on the standard style of the annotation to obtain a first sample dataset. The first sample dataset includes abnormal unit category styles and abnormal unit symbol styles. The abnormal category styles indicate that the categories of the first unit and the second unit are inconsistent, and the abnormal unit symbol styles indicate that the symbols of the first unit and the second unit are inconsistent.

[0126] Step S13: Randomly delete characters from the standard style of the annotation to obtain a second sample dataset. The second sample dataset includes a unit missing style, a value missing style, and a plus or minus sign missing style. The unit missing style indicates that the first unit and / or the second unit is missing. The value missing style indicates that the first value and / or the second value is missing. The plus or minus sign missing style indicates that the plus or minus sign is missing.

[0127] Step S14: Perform random character mixing on the standard style of the annotation to obtain a third sample dataset. The third sample dataset includes a first abnormal arrangement style, a second abnormal arrangement style and a third abnormal arrangement style. The first abnormal arrangement style indicates that the first value and the second value are misaligned. The second abnormal arrangement style indicates that the first value and the first unit are misaligned. The third abnormal arrangement style indicates that the second value and the second unit are misaligned.

[0128] Step S15: Construct a sample dataset based on the positive sample set and the negative sample set. The negative sample set includes the first sample dataset, the second sample dataset, and the third sample dataset. The positive sample set is the standard style set of the annotation.

[0129] In this embodiment, a standard style set for annotations, i.e., a set of positive examples, is constructed based on a consistent semantic tree. The semantic layer defines that the categories of the first unit and the second unit are consistent, the formal layer defines that the symbols of the first unit and the second unit are consistent, and the validation layer defines the logical relationships between the first value, the first unit, the second value, and the second unit. The standard style set for annotations covers common annotation types in automotive drawings, such as dimensions, temperature, pressure, and angles, exemplified by "25℃±1℃", "80mm±3mm", and "120kPa±5kPa", providing a compliance reference for the subsequent generation of abnormal samples.

[0130] After constructing the standard style set for annotations, the first sample dataset can be obtained by randomly replacing units in the standard styles. This breaks the consistency of the semantic or formal layers, generating unit category anomaly styles and unit symbol anomaly styles as negative examples. The unit category anomaly style indicates that the first and second unit categories are different; for example, "25℃±1℃" can be replaced with "25℃±1°". The unit symbol anomaly style indicates that the first and second unit symbols are different; for example, "80mm±3mm" can be replaced with "80mm±3cm". Although both belong to the length category, the symbols "mm" and "cm" are different. After setting corresponding labels for each unit category anomaly style and unit symbol anomaly style, the first sample dataset is formed. It should be understood that the first sample dataset can also include anomaly styles formed by combinations of unit category anomaly styles and unit symbol anomaly styles as negative examples, which will not be elaborated further here.

[0131] After constructing the standard style set for annotations, a second sample dataset can be obtained by randomly deleting characters from the standard styles. This disrupts the integrity of the annotations and generates three types of anomalous styles as negative examples: styles with missing units, styles with missing values, and styles with missing plus / minus signs. Specifically, a missing unit style indicates that the first and / or second unit is missing; for example, "80mm±3mm" can be changed to "80±3mm", "80mm±3", or "80±3". A missing value style indicates that the first and / or second value is missing; for example, "120kPa±5kPa" can be changed to "120kPa±kPa" or "±5kPa". A missing plus / minus sign style indicates that the "±" sign is missing; for example, "25℃±1℃" can be changed to "25℃1℃". After assigning corresponding labels to each missing unit style, missing value style, and missing plus / minus sign style, the second sample dataset is formed. It should be understood that the second sample dataset may also include anomalous patterns formed by at least two combinations of missing unit patterns, missing numerical patterns, and missing sign patterns, as counterexamples, which will not be elaborated here.

[0132] After constructing the standard style set for annotations, a third sample dataset can be obtained by randomly mixing characters in the standard styles. This disrupts the logical relationship of the validation layer, generating three types of abnormal styles—a first-order abnormal style, a second-order abnormal style, and a third-order abnormal style—as counterexamples. The first-order abnormal style indicates a misalignment between the first and second numerical values; for example, "80mm±3mm" can be mixed into "3mm±80mm," creating a logical contradiction in the magnitude of the values. The second-order abnormal style indicates a misalignment between the first numerical value and the first unit; for example, "25℃±1℃" can be mixed into "℃25±1℃," disrupting the correspondence between the first numerical value and the unit. The third-order abnormal style indicates a misalignment between the second numerical value and the second unit; for example, "120kPa±5kPa" can be mixed into "120kPa±kPa5," disrupting the correspondence between the second numerical value and the unit. After assigning corresponding labels to the first, second, and third-order abnormal styles, the third sample dataset is formed. It should be understood that the third sample dataset may also include anomalies formed by at least two combinations of the first, second, and third permutation anomalies, as counterexamples, which will not be elaborated here.

[0133] After generating the first, second, and third sample datasets, positive and negative samples can be mixed in a preset ratio (e.g., 3:1) to ensure a balanced distribution of categories across the datasets. The positive sample set is used by the model to learn the features of compliant annotations; the negative sample set, which combines the first, second, and third sample datasets and covers all anomaly types (unit, missing, and permutation), is used by the model to learn the features of abnormal annotations. The final constructed sample dataset includes both compliant references and covers all scenarios of anomalies in automotive drawing annotation consistency detection. Each sample has a clear label, providing comprehensive and accurate supervised learning data for subsequent model training and improving the model's ability to identify various anomalies.

[0134] In one embodiment, step S15 includes:

[0135] Step 151: Perform preset processing on each standard style in the positive sample set and each abnormal style in the negative sample set to obtain the sample data after preset processing. The preset processing includes at least one of adding scanning noise, blurring, grayscale deviation processing, tilting processing, and image overlay background elements. Adding scanning noise is to randomly apply optical character recognition (OCR) noise to the standard style or abnormal style.

[0136] Step 152: Set labels for each of the pre-processed sample data to obtain the sample dataset.

[0137] In this embodiment, interference factors from real car drawings are simulated and clearly labeled to make the sample dataset more closely resemble actual detection scenarios, providing the model with accurate and robust training data. Specifically, various preset processing methods are used to simulate non-ideal states that may occur in real car drawings during scanning and archiving, targeting standard patterns in the positive sample set and abnormal patterns in the negative sample set. This ensures that the model can adapt to complex environmental interference during training. These preset processing methods include at least one of the following: adding scanning noise, blurring, grayscale deviation processing, tilting, and overlaying background elements onto the image. Through these processing methods, the sample data automatically synthesized based on standard-driven methods is transformed from ideally pure labeled data into scenario-based samples containing real interference. The final sample dataset can more comprehensively cover the interference that may be encountered in actual detection. This ensures that the model not only learns the compliance rules of the labels during training but also adapts to interference factors in real scenarios, thereby improving the robustness of subsequent detection and avoiding detection failures caused by the disconnect between samples and actual scenarios.

[0138] For example, random OCR noise is applied to simulate the graininess caused by paper texture and insufficient scanner precision when scanning paper drawings, or the blurring of ink due to paper aging. For example, "°" is adjusted to "o", "℃" is adjusted to "°C", and noise appears on the edge of the "℃" symbol, and local noise is masked on the "±" symbol, so that the training model has robustness.

[0139] Blur processing can be achieved through Gaussian blur or motion blur to simulate the blurring effect of inaccurate focus during scanning or creases in drawings, such as blurring the edges of the "mm" character or making the value "50" difficult to distinguish due to blur.

[0140] Gray-scale deviation processing can adjust the gray-scale value distribution of a sample (locally brighter or darker) to simulate the differences in brightness caused by uneven illumination during scanning (such as local attenuation of the scanner's light source). For example, the contrast of a "kPa" unit area is reduced due to being darker.

[0141] Tilt processing can be achieved by randomly rotating the sample image by ±15° to simulate the skewed state when manually placing the paper during scanning, such as tilting "120kPa±5kPa" by 10° and arranging the characters non-horizontally.

[0142] Image overlay background elements can be simple part outlines, grid lines, or irrelevant text (such as "bolt number: B01") overlaid around annotations to simulate the complex background where annotations coexist with part graphics in real automotive drawings, avoiding the model learning only against a pure background.

[0143] Furthermore, each sample data point, after pre-processing, is assigned a precise label as a criterion for model training. If the sample originates from the positive example set, its label can indicate compliance with annotations and associate it with its corresponding annotation type; if the sample originates from the negative example set, its label must clearly indicate the specific anomaly type, such as annotating the unit category anomaly. Ultimately, a high-quality sample dataset is constructed that covers all anomaly types, adapts to real-world interference, and has clearly defined labels, laying a data foundation for the accuracy and robustness of model training.

[0144] Please see Figure 2 , Figure 2 This is a structural diagram of an automatic detection device for consistency of automotive drawing annotations using computing power, provided by an embodiment of the present invention. Figure 2 As shown, the intelligent computing cloud platform uses computing power to automatically detect the consistency of automotive drawing annotations. The device 200 includes:

[0145] The construction module 201 is used to construct a sample dataset based on a preset consistency semantic tree, and divide the sample dataset into a training set and a validation set. The consistency semantic tree is used to represent the standard style of annotations in automotive drawings. The annotations include a first value, a first unit corresponding to the first value, a plus or minus sign, a second value, and a second unit corresponding to the second value. The consistency semantic tree includes a semantic layer, a formal layer, and a validation layer. The semantic layer is used to define that the categories of the first unit and the second unit are consistent. The formal layer is used to define that the signs of the first unit and the second unit are consistent. The validation layer is used to define the logical relationship between the first value, the first unit, the second value, and the second unit.

[0146] Training module 202 is used by the intelligent computing cloud platform to perform N iterations of training on the initial model based on the training set using computing power, and to determine the initial model after the Mth iteration of training as the final model. The loss value of the initial model after the Mth iteration of training in the validation set is less than the loss value of the initial model after the Lth iteration of training in the validation set. N is a positive integer, M is a positive integer less than or equal to N, L is a positive integer less than or equal to N, and L is not equal to M. The Lth iteration of training is any one of the N iterations of training except for the Mth iteration of training.

[0147] The deployment module 203 is used to deploy the final model on the intelligent computing cloud platform to perform batch annotation consistency checks on the received automotive drawings.

[0148] In one embodiment, the construction module 201 is specifically used for:

[0149] Based on the consistent semantic tree, a standard style set for annotation is constructed;

[0150] The standard style of the annotation is subjected to random unit replacement processing to obtain a first sample dataset. The first sample dataset includes unit category abnormal styles and unit symbol abnormal styles. The category abnormal styles indicate that the categories of the first unit and the second unit are inconsistent, and the unit symbol abnormal styles indicate that the symbols of the first unit and the second unit are inconsistent.

[0151] The standard style of the annotation is subjected to random character deletion to obtain a second sample dataset. The second sample dataset includes a unit missing style, a value missing style, and a plus or minus sign missing style. The unit missing style indicates that the first unit and / or the second unit is missing. The value missing style indicates that the first value and / or the second value is missing. The plus or minus sign missing style indicates that the plus or minus sign is missing.

[0152] The standard style of the annotation is subjected to random character mixing to obtain a third sample dataset. The third sample dataset includes a first abnormal arrangement style, a second abnormal arrangement style and a third abnormal arrangement style. The first abnormal arrangement style indicates that the first value and the second value are misaligned. The second abnormal arrangement style indicates that the first value and the first unit are misaligned. The third abnormal arrangement style indicates that the second value and the second unit are misaligned.

[0153] A sample dataset is constructed based on the positive example sample set and the negative example sample set. The negative example sample set includes the first sample dataset, the second sample dataset, and the third sample dataset. The positive example sample set is the standard style set of the annotation.

[0154] In one embodiment, constructing the sample dataset based on the positive example set and the negative example set includes:

[0155] Pre-processing is performed on each standard style in the positive sample set and each abnormal style in the negative sample set to obtain sample data after pre-processing. The pre-processing includes at least one of adding scanning noise, blurring, grayscale deviation processing, tilting, and image overlaying background elements. Adding scanning noise is to randomly apply optical character recognition (OCR) noise to the standard style or abnormal style.

[0156] Label each of the pre-processed sample data to obtain the sample dataset.

[0157] In one embodiment, the training module 202 is specifically used for:

[0158] An initial model with a dual-branch structure is constructed. The initial model includes a detection branch and an inference branch. The detection branch is used to output the location information written in the car drawing, the detection results of each anomaly pattern, and the confidence level of the detection results. The inference branch is used to output the deviation degree of each anomaly pattern.

[0159] The intelligent computing cloud platform utilizes computing power and employs a distributed data parallel framework to input a preset number of sample data randomly selected from the training set into the initial model for feature extraction. After feature extraction, the detection branch outputs the predicted location information, the detection results of each anomaly pattern, and the confidence level of the detection results. The inference branch outputs the deviation degree of the predicted anomaly pattern. The detection results represent the qualitative detection results of each anomaly pattern, and the deviation degree represents the quantitative detection results of each anomaly pattern.

[0160] When the results of the quantitative detection match the results of the qualitative detection, the total loss value of the initial model is calculated based on the first loss value, the second loss value, and the third loss value. The first loss value represents the error between the location information of the sample data output by the detection branch and the true location information of the sample data. The second loss value represents the error between the detection result of the sample data output by the detection branch and the true result corresponding to the label of the sample data. The third loss value represents the confidence error of the detection result of the sample data output by the detection branch.

[0161] Based on the total loss value, the gradient of the model parameters corresponding to the initial model is calculated by backpropagation using the chain rule, and the model parameters are iterated N times according to the gradient and the learning rate to obtain the initial model after N iterations of training, where N is a preset value. The model parameters include convolution kernel weights, bias terms, and regression parameters, and the learning rate is reduced according to the cosine annealing algorithm in each iteration.

[0162] If the total loss value of the initial model after the Mth iteration of training is less than the total loss value of the initial model after the Lth iteration of training in the same validation set, the initial model after the Mth iteration of training is determined as the final model.

[0163] In one embodiment, deployment module 203 is specifically used for:

[0164] After converting the final model into an inference format adapted to the intelligent computing cloud platform, it is deployed on the GPU inference node of the intelligent computing cloud platform, and the final model is encapsulated into a callable API interface. The API interface supports receiving batches of automotive drawings and outputting the detection results of the consistency of the annotations of the automotive drawings.

[0165] In one embodiment, the apparatus further includes:

[0166] A generation module is used to generate a review report based on the detection results. The review report includes the abnormal style type, the location of the abnormal style in the car drawing, and a screenshot of the abnormal style.

[0167] The review module is used to review the test results based on the review report. If it is determined that there are abnormal results in the test results, the abnormal patterns in the car drawings corresponding to the abnormal results are used as new sample data.

[0168] The intelligent computing cloud platform provided in this embodiment of the invention provides an automatic detection device for consistency of automotive drawing annotations using computing power. This device can realize the various processes of the above-mentioned intelligent computing cloud platform automatic detection method for consistency of automotive drawing annotations using computing power. The technical features are one-to-one and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0169] It should be noted that the intelligent computing cloud platform in this embodiment of the invention can be a device, or it can be a component, integrated circuit, or chip in an electronic device.

[0170] This invention also provides an electronic device, see [link to relevant documentation]. Figure 3 , Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device includes a memory 301, a processor 302, and a program or instructions stored in the memory 301 that run on the memory. When the program or instructions are executed by the processor 302, they can achieve the following: Figure 1 The corresponding intelligent computing cloud platform uses computing power to perform any step in the embodiment of the automatic detection method for consistency of automotive drawing annotations and achieves the same beneficial effect, which will not be elaborated here.

[0171] The processor 302 can be a CPU, ASIC, FPGA or GPU.

[0172] Those skilled in the art will understand that all or part of the steps of the above-described intelligent computing cloud platform for automatically detecting the consistency of automotive drawing annotations using computing power can be implemented by hardware related to program instructions, and the program can be stored in a readable medium.

[0173] This invention also provides a readable storage medium storing a computer program, which, when executed by a processor, can perform the above-described functions. Figure 1The corresponding intelligent computing cloud platform can perform any step in the embodiment of the automatic detection method for consistency of automotive drawing annotations using computing power, and can achieve the same technical effect. To avoid repetition, it will not be described again here. The storage medium mentioned is such as read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, etc.

[0174] The present invention also provides a computer program product, including computer instructions that, when executed by a processor, implement the above-described... Figure 1 The corresponding intelligent computing cloud platform performs each process of the automatic detection method for consistency of automotive drawing annotations through computing power, and can achieve the same technical effect. To avoid repetition, it will not be described in detail here.

[0175] In the embodiments of this invention, the terms "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices. Additionally, the use of "and / or" in this invention indicates at least one of the connected objects, such as A and / or B and / or C, representing seven possibilities: A alone, B alone, C alone, both A and B present, both B and C present, both A and C present, and A, B, and C present.

[0176] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0177] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or second terminal device, etc.) to execute the methods of the various embodiments of the present invention.

[0178] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these modifications are within the protection scope of the present invention.

Claims

1. A method for automatically detecting the consistency of automotive drawing annotations using a smart computing cloud platform, characterized in that: The method includes: Step S1: Construct a sample dataset based on a preset consistency semantic tree, and divide the sample dataset into a training set and a validation set. The consistency semantic tree is used to represent the standard style of annotations in automotive drawings. The annotations include a first value, a first unit corresponding to the first value, a plus or minus sign, a second value, and a second unit corresponding to the second value. The consistency semantic tree includes a semantic layer, a formal layer, and a validation layer. The semantic layer is used to define that the categories of the first unit and the second unit are consistent. The formal layer is used to define that the symbols of the first unit and the second unit are consistent. The validation layer is used to define the logical relationship between the first value, the first unit, the second value, and the second unit. Step S2: The intelligent computing cloud platform uses computing power to perform N iterations of training on the initial model based on the training set. The initial model after the Mth iteration is determined as the final model. The loss value of the initial model after the Mth iteration in the validation set is less than the loss value of the initial model after the Lth iteration in the validation set. N is a positive integer, M is a positive integer less than or equal to N, L is a positive integer less than or equal to N, and L is not equal to M. The Lth iteration is any one of the N iterations except the Mth iteration. Step S3: Deploy the final model on the intelligent computing cloud platform to perform batch annotation consistency checks on the received automotive drawings.

2. The method as described in claim 1, characterized in that, Step S1 includes: Step S11: Based on the consistent semantic tree, construct a standard style set for annotation writing; Step S12: Perform random unit replacement processing on the standard style of the annotation to obtain a first sample dataset. The first sample dataset includes abnormal unit category styles and abnormal unit symbol styles. The abnormal category styles indicate that the categories of the first unit and the second unit are inconsistent, and the abnormal unit symbol styles indicate that the symbols of the first unit and the second unit are inconsistent. Step S13: Randomly delete characters from the standard style of the annotation to obtain a second sample dataset. The second sample dataset includes a unit missing style, a value missing style, and a plus or minus sign missing style. The unit missing style indicates that the first unit and / or the second unit is missing. The value missing style indicates that the first value and / or the second value is missing. The plus or minus sign missing style indicates that the plus or minus sign is missing. Step S14: Perform random character mixing on the standard style of the annotation to obtain a third sample dataset. The third sample dataset includes a first abnormal arrangement style, a second abnormal arrangement style and a third abnormal arrangement style. The first abnormal arrangement style indicates that the first value and the second value are misaligned. The second abnormal arrangement style indicates that the first value and the first unit are misaligned. The third abnormal arrangement style indicates that the second value and the second unit are misaligned. Step S15: Construct a sample dataset based on the positive sample set and the negative sample set. The negative sample set includes the first sample dataset, the second sample dataset, and the third sample dataset. The positive sample set is the standard style set of the annotation.

3. The method as described in claim 2, characterized in that, Step S15 includes: Step 151: Perform preset processing on each standard style in the positive sample set and each abnormal style in the negative sample set to obtain the preset processed sample data. The preset processing includes at least one of grayscale deviation processing, tilt processing, and image overlay background elements. Step 152: Set labels for each of the pre-processed sample data to obtain the sample dataset.

4. The method as described in claim 1, characterized in that, Step S2 includes: Step S21: Construct an initial model with a dual-branch structure. The initial model includes a detection branch and an inference branch. The detection branch is used to output the location information written in the car drawing, the detection results of each anomaly pattern, and the confidence level of the detection results. The inference branch is used to output the deviation degree of each anomaly pattern. Step S22: The intelligent computing cloud platform utilizes computing power and adopts a distributed data parallel framework to input a preset number of sample data randomly selected from the training set into the initial model for feature extraction. After that, the detection branch outputs the predicted location information, the detection results of each anomaly pattern, and the confidence level of the detection results. The inference branch outputs the deviation degree of the predicted anomaly pattern. The detection results represent the qualitative detection results of each anomaly pattern, and the deviation degree represents the quantitative detection results of each anomaly pattern. Step S23: If the quantitative detection result matches the qualitative detection result, calculate the total loss value of the initial model based on the first loss value, the second loss value, and the third loss value. The first loss value represents the error between the location information of the sample data output by the detection branch and the true location information of the sample data. The second loss value represents the error between the detection result of the sample data output by the detection branch and the true result corresponding to the label of the sample data. The third loss value represents the confidence error of the detection result of the sample data output by the detection branch. Step S24: Based on the total loss value, backpropagation is performed using the chain rule to calculate the gradient of the model parameters corresponding to the initial model, and the model parameters are iterated N times according to the gradient and the learning rate to obtain the initial model after N iterations of training, where N is a preset value, the model parameters include convolution kernel weights, bias terms, and regression parameters, and the learning rate is reduced according to the cosine annealing algorithm in each iteration. Step S25: If the total loss value of the initial model after the Mth iteration of training is less than the total loss value of the initial model after the Lth iteration of training in the validation set, the initial model after the Mth iteration of training is determined as the final model.

5. The method according to any one of claims 1 to 4, characterized in that, Step S3 includes: Step S31: After converting the final model into an inference format adapted to the intelligent computing cloud platform, deploy it on the GPU inference node of the intelligent computing cloud platform, and encapsulate the final model into a callable API interface. The API interface supports receiving batches of automotive drawings and outputting the detection results of the consistency of the annotations of the automotive drawings.

6. The method as described in claim 5, characterized in that, After step S3, the method further includes: Step S4: Generate a review report based on the detection results. The review report includes the abnormal style type, the location of the abnormal style in the car drawing, and a screenshot of the abnormal style. Step S5: Based on the review report, the test results are reviewed. If it is determined that there are abnormal results in the test results, the abnormal patterns in the car drawings corresponding to the abnormal results are used as new sample data.

7. An automatic detection device for consistency of automotive drawing annotations using computing power on an intelligent computing cloud platform, characterized in that, The device includes: The construction module is used to construct a sample dataset based on a preset consistency semantic tree, and divide the sample dataset into a training set and a validation set. The consistency semantic tree is used to represent the standard style of annotations in automotive drawings. The annotations include a first value, a first unit corresponding to the first value, a plus or minus sign, a second value, and a second unit corresponding to the second value. The consistency semantic tree includes a semantic layer, a formal layer, and a validation layer. The semantic layer is used to define that the categories of the first unit and the second unit are consistent. The formal layer is used to define that the signs of the first unit and the second unit are consistent. The validation layer is used to define the logical relationship between the first value, the first unit, the second value, and the second unit. The training module is used by the intelligent computing cloud platform to perform N iterations of training on the initial model based on the training set using computing power. The initial model after the Mth iteration of training is determined as the final model. The loss value of the initial model after the Mth iteration of training in the validation set is less than the loss value of the initial model after the Lth iteration of training in the validation set. N is a positive integer, M is a positive integer less than or equal to N, L is a positive integer less than or equal to N, and L is not equal to M. The Lth iteration of training is any one of the N iterations of training except for the Mth iteration of training. The deployment module is used to deploy the final model on the intelligent computing cloud platform to perform batch annotation consistency checks on the received automotive drawings.

8. An electronic device, characterized in that, include: A processor, a memory, and a program stored in the memory and executable on the processor, wherein the program, when executed by the processor, implements the steps of the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, Includes computer instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 6.

Images