Automobile part intelligent production control and quality early warning system based on AI model

Abstract

The invention discloses an automobile part intelligent production control and quality early warning system based on an AI model. Comprising a data acquisition module, a production process monitoring module, an AI analysis engine module, an adaptive control module, a quality early warning module, an equipment collaborative scheduling module, a data governance and model training module, a user permission and audit module and a remote operation and maintenance and knowledge precipitation module. According to the system and the method, an artificial intelligence algorithm and an industrial automation control technology are fused, and deep perception, dynamic regulation and control and defect pre-recognition of a production process are realized. The prediction precision of the method is superior to that of a traditional statistical process control method in a plurality of actual production scenes.

Description

Technical Field

[0001] This invention belongs to the field of automotive manufacturing technology and relates to an intelligent production control and quality early warning system for automotive parts based on an AI model. Background Technology

[0002] As a vital component of modern industry, automobiles involve a wide variety of parts, complex processes, and high precision requirements. The production process encompasses multiple stages, including stamping, welding, painting, and assembly. Quality fluctuations at any stage can impact the overall vehicle performance and safety. Currently, most automotive parts manufacturers still rely on manual experience or traditional automated control systems for production management and quality monitoring. This makes it difficult to cope with the data complexity and uncertainty brought about by large-scale customized production, leading to resource waste and decreased product consistency. Summary of the Invention

[0003] In order to overcome at least one deficiency of the prior art, the present invention provides an intelligent production control and quality early warning system for automotive parts based on an AI model.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent production control and quality early warning system for automotive parts based on an AI model, comprising a data acquisition module, a production process monitoring module, an AI analysis engine module, an adaptive control module, a quality early warning module, an equipment collaborative scheduling module, a data governance and model training module, a user permission and auditing module, and a remote operation and maintenance and knowledge accumulation module; the data acquisition module is connected to the production process monitoring module.

[0005] Furthermore, the data acquisition module interfaces with various sensor devices on the production line to continuously collect process parameters and equipment status information involved in the production process; the production process monitoring module receives the data collected by the data acquisition module and displays the changing trends of key process parameters through a visual interface, supporting multi-dimensional queries and historical backtracking; the AI ​​analysis engine module is used for model inference and intelligent judgment; the adaptive control module dynamically adjusts process parameters based on model suggestions; the quality early warning module is used for defect prevention and early intervention, constructing a multi-level early warning system; the equipment collaborative scheduling module is designed for multi-process joint operation scenarios to optimize resource allocation efficiency; the data governance and model training module regularly archives historical operating data, cleans noise samples, labels typical fault cases, and forms a training set; the user permission and auditing module is used to ensure system security and compliance; and the remote operation and maintenance and knowledge accumulation module is used to expand the service boundaries of the system.

[0006] Furthermore, the data acquisition module uses edge computing nodes for preliminary data filtering and compression, synchronously collects time-series data and timestamps it, stores it in a structured format in a time-series database, and caches unstructured image data in a distributed file system for subsequent use.

[0007] Furthermore, the data acquisition module uses edge computing nodes for preliminary data filtering and compression, synchronously collects time-series data and timestamps it, stores it in a structured format in a time-series database, and caches unstructured image data in a distributed file system for subsequent use.

[0008] Furthermore, the multi-level early warning system comprises three layers. The first layer is based on statistical process control, which implements X-bar R-chart monitoring for quantifiable indicators. The second layer introduces deep learning image recognition technology, which uses convolutional neural networks to automatically classify the appearance of parts and identify micro-defects. The third layer integrates multi-source information for root cause analysis, which uses graph neural networks to mine the implicit correlations between process parameters and trace the source of potential quality problems.

[0009] Furthermore, the data acquisition and preprocessing of the data acquisition module includes missing value handling, physical constraint verification mechanism, time synchronization, feature derivation, and data standardization.

[0010] Furthermore, the adaptive control module identifies the changing trends of input characteristics through an online learning mechanism and automatically adjusts the PID controller gain to maintain stability within the process window.

[0011] Furthermore, the quality early warning module integrates deep learning, unsupervised learning, and time series modeling algorithms to construct an anomaly detection architecture, capturing and providing early warnings of potential quality problems in the production process.

[0012] Furthermore, the hybrid anomaly detection architecture has a backbone model based on an autoencoder-based unsupervised learning framework. A long short-term memory network is introduced between the encoder and decoder to form an LSTM-AE structure, which captures the dynamic evolution patterns between processes and identifies progressive degradation trends.

[0013] Furthermore, the anomaly detection architecture is configured with an isolated forest and a support vector machine as auxiliary detectors to form an integrated discrimination mechanism. The isolated forest locates outliers in sparse regions by randomly segmenting the feature space, while the support vector machine constructs an optimal boundary in the feature space, encompassing the vast majority of normal samples and being sensitive to external intrusion points. The three models output their respective anomaly scores, and a comprehensive anomaly index is generated through a weighted fusion strategy. The weight allocation is dynamically adjusted based on the F1-score of each model on the validation set.

[0014] In summary, the advantages of this invention are: This invention integrates artificial intelligence algorithms with industrial automation control technology to achieve deep perception, dynamic control, and proactive defect identification of the production process. It demonstrates superior prediction accuracy compared to traditional statistical process control methods in multiple real-world production scenarios.

[0015] At the production control level, this invention utilizes an adaptive feedback adjustment module mechanism to dynamically adjust processing parameters based on real-time operating conditions, effectively reducing product deviations caused by delayed human intervention or improper settings. Coordinated scheduling between equipment is achieved through a decision-making module within a reinforcement learning framework, improving the overall response speed and resource utilization of the production line.

[0016] The quality early warning module of this invention can issue an alarm before defects form, supporting early intervention. By learning from historical defective product data, the model can accurately locate potential failure modes and push suggested remedial measures to the operation terminal, enhancing production transparency and decision-making agility. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system modules of the present invention. Detailed Implementation

[0018] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0019] Example: like Figure 1 As shown, an intelligent production control and quality early warning system for automotive parts based on an AI model includes a data acquisition module, a production process monitoring module, an AI analysis engine module, an adaptive control module, a quality early warning module, an equipment collaborative scheduling module, a data governance and model training module, a user permission and auditing module, and a remote operation and maintenance and knowledge accumulation module; the data acquisition module is connected to the production process monitoring module.

[0020] The data acquisition module interfaces with various sensor devices on the production line to continuously collect process parameters and equipment status information involved in the production process. These sensors include those for temperature, pressure, vibration, current, voltage, and visual imaging devices, enabling comprehensive monitoring of the operating status of machine tools, assembly robots, and conveyors. The visual imaging device acquires images of component surfaces to support subsequent quality defect identification tasks. The data acquisition module uses edge computing nodes for initial data filtering and compression, reducing network transmission load while ensuring real-time performance. Time-series data is synchronously acquired with millisecond-level precision and timestamped, stored in a structured time-series database, while unstructured image data is cached in a distributed file system for later retrieval.

[0021] Data processed at the edge is encapsulated into a unified message format (such as JSON or Protobuf) and pushed to the central data platform via the MQTT protocol. The platform internally uses a distributed message queue Kafka cluster to receive and buffer real-time data streams from different production lines, ensuring that critical information is not lost even during network fluctuations. Once data enters the platform, it is routed to different processing channels based on its category: process parameter data flows into the time-series database InfluxDB for storage; image data is temporarily stored in object storage MinIO; and equipment metadata and work order information are synchronously updated in the relational database PostgreSQL. The entire data flow process is visualized, orchestrated, and monitored by Apache NiFi, ensuring coordinated operation between components.

[0022] The production process monitoring module receives data from the data acquisition module and displays the changing trends of key process parameters through a visual interface, supporting multi-dimensional queries and historical backtracking. This module integrates dynamic dashboard technology, allowing configuration of KPI thresholds; any deviation from the normal range triggers an initial alarm signal. The monitoring interface is structured by workshop, workstation, and process level, allowing operators to quickly locate abnormal processes via touchscreen terminals. Simultaneously, this module is deeply integrated with the MES (Manufacturing Execution System), enabling the linked presentation of operational data such as production order progress, equipment utilization, and yield rate, improving on-site management transparency.

[0023] The AI ​​analysis engine module, as the core component of the system, is used for model inference and intelligent judgment. This module sets up trained quality prediction, fault identification, and energy consumption optimization models, receives data streams from the front end, and performs real-time inference. The models support a hot update mechanism, allowing version switching without affecting online services. To improve response speed, the inference process utilizes GPU-accelerated computation, and priority queues are set up for different tasks to ensure the timeliness of analysis on critical paths. The model output includes not only classification or regression values ​​but also confidence scores and feature contribution analyses, enhancing the credibility of decisions.

[0024] When a new batch of data reaches a specified threshold or a fixed time interval, a trigger initiates a data batch processing task, calling the feature extraction module to generate input vectors. These features encompass statistical features (mean, variance, kurtosis), frequency domain features (dominant frequency components after FFT transformation), temporal patterns (ARIMA residual sequences), and image texture features (GLCM matrix parameters). The model selects the appropriate inference path based on the current operating conditions—using a lightweight LSTM network for trend prediction under normal conditions, switching to the ensemble classifier XGBoost for risk scoring in the early stages of anomalies, and activating the deep convolutional network ResNet-34 for fine-grained defect localization when there are significant deviations.

[0025] The model output includes information from multiple dimensions: first, the predicted values ​​and confidence intervals of key parameters for the next time step, used to determine whether there is a risk of exceeding limits; second, the quality probability score of the current batch of products, marking potential defective products; and third, the equipment health degradation index, reflecting the degree of mechanical wear. These outputs are fed into the rule engine Drools for logical judgment. The rule base predefines hundreds of expert experience rules, such as "if the spindle vibration amplitude exceeds the historical 99th percentile for three consecutive minutes and the temperature rise rate is greater than 0.8℃ / min, it is judged as a risk of spindle bearing overheating." Based on this, the system generates different levels of warning signals (green - normal, yellow - attention, orange - alarm, red - emergency shutdown) and submits corresponding control suggestions to the decision center.

[0026] The decision-making center employs a state-action network trained using the reinforcement learning PPO algorithm. Combining factors such as current production line load, order priority, and maintenance resource availability, it automatically generates the optimal control strategy. For minor deviations, the system sends parameter adjustment commands to the PLC controller, automatically correcting process variables such as feed rate, coolant flow rate, or clamping torque. When moderate risks occur, the system notifies the MES system to insert a quality inspection node and replan the path for subsequent processes. In the face of major potential malfunctions, the system immediately cuts off power and isolates the affected area to prevent defect propagation. All operation records, along with contextual data, are packaged and stored in the audit log for later traceability and model retraining.

[0027] The adaptive control module dynamically adjusts process parameters based on model recommendations. For example, when an increasing tool wear trend is detected, it automatically reduces the cutting feed rate and increases the coolant flow rate; during welding, if a weld deviation risk is identified, a trajectory correction command is sent to the robotic arm controller. This module constructs functional block logic, supporting multiple industrial communication protocols such as OPCUA, Modbus TCP, and Profinet, ensuring compatibility with heterogeneous equipment. The control strategy employs a combination of fuzzy PID control and reinforcement learning, ensuring stability while providing continuous optimization capabilities.

[0028] The quality early warning module is used for defect prevention and early intervention, constructing a multi-level early warning system. This system comprises three layers: the first layer, based on Statistical Process Control (SPC), monitors quantifiable indicators such as dimensional tolerances and surface roughness using X-bar R-charts; the second layer introduces deep learning image recognition technology, utilizing convolutional neural networks to automatically classify part appearance and identify microscopic defects such as cracks, burrs, and deformation; the third layer integrates multi-source information for root cause analysis, using graph neural networks to uncover implicit correlations between process parameters and trace the source of potential quality problems. Early warning levels are divided into four categories: blue, yellow, orange, and red, corresponding to four response actions: observation, alert, pause, and shutdown, respectively.

[0029] The equipment collaborative scheduling module is designed for multi-process collaborative operations, optimizing resource allocation efficiency. This module receives production plan input and, considering factors such as current equipment health status, complexity of pending tasks, and material availability, uses an improved genetic algorithm to generate the optimal scheduling plan. In the event of sudden failures or emergency order insertions, it supports dynamic rescheduling, recalculating task priorities and path planning. The module includes a built-in digital twin simulation environment, allowing for virtual simulation of the scheduling plan's effectiveness, assessment of capacity bottlenecks and conflict points, and prevention of actual production interruptions.

[0030] The data governance and model training module regularly archives historical operational data, cleans noisy samples, and labels typical failure cases to form a high-quality training set. It supports automated machine learning (AutoML) workflows, completing the entire process of feature selection, hyperparameter tuning, and model selection. After training, the new model must undergo offline validation and shadow mode testing to confirm performance compliance before replacing the old version. This module also provides model interpretability tools to help engineers understand black-box decision-making logic and promote trust in human-machine collaboration.

[0031] The user permissions and auditing module ensures system security and compliance. Access permissions are set according to different roles. Administrators can configure process templates and system parameters, quality inspectors can only view relevant reports and warning records, and operators can only perform prescribed operations. All critical operations are logged, including operation time, IP address, and changes, meeting the audit requirements of the ISO / TS 16949 quality management system. The system supports two-factor authentication and SSL encrypted transmission to prevent unauthorized access and data leakage.

[0032] The remote operation and maintenance and knowledge accumulation module extends the system's service boundaries. By connecting to the enterprise's private cloud platform through a secure gateway, it allows expert teams to remotely diagnose system operation and push upgrade patches or optimization suggestions. Accumulated fault cases and handling experience are structured into a knowledge graph to assist in new employee training and the development of standardized operating procedures. This module can also generate periodic analysis reports, summarizing key indicators such as equipment OEE, early warning accuracy, and false alarm rate, providing strategic decision-making support for management. The various functional modules communicate decoupled through a message bus, using JSON Schema to define interface specifications, ensuring data semantic consistency and system maintainability.

[0033] In this embodiment, the data acquisition and preprocessing module plays a crucial role in the system as a fundamental step in AI model building. The goal of the data acquisition module is to obtain comprehensive, accurate, and timely production site data, and to improve data quality through standardization processing, providing reliable input for subsequent feature extraction and model training.

[0034] The data sources are diverse and heterogeneous, encompassing multiple channels such as production equipment sensors, vision inspection devices, MES (Manufacturing Execution System) databases, PLC control systems, and manually entered quality inspection records. Sensor data includes continuous time-series signals such as temperature, pressure, vibration frequency, current load, and rotational speed, with sampling frequencies typically ranging from tens to hundreds of times per second. Machine vision systems output images or point cloud data for surface defect identification and dimensional measurement. MES systems store structured data such as order information, process parameter settings, and equipment start-up and shutdown logs. Quality inspection reports are mostly in unstructured text format, requiring information extraction and processing. To address this complex data ecosystem, the system adopts a layered acquisition architecture, utilizing edge computing nodes to collect and initially filter data from underlying devices, avoiding network congestion and reducing transmission latency. The edge layer connects to various controllers via industrial communication protocols such as OPC UA and Modbus TCP, achieving millisecond-level data capture response, while also configuring a local caching mechanism to handle brief network outages. All raw data is uploaded to the cloud data center after timestamp alignment, forming a unified data lake storage system that supports the coexistence of structured and unstructured data.

[0035] To ensure the consistency and timeliness of the data required for modeling, the preprocessing process is divided into several key steps, including missing value handling, physical constraint verification mechanism, time synchronization, feature derivation, and data standardization. Missing value handling addresses the issue of sensor malfunctions or communication interruptions potentially leaving some fields empty. Different imputation strategies are employed based on different data types. For continuous variables such as temperature and pressure, a sliding window-based linear interpolation method is used to recover short-term breakpoints. If the missing value duration is long, the historical mean under similar operating conditions is introduced as a substitute, while simultaneously labeling the sample's confidence level. For discrete event data, such as equipment status codes, null values ​​are retained and converted into specific identifiers to prevent misjudgment of operating status. Outlier detection follows closely, utilizing statistical methods combined with domain knowledge for dual verification. The three-standard-deviation rule is used for initial screening of values ​​that deviate too far from the mean, followed by further confirmation using box plot boundary checks.

[0036] A physical constraint verification mechanism is introduced; for example, the spindle speed of a certain machining axis cannot exceed the maximum value marked on the equipment nameplate. Any reading exceeding this limit is considered abnormal and corrected or discarded. After basic cleaning, data smoothing is performed to reduce high-frequency noise interference. For data streams such as vibration signals, which are susceptible to electromagnetic interference, a Savitzky-Golay filter is applied for local polynomial fitting, effectively suppressing random fluctuations while preserving signal trends. For image data, grayscale normalization, noise reduction filtering, and geometric distortion correction are performed to ensure the accuracy of subsequent visual analysis.

[0037] The internal clocks of each subsystem have slight deviations, and direct splicing could cause causal inconsistencies. Therefore, the system establishes a global time reference for time synchronization, with all edge nodes periodically synchronizing with the NTP server, keeping the error within ±1 millisecond. During the data aggregation phase, the timestamp of the main control PLC is used as an anchor point to interpolate and resample information from other sources, ensuring all variables are aligned at the same time granularity. For example, temperature control data updated every 500 milliseconds and vibration signals collected every 200 milliseconds are uniformly reconstructed into a complete vector sequence with 200-millisecond intervals. This process leverages Apache Flink to achieve real-time stream processing capabilities, meeting the needs of online learning.

[0038] Feature derivation aims to uncover hidden process patterns. By mathematically transforming the original signal, new, more representative variables are generated. For example, calculating the standard deviation of tool cutting force reflects machining stability, extracting harmonic components of motor current to determine bearing wear, or combining multiple relevant parameters into a comprehensive health index. These derived features significantly enhance the model's ability to perceive complex degradation patterns.

[0039] Considering the vast differences in magnitude between different physical units—for example, temperature is measured in degrees Celsius while vibration amplitude may only be at the micrometer level—the scale effect must be eliminated. The system uses the Z-score normalization formula to transform continuous variables, ensuring their distribution has a mean of 0 and a variance of 1. For process parameters with clearly defined upper and lower limits, Min-Max normalization is used to map them to the [0, 1] interval. Categorical variables, such as material batches and mold numbers, are expanded into binary vectors using unique thermal encoding, facilitating neural network processing.

[0040] The entire preprocessing pipeline is encapsulated as reusable data pipeline components, supporting flexible configuration and version management. Each time a new production line is added or a process changes, only the corresponding module parameters need to be adjusted for rapid adaptation. Audit logs are maintained for all processing steps, recording the cleaning trajectory and transformation rules for each batch of data, ensuring traceability of results. The final high-quality output dataset not only serves current AI model training but also provides a solid foundation for the long-term accumulation of industrial knowledge. Through this rigorous data governance mechanism, the system achieves a leap from "raw data" to "usable information," providing a source of assurance for intelligent decision-making.

[0041] As a core component of the intelligent production control system, the production process monitoring module is responsible for the dynamic acquisition, transmission, and analysis of various key parameters in the manufacturing process. In the automotive parts manufacturing scenario, the production line involves multiple process stages such as stamping, welding, painting, and assembly, each with numerous factors affecting product quality and equipment operating status. Traditional monitoring methods rely on manual inspections or fixed-cycle data recording, which are insufficient to cope with complex and ever-changing production environments, easily leading to problems such as response delays and data gaps. By introducing an AI model-driven real-time monitoring system, the system can continuously acquire multi-dimensional physical signals such as temperature, pressure, vibration, current, and rotational speed through a sensor network deployed at each workstation, and combine this with visual recognition devices to capture product appearance features and assembly precision information. This data is uploaded to edge computing nodes at millisecond-level frequencies, where preliminary filtering and compression are performed locally to avoid communication delays caused by data flooding. The effective data stream, after timestamp alignment and noise removal, is sent to the AI ​​analysis module to construct a state profile of the production process.

[0042] For different types of production equipment, corresponding data acquisition terminals are configured. For example, PLC interface adapters are used to extract the operation logs of CNC machine tools, infrared thermal imagers are used to monitor the heat distribution in welding areas, and laser rangefinders are used to detect dimensional deviations of structural components. All acquisition units follow a unified communication protocol standard to ensure interoperability between equipment of different brands and models. To improve data availability, the system has a built-in self-diagnostic function that can automatically identify abnormal conditions such as sensor failure, signal drift, or network interruption, and trigger alarms to prompt maintenance personnel to intervene and troubleshoot. When the collected raw data enters the preprocessing stage, it is processed by segmented normalization using the sliding window method to eliminate modeling bias caused by differences in dimensions. At the same time, wavelet transform technology is used to separate low-frequency components reflecting equipment degradation trends and high-frequency abrupt change components characterizing sudden failures.

[0043] At the AI ​​model level, real-time monitoring relies on a fusion architecture of deep neural networks and time series prediction algorithms. Long Short-Term Memory (LSTM) networks are widely used to model the temporal dependencies between processes, accurately predicting the range of key process parameters for the next moment by learning from historical operational data. When the actual measured value deviates from the predicted range by more than the set tolerance, the system identifies it as a potential anomaly and initiates further diagnostic procedures. Convolutional Neural Networks (CNNs) are used to process image data from industrial cameras, enabling online identification of defects such as cracks, scratches, and deformations on the surface of parts. A transfer learning strategy is introduced during model training, using publicly available image datasets to pre-adjust weight parameters, and then fine-tuning them using internal labeled samples, significantly reducing the need for large-scale labeled data.

[0044] Real-time monitoring not only focuses on fluctuations in single-point data but also emphasizes tracking the status of the entire production chain. The system establishes a virtual production line model based on digital twin technology, fully mapping the equipment layout, material flow paths, and control logic in the physical world to the cloud platform. Each machining center has a unique physical agent in the digital space, and its operating status is continuously updated by the received real-time data. Operators can intuitively view the current load, cycle time matching degree, and bottleneck station distribution of each process through a 3D visualization interface. Once a robot enters a protective shutdown state due to servo motor overheating, the system immediately marks it with a red warning sign in the virtual model and simultaneously pushes alarm information to the mobile terminals of the relevant responsible persons. At the same time, the AI ​​inference engine begins to assess the impact of the event on subsequent processes, determining whether the scheduling plan needs to be adjusted or backup equipment needs to be activated.

[0045] The adaptive control module ensures stable and efficient manufacturing processes by dynamically adjusting control parameters in real time based on changes in the production environment. The system integrates a multi-source sensor network to continuously collect data on equipment operating status, process parameters, environmental conditions, and material flow, forming a high-frequency feedback loop. This data is initially processed by edge computing nodes and then analyzed and used for real-time decision-making by an AI model. Deep reinforcement learning algorithms are used to construct the control strategy model, which, without explicit prior rules, continuously optimizes action selection through interaction with the production environment, achieving precise control of key process variables such as temperature, pressure, and speed.

[0046] The adaptive control module, relying on an online learning mechanism, can identify the changing trends of input characteristics and automatically adjust the PID controller gain or other actuator parameters to maintain stability within the process window. For example, in the die-casting process, the fluidity of molten metal is significantly affected by small temperature changes. The AI ​​model combines historical defect data with current thermal imaging information to predict potential risks of porosity or shrinkage and adjusts the mold cooling rate and injection pressure curve in advance to avoid potential quality problems.

[0047] The quality early warning module integrates multiple advanced algorithms such as deep learning, unsupervised learning, and time series modeling to construct a multi-level, multi-modal anomaly detection architecture, enabling accurate capture and early warning of potential quality problems in the production process.

[0048] A hybrid anomaly detection architecture is employed. The backbone model is an unsupervised learning framework based on an autoencoder (AE), suitable for real-world industrial scenarios lacking sufficient labeled samples. This model learns a low-dimensional, compact representation of the data and reconstructs the input signal by training on a large amount of data under normal operating conditions. When the input sample deviates from the normal pattern, the reconstruction error increases significantly, allowing for the assessment of the anomaly's degree. To further enhance the modeling capability for time dependencies, a Long Short-Term Memory (LSTM) network is introduced between the encoder and decoder, forming an LSTM-AE structure. This structure effectively captures the dynamic evolution patterns between processes and identifies gradual degradation trends. For example, in a critical machining process, tool wear causes a slow increase in cutting force, which traditional thresholding methods struggle to detect in time. However, LSTM-AE can detect subtle trend shifts through historical sequence comparisons, issuing an early warning.

[0049] The anomaly detection architecture employs both Isolation Forest and One-Class SVM as auxiliary detectors, forming an ensemble discrimination mechanism. Isolation Forest quickly locates outliers in sparse regions by randomly partitioning the feature space, making it particularly suitable for handling high-dimensional sparse data. One-Class SVM, on the other hand, constructs optimal boundaries in the feature space, encompassing the vast majority of normal samples and being sensitive to external intrusion points. The three models output their respective anomaly scores, which are then weighted and fused to generate a comprehensive anomaly index. The weight allocation is dynamically adjusted based on each model's F1-score on the validation set, ensuring the overall robustness and accuracy of the discrimination.

[0050] The test results are fed back to the quality early warning platform with millisecond-level latency and are linked with the MES system. Each abnormal event is accompanied by a confidence score, occurrence timestamp, involved equipment number, process steps, and suggested handling measures, which can be visualized on the monitoring screen or pushed to mobile devices. For recurring anomalies of the same type, the system automatically clusters and merges them to generate trend reports to help management identify systemic risks. In addition, all test records are stored in the quality knowledge base for subsequent root cause analysis and process optimization closure.

[0051] The entire anomaly detection process spans the entire manufacturing chain, from raw material input to finished product delivery, covering core processes such as stamping, welding, machining, assembly, and painting. This significantly improves process controllability and product quality stability. Through continuous accumulation of operational data and iterative optimization of the model structure, the system possesses self-evolution capabilities, gradually approaching the ideal detection performance boundary.

[0052] Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort should fall within the scope of protection of the present invention.

Claims

1. An intelligent production control and quality early warning system for automotive parts based on an AI model, characterized in that: It includes a data acquisition module, a production process monitoring module, an AI analysis engine module, an adaptive control module, a quality early warning module, an equipment collaborative scheduling module, a data governance and model training module, a user permission and auditing module, and a remote operation and maintenance and knowledge accumulation module; The data acquisition module is connected to the production process monitoring module.

2. The intelligent production control and quality early warning system for automotive parts based on an AI model according to claim 1, characterized in that: The data acquisition module interfaces with various sensor devices on the production line to continuously collect process parameters and equipment status information involved in the production process; The production process monitoring module receives data collected by the data acquisition module and displays the changing trends of key process parameters through a visual interface, supporting multi-dimensional queries and historical backtracking; the AI ​​analysis engine module is used for model reasoning and intelligent judgment. The adaptive control module dynamically adjusts process parameters based on model recommendations; The quality early warning module is used for defect prevention and early intervention, building a multi-level early warning system; the equipment collaborative scheduling module is designed for multi-process joint operation scenarios, optimizing resource allocation efficiency; the data governance and model training module regularly archives historical operation data, cleans noise samples, labels typical failure cases, and forms a training set; the user permission and auditing module is used to ensure system security and compliance; and the remote operation and maintenance and knowledge accumulation module is used to expand the service boundaries of the system.

3. The intelligent production control and quality early warning system for automotive parts based on an AI model according to claim 2, characterized in that: The data acquisition module uses edge computing nodes for initial data filtering and compression. Time-series data is collected synchronously and timestamped, and stored in a structured time-series database. Unstructured image data is cached in a distributed file system for later use.

4. The intelligent production control and quality early warning system for automotive parts based on an AI model according to claim 2, characterized in that: The AI ​​analysis engine module is equipped with a trained quality prediction model, a fault identification model, and an energy consumption optimization model. It receives data streams from the front end and performs real-time inferences.

5. The intelligent production control and quality early warning system for automotive parts based on an AI model according to claim 4, characterized in that: The multi-level early warning system consists of three layers. The first layer is based on statistical process control, which implements X-bar R-chart monitoring for quantifiable indicators. The second layer introduces deep learning image recognition technology, which uses convolutional neural networks to automatically classify the appearance of parts and identify micro-defects. The third layer integrates multi-source information for root cause analysis, which uses graph neural networks to mine the implicit correlations between process parameters and trace the source of potential quality problems.

6. The intelligent production control and quality early warning system for automotive parts based on an AI model according to claim 2, characterized in that: The data acquisition and preprocessing of the data acquisition module includes missing value handling, physical constraint verification mechanism, time synchronization, feature derivation, and data standardization.

7. The intelligent production control and quality early warning system for automotive parts based on an AI model according to claim 4, characterized in that: The adaptive control module uses an online learning mechanism to identify the changing trends of input characteristics and automatically adjusts the PID controller gain to maintain stability within the process window.

8. The intelligent production control and quality early warning system for automotive parts based on an AI model according to claim 2, characterized in that: The quality early warning module integrates deep learning, unsupervised learning, and time series modeling algorithms to construct an anomaly detection architecture, capturing and providing early warnings of potential quality problems in the production process.

9. The intelligent production control and quality early warning system for automotive parts based on an AI model according to claim 8, characterized in that: The hybrid anomaly detection architecture has a backbone model based on an autoencoder-based unsupervised learning framework. A long short-term memory network is introduced between the encoder and decoder to form an LSTM-AE structure, which captures the dynamic evolution patterns between processes and identifies progressive degradation trends.

10. The intelligent production control and quality early warning system for automotive parts based on an AI model according to claim 5, characterized in that: The anomaly detection architecture is configured with an isolated forest and a support vector machine as auxiliary detectors to form an integrated discrimination mechanism. The isolated forest locates outliers in sparse regions by randomly segmenting the feature space. The support vector machine constructs an optimal boundary in the feature space, which encloses the vast majority of normal samples and is sensitive to external intrusion points. The three models output their respective anomaly scores, and a comprehensive anomaly index is generated through a weighted fusion strategy. The weight allocation is dynamically adjusted based on the F1-score of each model on the validation set.

Images